more easy lowp shader stages

This fills out a couple more matrix and gather stages.

Deletes a not particularly important unit test that was using a
scale matrix in a weird, non-lowp compatible way.

This will require guards for Blink layout tests.

Change-Id: I54cb228ff541f771e8f4758f07d26c5161d48af3
Reviewed-on: https://skia-review.googlesource.com/62520
Reviewed-by: Mike Reed <reed@google.com>
Commit-Queue: Mike Klein <mtklein@chromium.org>
diff --git a/src/jumper/SkJumper.cpp b/src/jumper/SkJumper.cpp
index e026287..a9166c1 100644
--- a/src/jumper/SkJumper.cpp
+++ b/src/jumper/SkJumper.cpp
@@ -228,7 +228,13 @@
     LOWP(exclusion)
     LOWP(hardlight)
     LOWP(overlay)
+#if defined(SK_LEGACY_LOWP_STAGES)
     LOWP(seed_shader) LOWP(matrix_2x3) LOWP(gather_8888)
+#else
+    LOWP(seed_shader)
+    LOWP(matrix_translate) LOWP(matrix_scale_translate) LOWP(matrix_2x3) LOWP(matrix_perspective)
+    LOWP(gather_8888) LOWP(gather_bgra) LOWP(gather_565) LOWP(gather_a8) LOWP(gather_g8)
+#endif
     #undef LOWP
 #endif
 
diff --git a/src/jumper/SkJumper_generated.S b/src/jumper/SkJumper_generated.S
index c4292fb..6dd791a 100644
--- a/src/jumper/SkJumper_generated.S
+++ b/src/jumper/SkJumper_generated.S
@@ -9465,9 +9465,9 @@
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
   .byte  98,242,125,40,124,201               // vpbroadcastd  %ecx,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  98,241,116,56,88,13,42,43,3,0       // vaddps        0x32b2a(%rip){1to8},%ymm1,%ymm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,241,116,56,88,13,218,65,3,0      // vaddps        0x341da(%rip){1to8},%ymm1,%ymm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,35,43,3,0         // vbroadcastss  0x32b23(%rip),%ymm2        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,21,211,65,3,0        // vbroadcastss  0x341d3(%rip),%ymm2        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,228,87,219                      // vxorps        %ymm3,%ymm3,%ymm3
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
@@ -9481,16 +9481,16 @@
 _sk_dither_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  98,114,125,40,124,194               // vpbroadcastd  %edx,%ymm8
-  .byte  197,61,254,5,173,44,3,0             // vpaddd        0x32cad(%rip),%ymm8,%ymm8        # 32de0 <_sk_srcover_rgba_8888_sse2_lowp+0x460>
+  .byte  197,61,254,5,77,67,3,0              // vpaddd        0x3434d(%rip),%ymm8,%ymm8        # 34480 <_sk_srcover_rgba_8888_sse2_lowp+0x450>
   .byte  98,114,125,40,124,201               // vpbroadcastd  %ecx,%ymm9
   .byte  196,65,53,239,200                   // vpxor         %ymm8,%ymm9,%ymm9
-  .byte  196,98,125,88,21,237,42,3,0         // vpbroadcastd  0x32aed(%rip),%ymm10        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,98,125,88,21,157,65,3,0         // vpbroadcastd  0x3419d(%rip),%ymm10        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  196,65,53,219,218                   // vpand         %ymm10,%ymm9,%ymm11
   .byte  196,193,37,114,243,5                // vpslld        $0x5,%ymm11,%ymm11
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,193,45,114,242,4                // vpslld        $0x4,%ymm10,%ymm10
-  .byte  196,98,125,88,37,210,42,3,0         // vpbroadcastd  0x32ad2(%rip),%ymm12        # 32c38 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
-  .byte  196,98,125,88,45,205,42,3,0         // vpbroadcastd  0x32acd(%rip),%ymm13        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,125,88,37,130,65,3,0         // vpbroadcastd  0x34182(%rip),%ymm12        # 342e8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
+  .byte  196,98,125,88,45,125,65,3,0         // vpbroadcastd  0x3417d(%rip),%ymm13        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,65,53,219,245                   // vpand         %ymm13,%ymm9,%ymm14
   .byte  196,193,13,114,246,2                // vpslld        $0x2,%ymm14,%ymm14
   .byte  196,65,37,235,222                   // vpor          %ymm14,%ymm11,%ymm11
@@ -9505,8 +9505,8 @@
   .byte  196,65,61,235,195                   // vpor          %ymm11,%ymm8,%ymm8
   .byte  196,65,61,235,193                   // vpor          %ymm9,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,127,42,3,0         // vbroadcastss  0x32a7f(%rip),%ymm9        # 32c40 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
-  .byte  98,114,61,56,168,13,121,42,3,0      // vfmadd213ps   0x32a79(%rip){1to8},%ymm8,%ymm9        # 32c44 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
+  .byte  196,98,125,24,13,47,65,3,0          // vbroadcastss  0x3412f(%rip),%ymm9        # 342f0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
+  .byte  98,114,61,56,168,13,41,65,3,0       // vfmadd213ps   0x34129(%rip){1to8},%ymm8,%ymm9        # 342f4 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
   .byte  196,98,125,24,0                     // vbroadcastss  (%rax),%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
   .byte  197,188,88,192                      // vaddps        %ymm0,%ymm8,%ymm0
@@ -9539,7 +9539,7 @@
 FUNCTION(_sk_black_color_skx)
 _sk_black_color_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,6,42,3,0          // vbroadcastss  0x32a06(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,182,64,3,0        // vbroadcastss  0x340b6(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
@@ -9550,7 +9550,7 @@
 FUNCTION(_sk_white_color_skx)
 _sk_white_color_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,237,41,3,0         // vbroadcastss  0x329ed(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,157,64,3,0         // vbroadcastss  0x3409d(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  197,252,40,216                      // vmovaps       %ymm0,%ymm3
@@ -9596,7 +9596,7 @@
 FUNCTION(_sk_srcatop_skx)
 _sk_srcatop_skx:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,140,41,3,0          // vbroadcastss  0x3298c(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,60,64,3,0           // vbroadcastss  0x3403c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,226,61,184,196                  // vfmadd231ps   %ymm4,%ymm8,%ymm0
   .byte  197,244,89,207                      // vmulps        %ymm7,%ymm1,%ymm1
@@ -9612,7 +9612,7 @@
 .globl _sk_dstatop_skx
 FUNCTION(_sk_dstatop_skx)
 _sk_dstatop_skx:
-  .byte  196,98,125,24,5,91,41,3,0           // vbroadcastss  0x3295b(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,11,64,3,0           // vbroadcastss  0x3400b(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,226,101,184,196                 // vfmadd231ps   %ymm4,%ymm3,%ymm0
@@ -9651,7 +9651,7 @@
 .globl _sk_srcout_skx
 FUNCTION(_sk_srcout_skx)
 _sk_srcout_skx:
-  .byte  196,98,125,24,5,254,40,3,0          // vbroadcastss  0x328fe(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,174,63,3,0          // vbroadcastss  0x33fae(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
@@ -9664,7 +9664,7 @@
 .globl _sk_dstout_skx
 FUNCTION(_sk_dstout_skx)
 _sk_dstout_skx:
-  .byte  196,226,125,24,5,221,40,3,0         // vbroadcastss  0x328dd(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,141,63,3,0         // vbroadcastss  0x33f8d(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,92,219                      // vsubps        %ymm3,%ymm0,%ymm3
   .byte  197,228,89,196                      // vmulps        %ymm4,%ymm3,%ymm0
   .byte  197,228,89,205                      // vmulps        %ymm5,%ymm3,%ymm1
@@ -9677,7 +9677,7 @@
 .globl _sk_srcover_skx
 FUNCTION(_sk_srcover_skx)
 _sk_srcover_skx:
-  .byte  196,98,125,24,5,188,40,3,0          // vbroadcastss  0x328bc(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,108,63,3,0          // vbroadcastss  0x33f6c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,93,184,192                  // vfmadd231ps   %ymm8,%ymm4,%ymm0
   .byte  196,194,85,184,200                  // vfmadd231ps   %ymm8,%ymm5,%ymm1
@@ -9690,7 +9690,7 @@
 .globl _sk_dstover_skx
 FUNCTION(_sk_dstover_skx)
 _sk_dstover_skx:
-  .byte  196,98,125,24,5,151,40,3,0          // vbroadcastss  0x32897(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,71,63,3,0           // vbroadcastss  0x33f47(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  196,226,61,168,196                  // vfmadd213ps   %ymm4,%ymm8,%ymm0
   .byte  196,226,61,168,205                  // vfmadd213ps   %ymm5,%ymm8,%ymm1
@@ -9714,7 +9714,7 @@
 .globl _sk_multiply_skx
 FUNCTION(_sk_multiply_skx)
 _sk_multiply_skx:
-  .byte  196,98,125,24,5,94,40,3,0           // vbroadcastss  0x3285e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,14,63,3,0           // vbroadcastss  0x33f0e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,208                       // vmulps        %ymm0,%ymm9,%ymm10
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -9737,7 +9737,7 @@
 FUNCTION(_sk_plus__skx)
 _sk_plus__skx:
   .byte  197,252,88,196                      // vaddps        %ymm4,%ymm0,%ymm0
-  .byte  196,98,125,24,5,13,40,3,0           // vbroadcastss  0x3280d(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,189,62,3,0          // vbroadcastss  0x33ebd(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  197,244,88,205                      // vaddps        %ymm5,%ymm1,%ymm1
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
@@ -9767,7 +9767,7 @@
 .globl _sk_xor__skx
 FUNCTION(_sk_xor__skx)
 _sk_xor__skx:
-  .byte  196,98,125,24,5,184,39,3,0          // vbroadcastss  0x327b8(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,104,62,3,0          // vbroadcastss  0x33e68(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -9801,7 +9801,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,95,209                  // vmaxps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,60,39,3,0           // vbroadcastss  0x3273c(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,236,61,3,0          // vbroadcastss  0x33dec(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -9826,7 +9826,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,231,38,3,0          // vbroadcastss  0x326e7(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,151,61,3,0          // vbroadcastss  0x33d97(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -9854,7 +9854,7 @@
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,134,38,3,0          // vbroadcastss  0x32686(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,54,61,3,0           // vbroadcastss  0x33d36(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -9876,7 +9876,7 @@
   .byte  197,236,89,214                      // vmulps        %ymm6,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,64,38,3,0           // vbroadcastss  0x32640(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,240,60,3,0          // vbroadcastss  0x33cf0(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -9888,7 +9888,7 @@
 _sk_colorburn_skx:
   .byte  98,241,92,40,194,199,0              // vcmpeqps      %ymm7,%ymm4,%k0
   .byte  98,114,126,40,56,192                // vpmovm2d      %k0,%ymm8
-  .byte  196,98,125,24,13,29,38,3,0          // vbroadcastss  0x3261d(%rip),%ymm9        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,13,205,60,3,0         // vbroadcastss  0x33ccd(%rip),%ymm9        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,52,92,215                       // vsubps        %ymm7,%ymm9,%ymm10
   .byte  197,44,89,216                       // vmulps        %ymm0,%ymm10,%ymm11
   .byte  197,36,88,228                       // vaddps        %ymm4,%ymm11,%ymm12
@@ -9952,7 +9952,7 @@
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  98,209,92,40,194,192,0              // vcmpeqps      %ymm8,%ymm4,%k0
   .byte  98,114,126,40,56,200                // vpmovm2d      %k0,%ymm9
-  .byte  196,98,125,24,21,238,36,3,0         // vbroadcastss  0x324ee(%rip),%ymm10        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,21,158,59,3,0         // vbroadcastss  0x33b9e(%rip),%ymm10        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,44,92,223                       // vsubps        %ymm7,%ymm10,%ymm11
   .byte  197,36,89,224                       // vmulps        %ymm0,%ymm11,%ymm12
   .byte  98,241,124,40,194,195,0             // vcmpeqps      %ymm3,%ymm0,%k0
@@ -10009,7 +10009,7 @@
 .globl _sk_hardlight_skx
 FUNCTION(_sk_hardlight_skx)
 _sk_hardlight_skx:
-  .byte  196,98,125,24,5,230,35,3,0          // vbroadcastss  0x323e6(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,150,58,3,0          // vbroadcastss  0x33a96(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,208                       // vmulps        %ymm0,%ymm9,%ymm10
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -10063,7 +10063,7 @@
 .globl _sk_overlay_skx
 FUNCTION(_sk_overlay_skx)
 _sk_overlay_skx:
-  .byte  196,98,125,24,5,2,35,3,0            // vbroadcastss  0x32302(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,178,57,3,0          // vbroadcastss  0x339b2(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,208                       // vmulps        %ymm0,%ymm9,%ymm10
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -10129,15 +10129,15 @@
   .byte  196,65,52,88,217                    // vaddps        %ymm9,%ymm9,%ymm11
   .byte  196,65,36,88,235                    // vaddps        %ymm11,%ymm11,%ymm13
   .byte  197,4,92,243                        // vsubps        %ymm3,%ymm15,%ymm14
-  .byte  196,98,125,24,29,224,33,3,0         // vbroadcastss  0x321e0(%rip),%ymm11        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,29,144,56,3,0         // vbroadcastss  0x33890(%rip),%ymm11        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,65,36,92,225                    // vsubps        %ymm9,%ymm11,%ymm12
   .byte  196,98,13,168,227                   // vfmadd213ps   %ymm3,%ymm14,%ymm12
   .byte  197,156,89,212                      // vmulps        %ymm4,%ymm12,%ymm2
   .byte  196,66,21,168,237                   // vfmadd213ps   %ymm13,%ymm13,%ymm13
-  .byte  98,226,125,40,24,21,219,33,3,0      // vbroadcastss  0x321db(%rip),%ymm18        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  98,226,125,40,24,21,139,56,3,0      // vbroadcastss  0x3388b(%rip),%ymm18        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  98,161,52,40,88,202                 // vaddps        %ymm18,%ymm9,%ymm17
   .byte  98,81,116,32,89,229                 // vmulps        %ymm13,%ymm17,%ymm12
-  .byte  98,226,125,40,24,29,201,33,3,0      // vbroadcastss  0x321c9(%rip),%ymm19        # 32c4c <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
+  .byte  98,226,125,40,24,29,121,56,3,0      // vbroadcastss  0x33879(%rip),%ymm19        # 342fc <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
   .byte  98,50,53,40,184,227                 // vfmadd231ps   %ymm19,%ymm9,%ymm12
   .byte  98,194,125,40,78,201                // vrsqrt14ps    %ymm9,%ymm17
   .byte  98,162,125,40,76,201                // vrcp14ps      %ymm17,%ymm17
@@ -10259,11 +10259,11 @@
   .byte  196,65,28,89,219                    // vmulps        %ymm11,%ymm12,%ymm11
   .byte  196,65,36,94,222                    // vdivps        %ymm14,%ymm11,%ymm11
   .byte  196,99,37,74,225,240                // vblendvps     %ymm15,%ymm1,%ymm11,%ymm12
-  .byte  196,98,125,24,45,100,31,3,0         // vbroadcastss  0x31f64(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,53,95,31,3,0          // vbroadcastss  0x31f5f(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,45,20,54,3,0          // vbroadcastss  0x33614(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,15,54,3,0          // vbroadcastss  0x3360f(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,254                    // vmulps        %ymm14,%ymm5,%ymm15
   .byte  196,66,93,184,253                   // vfmadd231ps   %ymm13,%ymm4,%ymm15
-  .byte  98,226,125,40,24,5,79,31,3,0        // vbroadcastss  0x31f4f(%rip),%ymm16        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  98,226,125,40,24,5,255,53,3,0       // vbroadcastss  0x335ff(%rip),%ymm16        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  98,50,77,40,184,248                 // vfmadd231ps   %ymm16,%ymm6,%ymm15
   .byte  196,65,44,89,222                    // vmulps        %ymm14,%ymm10,%ymm11
   .byte  196,66,53,184,221                   // vfmadd231ps   %ymm13,%ymm9,%ymm11
@@ -10320,7 +10320,7 @@
   .byte  196,65,12,88,219                    // vaddps        %ymm11,%ymm14,%ymm11
   .byte  196,67,45,74,203,144                // vblendvps     %ymm9,%ymm11,%ymm10,%ymm9
   .byte  197,52,95,201                       // vmaxps        %ymm1,%ymm9,%ymm9
-  .byte  196,226,125,24,13,235,29,3,0        // vbroadcastss  0x31deb(%rip),%ymm1        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,13,155,52,3,0        // vbroadcastss  0x3349b(%rip),%ymm1        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,116,92,215                      // vsubps        %ymm7,%ymm1,%ymm10
   .byte  197,172,89,192                      // vmulps        %ymm0,%ymm10,%ymm0
   .byte  197,116,92,219                      // vsubps        %ymm3,%ymm1,%ymm11
@@ -10372,11 +10372,11 @@
   .byte  196,65,28,89,219                    // vmulps        %ymm11,%ymm12,%ymm11
   .byte  196,65,36,94,222                    // vdivps        %ymm14,%ymm11,%ymm11
   .byte  196,99,37,74,225,240                // vblendvps     %ymm15,%ymm1,%ymm11,%ymm12
-  .byte  196,98,125,24,45,32,29,3,0          // vbroadcastss  0x31d20(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,53,27,29,3,0          // vbroadcastss  0x31d1b(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,45,208,51,3,0         // vbroadcastss  0x333d0(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,203,51,3,0         // vbroadcastss  0x333cb(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,254                    // vmulps        %ymm14,%ymm5,%ymm15
   .byte  196,66,93,184,253                   // vfmadd231ps   %ymm13,%ymm4,%ymm15
-  .byte  98,226,125,40,24,5,11,29,3,0        // vbroadcastss  0x31d0b(%rip),%ymm16        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  98,226,125,40,24,5,187,51,3,0       // vbroadcastss  0x333bb(%rip),%ymm16        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  98,50,77,40,184,248                 // vfmadd231ps   %ymm16,%ymm6,%ymm15
   .byte  196,65,44,89,222                    // vmulps        %ymm14,%ymm10,%ymm11
   .byte  196,66,53,184,221                   // vfmadd231ps   %ymm13,%ymm9,%ymm11
@@ -10433,7 +10433,7 @@
   .byte  196,65,12,88,219                    // vaddps        %ymm11,%ymm14,%ymm11
   .byte  196,67,45,74,203,144                // vblendvps     %ymm9,%ymm11,%ymm10,%ymm9
   .byte  197,52,95,201                       // vmaxps        %ymm1,%ymm9,%ymm9
-  .byte  196,226,125,24,13,167,27,3,0        // vbroadcastss  0x31ba7(%rip),%ymm1        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,13,87,50,3,0         // vbroadcastss  0x33257(%rip),%ymm1        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,116,92,215                      // vsubps        %ymm7,%ymm1,%ymm10
   .byte  197,172,89,192                      // vmulps        %ymm0,%ymm10,%ymm0
   .byte  197,116,92,219                      // vsubps        %ymm3,%ymm1,%ymm11
@@ -10457,11 +10457,11 @@
   .byte  197,124,89,199                      // vmulps        %ymm7,%ymm0,%ymm8
   .byte  197,116,89,207                      // vmulps        %ymm7,%ymm1,%ymm9
   .byte  197,108,89,223                      // vmulps        %ymm7,%ymm2,%ymm11
-  .byte  196,98,125,24,37,112,27,3,0         // vbroadcastss  0x31b70(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,45,107,27,3,0         // vbroadcastss  0x31b6b(%rip),%ymm13        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,37,32,50,3,0          // vbroadcastss  0x33220(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,45,27,50,3,0          // vbroadcastss  0x3321b(%rip),%ymm13        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,245                    // vmulps        %ymm13,%ymm5,%ymm14
   .byte  196,66,93,184,244                   // vfmadd231ps   %ymm12,%ymm4,%ymm14
-  .byte  196,98,125,24,61,92,27,3,0          // vbroadcastss  0x31b5c(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,12,50,3,0          // vbroadcastss  0x3320c(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,66,77,184,247                   // vfmadd231ps   %ymm15,%ymm6,%ymm14
   .byte  196,65,52,89,213                    // vmulps        %ymm13,%ymm9,%ymm10
   .byte  196,66,61,184,212                   // vfmadd231ps   %ymm12,%ymm8,%ymm10
@@ -10519,7 +10519,7 @@
   .byte  196,65,20,88,219                    // vaddps        %ymm11,%ymm13,%ymm11
   .byte  196,67,53,74,203,192                // vblendvps     %ymm12,%ymm11,%ymm9,%ymm9
   .byte  98,49,52,40,95,202                  // vmaxps        %ymm18,%ymm9,%ymm9
-  .byte  196,98,125,24,29,237,25,3,0         // vbroadcastss  0x319ed(%rip),%ymm11        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,29,157,48,3,0         // vbroadcastss  0x3309d(%rip),%ymm11        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,36,92,231                       // vsubps        %ymm7,%ymm11,%ymm12
   .byte  197,156,89,192                      // vmulps        %ymm0,%ymm12,%ymm0
   .byte  197,36,92,219                       // vsubps        %ymm3,%ymm11,%ymm11
@@ -10543,11 +10543,11 @@
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
   .byte  197,100,89,205                      // vmulps        %ymm5,%ymm3,%ymm9
   .byte  197,100,89,222                      // vmulps        %ymm6,%ymm3,%ymm11
-  .byte  196,98,125,24,37,184,25,3,0         // vbroadcastss  0x319b8(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,45,179,25,3,0         // vbroadcastss  0x319b3(%rip),%ymm13        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,37,104,48,3,0         // vbroadcastss  0x33068(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,45,99,48,3,0          // vbroadcastss  0x33063(%rip),%ymm13        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,116,89,245                   // vmulps        %ymm13,%ymm1,%ymm14
   .byte  196,66,125,184,244                  // vfmadd231ps   %ymm12,%ymm0,%ymm14
-  .byte  196,98,125,24,61,164,25,3,0         // vbroadcastss  0x319a4(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,84,48,3,0          // vbroadcastss  0x33054(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,66,109,184,247                  // vfmadd231ps   %ymm15,%ymm2,%ymm14
   .byte  196,65,52,89,213                    // vmulps        %ymm13,%ymm9,%ymm10
   .byte  196,66,61,184,212                   // vfmadd231ps   %ymm12,%ymm8,%ymm10
@@ -10605,7 +10605,7 @@
   .byte  196,65,20,88,219                    // vaddps        %ymm11,%ymm13,%ymm11
   .byte  196,67,53,74,203,192                // vblendvps     %ymm12,%ymm11,%ymm9,%ymm9
   .byte  98,49,52,40,95,202                  // vmaxps        %ymm18,%ymm9,%ymm9
-  .byte  196,98,125,24,29,53,24,3,0          // vbroadcastss  0x31835(%rip),%ymm11        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,29,229,46,3,0         // vbroadcastss  0x32ee5(%rip),%ymm11        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,36,92,231                       // vsubps        %ymm7,%ymm11,%ymm12
   .byte  197,156,89,192                      // vmulps        %ymm0,%ymm12,%ymm0
   .byte  197,36,92,219                       // vsubps        %ymm3,%ymm11,%ymm11
@@ -10648,9 +10648,9 @@
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
   .byte  197,197,114,215,24                  // vpsrld        $0x18,%ymm7,%ymm7
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  196,98,125,24,5,140,23,3,0          // vbroadcastss  0x3178c(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,60,46,3,0           // vbroadcastss  0x32e3c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
-  .byte  196,98,125,24,13,171,23,3,0         // vbroadcastss  0x317ab(%rip),%ymm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,13,91,46,3,0          // vbroadcastss  0x32e5b(%rip),%ymm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,194,93,184,192                  // vfmadd231ps   %ymm8,%ymm4,%ymm0
   .byte  196,193,116,89,201                  // vmulps        %ymm9,%ymm1,%ymm1
@@ -10782,7 +10782,7 @@
 .globl _sk_clamp_1_skx
 FUNCTION(_sk_clamp_1_skx)
 _sk_clamp_1_skx:
-  .byte  196,98,125,24,5,146,21,3,0          // vbroadcastss  0x31592(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,66,44,3,0           // vbroadcastss  0x32c42(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
@@ -10794,7 +10794,7 @@
 .globl _sk_clamp_a_skx
 FUNCTION(_sk_clamp_a_skx)
 _sk_clamp_a_skx:
-  .byte  98,241,100,56,93,29,112,21,3,0      // vminps        0x31570(%rip){1to8},%ymm3,%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,100,56,93,29,32,44,3,0       // vminps        0x32c20(%rip){1to8},%ymm3,%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,93,195                      // vminps        %ymm3,%ymm0,%ymm0
   .byte  197,244,93,203                      // vminps        %ymm3,%ymm1,%ymm1
   .byte  197,236,93,211                      // vminps        %ymm3,%ymm2,%ymm2
@@ -10805,7 +10805,7 @@
 .globl _sk_clamp_a_dst_skx
 FUNCTION(_sk_clamp_a_dst_skx)
 _sk_clamp_a_dst_skx:
-  .byte  98,241,68,56,93,61,86,21,3,0        // vminps        0x31556(%rip){1to8},%ymm7,%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,68,56,93,61,6,44,3,0         // vminps        0x32c06(%rip){1to8},%ymm7,%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,220,93,231                      // vminps        %ymm7,%ymm4,%ymm4
   .byte  197,212,93,239                      // vminps        %ymm7,%ymm5,%ymm5
   .byte  197,204,93,247                      // vminps        %ymm7,%ymm6,%ymm6
@@ -10837,7 +10837,7 @@
 .globl _sk_invert_skx
 FUNCTION(_sk_invert_skx)
 _sk_invert_skx:
-  .byte  196,98,125,24,5,22,21,3,0           // vbroadcastss  0x31516(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,198,43,3,0          // vbroadcastss  0x32bc6(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,188,92,192                      // vsubps        %ymm0,%ymm8,%ymm0
   .byte  197,188,92,201                      // vsubps        %ymm1,%ymm8,%ymm1
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
@@ -10891,9 +10891,9 @@
 .globl _sk_unpremul_skx
 FUNCTION(_sk_unpremul_skx)
 _sk_unpremul_skx:
-  .byte  196,98,125,24,5,177,20,3,0          // vbroadcastss  0x314b1(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,97,43,3,0           // vbroadcastss  0x32b61(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,94,195                       // vdivps        %ymm3,%ymm8,%ymm8
-  .byte  98,241,60,56,194,5,210,20,3,0,1     // vcmpltps      0x314d2(%rip){1to8},%ymm8,%k0        # 32c60 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
+  .byte  98,241,60,56,194,5,130,43,3,0,1     // vcmpltps      0x32b82(%rip){1to8},%ymm8,%k0        # 34310 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
   .byte  98,114,126,40,56,200                // vpmovm2d      %k0,%ymm9
   .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
   .byte  196,67,45,74,192,144                // vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
@@ -10907,16 +10907,16 @@
 .globl _sk_from_srgb_skx
 FUNCTION(_sk_from_srgb_skx)
 _sk_from_srgb_skx:
-  .byte  196,98,125,24,5,172,20,3,0          // vbroadcastss  0x314ac(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,92,43,3,0           // vbroadcastss  0x32b5c(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,158,20,3,0         // vbroadcastss  0x3149e(%rip),%ymm11        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  .byte  196,98,125,24,37,125,20,3,0         // vbroadcastss  0x3147d(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,78,43,3,0          // vbroadcastss  0x32b4e(%rip),%ymm11        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,37,45,43,3,0          // vbroadcastss  0x32b2d(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,125,168,235                  // vfmadd213ps   %ymm11,%ymm0,%ymm13
-  .byte  196,98,125,24,53,134,20,3,0         // vbroadcastss  0x31486(%rip),%ymm14        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,53,54,43,3,0          // vbroadcastss  0x32b36(%rip),%ymm14        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,124,20,3,0         // vbroadcastss  0x3147c(%rip),%ymm10        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,21,44,43,3,0          // vbroadcastss  0x32b2c(%rip),%ymm10        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  98,209,124,40,194,194,1             // vcmpltps      %ymm10,%ymm0,%k0
   .byte  98,242,126,40,56,192                // vpmovm2d      %k0,%ymm0
   .byte  196,195,21,74,193,0                 // vblendvps     %ymm0,%ymm9,%ymm13,%ymm0
@@ -10942,16 +10942,16 @@
 .globl _sk_from_srgb_dst_skx
 FUNCTION(_sk_from_srgb_dst_skx)
 _sk_from_srgb_dst_skx:
-  .byte  196,98,125,24,5,255,19,3,0          // vbroadcastss  0x313ff(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,175,42,3,0          // vbroadcastss  0x32aaf(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,92,89,200                    // vmulps        %ymm8,%ymm4,%ymm9
   .byte  197,92,89,212                       // vmulps        %ymm4,%ymm4,%ymm10
-  .byte  196,98,125,24,29,241,19,3,0         // vbroadcastss  0x313f1(%rip),%ymm11        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  .byte  196,98,125,24,37,208,19,3,0         // vbroadcastss  0x313d0(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,161,42,3,0         // vbroadcastss  0x32aa1(%rip),%ymm11        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,37,128,42,3,0         // vbroadcastss  0x32a80(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,93,168,235                   // vfmadd213ps   %ymm11,%ymm4,%ymm13
-  .byte  196,98,125,24,53,217,19,3,0         // vbroadcastss  0x313d9(%rip),%ymm14        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,53,137,42,3,0         // vbroadcastss  0x32a89(%rip),%ymm14        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,207,19,3,0         // vbroadcastss  0x313cf(%rip),%ymm10        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,21,127,42,3,0         // vbroadcastss  0x32a7f(%rip),%ymm10        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  98,209,92,40,194,194,1              // vcmpltps      %ymm10,%ymm4,%k0
   .byte  98,242,126,40,56,224                // vpmovm2d      %k0,%ymm4
   .byte  196,195,21,74,225,64                // vblendvps     %ymm4,%ymm9,%ymm13,%ymm4
@@ -10978,19 +10978,19 @@
 FUNCTION(_sk_to_srgb_skx)
 _sk_to_srgb_skx:
   .byte  98,114,125,40,78,192                // vrsqrt14ps    %ymm0,%ymm8
-  .byte  196,98,125,24,13,92,19,3,0          // vbroadcastss  0x3135c(%rip),%ymm9        # 32c74 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
+  .byte  196,98,125,24,13,12,42,3,0          // vbroadcastss  0x32a0c(%rip),%ymm9        # 34324 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
   .byte  196,65,124,89,209                   // vmulps        %ymm9,%ymm0,%ymm10
-  .byte  196,98,125,24,29,82,19,3,0          // vbroadcastss  0x31352(%rip),%ymm11        # 32c78 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
-  .byte  196,98,125,24,37,77,19,3,0          // vbroadcastss  0x3134d(%rip),%ymm12        # 32c7c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
+  .byte  196,98,125,24,29,2,42,3,0           // vbroadcastss  0x32a02(%rip),%ymm11        # 34328 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
+  .byte  196,98,125,24,37,253,41,3,0         // vbroadcastss  0x329fd(%rip),%ymm12        # 3432c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,61,168,235                   // vfmadd213ps   %ymm11,%ymm8,%ymm13
-  .byte  196,98,125,24,53,62,19,3,0          // vbroadcastss  0x3133e(%rip),%ymm14        # 32c80 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
+  .byte  196,98,125,24,53,238,41,3,0         // vbroadcastss  0x329ee(%rip),%ymm14        # 34330 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
   .byte  196,66,61,168,238                   // vfmadd213ps   %ymm14,%ymm8,%ymm13
-  .byte  196,98,125,24,61,52,19,3,0          // vbroadcastss  0x31334(%rip),%ymm15        # 32c84 <_sk_srcover_rgba_8888_sse2_lowp+0x304>
+  .byte  196,98,125,24,61,228,41,3,0         // vbroadcastss  0x329e4(%rip),%ymm15        # 34334 <_sk_srcover_rgba_8888_sse2_lowp+0x304>
   .byte  196,65,60,88,199                    // vaddps        %ymm15,%ymm8,%ymm8
   .byte  98,82,125,40,76,192                 // vrcp14ps      %ymm8,%ymm8
   .byte  196,65,20,89,192                    // vmulps        %ymm8,%ymm13,%ymm8
-  .byte  196,98,125,24,45,31,19,3,0          // vbroadcastss  0x3131f(%rip),%ymm13        # 32c88 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
+  .byte  196,98,125,24,45,207,41,3,0         // vbroadcastss  0x329cf(%rip),%ymm13        # 34338 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
   .byte  98,209,124,40,194,197,1             // vcmpltps      %ymm13,%ymm0,%k0
   .byte  98,242,126,40,56,192                // vpmovm2d      %k0,%ymm0
   .byte  196,195,61,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm8,%ymm0
@@ -11028,7 +11028,7 @@
   .byte  197,116,93,194                      // vminps        %ymm2,%ymm1,%ymm8
   .byte  196,65,124,93,208                   // vminps        %ymm8,%ymm0,%ymm10
   .byte  98,193,52,40,92,194                 // vsubps        %ymm10,%ymm9,%ymm16
-  .byte  196,98,125,24,5,16,18,3,0           // vbroadcastss  0x31210(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,192,40,3,0          // vbroadcastss  0x328c0(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  98,49,60,40,94,224                  // vdivps        %ymm16,%ymm8,%ymm12
   .byte  98,209,52,40,194,194,0              // vcmpeqps      %ymm10,%ymm9,%k0
   .byte  98,114,126,40,56,192                // vpmovm2d      %k0,%ymm8
@@ -11037,23 +11037,23 @@
   .byte  197,116,92,242                      // vsubps        %ymm2,%ymm1,%ymm14
   .byte  98,241,116,40,194,194,1             // vcmpltps      %ymm2,%ymm1,%k0
   .byte  98,114,126,40,56,248                // vpmovm2d      %k0,%ymm15
-  .byte  196,98,125,24,29,50,18,3,0          // vbroadcastss  0x31232(%rip),%ymm11        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,98,125,24,29,226,40,3,0         // vbroadcastss  0x328e2(%rip),%ymm11        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  197,196,87,255                      // vxorps        %ymm7,%ymm7,%ymm7
   .byte  196,67,69,74,219,240                // vblendvps     %ymm15,%ymm11,%ymm7,%ymm11
   .byte  196,66,29,168,243                   // vfmadd213ps   %ymm11,%ymm12,%ymm14
   .byte  98,241,52,40,194,193,0              // vcmpeqps      %ymm1,%ymm9,%k0
   .byte  98,114,126,40,56,216                // vpmovm2d      %k0,%ymm11
   .byte  197,236,92,208                      // vsubps        %ymm0,%ymm2,%ymm2
-  .byte  196,98,125,24,61,13,18,3,0          // vbroadcastss  0x3120d(%rip),%ymm15        # 32c90 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  .byte  196,98,125,24,61,189,40,3,0         // vbroadcastss  0x328bd(%rip),%ymm15        # 34340 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   .byte  196,194,29,168,215                  // vfmadd213ps   %ymm15,%ymm12,%ymm2
   .byte  197,252,92,193                      // vsubps        %ymm1,%ymm0,%ymm0
-  .byte  98,242,29,56,168,5,254,17,3,0       // vfmadd213ps   0x311fe(%rip){1to8},%ymm12,%ymm0        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  98,242,29,56,168,5,174,40,3,0       // vfmadd213ps   0x328ae(%rip){1to8},%ymm12,%ymm0        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  196,227,125,74,194,176              // vblendvps     %ymm11,%ymm2,%ymm0,%ymm0
   .byte  196,195,125,74,198,208              // vblendvps     %ymm13,%ymm14,%ymm0,%ymm0
   .byte  196,227,125,74,199,128              // vblendvps     %ymm8,%ymm7,%ymm0,%ymm0
-  .byte  98,241,124,56,89,5,230,17,3,0       // vmulps        0x311e6(%rip){1to8},%ymm0,%ymm0        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  98,241,124,56,89,5,150,40,3,0       // vmulps        0x32896(%rip){1to8},%ymm0,%ymm0        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  196,193,52,88,202                   // vaddps        %ymm10,%ymm9,%ymm1
-  .byte  196,98,125,24,29,108,17,3,0         // vbroadcastss  0x3116c(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,28,40,3,0          // vbroadcastss  0x3281c(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,116,89,211                  // vmulps        %ymm11,%ymm1,%ymm2
   .byte  98,241,36,40,194,194,1              // vcmpltps      %ymm2,%ymm11,%k0
   .byte  98,114,126,40,56,216                // vpmovm2d      %k0,%ymm11
@@ -11072,7 +11072,7 @@
 _sk_hsl_to_rgb_skx:
   .byte  98,225,124,40,40,215                // vmovaps       %ymm7,%ymm18
   .byte  98,225,124,40,40,230                // vmovaps       %ymm6,%ymm20
-  .byte  98,226,125,40,24,5,30,17,3,0        // vbroadcastss  0x3111e(%rip),%ymm16        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,226,125,40,24,5,206,39,3,0       // vbroadcastss  0x327ce(%rip),%ymm16        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  98,241,124,32,194,194,2             // vcmpleps      %ymm2,%ymm16,%k0
   .byte  98,114,126,40,56,192                // vpmovm2d      %k0,%ymm8
   .byte  197,116,89,202                      // vmulps        %ymm2,%ymm1,%ymm9
@@ -11080,27 +11080,27 @@
   .byte  196,67,53,74,194,128                // vblendvps     %ymm8,%ymm10,%ymm9,%ymm8
   .byte  197,60,88,218                       // vaddps        %ymm2,%ymm8,%ymm11
   .byte  196,65,124,40,203                   // vmovaps       %ymm11,%ymm9
-  .byte  98,114,109,56,186,13,83,17,3,0      // vfmsub231ps   0x31153(%rip){1to8},%ymm2,%ymm9        # 32c90 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  .byte  98,114,109,56,186,13,3,40,3,0       // vfmsub231ps   0x32803(%rip){1to8},%ymm2,%ymm9        # 34340 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  98,209,116,40,194,192,0             // vcmpeqps      %ymm8,%ymm1,%k0
   .byte  98,242,126,40,56,248                // vpmovm2d      %k0,%ymm7
-  .byte  98,241,124,56,88,13,67,17,3,0       // vaddps        0x31143(%rip){1to8},%ymm0,%ymm1        # 32c9c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
+  .byte  98,241,124,56,88,13,243,39,3,0      // vaddps        0x327f3(%rip){1to8},%ymm0,%ymm1        # 3434c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
   .byte  196,99,125,8,193,1                  // vroundps      $0x1,%ymm1,%ymm8
   .byte  196,193,116,92,200                  // vsubps        %ymm8,%ymm1,%ymm1
-  .byte  98,226,125,40,24,29,50,17,3,0       // vbroadcastss  0x31132(%rip),%ymm19        # 32ca0 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
+  .byte  98,226,125,40,24,29,226,39,3,0      // vbroadcastss  0x327e2(%rip),%ymm19        # 34350 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
   .byte  98,241,100,32,194,193,2             // vcmpleps      %ymm1,%ymm19,%k0
   .byte  98,114,126,40,56,192                // vpmovm2d      %k0,%ymm8
   .byte  196,65,36,92,241                    // vsubps        %ymm9,%ymm11,%ymm14
-  .byte  196,98,125,24,61,3,17,3,0           // vbroadcastss  0x31103(%rip),%ymm15        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,98,125,24,61,179,39,3,0         // vbroadcastss  0x327b3(%rip),%ymm15        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  196,65,116,89,231                   // vmulps        %ymm15,%ymm1,%ymm12
-  .byte  98,226,125,40,24,13,252,16,3,0      // vbroadcastss  0x310fc(%rip),%ymm17        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  98,226,125,40,24,13,172,39,3,0      // vbroadcastss  0x327ac(%rip),%ymm17        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  98,81,116,32,92,212                 // vsubps        %ymm12,%ymm17,%ymm10
   .byte  196,66,13,168,209                   // vfmadd213ps   %ymm9,%ymm14,%ymm10
   .byte  196,67,45,74,193,128                // vblendvps     %ymm8,%ymm9,%ymm10,%ymm8
   .byte  98,241,124,32,194,193,2             // vcmpleps      %ymm1,%ymm16,%k0
   .byte  98,114,126,40,56,208                // vpmovm2d      %k0,%ymm10
   .byte  196,67,37,74,192,160                // vblendvps     %ymm10,%ymm8,%ymm11,%ymm8
-  .byte  196,98,125,24,21,211,16,3,0         // vbroadcastss  0x310d3(%rip),%ymm10        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  196,98,125,24,21,131,39,3,0         // vbroadcastss  0x32783(%rip),%ymm10        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  98,241,44,40,194,193,2              // vcmpleps      %ymm1,%ymm10,%k0
   .byte  98,242,126,40,56,200                // vpmovm2d      %k0,%ymm1
   .byte  196,66,13,168,225                   // vfmadd213ps   %ymm9,%ymm14,%ymm12
@@ -11122,7 +11122,7 @@
   .byte  196,66,13,168,233                   // vfmadd213ps   %ymm9,%ymm14,%ymm13
   .byte  196,195,21,74,200,16                // vblendvps     %ymm1,%ymm8,%ymm13,%ymm1
   .byte  196,227,117,74,202,112              // vblendvps     %ymm7,%ymm2,%ymm1,%ymm1
-  .byte  98,241,124,56,88,5,89,16,3,0        // vaddps        0x31059(%rip){1to8},%ymm0,%ymm0        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  98,241,124,56,88,5,9,39,3,0         // vaddps        0x32709(%rip){1to8},%ymm0,%ymm0        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,99,125,8,192,1                  // vroundps      $0x1,%ymm0,%ymm8
   .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
   .byte  98,241,100,32,194,192,2             // vcmpleps      %ymm0,%ymm19,%k0
@@ -11171,10 +11171,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,54                              // jne           1d27 <_sk_scale_u8_skx+0x4e>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,113,22,3,0             // vpand         0x31671(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,81,45,3,0              // vpand         0x32d51(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  98,113,60,56,89,5,149,15,3,0        // vmulps        0x30f95(%rip){1to8},%ymm8,%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,113,60,56,89,5,69,38,3,0         // vmulps        0x32645(%rip){1to8},%ymm8,%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
   .byte  197,188,89,210                      // vmulps        %ymm2,%ymm8,%ymm2
@@ -11251,15 +11251,15 @@
   .byte  15,133,145,0,0,0                    // jne           1ea0 <_sk_scale_565_skx+0xb0>
   .byte  196,1,122,111,4,88                  // vmovdqu       (%r8,%r11,2),%xmm8
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
-  .byte  98,113,61,56,219,13,136,14,3,0      // vpandd        0x30e88(%rip){1to8},%ymm8,%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  98,113,61,56,219,13,56,37,3,0       // vpandd        0x32538(%rip){1to8},%ymm8,%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  98,113,52,56,89,13,125,14,3,0       // vmulps        0x30e7d(%rip){1to8},%ymm9,%ymm9        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
-  .byte  98,113,61,56,219,21,119,14,3,0      // vpandd        0x30e77(%rip){1to8},%ymm8,%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  98,113,52,56,89,13,45,37,3,0        // vmulps        0x3252d(%rip){1to8},%ymm9,%ymm9        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  98,113,61,56,219,21,39,37,3,0       // vpandd        0x32527(%rip){1to8},%ymm8,%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  98,113,44,56,89,21,108,14,3,0       // vmulps        0x30e6c(%rip){1to8},%ymm10,%ymm10        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
-  .byte  98,113,61,56,219,5,102,14,3,0       // vpandd        0x30e66(%rip){1to8},%ymm8,%ymm8        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  98,113,44,56,89,21,28,37,3,0        // vmulps        0x3251c(%rip){1to8},%ymm10,%ymm10        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  98,113,61,56,219,5,22,37,3,0        // vpandd        0x32516(%rip){1to8},%ymm8,%ymm8        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  98,113,60,56,89,5,91,14,3,0         // vmulps        0x30e5b(%rip){1to8},%ymm8,%ymm8        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  98,113,60,56,89,5,11,37,3,0         // vmulps        0x3250b(%rip){1to8},%ymm8,%ymm8        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  98,241,100,40,194,199,1             // vcmpltps      %ymm7,%ymm3,%k0
   .byte  98,114,126,40,56,216                // vpmovm2d      %k0,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -11355,10 +11355,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,74                              // jne           1ff5 <_sk_lerp_u8_skx+0x62>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,183,19,3,0             // vpand         0x313b7(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,151,42,3,0             // vpand         0x32a97(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  98,113,60,56,89,5,219,12,3,0        // vmulps        0x30cdb(%rip){1to8},%ymm8,%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,113,60,56,89,5,139,35,3,0        // vmulps        0x3238b(%rip){1to8},%ymm8,%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,92,196                      // vsubps        %ymm4,%ymm0,%ymm0
   .byte  196,226,61,168,196                  // vfmadd213ps   %ymm4,%ymm8,%ymm0
   .byte  197,244,92,205                      // vsubps        %ymm5,%ymm1,%ymm1
@@ -11438,15 +11438,15 @@
   .byte  15,133,165,0,0,0                    // jne           2180 <_sk_lerp_565_skx+0xc4>
   .byte  196,1,122,111,4,88                  // vmovdqu       (%r8,%r11,2),%xmm8
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
-  .byte  98,113,61,56,219,13,188,11,3,0      // vpandd        0x30bbc(%rip){1to8},%ymm8,%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  98,113,61,56,219,13,108,34,3,0      // vpandd        0x3226c(%rip){1to8},%ymm8,%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  98,113,52,56,89,13,177,11,3,0       // vmulps        0x30bb1(%rip){1to8},%ymm9,%ymm9        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
-  .byte  98,113,61,56,219,21,171,11,3,0      // vpandd        0x30bab(%rip){1to8},%ymm8,%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  98,113,52,56,89,13,97,34,3,0        // vmulps        0x32261(%rip){1to8},%ymm9,%ymm9        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  98,113,61,56,219,21,91,34,3,0       // vpandd        0x3225b(%rip){1to8},%ymm8,%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  98,113,44,56,89,21,160,11,3,0       // vmulps        0x30ba0(%rip){1to8},%ymm10,%ymm10        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
-  .byte  98,113,61,56,219,5,154,11,3,0       // vpandd        0x30b9a(%rip){1to8},%ymm8,%ymm8        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  98,113,44,56,89,21,80,34,3,0        // vmulps        0x32250(%rip){1to8},%ymm10,%ymm10        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  98,113,61,56,219,5,74,34,3,0        // vpandd        0x3224a(%rip){1to8},%ymm8,%ymm8        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  98,113,60,56,89,5,143,11,3,0        // vmulps        0x30b8f(%rip){1to8},%ymm8,%ymm8        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  98,113,60,56,89,5,63,34,3,0         // vmulps        0x3223f(%rip){1to8},%ymm8,%ymm8        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  98,241,100,40,194,199,1             // vcmpltps      %ymm7,%ymm3,%k0
   .byte  98,114,126,40,56,216                // vpmovm2d      %k0,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -11546,7 +11546,7 @@
   .byte  196,162,61,146,20,136               // vgatherdps    %ymm8,(%rax,%ymm9,4),%ymm2
   .byte  197,229,114,211,24                  // vpsrld        $0x18,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  98,241,100,56,89,29,226,9,3,0       // vmulps        0x309e2(%rip){1to8},%ymm3,%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,100,56,89,29,146,32,3,0      // vmulps        0x32092(%rip){1to8},%ymm3,%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -11625,7 +11625,7 @@
   .byte  197,113,105,219                     // vpunpckhwd    %xmm3,%xmm1,%xmm11
   .byte  197,185,108,200                     // vpunpcklqdq   %xmm0,%xmm8,%xmm1
   .byte  197,57,109,192                      // vpunpckhqdq   %xmm0,%xmm8,%xmm8
-  .byte  197,121,111,21,117,15,3,0           // vmovdqa       0x30f75(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,121,111,21,85,38,3,0            // vmovdqa       0x32655(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,193,113,219,194                 // vpand         %xmm10,%xmm1,%xmm0
   .byte  196,226,125,51,200                  // vpmovzxwd     %xmm0,%ymm1
   .byte  76,139,64,8                         // mov           0x8(%rax),%r8
@@ -11651,7 +11651,7 @@
   .byte  197,185,235,219                     // vpor          %xmm3,%xmm8,%xmm3
   .byte  196,226,125,51,219                  // vpmovzxwd     %xmm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  98,241,100,56,89,29,72,8,3,0        // vmulps        0x30848(%rip){1to8},%ymm3,%ymm3        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  98,241,100,56,89,29,248,30,3,0      // vmulps        0x31ef8(%rip){1to8},%ymm3,%ymm3        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,1,123,16,4,72                   // vmovsd        (%r8,%r9,2),%xmm8
@@ -11708,7 +11708,7 @@
   .byte  197,249,105,193                     // vpunpckhwd    %xmm1,%xmm0,%xmm0
   .byte  197,105,108,200                     // vpunpcklqdq   %xmm0,%xmm2,%xmm9
   .byte  197,185,108,195                     // vpunpcklqdq   %xmm3,%xmm8,%xmm0
-  .byte  197,121,111,21,254,13,3,0           // vmovdqa       0x30dfe(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,121,111,21,222,36,3,0           // vmovdqa       0x324de(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,193,121,219,194                 // vpand         %xmm10,%xmm0,%xmm0
   .byte  196,226,125,51,200                  // vpmovzxwd     %xmm0,%ymm1
   .byte  76,139,64,8                         // mov           0x8(%rax),%r8
@@ -11729,7 +11729,7 @@
   .byte  196,226,125,51,219                  // vpmovzxwd     %xmm3,%ymm3
   .byte  196,226,29,146,20,152               // vgatherdps    %ymm12,(%rax,%ymm3,4),%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,89,6,3,0          // vbroadcastss  0x30659(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,9,29,3,0          // vbroadcastss  0x31d09(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,129,121,110,4,72                // vmovd         (%r8,%r9,2),%xmm0
   .byte  196,65,49,87,201                    // vxorpd        %xmm9,%xmm9,%xmm9
@@ -11798,7 +11798,7 @@
 FUNCTION(_sk_byte_tables_skx)
 _sk_byte_tables_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,78,5,3,0            // vbroadcastss  0x3054e(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,254,27,3,0          // vbroadcastss  0x31bfe(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,125,91,200                      // vcvtps2dq     %ymm0,%ymm9
   .byte  196,65,249,126,200                  // vmovq         %xmm9,%r8
@@ -11928,7 +11928,7 @@
   .byte  67,15,182,4,17                      // movzbl        (%r9,%r10,1),%eax
   .byte  196,194,125,49,193                  // vpmovzxbd     %xmm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,32,3,3,0            // vbroadcastss  0x30320(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,208,25,3,0          // vbroadcastss  0x319d0(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,226,125,49,201                  // vpmovzxbd     %xmm1,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
@@ -12049,7 +12049,7 @@
   .byte  67,15,182,4,17                      // movzbl        (%r9,%r10,1),%eax
   .byte  196,194,125,49,193                  // vpmovzxbd     %xmm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,247,0,3,0           // vbroadcastss  0x300f7(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,167,23,3,0          // vbroadcastss  0x317a7(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,194,125,49,202                  // vpmovzxbd     %xmm10,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
@@ -12146,33 +12146,33 @@
   .byte  98,114,125,56,168,80,2              // vfmadd213ps   0x8(%rax){1to8},%ymm0,%ymm10
   .byte  196,226,125,24,0                    // vbroadcastss  (%rax),%ymm0
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,235,255,2,0        // vbroadcastss  0x2ffeb(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  98,114,37,56,168,37,233,255,2,0     // vfmadd213ps   0x2ffe9(%rip){1to8},%ymm11,%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
-  .byte  98,113,44,56,84,21,219,255,2,0      // vandps        0x2ffdb(%rip){1to8},%ymm10,%ymm10        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
-  .byte  98,113,44,56,86,21,49,255,2,0       // vorps         0x2ff31(%rip){1to8},%ymm10,%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,45,56,188,37,207,255,2,0     // vfnmadd231ps  0x2ffcf(%rip){1to8},%ymm10,%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
-  .byte  98,113,44,56,88,21,201,255,2,0      // vaddps        0x2ffc9(%rip){1to8},%ymm10,%ymm10        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
-  .byte  196,98,125,24,29,196,255,2,0        // vbroadcastss  0x2ffc4(%rip),%ymm11        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,155,22,3,0         // vbroadcastss  0x3169b(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  98,114,37,56,168,37,153,22,3,0      // vfmadd213ps   0x31699(%rip){1to8},%ymm11,%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  98,113,44,56,84,21,139,22,3,0       // vandps        0x3168b(%rip){1to8},%ymm10,%ymm10        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  98,113,44,56,86,21,225,21,3,0       // vorps         0x315e1(%rip){1to8},%ymm10,%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,45,56,188,37,127,22,3,0      // vfnmadd231ps  0x3167f(%rip){1to8},%ymm10,%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  98,113,44,56,88,21,121,22,3,0       // vaddps        0x31679(%rip){1to8},%ymm10,%ymm10        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,29,116,22,3,0         // vbroadcastss  0x31674(%rip),%ymm11        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,65,28,92,210                    // vsubps        %ymm10,%ymm12,%ymm10
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  196,99,125,8,208,1                  // vroundps      $0x1,%ymm0,%ymm10
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  98,241,124,56,88,5,164,255,2,0      // vaddps        0x2ffa4(%rip){1to8},%ymm0,%ymm0        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
-  .byte  98,242,45,56,188,5,158,255,2,0      // vfnmadd231ps  0x2ff9e(%rip){1to8},%ymm10,%ymm0        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
-  .byte  196,98,125,24,29,153,255,2,0        // vbroadcastss  0x2ff99(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  98,241,124,56,88,5,84,22,3,0        // vaddps        0x31654(%rip){1to8},%ymm0,%ymm0        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  98,242,45,56,188,5,78,22,3,0        // vfnmadd231ps  0x3164e(%rip){1to8},%ymm10,%ymm0        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,73,22,3,0          // vbroadcastss  0x31649(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,143,255,2,0        // vbroadcastss  0x2ff8f(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,63,22,3,0          // vbroadcastss  0x3163f(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,124,88,194                  // vaddps        %ymm10,%ymm0,%ymm0
-  .byte  98,241,124,56,89,5,127,255,2,0      // vmulps        0x2ff7f(%rip){1to8},%ymm0,%ymm0        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  98,241,124,56,89,5,47,22,3,0        // vmulps        0x3162f(%rip){1to8},%ymm0,%ymm0        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
   .byte  197,172,88,192                      // vaddps        %ymm0,%ymm10,%ymm0
   .byte  196,195,125,74,193,128              // vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
-  .byte  98,241,124,56,93,5,151,254,2,0      // vminps        0x2fe97(%rip){1to8},%ymm0,%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,93,5,71,21,3,0        // vminps        0x31547(%rip){1to8},%ymm0,%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12189,33 +12189,33 @@
   .byte  98,114,117,56,168,80,2              // vfmadd213ps   0x8(%rax){1to8},%ymm1,%ymm10
   .byte  196,226,125,24,8                    // vbroadcastss  (%rax),%ymm1
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,238,254,2,0        // vbroadcastss  0x2feee(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  98,114,37,56,168,37,236,254,2,0     // vfmadd213ps   0x2feec(%rip){1to8},%ymm11,%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
-  .byte  98,113,44,56,84,21,222,254,2,0      // vandps        0x2fede(%rip){1to8},%ymm10,%ymm10        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
-  .byte  98,113,44,56,86,21,52,254,2,0       // vorps         0x2fe34(%rip){1to8},%ymm10,%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,45,56,188,37,210,254,2,0     // vfnmadd231ps  0x2fed2(%rip){1to8},%ymm10,%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
-  .byte  98,113,44,56,88,21,204,254,2,0      // vaddps        0x2fecc(%rip){1to8},%ymm10,%ymm10        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
-  .byte  196,98,125,24,29,199,254,2,0        // vbroadcastss  0x2fec7(%rip),%ymm11        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,158,21,3,0         // vbroadcastss  0x3159e(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  98,114,37,56,168,37,156,21,3,0      // vfmadd213ps   0x3159c(%rip){1to8},%ymm11,%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  98,113,44,56,84,21,142,21,3,0       // vandps        0x3158e(%rip){1to8},%ymm10,%ymm10        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  98,113,44,56,86,21,228,20,3,0       // vorps         0x314e4(%rip){1to8},%ymm10,%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,45,56,188,37,130,21,3,0      // vfnmadd231ps  0x31582(%rip){1to8},%ymm10,%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  98,113,44,56,88,21,124,21,3,0       // vaddps        0x3157c(%rip){1to8},%ymm10,%ymm10        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,29,119,21,3,0         // vbroadcastss  0x31577(%rip),%ymm11        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,65,28,92,210                    // vsubps        %ymm10,%ymm12,%ymm10
   .byte  196,193,116,89,202                  // vmulps        %ymm10,%ymm1,%ymm1
   .byte  196,99,125,8,209,1                  // vroundps      $0x1,%ymm1,%ymm10
   .byte  196,65,116,92,210                   // vsubps        %ymm10,%ymm1,%ymm10
-  .byte  98,241,116,56,88,13,167,254,2,0     // vaddps        0x2fea7(%rip){1to8},%ymm1,%ymm1        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
-  .byte  98,242,45,56,188,13,161,254,2,0     // vfnmadd231ps  0x2fea1(%rip){1to8},%ymm10,%ymm1        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
-  .byte  196,98,125,24,29,156,254,2,0        // vbroadcastss  0x2fe9c(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  98,241,116,56,88,13,87,21,3,0       // vaddps        0x31557(%rip){1to8},%ymm1,%ymm1        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  98,242,45,56,188,13,81,21,3,0       // vfnmadd231ps  0x31551(%rip){1to8},%ymm10,%ymm1        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,76,21,3,0          // vbroadcastss  0x3154c(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,146,254,2,0        // vbroadcastss  0x2fe92(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,66,21,3,0          // vbroadcastss  0x31542(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,116,88,202                  // vaddps        %ymm10,%ymm1,%ymm1
-  .byte  98,241,116,56,89,13,130,254,2,0     // vmulps        0x2fe82(%rip){1to8},%ymm1,%ymm1        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  98,241,116,56,89,13,50,21,3,0       // vmulps        0x31532(%rip){1to8},%ymm1,%ymm1        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
   .byte  197,172,88,201                      // vaddps        %ymm1,%ymm10,%ymm1
   .byte  196,195,117,74,201,128              // vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,116,95,200                  // vmaxps        %ymm8,%ymm1,%ymm1
-  .byte  98,241,116,56,93,13,154,253,2,0     // vminps        0x2fd9a(%rip){1to8},%ymm1,%ymm1        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,116,56,93,13,74,20,3,0       // vminps        0x3144a(%rip){1to8},%ymm1,%ymm1        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12232,33 +12232,33 @@
   .byte  98,114,109,56,168,80,2              // vfmadd213ps   0x8(%rax){1to8},%ymm2,%ymm10
   .byte  196,226,125,24,16                   // vbroadcastss  (%rax),%ymm2
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,241,253,2,0        // vbroadcastss  0x2fdf1(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  98,114,37,56,168,37,239,253,2,0     // vfmadd213ps   0x2fdef(%rip){1to8},%ymm11,%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
-  .byte  98,113,44,56,84,21,225,253,2,0      // vandps        0x2fde1(%rip){1to8},%ymm10,%ymm10        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
-  .byte  98,113,44,56,86,21,55,253,2,0       // vorps         0x2fd37(%rip){1to8},%ymm10,%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,45,56,188,37,213,253,2,0     // vfnmadd231ps  0x2fdd5(%rip){1to8},%ymm10,%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
-  .byte  98,113,44,56,88,21,207,253,2,0      // vaddps        0x2fdcf(%rip){1to8},%ymm10,%ymm10        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
-  .byte  196,98,125,24,29,202,253,2,0        // vbroadcastss  0x2fdca(%rip),%ymm11        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,161,20,3,0         // vbroadcastss  0x314a1(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  98,114,37,56,168,37,159,20,3,0      // vfmadd213ps   0x3149f(%rip){1to8},%ymm11,%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  98,113,44,56,84,21,145,20,3,0       // vandps        0x31491(%rip){1to8},%ymm10,%ymm10        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  98,113,44,56,86,21,231,19,3,0       // vorps         0x313e7(%rip){1to8},%ymm10,%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,45,56,188,37,133,20,3,0      // vfnmadd231ps  0x31485(%rip){1to8},%ymm10,%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  98,113,44,56,88,21,127,20,3,0       // vaddps        0x3147f(%rip){1to8},%ymm10,%ymm10        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,29,122,20,3,0         // vbroadcastss  0x3147a(%rip),%ymm11        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,65,28,92,210                    // vsubps        %ymm10,%ymm12,%ymm10
   .byte  196,193,108,89,210                  // vmulps        %ymm10,%ymm2,%ymm2
   .byte  196,99,125,8,210,1                  // vroundps      $0x1,%ymm2,%ymm10
   .byte  196,65,108,92,210                   // vsubps        %ymm10,%ymm2,%ymm10
-  .byte  98,241,108,56,88,21,170,253,2,0     // vaddps        0x2fdaa(%rip){1to8},%ymm2,%ymm2        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
-  .byte  98,242,45,56,188,21,164,253,2,0     // vfnmadd231ps  0x2fda4(%rip){1to8},%ymm10,%ymm2        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
-  .byte  196,98,125,24,29,159,253,2,0        // vbroadcastss  0x2fd9f(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  98,241,108,56,88,21,90,20,3,0       // vaddps        0x3145a(%rip){1to8},%ymm2,%ymm2        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  98,242,45,56,188,21,84,20,3,0       // vfnmadd231ps  0x31454(%rip){1to8},%ymm10,%ymm2        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,79,20,3,0          // vbroadcastss  0x3144f(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,149,253,2,0        // vbroadcastss  0x2fd95(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,69,20,3,0          // vbroadcastss  0x31445(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,108,88,210                  // vaddps        %ymm10,%ymm2,%ymm2
-  .byte  98,241,108,56,89,21,133,253,2,0     // vmulps        0x2fd85(%rip){1to8},%ymm2,%ymm2        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  98,241,108,56,89,21,53,20,3,0       // vmulps        0x31435(%rip){1to8},%ymm2,%ymm2        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,253,91,210                      // vcvtps2dq     %ymm2,%ymm2
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
   .byte  197,172,88,210                      // vaddps        %ymm2,%ymm10,%ymm2
   .byte  196,195,109,74,209,128              // vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
-  .byte  98,241,108,56,93,21,157,252,2,0     // vminps        0x2fc9d(%rip){1to8},%ymm2,%ymm2        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,108,56,93,21,77,19,3,0       // vminps        0x3134d(%rip){1to8},%ymm2,%ymm2        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12275,33 +12275,33 @@
   .byte  98,114,101,56,168,80,2              // vfmadd213ps   0x8(%rax){1to8},%ymm3,%ymm10
   .byte  196,226,125,24,24                   // vbroadcastss  (%rax),%ymm3
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,244,252,2,0        // vbroadcastss  0x2fcf4(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  98,114,37,56,168,37,242,252,2,0     // vfmadd213ps   0x2fcf2(%rip){1to8},%ymm11,%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
-  .byte  98,113,44,56,84,21,228,252,2,0      // vandps        0x2fce4(%rip){1to8},%ymm10,%ymm10        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
-  .byte  98,113,44,56,86,21,58,252,2,0       // vorps         0x2fc3a(%rip){1to8},%ymm10,%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,45,56,188,37,216,252,2,0     // vfnmadd231ps  0x2fcd8(%rip){1to8},%ymm10,%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
-  .byte  98,113,44,56,88,21,210,252,2,0      // vaddps        0x2fcd2(%rip){1to8},%ymm10,%ymm10        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
-  .byte  196,98,125,24,29,205,252,2,0        // vbroadcastss  0x2fccd(%rip),%ymm11        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,164,19,3,0         // vbroadcastss  0x313a4(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  98,114,37,56,168,37,162,19,3,0      // vfmadd213ps   0x313a2(%rip){1to8},%ymm11,%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  98,113,44,56,84,21,148,19,3,0       // vandps        0x31394(%rip){1to8},%ymm10,%ymm10        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  98,113,44,56,86,21,234,18,3,0       // vorps         0x312ea(%rip){1to8},%ymm10,%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,45,56,188,37,136,19,3,0      // vfnmadd231ps  0x31388(%rip){1to8},%ymm10,%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  98,113,44,56,88,21,130,19,3,0       // vaddps        0x31382(%rip){1to8},%ymm10,%ymm10        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,29,125,19,3,0         // vbroadcastss  0x3137d(%rip),%ymm11        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,65,28,92,210                    // vsubps        %ymm10,%ymm12,%ymm10
   .byte  196,193,100,89,218                  // vmulps        %ymm10,%ymm3,%ymm3
   .byte  196,99,125,8,211,1                  // vroundps      $0x1,%ymm3,%ymm10
   .byte  196,65,100,92,210                   // vsubps        %ymm10,%ymm3,%ymm10
-  .byte  98,241,100,56,88,29,173,252,2,0     // vaddps        0x2fcad(%rip){1to8},%ymm3,%ymm3        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
-  .byte  98,242,45,56,188,29,167,252,2,0     // vfnmadd231ps  0x2fca7(%rip){1to8},%ymm10,%ymm3        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
-  .byte  196,98,125,24,29,162,252,2,0        // vbroadcastss  0x2fca2(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  98,241,100,56,88,29,93,19,3,0       // vaddps        0x3135d(%rip){1to8},%ymm3,%ymm3        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  98,242,45,56,188,29,87,19,3,0       // vfnmadd231ps  0x31357(%rip){1to8},%ymm10,%ymm3        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,82,19,3,0          // vbroadcastss  0x31352(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,152,252,2,0        // vbroadcastss  0x2fc98(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,72,19,3,0          // vbroadcastss  0x31348(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,100,88,218                  // vaddps        %ymm10,%ymm3,%ymm3
-  .byte  98,241,100,56,89,29,136,252,2,0     // vmulps        0x2fc88(%rip){1to8},%ymm3,%ymm3        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  98,241,100,56,89,29,56,19,3,0       // vmulps        0x31338(%rip){1to8},%ymm3,%ymm3        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,253,91,219                      // vcvtps2dq     %ymm3,%ymm3
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
   .byte  197,172,88,219                      // vaddps        %ymm3,%ymm10,%ymm3
   .byte  196,195,101,74,217,128              // vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,100,95,216                  // vmaxps        %ymm8,%ymm3,%ymm3
-  .byte  98,241,100,56,93,29,160,251,2,0     // vminps        0x2fba0(%rip){1to8},%ymm3,%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,100,56,93,29,80,18,3,0       // vminps        0x31250(%rip){1to8},%ymm3,%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12312,33 +12312,33 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  98,226,125,40,24,48                 // vbroadcastss  (%rax),%ymm22
   .byte  197,124,91,200                      // vcvtdq2ps     %ymm0,%ymm9
-  .byte  196,98,125,24,21,31,252,2,0         // vbroadcastss  0x2fc1f(%rip),%ymm10        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,98,125,24,29,26,252,2,0         // vbroadcastss  0x2fc1a(%rip),%ymm11        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,21,207,18,3,0         // vbroadcastss  0x312cf(%rip),%ymm10        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,29,202,18,3,0         // vbroadcastss  0x312ca(%rip),%ymm11        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,124,84,195                  // vandps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,37,108,251,2,0        // vbroadcastss  0x2fb6c(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,28,18,3,0          // vbroadcastss  0x3121c(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,124,86,196                  // vorps         %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,45,2,252,2,0          // vbroadcastss  0x2fc02(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,178,18,3,0         // vbroadcastss  0x312b2(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,45,168,205                   // vfmadd213ps   %ymm13,%ymm10,%ymm9
-  .byte  196,98,125,24,53,248,251,2,0        // vbroadcastss  0x2fbf8(%rip),%ymm14        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,53,168,18,3,0         // vbroadcastss  0x312a8(%rip),%ymm14        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,125,188,206                  // vfnmadd231ps  %ymm14,%ymm0,%ymm9
-  .byte  196,98,125,24,61,238,251,2,0        // vbroadcastss  0x2fbee(%rip),%ymm15        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,61,158,18,3,0         // vbroadcastss  0x3129e(%rip),%ymm15        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,124,88,199                  // vaddps        %ymm15,%ymm0,%ymm0
-  .byte  98,226,125,40,24,5,227,251,2,0      // vbroadcastss  0x2fbe3(%rip),%ymm16        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  98,226,125,40,24,5,147,18,3,0       // vbroadcastss  0x31293(%rip),%ymm16        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  98,241,124,32,94,192                // vdivps        %ymm0,%ymm16,%ymm0
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
   .byte  98,177,124,40,89,198                // vmulps        %ymm22,%ymm0,%ymm0
   .byte  196,99,125,8,200,1                  // vroundps      $0x1,%ymm0,%ymm9
   .byte  196,65,124,92,201                   // vsubps        %ymm9,%ymm0,%ymm9
-  .byte  98,226,125,40,24,13,194,251,2,0     // vbroadcastss  0x2fbc2(%rip),%ymm17        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  98,226,125,40,24,13,114,18,3,0      // vbroadcastss  0x31272(%rip),%ymm17        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  98,177,124,40,88,193                // vaddps        %ymm17,%ymm0,%ymm0
-  .byte  98,226,125,40,24,21,182,251,2,0     // vbroadcastss  0x2fbb6(%rip),%ymm18        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  98,226,125,40,24,21,102,18,3,0      // vbroadcastss  0x31266(%rip),%ymm18        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  98,178,53,40,188,194                // vfnmadd231ps  %ymm18,%ymm9,%ymm0
-  .byte  98,226,125,40,24,29,170,251,2,0     // vbroadcastss  0x2fbaa(%rip),%ymm19        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  98,226,125,40,24,29,90,18,3,0       // vbroadcastss  0x3125a(%rip),%ymm19        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  98,81,100,32,92,201                 // vsubps        %ymm9,%ymm19,%ymm9
-  .byte  98,226,125,40,24,37,158,251,2,0     // vbroadcastss  0x2fb9e(%rip),%ymm20        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  98,226,125,40,24,37,78,18,3,0       // vbroadcastss  0x3124e(%rip),%ymm20        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  98,81,92,32,94,201                  // vdivps        %ymm9,%ymm20,%ymm9
   .byte  196,193,124,88,193                  // vaddps        %ymm9,%ymm0,%ymm0
-  .byte  196,98,125,24,13,142,251,2,0        // vbroadcastss  0x2fb8e(%rip),%ymm9        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,13,62,18,3,0          // vbroadcastss  0x3123e(%rip),%ymm9        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  98,225,124,40,91,233                // vcvtdq2ps     %ymm1,%ymm21
@@ -12384,23 +12384,23 @@
 .globl _sk_lab_to_xyz_skx
 FUNCTION(_sk_lab_to_xyz_skx)
 _sk_lab_to_xyz_skx:
-  .byte  196,98,125,24,5,186,250,2,0         // vbroadcastss  0x2faba(%rip),%ymm8        # 32cf4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
-  .byte  196,98,125,24,13,25,250,2,0         // vbroadcastss  0x2fa19(%rip),%ymm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  196,98,125,24,21,172,250,2,0        // vbroadcastss  0x2faac(%rip),%ymm10        # 32cf8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
+  .byte  196,98,125,24,5,106,17,3,0          // vbroadcastss  0x3116a(%rip),%ymm8        # 343a4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
+  .byte  196,98,125,24,13,201,16,3,0         // vbroadcastss  0x310c9(%rip),%ymm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,21,92,17,3,0          // vbroadcastss  0x3115c(%rip),%ymm10        # 343a8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
   .byte  196,194,53,168,202                  // vfmadd213ps   %ymm10,%ymm9,%ymm1
   .byte  196,194,53,168,210                  // vfmadd213ps   %ymm10,%ymm9,%ymm2
-  .byte  98,114,125,56,168,5,156,250,2,0     // vfmadd213ps   0x2fa9c(%rip){1to8},%ymm0,%ymm8        # 32cfc <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
-  .byte  98,241,60,56,89,5,150,250,2,0       // vmulps        0x2fa96(%rip){1to8},%ymm8,%ymm0        # 32d00 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
-  .byte  98,242,125,56,152,13,144,250,2,0    // vfmadd132ps   0x2fa90(%rip){1to8},%ymm0,%ymm1        # 32d04 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
-  .byte  98,242,125,56,156,21,138,250,2,0    // vfnmadd132ps  0x2fa8a(%rip){1to8},%ymm0,%ymm2        # 32d08 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
+  .byte  98,114,125,56,168,5,76,17,3,0       // vfmadd213ps   0x3114c(%rip){1to8},%ymm0,%ymm8        # 343ac <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
+  .byte  98,241,60,56,89,5,70,17,3,0         // vmulps        0x31146(%rip){1to8},%ymm8,%ymm0        # 343b0 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
+  .byte  98,242,125,56,152,13,64,17,3,0      // vfmadd132ps   0x31140(%rip){1to8},%ymm0,%ymm1        # 343b4 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
+  .byte  98,242,125,56,156,21,58,17,3,0      // vfnmadd132ps  0x3113a(%rip){1to8},%ymm0,%ymm2        # 343b8 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
   .byte  197,116,89,193                      // vmulps        %ymm1,%ymm1,%ymm8
   .byte  196,65,116,89,192                   // vmulps        %ymm8,%ymm1,%ymm8
-  .byte  196,98,125,24,13,124,250,2,0        // vbroadcastss  0x2fa7c(%rip),%ymm9        # 32d0c <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
+  .byte  196,98,125,24,13,44,17,3,0          // vbroadcastss  0x3112c(%rip),%ymm9        # 343bc <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
   .byte  98,209,52,40,194,192,1              // vcmpltps      %ymm8,%ymm9,%k0
   .byte  98,114,126,40,56,208                // vpmovm2d      %k0,%ymm10
-  .byte  196,98,125,24,29,106,250,2,0        // vbroadcastss  0x2fa6a(%rip),%ymm11        # 32d10 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
+  .byte  196,98,125,24,29,26,17,3,0          // vbroadcastss  0x3111a(%rip),%ymm11        # 343c0 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,37,96,250,2,0         // vbroadcastss  0x2fa60(%rip),%ymm12        # 32d14 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
+  .byte  196,98,125,24,37,16,17,3,0          // vbroadcastss  0x31110(%rip),%ymm12        # 343c4 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
   .byte  196,193,116,89,204                  // vmulps        %ymm12,%ymm1,%ymm1
   .byte  196,67,117,74,192,160               // vblendvps     %ymm10,%ymm8,%ymm1,%ymm8
   .byte  197,252,89,200                      // vmulps        %ymm0,%ymm0,%ymm1
@@ -12417,8 +12417,8 @@
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,193,108,89,212                  // vmulps        %ymm12,%ymm2,%ymm2
   .byte  196,227,109,74,208,144              // vblendvps     %ymm9,%ymm0,%ymm2,%ymm2
-  .byte  98,241,60,56,89,5,5,250,2,0         // vmulps        0x2fa05(%rip){1to8},%ymm8,%ymm0        # 32d18 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
-  .byte  98,241,108,56,89,21,255,249,2,0     // vmulps        0x2f9ff(%rip){1to8},%ymm2,%ymm2        # 32d1c <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
+  .byte  98,241,60,56,89,5,181,16,3,0        // vmulps        0x310b5(%rip){1to8},%ymm8,%ymm0        # 343c8 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
+  .byte  98,241,108,56,89,21,175,16,3,0      // vmulps        0x310af(%rip){1to8},%ymm2,%ymm2        # 343cc <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -12435,10 +12435,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,49                              // jne           336a <_sk_load_a8_skx+0x49>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,41,0,3,0              // vpand         0x30029(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,9,23,3,0              // vpand         0x31709(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,29,78,249,2,0      // vmulps        0x2f94e(%rip){1to8},%ymm0,%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,124,56,89,29,254,15,3,0      // vmulps        0x30ffe(%rip){1to8},%ymm0,%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -12512,10 +12512,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,49                              // jne           3475 <_sk_load_a8_dst_skx+0x49>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,30,255,2,0           // vpand         0x2ff1e(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,254,21,3,0           // vpand         0x315fe(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,228                  // vpmovzxwd     %xmm4,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  98,241,92,56,89,61,67,248,2,0       // vmulps        0x2f843(%rip){1to8},%ymm4,%ymm7        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,92,56,89,61,243,14,3,0       // vmulps        0x30ef3(%rip){1to8},%ymm4,%ymm7        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -12631,7 +12631,7 @@
   .byte  196,227,121,32,192,7                // vpinsrb       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,29,134,246,2,0     // vmulps        0x2f686(%rip){1to8},%ymm0,%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,124,56,89,29,54,13,3,0       // vmulps        0x30d36(%rip){1to8},%ymm0,%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -12652,7 +12652,7 @@
   .byte  77,15,175,193                       // imul          %r9,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  98,113,100,56,89,5,6,246,2,0        // vmulps        0x2f606(%rip){1to8},%ymm3,%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  98,113,100,56,89,5,182,12,3,0       // vmulps        0x30cb6(%rip){1to8},%ymm3,%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,125,91,192                   // vcvtps2dq     %ymm8,%ymm8
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,66,57,43,193                    // vpackusdw     %xmm9,%xmm8,%xmm8
@@ -12719,12 +12719,12 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,54                              // jne           3752 <_sk_load_g8_skx+0x4e>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,70,252,2,0            // vpand         0x2fc46(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,38,19,3,0             // vpand         0x31326(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,107,245,2,0      // vmulps        0x2f56b(%rip){1to8},%ymm0,%ymm0        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,124,56,89,5,27,12,3,0        // vmulps        0x30c1b(%rip){1to8},%ymm0,%ymm0        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,232,244,2,0       // vbroadcastss  0x2f4e8(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,152,11,3,0        // vbroadcastss  0x30b98(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  255,224                             // jmpq          *%rax
@@ -12796,12 +12796,12 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,54                              // jne           3862 <_sk_load_g8_dst_skx+0x4e>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,54,251,2,0           // vpand         0x2fb36(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,22,18,3,0            // vpand         0x31216(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,228                  // vpmovzxwd     %xmm4,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  98,241,92,56,89,37,91,244,2,0       // vmulps        0x2f45b(%rip){1to8},%ymm4,%ymm4        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,92,56,89,37,11,11,3,0        // vmulps        0x30b0b(%rip){1to8},%ymm4,%ymm4        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,216,243,2,0       // vbroadcastss  0x2f3d8(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,136,10,3,0        // vbroadcastss  0x30a88(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,236                      // vmovaps       %ymm4,%ymm5
   .byte  197,252,40,244                      // vmovaps       %ymm4,%ymm6
   .byte  255,224                             // jmpq          *%rax
@@ -12915,9 +12915,9 @@
   .byte  196,227,121,32,192,7                // vpinsrb       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,150,242,2,0      // vmulps        0x2f296(%rip){1to8},%ymm0,%ymm0        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  98,241,124,56,89,5,70,9,3,0         // vmulps        0x30946(%rip){1to8},%ymm0,%ymm0        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,19,242,2,0        // vbroadcastss  0x2f213(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,195,8,3,0         // vbroadcastss  0x308c3(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -12941,17 +12941,17 @@
   .byte  117,96                              // jne           3aa9 <_sk_load_565_skx+0x7b>
   .byte  196,129,122,111,4,88                // vmovdqu       (%r8,%r11,2),%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  98,241,109,56,219,5,78,242,2,0      // vpandd        0x2f24e(%rip){1to8},%ymm2,%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  98,241,109,56,219,5,254,8,3,0       // vpandd        0x308fe(%rip){1to8},%ymm2,%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,68,242,2,0       // vmulps        0x2f244(%rip){1to8},%ymm0,%ymm0        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
-  .byte  98,241,109,56,219,13,62,242,2,0     // vpandd        0x2f23e(%rip){1to8},%ymm2,%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  98,241,124,56,89,5,244,8,3,0        // vmulps        0x308f4(%rip){1to8},%ymm0,%ymm0        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  98,241,109,56,219,13,238,8,3,0      // vpandd        0x308ee(%rip){1to8},%ymm2,%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  98,241,116,56,89,13,52,242,2,0      // vmulps        0x2f234(%rip){1to8},%ymm1,%ymm1        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
-  .byte  98,241,109,56,219,21,46,242,2,0     // vpandd        0x2f22e(%rip){1to8},%ymm2,%ymm2        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  98,241,116,56,89,13,228,8,3,0       // vmulps        0x308e4(%rip){1to8},%ymm1,%ymm1        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  98,241,109,56,219,21,222,8,3,0      // vpandd        0x308de(%rip){1to8},%ymm2,%ymm2        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  98,241,108,56,89,21,36,242,2,0      // vmulps        0x2f224(%rip){1to8},%ymm2,%ymm2        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  98,241,108,56,89,21,212,8,3,0       // vmulps        0x308d4(%rip){1to8},%ymm2,%ymm2        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,137,241,2,0       // vbroadcastss  0x2f189(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,57,8,3,0          // vbroadcastss  0x30839(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -13019,17 +13019,17 @@
   .byte  117,96                              // jne           3bdb <_sk_load_565_dst_skx+0x7b>
   .byte  196,129,122,111,36,88               // vmovdqu       (%r8,%r11,2),%xmm4
   .byte  196,226,125,51,244                  // vpmovzxwd     %xmm4,%ymm6
-  .byte  98,241,77,56,219,37,28,241,2,0      // vpandd        0x2f11c(%rip){1to8},%ymm6,%ymm4        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  98,241,77,56,219,37,204,7,3,0       // vpandd        0x307cc(%rip){1to8},%ymm6,%ymm4        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  98,241,92,56,89,37,18,241,2,0       // vmulps        0x2f112(%rip){1to8},%ymm4,%ymm4        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
-  .byte  98,241,77,56,219,45,12,241,2,0      // vpandd        0x2f10c(%rip){1to8},%ymm6,%ymm5        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  98,241,92,56,89,37,194,7,3,0        // vmulps        0x307c2(%rip){1to8},%ymm4,%ymm4        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  98,241,77,56,219,45,188,7,3,0       // vpandd        0x307bc(%rip){1to8},%ymm6,%ymm5        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  98,241,84,56,89,45,2,241,2,0        // vmulps        0x2f102(%rip){1to8},%ymm5,%ymm5        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
-  .byte  98,241,77,56,219,53,252,240,2,0     // vpandd        0x2f0fc(%rip){1to8},%ymm6,%ymm6        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  98,241,84,56,89,45,178,7,3,0        // vmulps        0x307b2(%rip){1to8},%ymm5,%ymm5        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  98,241,77,56,219,53,172,7,3,0       // vpandd        0x307ac(%rip){1to8},%ymm6,%ymm6        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  98,241,76,56,89,53,242,240,2,0      // vmulps        0x2f0f2(%rip){1to8},%ymm6,%ymm6        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  98,241,76,56,89,53,162,7,3,0        // vmulps        0x307a2(%rip){1to8},%ymm6,%ymm6        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,87,240,2,0        // vbroadcastss  0x2f057(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,7,7,3,0           // vbroadcastss  0x30707(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -13137,17 +13137,17 @@
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  98,241,109,56,219,5,55,239,2,0      // vpandd        0x2ef37(%rip){1to8},%ymm2,%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  98,241,109,56,219,5,231,5,3,0       // vpandd        0x305e7(%rip){1to8},%ymm2,%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,45,239,2,0       // vmulps        0x2ef2d(%rip){1to8},%ymm0,%ymm0        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
-  .byte  98,241,109,56,219,13,39,239,2,0     // vpandd        0x2ef27(%rip){1to8},%ymm2,%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  98,241,124,56,89,5,221,5,3,0        // vmulps        0x305dd(%rip){1to8},%ymm0,%ymm0        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  98,241,109,56,219,13,215,5,3,0      // vpandd        0x305d7(%rip){1to8},%ymm2,%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  98,241,116,56,89,13,29,239,2,0      // vmulps        0x2ef1d(%rip){1to8},%ymm1,%ymm1        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
-  .byte  98,241,109,56,219,21,23,239,2,0     // vpandd        0x2ef17(%rip){1to8},%ymm2,%ymm2        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  98,241,116,56,89,13,205,5,3,0       // vmulps        0x305cd(%rip){1to8},%ymm1,%ymm1        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  98,241,109,56,219,21,199,5,3,0      // vpandd        0x305c7(%rip){1to8},%ymm2,%ymm2        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  98,241,108,56,89,21,13,239,2,0      // vmulps        0x2ef0d(%rip){1to8},%ymm2,%ymm2        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  98,241,108,56,89,21,189,5,3,0       // vmulps        0x305bd(%rip){1to8},%ymm2,%ymm2        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,114,238,2,0       // vbroadcastss  0x2ee72(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,34,5,3,0          // vbroadcastss  0x30522(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -13165,11 +13165,11 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,58,239,2,0          // vbroadcastss  0x2ef3a(%rip),%ymm8        # 32d20 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
+  .byte  196,98,125,24,5,234,5,3,0           // vbroadcastss  0x305ea(%rip),%ymm8        # 343d0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,53,114,241,11               // vpslld        $0xb,%ymm9,%ymm9
-  .byte  98,113,116,56,89,21,36,239,2,0      // vmulps        0x2ef24(%rip){1to8},%ymm1,%ymm10        # 32d24 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
+  .byte  98,113,116,56,89,21,212,5,3,0       // vmulps        0x305d4(%rip){1to8},%ymm1,%ymm10        # 343d4 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
   .byte  196,65,125,91,210                   // vcvtps2dq     %ymm10,%ymm10
   .byte  196,193,45,114,242,5                // vpslld        $0x5,%ymm10,%ymm10
   .byte  196,65,45,235,201                   // vpor          %ymm9,%ymm10,%ymm9
@@ -13216,7 +13216,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,225                 // jmpq          ffffffffe2003eb8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe1fd1538>
+  .byte  233,255,255,255,225                 // jmpq          ffffffffe2003eb8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe1fcfe88>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
@@ -13239,18 +13239,18 @@
   .byte  117,111                             // jne           3f4a <_sk_load_4444_skx+0x8a>
   .byte  196,129,122,111,4,88                // vmovdqu       (%r8,%r11,2),%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  98,241,101,56,219,5,56,238,2,0      // vpandd        0x2ee38(%rip){1to8},%ymm3,%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  98,241,101,56,219,5,232,4,3,0       // vpandd        0x304e8(%rip){1to8},%ymm3,%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,46,238,2,0       // vmulps        0x2ee2e(%rip){1to8},%ymm0,%ymm0        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
-  .byte  98,241,101,56,219,13,40,238,2,0     // vpandd        0x2ee28(%rip){1to8},%ymm3,%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  98,241,124,56,89,5,222,4,3,0        // vmulps        0x304de(%rip){1to8},%ymm0,%ymm0        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  98,241,101,56,219,13,216,4,3,0      // vpandd        0x304d8(%rip){1to8},%ymm3,%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  98,241,116,56,89,13,30,238,2,0      // vmulps        0x2ee1e(%rip){1to8},%ymm1,%ymm1        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
-  .byte  98,241,101,56,219,21,24,238,2,0     // vpandd        0x2ee18(%rip){1to8},%ymm3,%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  98,241,116,56,89,13,206,4,3,0       // vmulps        0x304ce(%rip){1to8},%ymm1,%ymm1        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  98,241,101,56,219,21,200,4,3,0      // vpandd        0x304c8(%rip){1to8},%ymm3,%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  98,241,108,56,89,21,14,238,2,0      // vmulps        0x2ee0e(%rip){1to8},%ymm2,%ymm2        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
-  .byte  98,241,101,56,219,29,8,238,2,0      // vpandd        0x2ee08(%rip){1to8},%ymm3,%ymm3        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  98,241,108,56,89,21,190,4,3,0       // vmulps        0x304be(%rip){1to8},%ymm2,%ymm2        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  98,241,101,56,219,29,184,4,3,0      // vpandd        0x304b8(%rip){1to8},%ymm3,%ymm3        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  98,241,100,56,89,29,254,237,2,0     // vmulps        0x2edfe(%rip){1to8},%ymm3,%ymm3        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  98,241,100,56,89,29,174,4,3,0       // vmulps        0x304ae(%rip){1to8},%ymm3,%ymm3        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -13320,18 +13320,18 @@
   .byte  117,111                             // jne           408a <_sk_load_4444_dst_skx+0x8a>
   .byte  196,129,122,111,36,88               // vmovdqu       (%r8,%r11,2),%xmm4
   .byte  196,226,125,51,252                  // vpmovzxwd     %xmm4,%ymm7
-  .byte  98,241,69,56,219,37,248,236,2,0     // vpandd        0x2ecf8(%rip){1to8},%ymm7,%ymm4        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  98,241,69,56,219,37,168,3,3,0       // vpandd        0x303a8(%rip){1to8},%ymm7,%ymm4        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  98,241,92,56,89,37,238,236,2,0      // vmulps        0x2ecee(%rip){1to8},%ymm4,%ymm4        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
-  .byte  98,241,69,56,219,45,232,236,2,0     // vpandd        0x2ece8(%rip){1to8},%ymm7,%ymm5        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  98,241,92,56,89,37,158,3,3,0        // vmulps        0x3039e(%rip){1to8},%ymm4,%ymm4        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  98,241,69,56,219,45,152,3,3,0       // vpandd        0x30398(%rip){1to8},%ymm7,%ymm5        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  98,241,84,56,89,45,222,236,2,0      // vmulps        0x2ecde(%rip){1to8},%ymm5,%ymm5        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
-  .byte  98,241,69,56,219,53,216,236,2,0     // vpandd        0x2ecd8(%rip){1to8},%ymm7,%ymm6        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  98,241,84,56,89,45,142,3,3,0        // vmulps        0x3038e(%rip){1to8},%ymm5,%ymm5        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  98,241,69,56,219,53,136,3,3,0       // vpandd        0x30388(%rip){1to8},%ymm7,%ymm6        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  98,241,76,56,89,53,206,236,2,0      // vmulps        0x2ecce(%rip){1to8},%ymm6,%ymm6        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
-  .byte  98,241,69,56,219,61,200,236,2,0     // vpandd        0x2ecc8(%rip){1to8},%ymm7,%ymm7        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  98,241,76,56,89,53,126,3,3,0        // vmulps        0x3037e(%rip){1to8},%ymm6,%ymm6        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  98,241,69,56,219,61,120,3,3,0       // vpandd        0x30378(%rip){1to8},%ymm7,%ymm7        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  98,241,68,56,89,61,190,236,2,0      // vmulps        0x2ecbe(%rip){1to8},%ymm7,%ymm7        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  98,241,68,56,89,61,110,3,3,0        // vmulps        0x3036e(%rip){1to8},%ymm7,%ymm7        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -13440,18 +13440,18 @@
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  98,241,101,56,219,5,3,235,2,0       // vpandd        0x2eb03(%rip){1to8},%ymm3,%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  98,241,101,56,219,5,179,1,3,0       // vpandd        0x301b3(%rip){1to8},%ymm3,%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  98,241,124,56,89,5,249,234,2,0      // vmulps        0x2eaf9(%rip){1to8},%ymm0,%ymm0        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
-  .byte  98,241,101,56,219,13,243,234,2,0    // vpandd        0x2eaf3(%rip){1to8},%ymm3,%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  98,241,124,56,89,5,169,1,3,0        // vmulps        0x301a9(%rip){1to8},%ymm0,%ymm0        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  98,241,101,56,219,13,163,1,3,0      // vpandd        0x301a3(%rip){1to8},%ymm3,%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  98,241,116,56,89,13,233,234,2,0     // vmulps        0x2eae9(%rip){1to8},%ymm1,%ymm1        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
-  .byte  98,241,101,56,219,21,227,234,2,0    // vpandd        0x2eae3(%rip){1to8},%ymm3,%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  98,241,116,56,89,13,153,1,3,0       // vmulps        0x30199(%rip){1to8},%ymm1,%ymm1        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  98,241,101,56,219,21,147,1,3,0      // vpandd        0x30193(%rip){1to8},%ymm3,%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  98,241,108,56,89,21,217,234,2,0     // vmulps        0x2ead9(%rip){1to8},%ymm2,%ymm2        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
-  .byte  98,241,101,56,219,29,211,234,2,0    // vpandd        0x2ead3(%rip){1to8},%ymm3,%ymm3        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  98,241,108,56,89,21,137,1,3,0       // vmulps        0x30189(%rip){1to8},%ymm2,%ymm2        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  98,241,101,56,219,29,131,1,3,0      // vpandd        0x30183(%rip){1to8},%ymm3,%ymm3        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  98,241,100,56,89,29,201,234,2,0     // vmulps        0x2eac9(%rip){1to8},%ymm3,%ymm3        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  98,241,100,56,89,29,121,1,3,0       // vmulps        0x30179(%rip){1to8},%ymm3,%ymm3        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
@@ -13470,7 +13470,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,163,234,2,0         // vbroadcastss  0x2eaa3(%rip),%ymm8        # 32d48 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
+  .byte  196,98,125,24,5,83,1,3,0            // vbroadcastss  0x30153(%rip),%ymm8        # 343f8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,53,114,241,12               // vpslld        $0xc,%ymm9,%ymm9
@@ -13525,7 +13525,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,224                 // callq         ffffffffe1004388 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd1a08>
+  .byte  232,255,255,255,224                 // callq         ffffffffe1004388 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd0358>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
@@ -13551,7 +13551,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,241,127,169,111,195              // vmovdqu8      %ymm3,%ymm0{%k1}{z}
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,218,232,2,0         // vbroadcastss  0x2e8da(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,138,255,2,0         // vbroadcastss  0x2ff8a(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,245,114,211,8                   // vpsrld        $0x8,%ymm3,%ymm1
   .byte  98,241,127,169,111,201              // vmovdqu8      %ymm1,%ymm1{%k1}{z}
@@ -13638,7 +13638,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,241,127,169,111,231              // vmovdqu8      %ymm7,%ymm4{%k1}{z}
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,98,125,24,5,150,231,2,0         // vbroadcastss  0x2e796(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,70,254,2,0          // vbroadcastss  0x2fe46(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,92,89,224                   // vmulps        %ymm8,%ymm4,%ymm4
   .byte  197,213,114,215,8                   // vpsrld        $0x8,%ymm7,%ymm5
   .byte  98,241,127,169,111,237              // vmovdqu8      %ymm5,%ymm5{%k1}{z}
@@ -13732,7 +13732,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,209,127,169,111,192              // vmovdqu8      %ymm8,%ymm0{%k1}{z}
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,29,41,230,2,0        // vbroadcastss  0x2e629(%rip),%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,29,217,252,2,0       // vbroadcastss  0x2fcd9(%rip),%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
   .byte  196,193,117,114,208,8               // vpsrld        $0x8,%ymm8,%ymm1
   .byte  98,241,127,169,111,201              // vmovdqu8      %ymm1,%ymm1{%k1}{z}
@@ -13759,7 +13759,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,126,229,2,0         // vbroadcastss  0x2e57e(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,46,252,2,0          // vbroadcastss  0x2fc2e(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -13815,7 +13815,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,219                 // jmpq          ffffffffdc0047c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd1e40>
+  .byte  233,255,255,255,219                 // jmpq          ffffffffdc0047c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd0790>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,205                             // dec           %ebp
@@ -13841,7 +13841,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,241,127,169,111,195              // vmovdqu8      %ymm3,%ymm0{%k1}{z}
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,162,228,2,0         // vbroadcastss  0x2e4a2(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,82,251,2,0          // vbroadcastss  0x2fb52(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,208                  // vmulps        %ymm8,%ymm0,%ymm2
   .byte  197,253,114,211,8                   // vpsrld        $0x8,%ymm3,%ymm0
   .byte  98,241,127,169,111,192              // vmovdqu8      %ymm0,%ymm0{%k1}{z}
@@ -13928,7 +13928,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,241,127,169,111,231              // vmovdqu8      %ymm7,%ymm4{%k1}{z}
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,98,125,24,5,94,227,2,0          // vbroadcastss  0x2e35e(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,14,250,2,0          // vbroadcastss  0x2fa0e(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,92,89,240                   // vmulps        %ymm8,%ymm4,%ymm6
   .byte  197,221,114,215,8                   // vpsrld        $0x8,%ymm7,%ymm4
   .byte  98,241,127,169,111,228              // vmovdqu8      %ymm4,%ymm4{%k1}{z}
@@ -14022,7 +14022,7 @@
   .byte  197,251,146,200                     // kmovd         %eax,%k1
   .byte  98,209,127,169,111,192              // vmovdqu8      %ymm8,%ymm0{%k1}{z}
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,29,241,225,2,0       // vbroadcastss  0x2e1f1(%rip),%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,29,161,248,2,0       // vbroadcastss  0x2f8a1(%rip),%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,211                      // vmulps        %ymm3,%ymm0,%ymm2
   .byte  196,193,125,114,208,8               // vpsrld        $0x8,%ymm8,%ymm0
   .byte  98,241,127,169,111,192              // vmovdqu8      %ymm0,%ymm0{%k1}{z}
@@ -14049,7 +14049,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,70,225,2,0          // vbroadcastss  0x2e146(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,246,247,2,0         // vbroadcastss  0x2f7f6(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,108,89,200                   // vmulps        %ymm8,%ymm2,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -14105,7 +14105,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,219                 // jmpq          ffffffffdc004bf8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd2278>
+  .byte  233,255,255,255,219                 // jmpq          ffffffffdc004bf8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd0bc8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,205                             // dec           %ebp
@@ -14368,7 +14368,7 @@
   .byte  197,241,235,192                     // vpor          %xmm0,%xmm1,%xmm0
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,21,175,220,2,0        // vbroadcastss  0x2dcaf(%rip),%ymm10        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,21,95,243,2,0         // vbroadcastss  0x2f35f(%rip),%ymm10        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  197,185,109,202                     // vpunpckhqdq   %xmm2,%xmm8,%xmm1
   .byte  197,233,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm2
@@ -14459,7 +14459,7 @@
   .byte  197,241,235,192                     // vpor          %xmm0,%xmm1,%xmm0
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,33,219,2,0         // vbroadcastss  0x2db21(%rip),%ymm9        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,13,209,241,2,0        // vbroadcastss  0x2f1d1(%rip),%ymm9        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,185,109,203                     // vpunpckhqdq   %xmm3,%xmm8,%xmm1
   .byte  197,225,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm3
@@ -14475,7 +14475,7 @@
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  196,193,108,89,209                  // vmulps        %ymm9,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,65,218,2,0        // vbroadcastss  0x2da41(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,241,240,2,0       // vbroadcastss  0x2f0f1(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,4,64                // vmovd         (%r8,%rax,2),%xmm0
   .byte  196,65,49,87,201                    // vxorpd        %xmm9,%xmm9,%xmm9
@@ -14551,7 +14551,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  73,99,193                           // movslq        %r9d,%rax
-  .byte  196,98,125,24,5,10,218,2,0          // vbroadcastss  0x2da0a(%rip),%ymm8        # 32d4c <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
+  .byte  196,98,125,24,5,186,240,2,0         // vbroadcastss  0x2f0ba(%rip),%ymm8        # 343fc <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
@@ -14893,7 +14893,7 @@
   .byte  196,65,124,92,218                   // vsubps        %ymm10,%ymm0,%ymm11
   .byte  196,193,58,88,192                   // vaddss        %xmm8,%xmm8,%xmm0
   .byte  196,98,125,24,192                   // vbroadcastss  %xmm0,%ymm8
-  .byte  197,178,89,5,119,211,2,0            // vmulss        0x2d377(%rip),%xmm9,%xmm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,5,39,234,2,0             // vmulss        0x2ea27(%rip),%xmm9,%xmm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,89,192                      // vmulps        %ymm0,%ymm11,%ymm0
   .byte  196,227,125,8,192,1                 // vroundps      $0x1,%ymm0,%ymm0
@@ -14916,7 +14916,7 @@
   .byte  196,65,116,92,218                   // vsubps        %ymm10,%ymm1,%ymm11
   .byte  196,193,58,88,200                   // vaddss        %xmm8,%xmm8,%xmm1
   .byte  196,98,125,24,193                   // vbroadcastss  %xmm1,%ymm8
-  .byte  197,178,89,13,38,211,2,0            // vmulss        0x2d326(%rip),%xmm9,%xmm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,13,214,233,2,0           // vmulss        0x2e9d6(%rip),%xmm9,%xmm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,226,125,24,201                  // vbroadcastss  %xmm1,%ymm1
   .byte  197,164,89,201                      // vmulps        %ymm1,%ymm11,%ymm1
   .byte  196,227,125,8,201,1                 // vroundps      $0x1,%ymm1,%ymm1
@@ -14934,7 +14934,7 @@
 _sk_clamp_x_1_skx:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  98,241,124,56,93,5,237,210,2,0      // vminps        0x2d2ed(%rip){1to8},%ymm0,%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,93,5,157,233,2,0      // vminps        0x2e99d(%rip){1to8},%ymm0,%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -14946,7 +14946,7 @@
   .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  98,241,124,56,93,5,203,210,2,0      // vminps        0x2d2cb(%rip){1to8},%ymm0,%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,93,5,123,233,2,0      // vminps        0x2e97b(%rip){1to8},%ymm0,%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -14954,9 +14954,9 @@
 .globl _sk_mirror_x_1_skx
 FUNCTION(_sk_mirror_x_1_skx)
 _sk_mirror_x_1_skx:
-  .byte  196,98,125,24,5,214,210,2,0         // vbroadcastss  0x2d2d6(%rip),%ymm8        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  196,98,125,24,5,134,233,2,0         // vbroadcastss  0x2e986(%rip),%ymm8        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  98,113,124,56,89,13,171,210,2,0     // vmulps        0x2d2ab(%rip){1to8},%ymm0,%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,113,124,56,89,13,91,233,2,0      // vmulps        0x2e95b(%rip){1to8},%ymm0,%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,67,125,8,201,1                  // vroundps      $0x1,%ymm9,%ymm9
   .byte  196,65,52,88,201                    // vaddps        %ymm9,%ymm9,%ymm9
   .byte  196,193,124,92,193                  // vsubps        %ymm9,%ymm0,%ymm0
@@ -14965,7 +14965,7 @@
   .byte  197,60,92,200                       // vsubps        %ymm0,%ymm8,%ymm9
   .byte  197,180,84,192                      // vandps        %ymm0,%ymm9,%ymm0
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  98,241,124,56,93,5,127,210,2,0      // vminps        0x2d27f(%rip){1to8},%ymm0,%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,93,5,47,233,2,0       // vminps        0x2e92f(%rip){1to8},%ymm0,%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -14973,9 +14973,9 @@
 .globl _sk_luminance_to_alpha_skx
 FUNCTION(_sk_luminance_to_alpha_skx)
 _sk_luminance_to_alpha_skx:
-  .byte  98,241,116,56,89,29,145,211,2,0     // vmulps        0x2d391(%rip){1to8},%ymm1,%ymm3        # 32d50 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
-  .byte  98,242,125,56,184,29,139,211,2,0    // vfmadd231ps   0x2d38b(%rip){1to8},%ymm0,%ymm3        # 32d54 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
-  .byte  98,242,109,56,184,29,133,211,2,0    // vfmadd231ps   0x2d385(%rip){1to8},%ymm2,%ymm3        # 32d58 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
+  .byte  98,241,116,56,89,29,65,234,2,0      // vmulps        0x2ea41(%rip){1to8},%ymm1,%ymm3        # 34400 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
+  .byte  98,242,125,56,184,29,59,234,2,0     // vfmadd231ps   0x2ea3b(%rip){1to8},%ymm0,%ymm3        # 34404 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
+  .byte  98,242,109,56,184,29,53,234,2,0     // vfmadd231ps   0x2ea35(%rip){1to8},%ymm2,%ymm3        # 34408 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
@@ -15217,11 +15217,11 @@
 .globl _sk_gauss_a_to_rgba_skx
 FUNCTION(_sk_gauss_a_to_rgba_skx)
 _sk_gauss_a_to_rgba_skx:
-  .byte  196,226,125,24,5,211,207,2,0        // vbroadcastss  0x2cfd3(%rip),%ymm0        # 32d5c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
-  .byte  98,242,101,56,168,5,205,207,2,0     // vfmadd213ps   0x2cfcd(%rip){1to8},%ymm3,%ymm0        # 32d60 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
-  .byte  98,242,101,56,168,5,199,207,2,0     // vfmadd213ps   0x2cfc7(%rip){1to8},%ymm3,%ymm0        # 32d64 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
-  .byte  98,242,101,56,168,5,193,207,2,0     // vfmadd213ps   0x2cfc1(%rip){1to8},%ymm3,%ymm0        # 32d68 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
-  .byte  98,242,101,56,168,5,187,207,2,0     // vfmadd213ps   0x2cfbb(%rip){1to8},%ymm3,%ymm0        # 32d6c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
+  .byte  196,226,125,24,5,131,230,2,0        // vbroadcastss  0x2e683(%rip),%ymm0        # 3440c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
+  .byte  98,242,101,56,168,5,125,230,2,0     // vfmadd213ps   0x2e67d(%rip){1to8},%ymm3,%ymm0        # 34410 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
+  .byte  98,242,101,56,168,5,119,230,2,0     // vfmadd213ps   0x2e677(%rip){1to8},%ymm3,%ymm0        # 34414 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
+  .byte  98,242,101,56,168,5,113,230,2,0     // vfmadd213ps   0x2e671(%rip){1to8},%ymm3,%ymm0        # 34418 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
+  .byte  98,242,101,56,168,5,107,230,2,0     // vfmadd213ps   0x2e66b(%rip){1to8},%ymm3,%ymm0        # 3441c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
@@ -15239,7 +15239,7 @@
   .byte  76,139,72,72                        // mov           0x48(%rax),%r9
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  65,186,1,0,0,0                      // mov           $0x1,%r10d
-  .byte  196,226,125,24,21,77,206,2,0        // vbroadcastss  0x2ce4d(%rip),%ymm2        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,226,125,24,21,253,228,2,0       // vbroadcastss  0x2e4fd(%rip),%ymm2        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  196,65,53,239,201                   // vpxor         %ymm9,%ymm9,%ymm9
   .byte  196,130,125,24,28,145               // vbroadcastss  (%r9,%r10,4),%ymm3
   .byte  98,241,100,40,194,192,2             // vcmpleps      %ymm0,%ymm3,%k0
@@ -15339,24 +15339,24 @@
   .byte  196,65,52,95,226                    // vmaxps        %ymm10,%ymm9,%ymm12
   .byte  196,65,36,94,220                    // vdivps        %ymm12,%ymm11,%ymm11
   .byte  196,65,36,89,227                    // vmulps        %ymm11,%ymm11,%ymm12
-  .byte  196,98,125,24,45,227,205,2,0        // vbroadcastss  0x2cde3(%rip),%ymm13        # 32d70 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
-  .byte  98,114,29,56,168,45,221,205,2,0     // vfmadd213ps   0x2cddd(%rip){1to8},%ymm12,%ymm13        # 32d74 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
-  .byte  98,114,29,56,168,45,215,205,2,0     // vfmadd213ps   0x2cdd7(%rip){1to8},%ymm12,%ymm13        # 32d78 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
-  .byte  98,114,29,56,168,45,209,205,2,0     // vfmadd213ps   0x2cdd1(%rip){1to8},%ymm12,%ymm13        # 32d7c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
+  .byte  196,98,125,24,45,147,228,2,0        // vbroadcastss  0x2e493(%rip),%ymm13        # 34420 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
+  .byte  98,114,29,56,168,45,141,228,2,0     // vfmadd213ps   0x2e48d(%rip){1to8},%ymm12,%ymm13        # 34424 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
+  .byte  98,114,29,56,168,45,135,228,2,0     // vfmadd213ps   0x2e487(%rip){1to8},%ymm12,%ymm13        # 34428 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
+  .byte  98,114,29,56,168,45,129,228,2,0     // vfmadd213ps   0x2e481(%rip){1to8},%ymm12,%ymm13        # 3442c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
   .byte  196,65,36,89,221                    // vmulps        %ymm13,%ymm11,%ymm11
   .byte  98,209,52,40,194,194,1              // vcmpltps      %ymm10,%ymm9,%k0
   .byte  98,114,126,40,56,200                // vpmovm2d      %k0,%ymm9
-  .byte  196,98,125,24,21,186,205,2,0        // vbroadcastss  0x2cdba(%rip),%ymm10        # 32d80 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
+  .byte  196,98,125,24,21,106,228,2,0        // vbroadcastss  0x2e46a(%rip),%ymm10        # 34430 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
   .byte  196,65,44,92,211                    // vsubps        %ymm11,%ymm10,%ymm10
   .byte  196,67,37,74,202,144                // vblendvps     %ymm9,%ymm10,%ymm11,%ymm9
   .byte  98,209,124,40,194,192,1             // vcmpltps      %ymm8,%ymm0,%k0
   .byte  98,242,126,40,56,192                // vpmovm2d      %k0,%ymm0
-  .byte  196,98,125,24,21,69,204,2,0         // vbroadcastss  0x2cc45(%rip),%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,21,245,226,2,0        // vbroadcastss  0x2e2f5(%rip),%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,92,209                    // vsubps        %ymm9,%ymm10,%ymm10
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  98,209,116,40,194,192,1             // vcmpltps      %ymm8,%ymm1,%k0
   .byte  98,114,126,40,56,200                // vpmovm2d      %k0,%ymm9
-  .byte  196,98,125,24,21,40,204,2,0         // vbroadcastss  0x2cc28(%rip),%ymm10        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,21,216,226,2,0        // vbroadcastss  0x2e2d8(%rip),%ymm10        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,44,92,208                       // vsubps        %ymm0,%ymm10,%ymm10
   .byte  196,195,125,74,194,144              // vblendvps     %ymm9,%ymm10,%ymm0,%ymm0
   .byte  98,209,124,40,194,192,3             // vcmpunordps   %ymm8,%ymm0,%k0
@@ -15385,20 +15385,20 @@
   .byte  197,50,89,80,76                     // vmulss        0x4c(%rax),%xmm9,%xmm10
   .byte  196,66,125,24,210                   // vbroadcastss  %xmm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  98,113,44,56,89,21,36,205,2,0       // vmulps        0x2cd24(%rip){1to8},%ymm10,%ymm10        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  98,113,44,56,89,21,212,227,2,0      // vmulps        0x2e3d4(%rip){1to8},%ymm10,%ymm10        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
   .byte  196,98,125,184,216                  // vfmadd231ps   %ymm0,%ymm0,%ymm11
   .byte  196,193,50,89,193                   // vmulss        %xmm9,%xmm9,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  98,113,60,56,89,5,7,205,2,0         // vmulps        0x2cd07(%rip){1to8},%ymm8,%ymm8        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  98,113,60,56,89,5,183,227,2,0       // vmulps        0x2e3b7(%rip){1to8},%ymm8,%ymm8        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,194,45,184,194                  // vfmadd231ps   %ymm10,%ymm10,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  98,113,44,56,87,13,238,204,2,0      // vxorps        0x2ccee(%rip){1to8},%ymm10,%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  98,113,44,56,87,13,158,227,2,0      // vxorps        0x2e39e(%rip){1to8},%ymm10,%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  98,113,60,56,89,5,127,203,2,0       // vmulps        0x2cb7f(%rip){1to8},%ymm8,%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,113,60,56,89,5,47,226,2,0        // vmulps        0x2e22f(%rip){1to8},%ymm8,%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,89,208                    // vmulps        %ymm8,%ymm10,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
@@ -15416,20 +15416,20 @@
   .byte  197,50,89,80,76                     // vmulss        0x4c(%rax),%xmm9,%xmm10
   .byte  196,66,125,24,210                   // vbroadcastss  %xmm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  98,113,44,56,89,21,156,204,2,0      // vmulps        0x2cc9c(%rip){1to8},%ymm10,%ymm10        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  98,113,44,56,89,21,76,227,2,0       // vmulps        0x2e34c(%rip){1to8},%ymm10,%ymm10        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
   .byte  196,98,125,184,216                  // vfmadd231ps   %ymm0,%ymm0,%ymm11
   .byte  196,193,50,89,193                   // vmulss        %xmm9,%xmm9,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  98,113,60,56,89,5,127,204,2,0       // vmulps        0x2cc7f(%rip){1to8},%ymm8,%ymm8        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  98,113,60,56,89,5,47,227,2,0        // vmulps        0x2e32f(%rip){1to8},%ymm8,%ymm8        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,194,45,184,194                  // vfmadd231ps   %ymm10,%ymm10,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  98,113,44,56,87,13,102,204,2,0      // vxorps        0x2cc66(%rip){1to8},%ymm10,%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  98,113,44,56,87,13,22,227,2,0       // vxorps        0x2e316(%rip){1to8},%ymm10,%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  98,113,60,56,89,5,247,202,2,0       // vmulps        0x2caf7(%rip){1to8},%ymm8,%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,113,60,56,89,5,167,225,2,0       // vmulps        0x2e1a7(%rip){1to8},%ymm8,%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,89,208                    // vmulps        %ymm8,%ymm10,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
@@ -15446,13 +15446,13 @@
   .byte  197,58,89,72,76                     // vmulss        0x4c(%rax),%xmm8,%xmm9
   .byte  196,66,125,24,201                   // vbroadcastss  %xmm9,%ymm9
   .byte  197,52,88,200                       // vaddps        %ymm0,%ymm9,%ymm9
-  .byte  98,113,52,56,89,13,26,204,2,0       // vmulps        0x2cc1a(%rip){1to8},%ymm9,%ymm9        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  98,113,52,56,89,13,202,226,2,0      // vmulps        0x2e2ca(%rip){1to8},%ymm9,%ymm9        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  197,116,89,209                      // vmulps        %ymm1,%ymm1,%ymm10
   .byte  196,98,125,184,208                  // vfmadd231ps   %ymm0,%ymm0,%ymm10
   .byte  196,193,58,89,192                   // vmulss        %xmm8,%xmm8,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,172,92,192                      // vsubps        %ymm0,%ymm10,%ymm0
-  .byte  98,241,124,56,87,5,1,204,2,0        // vxorps        0x2cc01(%rip){1to8},%ymm0,%ymm0        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  98,241,124,56,87,5,177,226,2,0      // vxorps        0x2e2b1(%rip){1to8},%ymm0,%ymm0        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,193,124,94,193                  // vdivps        %ymm9,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15494,7 +15494,7 @@
 FUNCTION(_sk_save_xy_skx)
 _sk_save_xy_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,46,202,2,0          // vbroadcastss  0x2ca2e(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,222,224,2,0         // vbroadcastss  0x2e0de(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,124,88,200                   // vaddps        %ymm8,%ymm0,%ymm9
   .byte  196,67,125,8,209,1                  // vroundps      $0x1,%ymm9,%ymm10
   .byte  196,65,52,92,202                    // vsubps        %ymm10,%ymm9,%ymm9
@@ -15529,8 +15529,8 @@
 _sk_bilinear_nx_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,0                        // vmovups       (%rax),%ymm0
-  .byte  98,241,124,56,88,5,23,203,2,0       // vaddps        0x2cb17(%rip){1to8},%ymm0,%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
-  .byte  196,98,125,24,5,174,201,2,0         // vbroadcastss  0x2c9ae(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,88,5,199,225,2,0      // vaddps        0x2e1c7(%rip){1to8},%ymm0,%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,98,125,24,5,94,224,2,0          // vbroadcastss  0x2e05e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15543,7 +15543,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,0                        // vmovups       (%rax),%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
-  .byte  98,241,124,56,88,5,126,201,2,0      // vaddps        0x2c97e(%rip){1to8},%ymm0,%ymm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,241,124,56,88,5,46,224,2,0       // vaddps        0x2e02e(%rip){1to8},%ymm0,%ymm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15554,8 +15554,8 @@
 _sk_bilinear_ny_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,72,64                    // vmovups       0x40(%rax),%ymm1
-  .byte  98,241,116,56,88,13,197,202,2,0     // vaddps        0x2cac5(%rip){1to8},%ymm1,%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
-  .byte  196,98,125,24,5,92,201,2,0          // vbroadcastss  0x2c95c(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,116,56,88,13,117,225,2,0     // vaddps        0x2e175(%rip){1to8},%ymm1,%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,98,125,24,5,12,224,2,0          // vbroadcastss  0x2e00c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15568,7 +15568,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,72,64                    // vmovups       0x40(%rax),%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
-  .byte  98,241,116,56,88,13,43,201,2,0      // vaddps        0x2c92b(%rip){1to8},%ymm1,%ymm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,241,116,56,88,13,219,223,2,0     // vaddps        0x2dfdb(%rip){1to8},%ymm1,%ymm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15579,12 +15579,12 @@
 _sk_bicubic_n3x_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,0                        // vmovups       (%rax),%ymm0
-  .byte  98,241,124,56,88,5,119,202,2,0      // vaddps        0x2ca77(%rip){1to8},%ymm0,%ymm0        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
-  .byte  196,98,125,24,5,10,201,2,0          // vbroadcastss  0x2c90a(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,88,5,39,225,2,0       // vaddps        0x2e127(%rip){1to8},%ymm0,%ymm0        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,98,125,24,5,186,223,2,0         // vbroadcastss  0x2dfba(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,92,202,2,0         // vbroadcastss  0x2ca5c(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
-  .byte  98,114,61,56,168,21,94,201,2,0      // vfmadd213ps   0x2c95e(%rip){1to8},%ymm8,%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,12,225,2,0         // vbroadcastss  0x2e10c(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  98,114,61,56,168,21,14,224,2,0      // vfmadd213ps   0x2e00e(%rip){1to8},%ymm8,%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,44,89,193                    // vmulps        %ymm9,%ymm10,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15596,13 +15596,13 @@
 _sk_bicubic_n1x_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,0                        // vmovups       (%rax),%ymm0
-  .byte  98,241,124,56,88,5,41,202,2,0       // vaddps        0x2ca29(%rip){1to8},%ymm0,%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
-  .byte  196,98,125,24,5,192,200,2,0         // vbroadcastss  0x2c8c0(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,124,56,88,5,217,224,2,0      // vaddps        0x2e0d9(%rip){1to8},%ymm0,%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,98,125,24,5,112,223,2,0         // vbroadcastss  0x2df70(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,27,202,2,0         // vbroadcastss  0x2ca1b(%rip),%ymm9        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  .byte  98,114,61,56,168,13,21,202,2,0      // vfmadd213ps   0x2ca15(%rip){1to8},%ymm8,%ymm9        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  98,114,61,56,168,13,151,200,2,0     // vfmadd213ps   0x2c897(%rip){1to8},%ymm8,%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,61,56,168,13,5,202,2,0       // vfmadd213ps   0x2ca05(%rip){1to8},%ymm8,%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,203,224,2,0        // vbroadcastss  0x2e0cb(%rip),%ymm9        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  98,114,61,56,168,13,197,224,2,0     // vfmadd213ps   0x2e0c5(%rip){1to8},%ymm8,%ymm9        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  98,114,61,56,168,13,71,223,2,0      // vfmadd213ps   0x2df47(%rip){1to8},%ymm8,%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,61,56,168,13,181,224,2,0     // vfmadd213ps   0x2e0b5(%rip){1to8},%ymm8,%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  197,124,17,136,0,1,0,0              // vmovups       %ymm9,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15612,13 +15612,13 @@
 FUNCTION(_sk_bicubic_p1x_skx)
 _sk_bicubic_p1x_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,118,200,2,0         // vbroadcastss  0x2c876(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,38,223,2,0          // vbroadcastss  0x2df26(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,136,128,0,0,0            // vmovups       0x80(%rax),%ymm9
-  .byte  196,98,125,24,21,209,201,2,0        // vbroadcastss  0x2c9d1(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  .byte  98,114,53,56,168,21,203,201,2,0     // vfmadd213ps   0x2c9cb(%rip){1to8},%ymm9,%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,129,224,2,0        // vbroadcastss  0x2e081(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  98,114,53,56,168,21,123,224,2,0     // vfmadd213ps   0x2e07b(%rip){1to8},%ymm9,%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,66,53,168,208                   // vfmadd213ps   %ymm8,%ymm9,%ymm10
-  .byte  98,114,53,56,168,21,192,201,2,0     // vfmadd213ps   0x2c9c0(%rip){1to8},%ymm9,%ymm10        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  98,114,53,56,168,21,112,224,2,0     // vfmadd213ps   0x2e070(%rip){1to8},%ymm9,%ymm10        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  197,124,17,144,0,1,0,0              // vmovups       %ymm10,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15630,10 +15630,10 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,0                        // vmovups       (%rax),%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
-  .byte  98,241,124,56,88,5,152,201,2,0      // vaddps        0x2c998(%rip){1to8},%ymm0,%ymm0        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  98,241,124,56,88,5,72,224,2,0       // vaddps        0x2e048(%rip){1to8},%ymm0,%ymm0        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,130,201,2,0        // vbroadcastss  0x2c982(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
-  .byte  98,114,61,56,168,21,132,200,2,0     // vfmadd213ps   0x2c884(%rip){1to8},%ymm8,%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,50,224,2,0         // vbroadcastss  0x2e032(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  98,114,61,56,168,21,52,223,2,0      // vfmadd213ps   0x2df34(%rip){1to8},%ymm8,%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,52,89,194                    // vmulps        %ymm10,%ymm9,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15645,12 +15645,12 @@
 _sk_bicubic_n3y_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,72,64                    // vmovups       0x40(%rax),%ymm1
-  .byte  98,241,116,56,88,13,82,201,2,0      // vaddps        0x2c952(%rip){1to8},%ymm1,%ymm1        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
-  .byte  196,98,125,24,5,229,199,2,0         // vbroadcastss  0x2c7e5(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,116,56,88,13,2,224,2,0       // vaddps        0x2e002(%rip){1to8},%ymm1,%ymm1        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,98,125,24,5,149,222,2,0         // vbroadcastss  0x2de95(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,55,201,2,0         // vbroadcastss  0x2c937(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
-  .byte  98,114,61,56,168,21,57,200,2,0      // vfmadd213ps   0x2c839(%rip){1to8},%ymm8,%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,231,223,2,0        // vbroadcastss  0x2dfe7(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  98,114,61,56,168,21,233,222,2,0     // vfmadd213ps   0x2dee9(%rip){1to8},%ymm8,%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,44,89,193                    // vmulps        %ymm9,%ymm10,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15662,13 +15662,13 @@
 _sk_bicubic_n1y_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,72,64                    // vmovups       0x40(%rax),%ymm1
-  .byte  98,241,116,56,88,13,3,201,2,0       // vaddps        0x2c903(%rip){1to8},%ymm1,%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
-  .byte  196,98,125,24,5,154,199,2,0         // vbroadcastss  0x2c79a(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  98,241,116,56,88,13,179,223,2,0     // vaddps        0x2dfb3(%rip){1to8},%ymm1,%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,98,125,24,5,74,222,2,0          // vbroadcastss  0x2de4a(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,245,200,2,0        // vbroadcastss  0x2c8f5(%rip),%ymm9        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  .byte  98,114,61,56,168,13,239,200,2,0     // vfmadd213ps   0x2c8ef(%rip){1to8},%ymm8,%ymm9        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  98,114,61,56,168,13,113,199,2,0     // vfmadd213ps   0x2c771(%rip){1to8},%ymm8,%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  98,114,61,56,168,13,223,200,2,0     // vfmadd213ps   0x2c8df(%rip){1to8},%ymm8,%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,165,223,2,0        // vbroadcastss  0x2dfa5(%rip),%ymm9        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  98,114,61,56,168,13,159,223,2,0     // vfmadd213ps   0x2df9f(%rip){1to8},%ymm8,%ymm9        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  98,114,61,56,168,13,33,222,2,0      // vfmadd213ps   0x2de21(%rip){1to8},%ymm8,%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  98,114,61,56,168,13,143,223,2,0     // vfmadd213ps   0x2df8f(%rip){1to8},%ymm8,%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  197,124,17,136,64,1,0,0             // vmovups       %ymm9,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15678,13 +15678,13 @@
 FUNCTION(_sk_bicubic_p1y_skx)
 _sk_bicubic_p1y_skx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,80,199,2,0          // vbroadcastss  0x2c750(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,0,222,2,0           // vbroadcastss  0x2de00(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,72,64                    // vaddps        0x40(%rax),%ymm8,%ymm1
   .byte  197,124,16,136,192,0,0,0            // vmovups       0xc0(%rax),%ymm9
-  .byte  196,98,125,24,21,170,200,2,0        // vbroadcastss  0x2c8aa(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  .byte  98,114,53,56,168,21,164,200,2,0     // vfmadd213ps   0x2c8a4(%rip){1to8},%ymm9,%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,90,223,2,0         // vbroadcastss  0x2df5a(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  98,114,53,56,168,21,84,223,2,0      // vfmadd213ps   0x2df54(%rip){1to8},%ymm9,%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,66,53,168,208                   // vfmadd213ps   %ymm8,%ymm9,%ymm10
-  .byte  98,114,53,56,168,21,153,200,2,0     // vfmadd213ps   0x2c899(%rip){1to8},%ymm9,%ymm10        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  98,114,53,56,168,21,73,223,2,0      // vfmadd213ps   0x2df49(%rip){1to8},%ymm9,%ymm10        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  197,124,17,144,64,1,0,0             // vmovups       %ymm10,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -15696,10 +15696,10 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,72,64                    // vmovups       0x40(%rax),%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
-  .byte  98,241,116,56,88,13,112,200,2,0     // vaddps        0x2c870(%rip){1to8},%ymm1,%ymm1        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  98,241,116,56,88,13,32,223,2,0      // vaddps        0x2df20(%rip){1to8},%ymm1,%ymm1        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,90,200,2,0         // vbroadcastss  0x2c85a(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
-  .byte  98,114,61,56,168,21,92,199,2,0      // vfmadd213ps   0x2c75c(%rip){1to8},%ymm8,%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,10,223,2,0         // vbroadcastss  0x2df0a(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  98,114,61,56,168,21,12,222,2,0      // vfmadd213ps   0x2de0c(%rip){1to8},%ymm8,%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,52,89,194                    // vmulps        %ymm10,%ymm9,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -15825,7 +15825,7 @@
   .byte  98,226,61,40,64,224                 // vpmulld       %ymm0,%ymm8,%ymm20
   .byte  98,209,93,32,254,193                // vpaddd        %ymm9,%ymm20,%ymm0
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  98,98,125,40,88,29,100,198,2,0      // vpbroadcastd  0x2c664(%rip),%ymm27        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  98,98,125,40,88,29,20,221,2,0       // vpbroadcastd  0x2dd14(%rip),%ymm27        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  98,146,125,40,64,195                // vpmulld       %ymm27,%ymm0,%ymm0
   .byte  196,65,45,239,210                   // vpxor         %ymm10,%ymm10,%ymm10
   .byte  197,245,118,201                     // vpcmpeqd      %ymm1,%ymm1,%ymm1
@@ -15838,9 +15838,9 @@
   .byte  196,65,37,118,219                   // vpcmpeqd      %ymm11,%ymm11,%ymm11
   .byte  196,162,37,146,12,144               // vgatherdps    %ymm11,(%rax,%ymm10,4),%ymm1
   .byte  98,97,124,40,40,233                 // vmovaps       %ymm1,%ymm29
-  .byte  196,98,125,24,21,26,198,2,0         // vbroadcastss  0x2c61a(%rip),%ymm10        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,98,125,24,21,202,220,2,0        // vbroadcastss  0x2dcca(%rip),%ymm10        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  98,81,60,32,88,218                  // vaddps        %ymm10,%ymm24,%ymm11
-  .byte  98,226,125,40,88,5,158,196,2,0      // vpbroadcastd  0x2c49e(%rip),%ymm16        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  98,226,125,40,88,5,78,219,2,0       // vpbroadcastd  0x2db4e(%rip),%ymm16        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  98,177,125,40,254,192               // vpaddd        %ymm16,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  196,65,29,118,228                   // vpcmpeqd      %ymm12,%ymm12,%ymm12
@@ -16080,7 +16080,7 @@
   .byte  98,162,45,40,64,229                 // vpmulld       %ymm21,%ymm10,%ymm20
   .byte  98,241,93,32,254,193                // vpaddd        %ymm1,%ymm20,%ymm0
   .byte  72,139,0                            // mov           (%rax),%rax
-  .byte  98,226,125,40,88,5,246,192,2,0      // vpbroadcastd  0x2c0f6(%rip),%ymm16        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  98,226,125,40,88,5,166,215,2,0      // vpbroadcastd  0x2d7a6(%rip),%ymm16        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  98,50,125,40,64,216                 // vpmulld       %ymm16,%ymm0,%ymm11
   .byte  196,65,28,87,228                    // vxorps        %ymm12,%ymm12,%ymm12
   .byte  197,253,118,192                     // vpcmpeqd      %ymm0,%ymm0,%ymm0
@@ -16092,9 +16092,9 @@
   .byte  196,65,29,118,228                   // vpcmpeqd      %ymm12,%ymm12,%ymm12
   .byte  196,98,29,146,44,128                // vgatherdps    %ymm12,(%rax,%ymm0,4),%ymm13
   .byte  197,124,17,172,36,192,4,0,0         // vmovups       %ymm13,0x4c0(%rsp)
-  .byte  196,226,125,24,5,168,192,2,0        // vbroadcastss  0x2c0a8(%rip),%ymm0        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,226,125,24,5,88,215,2,0         // vbroadcastss  0x2d758(%rip),%ymm0        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  98,113,28,32,88,224                 // vaddps        %ymm0,%ymm28,%ymm12
-  .byte  98,226,125,40,88,13,44,191,2,0      // vpbroadcastd  0x2bf2c(%rip),%ymm17        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  98,226,125,40,88,13,220,213,2,0     // vpbroadcastd  0x2d5dc(%rip),%ymm17        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  98,49,37,40,254,217                 // vpaddd        %ymm17,%ymm11,%ymm11
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
   .byte  196,65,21,118,237                   // vpcmpeqd      %ymm13,%ymm13,%ymm13
@@ -16499,7 +16499,7 @@
   .byte  197,228,92,214                      // vsubps        %ymm6,%ymm3,%ymm2
   .byte  196,226,93,168,214                  // vfmadd213ps   %ymm6,%ymm4,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,202,181,2,0       // vbroadcastss  0x2b5ca(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,122,204,2,0       // vbroadcastss  0x2cc7a(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  98,145,124,40,40,227                // vmovaps       %ymm27,%ymm4
   .byte  197,252,16,108,36,128               // vmovups       -0x80(%rsp),%ymm5
   .byte  197,252,16,116,36,160               // vmovups       -0x60(%rsp),%ymm6
@@ -16600,10 +16600,10 @@
   .byte  197,249,110,201                     // vmovd         %ecx,%xmm1
   .byte  196,226,125,88,201                  // vpbroadcastd  %xmm1,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,158,180,2,0       // vbroadcastss  0x2b49e(%rip),%ymm2        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,21,78,203,2,0        // vbroadcastss  0x2cb4e(%rip),%ymm2        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,244,88,202                      // vaddps        %ymm2,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,147,180,2,0       // vbroadcastss  0x2b493(%rip),%ymm2        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,21,67,203,2,0        // vbroadcastss  0x2cb43(%rip),%ymm2        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,228,87,219                      // vxorps        %ymm3,%ymm3,%ymm3
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
@@ -16618,17 +16618,17 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,121,110,194                     // vmovd         %edx,%xmm8
   .byte  196,66,125,88,192                   // vpbroadcastd  %xmm8,%ymm8
-  .byte  197,61,254,5,58,182,2,0             // vpaddd        0x2b63a(%rip),%ymm8,%ymm8        # 32e00 <_sk_srcover_rgba_8888_sse2_lowp+0x480>
+  .byte  197,61,254,5,218,204,2,0            // vpaddd        0x2ccda(%rip),%ymm8,%ymm8        # 344a0 <_sk_srcover_rgba_8888_sse2_lowp+0x470>
   .byte  197,121,110,201                     // vmovd         %ecx,%xmm9
   .byte  196,66,125,88,201                   // vpbroadcastd  %xmm9,%ymm9
   .byte  196,65,53,239,200                   // vpxor         %ymm8,%ymm9,%ymm9
-  .byte  196,98,125,88,21,87,180,2,0         // vpbroadcastd  0x2b457(%rip),%ymm10        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,98,125,88,21,7,203,2,0          // vpbroadcastd  0x2cb07(%rip),%ymm10        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  196,65,53,219,218                   // vpand         %ymm10,%ymm9,%ymm11
   .byte  196,193,37,114,243,5                // vpslld        $0x5,%ymm11,%ymm11
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,193,45,114,242,4                // vpslld        $0x4,%ymm10,%ymm10
-  .byte  196,98,125,88,37,60,180,2,0         // vpbroadcastd  0x2b43c(%rip),%ymm12        # 32c38 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
-  .byte  196,98,125,88,45,55,180,2,0         // vpbroadcastd  0x2b437(%rip),%ymm13        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,125,88,37,236,202,2,0        // vpbroadcastd  0x2caec(%rip),%ymm12        # 342e8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
+  .byte  196,98,125,88,45,231,202,2,0        // vpbroadcastd  0x2cae7(%rip),%ymm13        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,65,53,219,245                   // vpand         %ymm13,%ymm9,%ymm14
   .byte  196,193,13,114,246,2                // vpslld        $0x2,%ymm14,%ymm14
   .byte  196,65,37,235,222                   // vpor          %ymm14,%ymm11,%ymm11
@@ -16643,8 +16643,8 @@
   .byte  196,65,61,235,195                   // vpor          %ymm11,%ymm8,%ymm8
   .byte  196,65,61,235,193                   // vpor          %ymm9,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,233,179,2,0        // vbroadcastss  0x2b3e9(%rip),%ymm9        # 32c40 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
-  .byte  196,98,125,24,21,228,179,2,0        // vbroadcastss  0x2b3e4(%rip),%ymm10        # 32c44 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
+  .byte  196,98,125,24,13,153,202,2,0        // vbroadcastss  0x2ca99(%rip),%ymm9        # 342f0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
+  .byte  196,98,125,24,21,148,202,2,0        // vbroadcastss  0x2ca94(%rip),%ymm10        # 342f4 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
   .byte  196,66,61,184,209                   // vfmadd231ps   %ymm9,%ymm8,%ymm10
   .byte  196,98,125,24,0                     // vbroadcastss  (%rax),%ymm8
   .byte  196,65,44,89,192                    // vmulps        %ymm8,%ymm10,%ymm8
@@ -16678,7 +16678,7 @@
 FUNCTION(_sk_black_color_hsw)
 _sk_black_color_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,108,179,2,0       // vbroadcastss  0x2b36c(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,28,202,2,0        // vbroadcastss  0x2ca1c(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
@@ -16689,7 +16689,7 @@
 FUNCTION(_sk_white_color_hsw)
 _sk_white_color_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,83,179,2,0         // vbroadcastss  0x2b353(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,3,202,2,0          // vbroadcastss  0x2ca03(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  197,252,40,216                      // vmovaps       %ymm0,%ymm3
@@ -16735,7 +16735,7 @@
 FUNCTION(_sk_srcatop_hsw)
 _sk_srcatop_hsw:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,242,178,2,0         // vbroadcastss  0x2b2f2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,162,201,2,0         // vbroadcastss  0x2c9a2(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,226,61,184,196                  // vfmadd231ps   %ymm4,%ymm8,%ymm0
   .byte  197,244,89,207                      // vmulps        %ymm7,%ymm1,%ymm1
@@ -16751,7 +16751,7 @@
 .globl _sk_dstatop_hsw
 FUNCTION(_sk_dstatop_hsw)
 _sk_dstatop_hsw:
-  .byte  196,98,125,24,5,193,178,2,0         // vbroadcastss  0x2b2c1(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,113,201,2,0         // vbroadcastss  0x2c971(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,226,101,184,196                 // vfmadd231ps   %ymm4,%ymm3,%ymm0
@@ -16790,7 +16790,7 @@
 .globl _sk_srcout_hsw
 FUNCTION(_sk_srcout_hsw)
 _sk_srcout_hsw:
-  .byte  196,98,125,24,5,100,178,2,0         // vbroadcastss  0x2b264(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,20,201,2,0          // vbroadcastss  0x2c914(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
@@ -16803,7 +16803,7 @@
 .globl _sk_dstout_hsw
 FUNCTION(_sk_dstout_hsw)
 _sk_dstout_hsw:
-  .byte  196,226,125,24,5,67,178,2,0         // vbroadcastss  0x2b243(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,243,200,2,0        // vbroadcastss  0x2c8f3(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,92,219                      // vsubps        %ymm3,%ymm0,%ymm3
   .byte  197,228,89,196                      // vmulps        %ymm4,%ymm3,%ymm0
   .byte  197,228,89,205                      // vmulps        %ymm5,%ymm3,%ymm1
@@ -16816,7 +16816,7 @@
 .globl _sk_srcover_hsw
 FUNCTION(_sk_srcover_hsw)
 _sk_srcover_hsw:
-  .byte  196,98,125,24,5,34,178,2,0          // vbroadcastss  0x2b222(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,210,200,2,0         // vbroadcastss  0x2c8d2(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,93,184,192                  // vfmadd231ps   %ymm8,%ymm4,%ymm0
   .byte  196,194,85,184,200                  // vfmadd231ps   %ymm8,%ymm5,%ymm1
@@ -16829,7 +16829,7 @@
 .globl _sk_dstover_hsw
 FUNCTION(_sk_dstover_hsw)
 _sk_dstover_hsw:
-  .byte  196,98,125,24,5,253,177,2,0         // vbroadcastss  0x2b1fd(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,173,200,2,0         // vbroadcastss  0x2c8ad(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  196,226,61,168,196                  // vfmadd213ps   %ymm4,%ymm8,%ymm0
   .byte  196,226,61,168,205                  // vfmadd213ps   %ymm5,%ymm8,%ymm1
@@ -16853,7 +16853,7 @@
 .globl _sk_multiply_hsw
 FUNCTION(_sk_multiply_hsw)
 _sk_multiply_hsw:
-  .byte  196,98,125,24,5,196,177,2,0         // vbroadcastss  0x2b1c4(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,116,200,2,0         // vbroadcastss  0x2c874(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,208                       // vmulps        %ymm0,%ymm9,%ymm10
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -16876,7 +16876,7 @@
 FUNCTION(_sk_plus__hsw)
 _sk_plus__hsw:
   .byte  197,252,88,196                      // vaddps        %ymm4,%ymm0,%ymm0
-  .byte  196,98,125,24,5,115,177,2,0         // vbroadcastss  0x2b173(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,35,200,2,0          // vbroadcastss  0x2c823(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  197,244,88,205                      // vaddps        %ymm5,%ymm1,%ymm1
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
@@ -16906,7 +16906,7 @@
 .globl _sk_xor__hsw
 FUNCTION(_sk_xor__hsw)
 _sk_xor__hsw:
-  .byte  196,98,125,24,5,30,177,2,0          // vbroadcastss  0x2b11e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,206,199,2,0         // vbroadcastss  0x2c7ce(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -16940,7 +16940,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,95,209                  // vmaxps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,162,176,2,0         // vbroadcastss  0x2b0a2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,82,199,2,0          // vbroadcastss  0x2c752(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -16965,7 +16965,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,77,176,2,0          // vbroadcastss  0x2b04d(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,253,198,2,0         // vbroadcastss  0x2c6fd(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -16993,7 +16993,7 @@
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,236,175,2,0         // vbroadcastss  0x2afec(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,156,198,2,0         // vbroadcastss  0x2c69c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -17015,7 +17015,7 @@
   .byte  197,236,89,214                      // vmulps        %ymm6,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,166,175,2,0         // vbroadcastss  0x2afa6(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,86,198,2,0          // vbroadcastss  0x2c656(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  196,194,69,184,216                  // vfmadd231ps   %ymm8,%ymm7,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -17025,7 +17025,7 @@
 .globl _sk_colorburn_hsw
 FUNCTION(_sk_colorburn_hsw)
 _sk_colorburn_hsw:
-  .byte  196,98,125,24,5,144,175,2,0         // vbroadcastss  0x2af90(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,64,198,2,0          // vbroadcastss  0x2c640(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,216                       // vmulps        %ymm0,%ymm9,%ymm11
   .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
@@ -17083,7 +17083,7 @@
 FUNCTION(_sk_colordodge_hsw)
 _sk_colordodge_hsw:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,160,174,2,0        // vbroadcastss  0x2aea0(%rip),%ymm9        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,13,80,197,2,0         // vbroadcastss  0x2c550(%rip),%ymm9        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,52,92,215                       // vsubps        %ymm7,%ymm9,%ymm10
   .byte  197,44,89,216                       // vmulps        %ymm0,%ymm10,%ymm11
   .byte  197,52,92,203                       // vsubps        %ymm3,%ymm9,%ymm9
@@ -17136,7 +17136,7 @@
 .globl _sk_hardlight_hsw
 FUNCTION(_sk_hardlight_hsw)
 _sk_hardlight_hsw:
-  .byte  196,98,125,24,5,186,173,2,0         // vbroadcastss  0x2adba(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,106,196,2,0         // vbroadcastss  0x2c46a(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,215                       // vsubps        %ymm7,%ymm8,%ymm10
   .byte  197,44,89,216                       // vmulps        %ymm0,%ymm10,%ymm11
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -17187,7 +17187,7 @@
 .globl _sk_overlay_hsw
 FUNCTION(_sk_overlay_hsw)
 _sk_overlay_hsw:
-  .byte  196,98,125,24,5,238,172,2,0         // vbroadcastss  0x2acee(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,158,195,2,0         // vbroadcastss  0x2c39e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,215                       // vsubps        %ymm7,%ymm8,%ymm10
   .byte  197,44,89,216                       // vmulps        %ymm0,%ymm10,%ymm11
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -17248,10 +17248,10 @@
   .byte  196,65,20,88,197                    // vaddps        %ymm13,%ymm13,%ymm8
   .byte  196,65,60,88,192                    // vaddps        %ymm8,%ymm8,%ymm8
   .byte  196,66,61,168,192                   // vfmadd213ps   %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,29,9,172,2,0          // vbroadcastss  0x2ac09(%rip),%ymm11        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  196,98,125,24,29,185,194,2,0        // vbroadcastss  0x2c2b9(%rip),%ymm11        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  196,65,20,88,227                    // vaddps        %ymm11,%ymm13,%ymm12
   .byte  196,65,28,89,192                    // vmulps        %ymm8,%ymm12,%ymm8
-  .byte  196,98,125,24,37,250,171,2,0        // vbroadcastss  0x2abfa(%rip),%ymm12        # 32c4c <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
+  .byte  196,98,125,24,37,170,194,2,0        // vbroadcastss  0x2c2aa(%rip),%ymm12        # 342fc <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
   .byte  196,66,21,184,196                   // vfmadd231ps   %ymm12,%ymm13,%ymm8
   .byte  196,65,124,82,245                   // vrsqrtps      %ymm13,%ymm14
   .byte  196,65,124,83,246                   // vrcpps        %ymm14,%ymm14
@@ -17261,7 +17261,7 @@
   .byte  197,4,194,255,2                     // vcmpleps      %ymm7,%ymm15,%ymm15
   .byte  196,67,13,74,240,240                // vblendvps     %ymm15,%ymm8,%ymm14,%ymm14
   .byte  197,116,88,249                      // vaddps        %ymm1,%ymm1,%ymm15
-  .byte  196,98,125,24,5,169,171,2,0         // vbroadcastss  0x2aba9(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,89,194,2,0          // vbroadcastss  0x2c259(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,65,60,92,237                    // vsubps        %ymm13,%ymm8,%ymm13
   .byte  197,132,92,195                      // vsubps        %ymm3,%ymm15,%ymm0
   .byte  196,98,125,168,235                  // vfmadd213ps   %ymm3,%ymm0,%ymm13
@@ -17374,11 +17374,11 @@
   .byte  196,65,28,89,210                    // vmulps        %ymm10,%ymm12,%ymm10
   .byte  196,65,44,94,214                    // vdivps        %ymm14,%ymm10,%ymm10
   .byte  196,67,45,74,224,240                // vblendvps     %ymm15,%ymm8,%ymm10,%ymm12
-  .byte  196,98,125,24,53,193,169,2,0        // vbroadcastss  0x2a9c1(%rip),%ymm14        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,61,188,169,2,0        // vbroadcastss  0x2a9bc(%rip),%ymm15        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,53,113,192,2,0        // vbroadcastss  0x2c071(%rip),%ymm14        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,61,108,192,2,0        // vbroadcastss  0x2c06c(%rip),%ymm15        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,173,169,2,0        // vbroadcastss  0x2a9ad(%rip),%ymm0        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,226,125,24,5,93,192,2,0         // vbroadcastss  0x2c05d(%rip),%ymm0        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,98,77,184,232                   // vfmadd231ps   %ymm0,%ymm6,%ymm13
   .byte  196,65,116,89,215                   // vmulps        %ymm15,%ymm1,%ymm10
   .byte  196,66,53,184,214                   // vfmadd231ps   %ymm14,%ymm9,%ymm10
@@ -17433,7 +17433,7 @@
   .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
   .byte  196,65,36,95,200                    // vmaxps        %ymm8,%ymm11,%ymm9
   .byte  196,65,116,95,192                   // vmaxps        %ymm8,%ymm1,%ymm8
-  .byte  196,226,125,24,13,110,168,2,0       // vbroadcastss  0x2a86e(%rip),%ymm1        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,13,30,191,2,0        // vbroadcastss  0x2bf1e(%rip),%ymm1        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,116,92,215                      // vsubps        %ymm7,%ymm1,%ymm10
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  197,116,92,219                      // vsubps        %ymm3,%ymm1,%ymm11
@@ -17487,11 +17487,11 @@
   .byte  196,65,28,89,210                    // vmulps        %ymm10,%ymm12,%ymm10
   .byte  196,65,44,94,214                    // vdivps        %ymm14,%ymm10,%ymm10
   .byte  196,67,45,74,224,240                // vblendvps     %ymm15,%ymm8,%ymm10,%ymm12
-  .byte  196,98,125,24,53,161,167,2,0        // vbroadcastss  0x2a7a1(%rip),%ymm14        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,61,156,167,2,0        // vbroadcastss  0x2a79c(%rip),%ymm15        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,53,81,190,2,0         // vbroadcastss  0x2be51(%rip),%ymm14        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,61,76,190,2,0         // vbroadcastss  0x2be4c(%rip),%ymm15        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,66,93,184,238                   // vfmadd231ps   %ymm14,%ymm4,%ymm13
-  .byte  196,226,125,24,5,141,167,2,0        // vbroadcastss  0x2a78d(%rip),%ymm0        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,226,125,24,5,61,190,2,0         // vbroadcastss  0x2be3d(%rip),%ymm0        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,98,77,184,232                   // vfmadd231ps   %ymm0,%ymm6,%ymm13
   .byte  196,65,116,89,215                   // vmulps        %ymm15,%ymm1,%ymm10
   .byte  196,66,53,184,214                   // vfmadd231ps   %ymm14,%ymm9,%ymm10
@@ -17546,7 +17546,7 @@
   .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
   .byte  196,65,36,95,200                    // vmaxps        %ymm8,%ymm11,%ymm9
   .byte  196,65,116,95,192                   // vmaxps        %ymm8,%ymm1,%ymm8
-  .byte  196,226,125,24,13,78,166,2,0        // vbroadcastss  0x2a64e(%rip),%ymm1        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,13,254,188,2,0       // vbroadcastss  0x2bcfe(%rip),%ymm1        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,116,92,215                      // vsubps        %ymm7,%ymm1,%ymm10
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  197,116,92,219                      // vsubps        %ymm3,%ymm1,%ymm11
@@ -17574,11 +17574,11 @@
   .byte  197,108,89,199                      // vmulps        %ymm7,%ymm2,%ymm8
   .byte  197,116,89,215                      // vmulps        %ymm7,%ymm1,%ymm10
   .byte  197,52,89,223                       // vmulps        %ymm7,%ymm9,%ymm11
-  .byte  196,98,125,24,45,3,166,2,0          // vbroadcastss  0x2a603(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,53,254,165,2,0        // vbroadcastss  0x2a5fe(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,45,179,188,2,0        // vbroadcastss  0x2bcb3(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,174,188,2,0        // vbroadcastss  0x2bcae(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,230                    // vmulps        %ymm14,%ymm5,%ymm12
   .byte  196,66,93,184,229                   // vfmadd231ps   %ymm13,%ymm4,%ymm12
-  .byte  196,98,125,24,61,239,165,2,0        // vbroadcastss  0x2a5ef(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,159,188,2,0        // vbroadcastss  0x2bc9f(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,66,77,184,231                   // vfmadd231ps   %ymm15,%ymm6,%ymm12
   .byte  196,65,44,89,206                    // vmulps        %ymm14,%ymm10,%ymm9
   .byte  196,66,61,184,205                   // vfmadd231ps   %ymm13,%ymm8,%ymm9
@@ -17634,7 +17634,7 @@
   .byte  196,193,116,95,206                  // vmaxps        %ymm14,%ymm1,%ymm1
   .byte  196,65,44,95,198                    // vmaxps        %ymm14,%ymm10,%ymm8
   .byte  196,65,124,95,206                   // vmaxps        %ymm14,%ymm0,%ymm9
-  .byte  196,226,125,24,5,165,164,2,0        // vbroadcastss  0x2a4a5(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,85,187,2,0         // vbroadcastss  0x2bb55(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,124,92,215                      // vsubps        %ymm7,%ymm0,%ymm10
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  197,124,92,219                      // vsubps        %ymm3,%ymm0,%ymm11
@@ -17662,11 +17662,11 @@
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
   .byte  197,100,89,213                      // vmulps        %ymm5,%ymm3,%ymm10
   .byte  197,100,89,222                      // vmulps        %ymm6,%ymm3,%ymm11
-  .byte  196,98,125,24,45,90,164,2,0         // vbroadcastss  0x2a45a(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  .byte  196,98,125,24,53,85,164,2,0         // vbroadcastss  0x2a455(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,45,10,187,2,0         // vbroadcastss  0x2bb0a(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,5,187,2,0          // vbroadcastss  0x2bb05(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,116,89,230                   // vmulps        %ymm14,%ymm1,%ymm12
   .byte  196,66,109,184,229                  // vfmadd231ps   %ymm13,%ymm2,%ymm12
-  .byte  196,98,125,24,61,70,164,2,0         // vbroadcastss  0x2a446(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,246,186,2,0        // vbroadcastss  0x2baf6(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,66,53,184,231                   // vfmadd231ps   %ymm15,%ymm9,%ymm12
   .byte  196,65,44,89,206                    // vmulps        %ymm14,%ymm10,%ymm9
   .byte  196,66,61,184,205                   // vfmadd231ps   %ymm13,%ymm8,%ymm9
@@ -17722,7 +17722,7 @@
   .byte  196,193,116,95,206                  // vmaxps        %ymm14,%ymm1,%ymm1
   .byte  196,65,44,95,198                    // vmaxps        %ymm14,%ymm10,%ymm8
   .byte  196,65,124,95,206                   // vmaxps        %ymm14,%ymm0,%ymm9
-  .byte  196,226,125,24,5,252,162,2,0        // vbroadcastss  0x2a2fc(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,172,185,2,0        // vbroadcastss  0x2b9ac(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,124,92,215                      // vsubps        %ymm7,%ymm0,%ymm10
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  197,124,92,219                      // vsubps        %ymm3,%ymm0,%ymm11
@@ -17753,17 +17753,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,179,0,0,0                    // jne           8a49 <_sk_srcover_rgba_8888_hsw+0xd3>
   .byte  196,129,126,111,60,152              // vmovdqu       (%r8,%r11,4),%ymm7
-  .byte  197,197,219,37,124,164,2,0          // vpand         0x2a47c(%rip),%ymm7,%ymm4        # 32e20 <_sk_srcover_rgba_8888_sse2_lowp+0x4a0>
+  .byte  197,197,219,37,28,187,2,0           // vpand         0x2bb1c(%rip),%ymm7,%ymm4        # 344c0 <_sk_srcover_rgba_8888_sse2_lowp+0x490>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,69,0,45,143,164,2,0         // vpshufb       0x2a48f(%rip),%ymm7,%ymm5        # 32e40 <_sk_srcover_rgba_8888_sse2_lowp+0x4c0>
+  .byte  196,226,69,0,45,47,187,2,0          // vpshufb       0x2bb2f(%rip),%ymm7,%ymm5        # 344e0 <_sk_srcover_rgba_8888_sse2_lowp+0x4b0>
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,226,69,0,53,162,164,2,0         // vpshufb       0x2a4a2(%rip),%ymm7,%ymm6        # 32e60 <_sk_srcover_rgba_8888_sse2_lowp+0x4e0>
+  .byte  196,226,69,0,53,66,187,2,0          // vpshufb       0x2bb42(%rip),%ymm7,%ymm6        # 34500 <_sk_srcover_rgba_8888_sse2_lowp+0x4d0>
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
   .byte  197,197,114,215,24                  // vpsrld        $0x18,%ymm7,%ymm7
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  196,98,125,24,5,92,162,2,0          // vbroadcastss  0x2a25c(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,12,185,2,0          // vbroadcastss  0x2b90c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
-  .byte  196,98,125,24,13,123,162,2,0        // vbroadcastss  0x2a27b(%rip),%ymm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,13,43,185,2,0         // vbroadcastss  0x2b92b(%rip),%ymm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,194,93,184,192                  // vfmadd231ps   %ymm8,%ymm4,%ymm0
   .byte  196,193,116,89,201                  // vmulps        %ymm9,%ymm1,%ymm1
@@ -17891,7 +17891,7 @@
 .globl _sk_clamp_1_hsw
 FUNCTION(_sk_clamp_1_hsw)
 _sk_clamp_1_hsw:
-  .byte  196,98,125,24,5,102,160,2,0         // vbroadcastss  0x2a066(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,22,183,2,0          // vbroadcastss  0x2b716(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
@@ -17903,7 +17903,7 @@
 .globl _sk_clamp_a_hsw
 FUNCTION(_sk_clamp_a_hsw)
 _sk_clamp_a_hsw:
-  .byte  196,98,125,24,5,69,160,2,0          // vbroadcastss  0x2a045(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,245,182,2,0         // vbroadcastss  0x2b6f5(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  197,252,93,195                      // vminps        %ymm3,%ymm0,%ymm0
   .byte  197,244,93,203                      // vminps        %ymm3,%ymm1,%ymm1
@@ -17915,7 +17915,7 @@
 .globl _sk_clamp_a_dst_hsw
 FUNCTION(_sk_clamp_a_dst_hsw)
 _sk_clamp_a_dst_hsw:
-  .byte  196,98,125,24,5,39,160,2,0          // vbroadcastss  0x2a027(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,215,182,2,0         // vbroadcastss  0x2b6d7(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,68,93,248                   // vminps        %ymm8,%ymm7,%ymm7
   .byte  197,220,93,231                      // vminps        %ymm7,%ymm4,%ymm4
   .byte  197,212,93,239                      // vminps        %ymm7,%ymm5,%ymm5
@@ -17948,7 +17948,7 @@
 .globl _sk_invert_hsw
 FUNCTION(_sk_invert_hsw)
 _sk_invert_hsw:
-  .byte  196,98,125,24,5,226,159,2,0         // vbroadcastss  0x29fe2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,146,182,2,0         // vbroadcastss  0x2b692(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,188,92,192                      // vsubps        %ymm0,%ymm8,%ymm0
   .byte  197,188,92,201                      // vsubps        %ymm1,%ymm8,%ymm1
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
@@ -18002,9 +18002,9 @@
 .globl _sk_unpremul_hsw
 FUNCTION(_sk_unpremul_hsw)
 _sk_unpremul_hsw:
-  .byte  196,98,125,24,5,125,159,2,0         // vbroadcastss  0x29f7d(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,45,182,2,0          // vbroadcastss  0x2b62d(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,94,195                       // vdivps        %ymm3,%ymm8,%ymm8
-  .byte  196,98,125,24,13,160,159,2,0        // vbroadcastss  0x29fa0(%rip),%ymm9        # 32c60 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
+  .byte  196,98,125,24,13,80,182,2,0         // vbroadcastss  0x2b650(%rip),%ymm9        # 34310 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
   .byte  196,65,60,194,201,1                 // vcmpltps      %ymm9,%ymm8,%ymm9
   .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
   .byte  196,67,45,74,192,144                // vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
@@ -18018,16 +18018,16 @@
 .globl _sk_from_srgb_hsw
 FUNCTION(_sk_from_srgb_hsw)
 _sk_from_srgb_hsw:
-  .byte  196,98,125,24,5,122,159,2,0         // vbroadcastss  0x29f7a(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,42,182,2,0          // vbroadcastss  0x2b62a(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,108,159,2,0        // vbroadcastss  0x29f6c(%rip),%ymm11        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  .byte  196,98,125,24,37,75,159,2,0         // vbroadcastss  0x29f4b(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,28,182,2,0         // vbroadcastss  0x2b61c(%rip),%ymm11        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,37,251,181,2,0        // vbroadcastss  0x2b5fb(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,125,168,235                  // vfmadd213ps   %ymm11,%ymm0,%ymm13
-  .byte  196,98,125,24,53,84,159,2,0         // vbroadcastss  0x29f54(%rip),%ymm14        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,53,4,182,2,0          // vbroadcastss  0x2b604(%rip),%ymm14        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,74,159,2,0         // vbroadcastss  0x29f4a(%rip),%ymm10        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,21,250,181,2,0        // vbroadcastss  0x2b5fa(%rip),%ymm10        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  196,193,124,194,194,1               // vcmpltps      %ymm10,%ymm0,%ymm0
   .byte  196,195,21,74,193,0                 // vblendvps     %ymm0,%ymm9,%ymm13,%ymm0
   .byte  196,65,116,89,200                   // vmulps        %ymm8,%ymm1,%ymm9
@@ -18050,16 +18050,16 @@
 .globl _sk_from_srgb_dst_hsw
 FUNCTION(_sk_from_srgb_dst_hsw)
 _sk_from_srgb_dst_hsw:
-  .byte  196,98,125,24,5,226,158,2,0         // vbroadcastss  0x29ee2(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,146,181,2,0         // vbroadcastss  0x2b592(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,92,89,200                    // vmulps        %ymm8,%ymm4,%ymm9
   .byte  197,92,89,212                       // vmulps        %ymm4,%ymm4,%ymm10
-  .byte  196,98,125,24,29,212,158,2,0        // vbroadcastss  0x29ed4(%rip),%ymm11        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  .byte  196,98,125,24,37,179,158,2,0        // vbroadcastss  0x29eb3(%rip),%ymm12        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,132,181,2,0        // vbroadcastss  0x2b584(%rip),%ymm11        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,37,99,181,2,0         // vbroadcastss  0x2b563(%rip),%ymm12        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,93,168,235                   // vfmadd213ps   %ymm11,%ymm4,%ymm13
-  .byte  196,98,125,24,53,188,158,2,0        // vbroadcastss  0x29ebc(%rip),%ymm14        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,53,108,181,2,0        // vbroadcastss  0x2b56c(%rip),%ymm14        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,66,45,168,238                   // vfmadd213ps   %ymm14,%ymm10,%ymm13
-  .byte  196,98,125,24,21,178,158,2,0        // vbroadcastss  0x29eb2(%rip),%ymm10        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,21,98,181,2,0         // vbroadcastss  0x2b562(%rip),%ymm10        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  196,193,92,194,226,1                // vcmpltps      %ymm10,%ymm4,%ymm4
   .byte  196,195,21,74,225,64                // vblendvps     %ymm4,%ymm9,%ymm13,%ymm4
   .byte  196,65,84,89,200                    // vmulps        %ymm8,%ymm5,%ymm9
@@ -18083,19 +18083,19 @@
 FUNCTION(_sk_to_srgb_hsw)
 _sk_to_srgb_hsw:
   .byte  197,124,82,200                      // vrsqrtps      %ymm0,%ymm9
-  .byte  196,98,125,24,5,86,158,2,0          // vbroadcastss  0x29e56(%rip),%ymm8        # 32c74 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
+  .byte  196,98,125,24,5,6,181,2,0           // vbroadcastss  0x2b506(%rip),%ymm8        # 34324 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
   .byte  196,65,124,89,208                   // vmulps        %ymm8,%ymm0,%ymm10
-  .byte  196,98,125,24,29,76,158,2,0         // vbroadcastss  0x29e4c(%rip),%ymm11        # 32c78 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
-  .byte  196,98,125,24,37,71,158,2,0         // vbroadcastss  0x29e47(%rip),%ymm12        # 32c7c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
+  .byte  196,98,125,24,29,252,180,2,0        // vbroadcastss  0x2b4fc(%rip),%ymm11        # 34328 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
+  .byte  196,98,125,24,37,247,180,2,0        // vbroadcastss  0x2b4f7(%rip),%ymm12        # 3432c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
   .byte  196,65,124,40,236                   // vmovaps       %ymm12,%ymm13
   .byte  196,66,53,168,235                   // vfmadd213ps   %ymm11,%ymm9,%ymm13
-  .byte  196,98,125,24,53,104,159,2,0        // vbroadcastss  0x29f68(%rip),%ymm14        # 32db0 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
+  .byte  196,98,125,24,53,24,182,2,0         // vbroadcastss  0x2b618(%rip),%ymm14        # 34460 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
   .byte  196,66,53,168,238                   // vfmadd213ps   %ymm14,%ymm9,%ymm13
-  .byte  196,98,125,24,61,94,159,2,0         // vbroadcastss  0x29f5e(%rip),%ymm15        # 32db4 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
+  .byte  196,98,125,24,61,14,182,2,0         // vbroadcastss  0x2b60e(%rip),%ymm15        # 34464 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
   .byte  196,65,52,88,207                    // vaddps        %ymm15,%ymm9,%ymm9
   .byte  196,65,124,83,201                   // vrcpps        %ymm9,%ymm9
   .byte  196,65,20,89,201                    // vmulps        %ymm9,%ymm13,%ymm9
-  .byte  196,98,125,24,45,26,158,2,0         // vbroadcastss  0x29e1a(%rip),%ymm13        # 32c88 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
+  .byte  196,98,125,24,45,202,180,2,0        // vbroadcastss  0x2b4ca(%rip),%ymm13        # 34338 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
   .byte  196,193,124,194,197,1               // vcmpltps      %ymm13,%ymm0,%ymm0
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -18129,26 +18129,26 @@
   .byte  197,116,93,202                      // vminps        %ymm2,%ymm1,%ymm9
   .byte  196,65,124,93,201                   // vminps        %ymm9,%ymm0,%ymm9
   .byte  196,65,60,92,209                    // vsubps        %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,29,49,157,2,0         // vbroadcastss  0x29d31(%rip),%ymm11        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,29,225,179,2,0        // vbroadcastss  0x2b3e1(%rip),%ymm11        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,65,36,94,218                    // vdivps        %ymm10,%ymm11,%ymm11
   .byte  197,116,92,226                      // vsubps        %ymm2,%ymm1,%ymm12
   .byte  197,116,194,234,1                   // vcmpltps      %ymm2,%ymm1,%ymm13
-  .byte  196,98,125,24,53,118,157,2,0        // vbroadcastss  0x29d76(%rip),%ymm14        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,98,125,24,53,38,180,2,0         // vbroadcastss  0x2b426(%rip),%ymm14        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  196,65,4,87,255                     // vxorps        %ymm15,%ymm15,%ymm15
   .byte  196,67,5,74,238,208                 // vblendvps     %ymm13,%ymm14,%ymm15,%ymm13
   .byte  196,66,37,168,229                   // vfmadd213ps   %ymm13,%ymm11,%ymm12
   .byte  197,236,92,208                      // vsubps        %ymm0,%ymm2,%ymm2
   .byte  197,124,92,233                      // vsubps        %ymm1,%ymm0,%ymm13
-  .byte  196,98,125,24,53,93,157,2,0         // vbroadcastss  0x29d5d(%rip),%ymm14        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  196,98,125,24,53,13,180,2,0         // vbroadcastss  0x2b40d(%rip),%ymm14        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  196,66,37,168,238                   // vfmadd213ps   %ymm14,%ymm11,%ymm13
-  .byte  196,98,125,24,53,75,157,2,0         // vbroadcastss  0x29d4b(%rip),%ymm14        # 32c90 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  .byte  196,98,125,24,53,251,179,2,0        // vbroadcastss  0x2b3fb(%rip),%ymm14        # 34340 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   .byte  196,194,37,168,214                  // vfmadd213ps   %ymm14,%ymm11,%ymm2
   .byte  197,188,194,201,0                   // vcmpeqps      %ymm1,%ymm8,%ymm1
   .byte  196,227,21,74,202,16                // vblendvps     %ymm1,%ymm2,%ymm13,%ymm1
   .byte  197,188,194,192,0                   // vcmpeqps      %ymm0,%ymm8,%ymm0
   .byte  196,195,117,74,196,0                // vblendvps     %ymm0,%ymm12,%ymm1,%ymm0
   .byte  196,193,60,88,201                   // vaddps        %ymm9,%ymm8,%ymm1
-  .byte  196,98,125,24,29,190,156,2,0        // vbroadcastss  0x29cbe(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,110,179,2,0        // vbroadcastss  0x2b36e(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,116,89,211                  // vmulps        %ymm11,%ymm1,%ymm2
   .byte  197,36,194,218,1                    // vcmpltps      %ymm2,%ymm11,%ymm11
   .byte  196,65,12,92,224                    // vsubps        %ymm8,%ymm14,%ymm12
@@ -18158,7 +18158,7 @@
   .byte  197,172,94,201                      // vdivps        %ymm1,%ymm10,%ymm1
   .byte  196,195,125,74,199,128              // vblendvps     %ymm8,%ymm15,%ymm0,%ymm0
   .byte  196,195,117,74,207,128              // vblendvps     %ymm8,%ymm15,%ymm1,%ymm1
-  .byte  196,98,125,24,5,241,156,2,0         // vbroadcastss  0x29cf1(%rip),%ymm8        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  196,98,125,24,5,161,179,2,0         // vbroadcastss  0x2b3a1(%rip),%ymm8        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -18175,30 +18175,30 @@
   .byte  197,252,17,92,36,128                // vmovups       %ymm3,-0x80(%rsp)
   .byte  197,252,40,233                      // vmovaps       %ymm1,%ymm5
   .byte  197,252,40,224                      // vmovaps       %ymm0,%ymm4
-  .byte  196,98,125,24,5,74,156,2,0          // vbroadcastss  0x29c4a(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,250,178,2,0         // vbroadcastss  0x2b2fa(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,60,194,202,2                    // vcmpleps      %ymm2,%ymm8,%ymm9
   .byte  197,84,89,210                       // vmulps        %ymm2,%ymm5,%ymm10
   .byte  196,65,84,92,218                    // vsubps        %ymm10,%ymm5,%ymm11
   .byte  196,67,45,74,203,144                // vblendvps     %ymm9,%ymm11,%ymm10,%ymm9
   .byte  197,52,88,210                       // vaddps        %ymm2,%ymm9,%ymm10
-  .byte  196,98,125,24,13,141,156,2,0        // vbroadcastss  0x29c8d(%rip),%ymm9        # 32c90 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  .byte  196,98,125,24,13,61,179,2,0         // vbroadcastss  0x2b33d(%rip),%ymm9        # 34340 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   .byte  196,66,109,170,202                  // vfmsub213ps   %ymm10,%ymm2,%ymm9
-  .byte  196,98,125,24,29,139,156,2,0        // vbroadcastss  0x29c8b(%rip),%ymm11        # 32c9c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
+  .byte  196,98,125,24,29,59,179,2,0         // vbroadcastss  0x2b33b(%rip),%ymm11        # 3434c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
   .byte  196,65,92,88,219                    // vaddps        %ymm11,%ymm4,%ymm11
   .byte  196,67,125,8,227,1                  // vroundps      $0x1,%ymm11,%ymm12
   .byte  196,65,36,92,252                    // vsubps        %ymm12,%ymm11,%ymm15
   .byte  196,65,44,92,217                    // vsubps        %ymm9,%ymm10,%ymm11
-  .byte  196,98,125,24,45,93,156,2,0         // vbroadcastss  0x29c5d(%rip),%ymm13        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,98,125,24,45,13,179,2,0         // vbroadcastss  0x2b30d(%rip),%ymm13        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  196,193,4,89,197                    // vmulps        %ymm13,%ymm15,%ymm0
-  .byte  196,98,125,24,53,87,156,2,0         // vbroadcastss  0x29c57(%rip),%ymm14        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  196,98,125,24,53,7,179,2,0          // vbroadcastss  0x2b307(%rip),%ymm14        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  197,12,92,224                       // vsubps        %ymm0,%ymm14,%ymm12
   .byte  196,66,37,168,225                   // vfmadd213ps   %ymm9,%ymm11,%ymm12
-  .byte  196,226,125,24,29,81,156,2,0        // vbroadcastss  0x29c51(%rip),%ymm3        # 32ca0 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
+  .byte  196,226,125,24,29,1,179,2,0         // vbroadcastss  0x2b301(%rip),%ymm3        # 34350 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
   .byte  196,193,100,194,255,2               // vcmpleps      %ymm15,%ymm3,%ymm7
   .byte  196,195,29,74,249,112               // vblendvps     %ymm7,%ymm9,%ymm12,%ymm7
   .byte  196,65,60,194,231,2                 // vcmpleps      %ymm15,%ymm8,%ymm12
   .byte  196,227,45,74,255,192               // vblendvps     %ymm12,%ymm7,%ymm10,%ymm7
-  .byte  196,98,125,24,37,40,156,2,0         // vbroadcastss  0x29c28(%rip),%ymm12        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  196,98,125,24,37,216,178,2,0        // vbroadcastss  0x2b2d8(%rip),%ymm12        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  196,65,28,194,255,2                 // vcmpleps      %ymm15,%ymm12,%ymm15
   .byte  196,194,37,168,193                  // vfmadd213ps   %ymm9,%ymm11,%ymm0
   .byte  196,99,125,74,255,240               // vblendvps     %ymm15,%ymm7,%ymm0,%ymm15
@@ -18214,7 +18214,7 @@
   .byte  197,156,194,192,2                   // vcmpleps      %ymm0,%ymm12,%ymm0
   .byte  196,194,37,168,249                  // vfmadd213ps   %ymm9,%ymm11,%ymm7
   .byte  196,227,69,74,201,0                 // vblendvps     %ymm0,%ymm1,%ymm7,%ymm1
-  .byte  196,226,125,24,5,220,155,2,0        // vbroadcastss  0x29bdc(%rip),%ymm0        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,226,125,24,5,140,178,2,0        // vbroadcastss  0x2b28c(%rip),%ymm0        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  197,220,88,192                      // vaddps        %ymm0,%ymm4,%ymm0
   .byte  196,227,125,8,224,1                 // vroundps      $0x1,%ymm0,%ymm4
   .byte  197,252,92,196                      // vsubps        %ymm4,%ymm0,%ymm0
@@ -18268,10 +18268,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,58                              // jne           91b7 <_sk_scale_u8_hsw+0x52>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,229,161,2,0            // vpand         0x2a1e5(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,197,184,2,0            // vpand         0x2b8c5(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,10,155,2,0         // vbroadcastss  0x29b0a(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,186,177,2,0        // vbroadcastss  0x2b1ba(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
@@ -18348,20 +18348,20 @@
   .byte  15,133,161,0,0,0                    // jne           932c <_sk_scale_565_hsw+0xc0>
   .byte  196,1,122,111,4,88                  // vmovdqu       (%r8,%r11,2),%xmm8
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
-  .byte  196,98,125,88,13,13,154,2,0         // vpbroadcastd  0x29a0d(%rip),%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,98,125,88,13,189,176,2,0        // vpbroadcastd  0x2b0bd(%rip),%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,61,219,201                   // vpand         %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,254,153,2,0        // vbroadcastss  0x299fe(%rip),%ymm10        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,98,125,24,21,174,176,2,0        // vbroadcastss  0x2b0ae(%rip),%ymm10        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,88,21,244,153,2,0        // vpbroadcastd  0x299f4(%rip),%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,98,125,88,21,164,176,2,0        // vpbroadcastd  0x2b0a4(%rip),%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,229,153,2,0        // vbroadcastss  0x299e5(%rip),%ymm11        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,98,125,24,29,149,176,2,0        // vbroadcastss  0x2b095(%rip),%ymm11        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,88,29,219,153,2,0        // vpbroadcastd  0x299db(%rip),%ymm11        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,98,125,88,29,139,176,2,0        // vpbroadcastd  0x2b08b(%rip),%ymm11        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,61,219,195                   // vpand         %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,204,153,2,0        // vbroadcastss  0x299cc(%rip),%ymm11        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,98,125,24,29,124,176,2,0        // vbroadcastss  0x2b07c(%rip),%ymm11        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  197,100,194,223,1                   // vcmpltps      %ymm7,%ymm3,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -18452,10 +18452,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,78                              // jne           9465 <_sk_lerp_u8_hsw+0x66>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,75,159,2,0             // vpand         0x29f4b(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,43,182,2,0             // vpand         0x2b62b(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,112,152,2,0        // vbroadcastss  0x29870(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,32,175,2,0         // vbroadcastss  0x2af20(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,252,92,196                      // vsubps        %ymm4,%ymm0,%ymm0
   .byte  196,226,61,168,196                  // vfmadd213ps   %ymm4,%ymm8,%ymm0
@@ -18507,7 +18507,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,221                 // callq         ffffffffde009510 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd6b90>
+  .byte  232,255,255,255,221                 // callq         ffffffffde009510 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd54e0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,210                             // callq         *%rdx
@@ -18533,20 +18533,20 @@
   .byte  15,133,181,0,0,0                    // jne           95f0 <_sk_lerp_565_hsw+0xd4>
   .byte  196,1,122,111,4,88                  // vmovdqu       (%r8,%r11,2),%xmm8
   .byte  196,66,125,51,192                   // vpmovzxwd     %xmm8,%ymm8
-  .byte  196,98,125,88,13,93,151,2,0         // vpbroadcastd  0x2975d(%rip),%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,98,125,88,13,13,174,2,0         // vpbroadcastd  0x2ae0d(%rip),%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,61,219,201                   // vpand         %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,78,151,2,0         // vbroadcastss  0x2974e(%rip),%ymm10        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,98,125,24,21,254,173,2,0        // vbroadcastss  0x2adfe(%rip),%ymm10        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,88,21,68,151,2,0         // vpbroadcastd  0x29744(%rip),%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,98,125,88,21,244,173,2,0        // vpbroadcastd  0x2adf4(%rip),%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,61,219,210                   // vpand         %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,53,151,2,0         // vbroadcastss  0x29735(%rip),%ymm11        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,98,125,24,29,229,173,2,0        // vbroadcastss  0x2ade5(%rip),%ymm11        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,88,29,43,151,2,0         // vpbroadcastd  0x2972b(%rip),%ymm11        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,98,125,88,29,219,173,2,0        // vpbroadcastd  0x2addb(%rip),%ymm11        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,61,219,195                   // vpand         %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,28,151,2,0         // vbroadcastss  0x2971c(%rip),%ymm11        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,98,125,24,29,204,173,2,0        // vbroadcastss  0x2adcc(%rip),%ymm11        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  197,100,194,223,1                   // vcmpltps      %ymm7,%ymm3,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -18620,7 +18620,7 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,116                             // jne           9712 <_sk_load_tables_hsw+0x7e>
   .byte  196,193,126,111,28,144              // vmovdqu       (%r8,%rdx,4),%ymm3
-  .byte  197,229,219,13,212,151,2,0          // vpand         0x297d4(%rip),%ymm3,%ymm1        # 32e80 <_sk_srcover_rgba_8888_sse2_lowp+0x500>
+  .byte  197,229,219,13,116,174,2,0          // vpand         0x2ae74(%rip),%ymm3,%ymm1        # 34520 <_sk_srcover_rgba_8888_sse2_lowp+0x4f0>
   .byte  196,65,61,118,192                   // vpcmpeqd      %ymm8,%ymm8,%ymm8
   .byte  76,139,64,8                         // mov           0x8(%rax),%r8
   .byte  76,139,72,16                        // mov           0x10(%rax),%r9
@@ -18628,16 +18628,16 @@
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
   .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
   .byte  196,194,53,146,4,136                // vgatherdps    %ymm9,(%r8,%ymm1,4),%ymm0
-  .byte  196,98,101,0,13,203,151,2,0         // vpshufb       0x297cb(%rip),%ymm3,%ymm9        # 32ea0 <_sk_srcover_rgba_8888_sse2_lowp+0x520>
+  .byte  196,98,101,0,13,107,174,2,0         // vpshufb       0x2ae6b(%rip),%ymm3,%ymm9        # 34540 <_sk_srcover_rgba_8888_sse2_lowp+0x510>
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  196,65,45,118,210                   // vpcmpeqd      %ymm10,%ymm10,%ymm10
   .byte  196,130,45,146,12,137               // vgatherdps    %ymm10,(%r9,%ymm9,4),%ymm1
   .byte  72,139,64,24                        // mov           0x18(%rax),%rax
-  .byte  196,98,101,0,13,207,151,2,0         // vpshufb       0x297cf(%rip),%ymm3,%ymm9        # 32ec0 <_sk_srcover_rgba_8888_sse2_lowp+0x540>
+  .byte  196,98,101,0,13,111,174,2,0         // vpshufb       0x2ae6f(%rip),%ymm3,%ymm9        # 34560 <_sk_srcover_rgba_8888_sse2_lowp+0x530>
   .byte  196,162,61,146,20,136               // vgatherdps    %ymm8,(%rax,%ymm9,4),%ymm2
   .byte  197,229,114,211,24                  // vpsrld        $0x18,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,159,149,2,0         // vbroadcastss  0x2959f(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,79,172,2,0          // vbroadcastss  0x2ac4f(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -18718,7 +18718,7 @@
   .byte  197,185,108,200                     // vpunpcklqdq   %xmm0,%xmm8,%xmm1
   .byte  197,57,109,192                      // vpunpckhqdq   %xmm0,%xmm8,%xmm8
   .byte  196,65,49,108,214                   // vpunpcklqdq   %xmm14,%xmm9,%xmm10
-  .byte  197,121,111,29,44,155,2,0           // vmovdqa       0x29b2c(%rip),%xmm11        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,121,111,29,12,178,2,0           // vmovdqa       0x2b20c(%rip),%xmm11        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,193,113,219,195                 // vpand         %xmm11,%xmm1,%xmm0
   .byte  196,226,125,51,200                  // vpmovzxwd     %xmm0,%ymm1
   .byte  196,65,29,118,228                   // vpcmpeqd      %ymm12,%ymm12,%ymm12
@@ -18743,7 +18743,7 @@
   .byte  197,185,235,219                     // vpor          %xmm3,%xmm8,%xmm3
   .byte  196,226,125,51,219                  // vpmovzxwd     %xmm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,4,148,2,0           // vbroadcastss  0x29404(%rip),%ymm8        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,5,180,170,2,0         // vbroadcastss  0x2aab4(%rip),%ymm8        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -18803,7 +18803,7 @@
   .byte  197,105,108,192                     // vpunpcklqdq   %xmm0,%xmm2,%xmm8
   .byte  197,241,108,195                     // vpunpcklqdq   %xmm3,%xmm1,%xmm0
   .byte  197,241,109,203                     // vpunpckhqdq   %xmm3,%xmm1,%xmm1
-  .byte  197,249,111,29,172,153,2,0          // vmovdqa       0x299ac(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,111,29,140,176,2,0          // vmovdqa       0x2b08c(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  197,249,219,195                     // vpand         %xmm3,%xmm0,%xmm0
   .byte  196,98,125,51,200                   // vpmovzxwd     %xmm0,%ymm9
   .byte  196,65,45,118,210                   // vpcmpeqd      %ymm10,%ymm10,%ymm10
@@ -18823,7 +18823,7 @@
   .byte  196,226,125,51,219                  // vpmovzxwd     %xmm3,%ymm3
   .byte  196,226,45,146,20,152               // vgatherdps    %ymm10,(%rax,%ymm3,4),%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,13,146,2,0        // vbroadcastss  0x2920d(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,189,168,2,0       // vbroadcastss  0x2a8bd(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,129,121,110,4,72                // vmovd         (%r8,%r9,2),%xmm0
   .byte  196,65,49,239,201                   // vpxor         %xmm9,%xmm9,%xmm9
@@ -18892,7 +18892,7 @@
 FUNCTION(_sk_byte_tables_hsw)
 _sk_byte_tables_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,2,145,2,0           // vbroadcastss  0x29102(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,178,167,2,0         // vbroadcastss  0x2a7b2(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,125,91,200                      // vcvtps2dq     %ymm0,%ymm9
   .byte  196,65,249,126,200                  // vmovq         %xmm9,%r8
@@ -19022,7 +19022,7 @@
   .byte  67,15,182,4,17                      // movzbl        (%r9,%r10,1),%eax
   .byte  196,194,125,49,193                  // vpmovzxbd     %xmm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,212,142,2,0         // vbroadcastss  0x28ed4(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,132,165,2,0         // vbroadcastss  0x2a584(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,226,125,49,201                  // vpmovzxbd     %xmm1,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
@@ -19144,7 +19144,7 @@
   .byte  67,15,182,4,17                      // movzbl        (%r9,%r10,1),%eax
   .byte  196,194,125,49,193                  // vpmovzxbd     %xmm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,168,140,2,0         // vbroadcastss  0x28ca8(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,88,163,2,0          // vbroadcastss  0x2a358(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,226,125,49,201                  // vpmovzxbd     %xmm1,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
@@ -19247,33 +19247,33 @@
   .byte  196,66,125,168,211                  // vfmadd213ps   %ymm11,%ymm0,%ymm10
   .byte  196,226,125,24,0                    // vbroadcastss  (%rax),%ymm0
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,142,139,2,0        // vbroadcastss  0x28b8e(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,98,125,24,45,137,139,2,0        // vbroadcastss  0x28b89(%rip),%ymm13        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,62,162,2,0         // vbroadcastss  0x2a23e(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,45,57,162,2,0         // vbroadcastss  0x2a239(%rip),%ymm13        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,219,138,2,0        // vbroadcastss  0x28adb(%rip),%ymm13        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,45,139,161,2,0        // vbroadcastss  0x2a18b(%rip),%ymm13        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,113,139,2,0        // vbroadcastss  0x28b71(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,33,162,2,0         // vbroadcastss  0x2a221(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,103,139,2,0        // vbroadcastss  0x28b67(%rip),%ymm11        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,29,23,162,2,0         // vbroadcastss  0x2a217(%rip),%ymm11        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,93,139,2,0         // vbroadcastss  0x28b5d(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,13,162,2,0         // vbroadcastss  0x2a20d(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,83,139,2,0         // vbroadcastss  0x28b53(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,3,162,2,0          // vbroadcastss  0x2a203(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,28,94,210                    // vdivps        %ymm10,%ymm12,%ymm10
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  196,99,125,8,208,1                  // vroundps      $0x1,%ymm0,%ymm10
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,52,139,2,0         // vbroadcastss  0x28b34(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,228,161,2,0        // vbroadcastss  0x2a1e4(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,42,139,2,0         // vbroadcastss  0x28b2a(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,218,161,2,0        // vbroadcastss  0x2a1da(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,98,45,172,216                   // vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  .byte  196,226,125,24,5,32,139,2,0         // vbroadcastss  0x28b20(%rip),%ymm0        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,5,208,161,2,0        // vbroadcastss  0x2a1d0(%rip),%ymm0        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,193,124,92,194                  // vsubps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,22,139,2,0         // vbroadcastss  0x28b16(%rip),%ymm10        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,21,198,161,2,0        // vbroadcastss  0x2a1c6(%rip),%ymm10        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  197,172,94,192                      // vdivps        %ymm0,%ymm10,%ymm0
   .byte  197,164,88,192                      // vaddps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,21,9,139,2,0          // vbroadcastss  0x28b09(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,185,161,2,0        // vbroadcastss  0x2a1b9(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -19281,7 +19281,7 @@
   .byte  196,195,125,74,193,128              // vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,29,138,2,0          // vbroadcastss  0x28a1d(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,205,160,2,0         // vbroadcastss  0x2a0cd(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -19301,33 +19301,33 @@
   .byte  196,66,117,168,211                  // vfmadd213ps   %ymm11,%ymm1,%ymm10
   .byte  196,226,125,24,8                    // vbroadcastss  (%rax),%ymm1
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,105,138,2,0        // vbroadcastss  0x28a69(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,98,125,24,45,100,138,2,0        // vbroadcastss  0x28a64(%rip),%ymm13        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,25,161,2,0         // vbroadcastss  0x2a119(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,45,20,161,2,0         // vbroadcastss  0x2a114(%rip),%ymm13        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,182,137,2,0        // vbroadcastss  0x289b6(%rip),%ymm13        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,45,102,160,2,0        // vbroadcastss  0x2a066(%rip),%ymm13        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,76,138,2,0         // vbroadcastss  0x28a4c(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,252,160,2,0        // vbroadcastss  0x2a0fc(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,66,138,2,0         // vbroadcastss  0x28a42(%rip),%ymm11        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,29,242,160,2,0        // vbroadcastss  0x2a0f2(%rip),%ymm11        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,56,138,2,0         // vbroadcastss  0x28a38(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,232,160,2,0        // vbroadcastss  0x2a0e8(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,46,138,2,0         // vbroadcastss  0x28a2e(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,222,160,2,0        // vbroadcastss  0x2a0de(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,28,94,210                    // vdivps        %ymm10,%ymm12,%ymm10
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
   .byte  196,193,116,89,202                  // vmulps        %ymm10,%ymm1,%ymm1
   .byte  196,99,125,8,209,1                  // vroundps      $0x1,%ymm1,%ymm10
   .byte  196,65,116,92,210                   // vsubps        %ymm10,%ymm1,%ymm10
-  .byte  196,98,125,24,29,15,138,2,0         // vbroadcastss  0x28a0f(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,191,160,2,0        // vbroadcastss  0x2a0bf(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,5,138,2,0          // vbroadcastss  0x28a05(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,181,160,2,0        // vbroadcastss  0x2a0b5(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,98,45,172,217                   // vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  .byte  196,226,125,24,13,251,137,2,0       // vbroadcastss  0x289fb(%rip),%ymm1        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,13,171,160,2,0       // vbroadcastss  0x2a0ab(%rip),%ymm1        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,193,116,92,202                  // vsubps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,241,137,2,0        // vbroadcastss  0x289f1(%rip),%ymm10        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,21,161,160,2,0        // vbroadcastss  0x2a0a1(%rip),%ymm10        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  197,172,94,201                      // vdivps        %ymm1,%ymm10,%ymm1
   .byte  197,164,88,201                      // vaddps        %ymm1,%ymm11,%ymm1
-  .byte  196,98,125,24,21,228,137,2,0        // vbroadcastss  0x289e4(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,148,160,2,0        // vbroadcastss  0x2a094(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,116,89,202                  // vmulps        %ymm10,%ymm1,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -19335,7 +19335,7 @@
   .byte  196,195,117,74,201,128              // vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,116,95,200                  // vmaxps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,5,248,136,2,0         // vbroadcastss  0x288f8(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,168,159,2,0         // vbroadcastss  0x29fa8(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -19355,33 +19355,33 @@
   .byte  196,66,109,168,211                  // vfmadd213ps   %ymm11,%ymm2,%ymm10
   .byte  196,226,125,24,16                   // vbroadcastss  (%rax),%ymm2
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,68,137,2,0         // vbroadcastss  0x28944(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,98,125,24,45,63,137,2,0         // vbroadcastss  0x2893f(%rip),%ymm13        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,244,159,2,0        // vbroadcastss  0x29ff4(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,45,239,159,2,0        // vbroadcastss  0x29fef(%rip),%ymm13        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,145,136,2,0        // vbroadcastss  0x28891(%rip),%ymm13        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,45,65,159,2,0         // vbroadcastss  0x29f41(%rip),%ymm13        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,39,137,2,0         // vbroadcastss  0x28927(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,215,159,2,0        // vbroadcastss  0x29fd7(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,29,137,2,0         // vbroadcastss  0x2891d(%rip),%ymm11        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,29,205,159,2,0        // vbroadcastss  0x29fcd(%rip),%ymm11        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,19,137,2,0         // vbroadcastss  0x28913(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,195,159,2,0        // vbroadcastss  0x29fc3(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,9,137,2,0          // vbroadcastss  0x28909(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,185,159,2,0        // vbroadcastss  0x29fb9(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,28,94,210                    // vdivps        %ymm10,%ymm12,%ymm10
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
   .byte  196,193,108,89,210                  // vmulps        %ymm10,%ymm2,%ymm2
   .byte  196,99,125,8,210,1                  // vroundps      $0x1,%ymm2,%ymm10
   .byte  196,65,108,92,210                   // vsubps        %ymm10,%ymm2,%ymm10
-  .byte  196,98,125,24,29,234,136,2,0        // vbroadcastss  0x288ea(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,154,159,2,0        // vbroadcastss  0x29f9a(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,224,136,2,0        // vbroadcastss  0x288e0(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,144,159,2,0        // vbroadcastss  0x29f90(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,98,45,172,218                   // vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  .byte  196,226,125,24,21,214,136,2,0       // vbroadcastss  0x288d6(%rip),%ymm2        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,21,134,159,2,0       // vbroadcastss  0x29f86(%rip),%ymm2        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,193,108,92,210                  // vsubps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,204,136,2,0        // vbroadcastss  0x288cc(%rip),%ymm10        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,21,124,159,2,0        // vbroadcastss  0x29f7c(%rip),%ymm10        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  197,172,94,210                      // vdivps        %ymm2,%ymm10,%ymm2
   .byte  197,164,88,210                      // vaddps        %ymm2,%ymm11,%ymm2
-  .byte  196,98,125,24,21,191,136,2,0        // vbroadcastss  0x288bf(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,111,159,2,0        // vbroadcastss  0x29f6f(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,108,89,210                  // vmulps        %ymm10,%ymm2,%ymm2
   .byte  197,253,91,210                      // vcvtps2dq     %ymm2,%ymm2
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -19389,7 +19389,7 @@
   .byte  196,195,109,74,209,128              // vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,211,135,2,0         // vbroadcastss  0x287d3(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,131,158,2,0         // vbroadcastss  0x29e83(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -19409,33 +19409,33 @@
   .byte  196,66,101,168,211                  // vfmadd213ps   %ymm11,%ymm3,%ymm10
   .byte  196,226,125,24,24                   // vbroadcastss  (%rax),%ymm3
   .byte  196,65,124,91,218                   // vcvtdq2ps     %ymm10,%ymm11
-  .byte  196,98,125,24,37,31,136,2,0         // vbroadcastss  0x2881f(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,98,125,24,45,26,136,2,0         // vbroadcastss  0x2881a(%rip),%ymm13        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,207,158,2,0        // vbroadcastss  0x29ecf(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,45,202,158,2,0        // vbroadcastss  0x29eca(%rip),%ymm13        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,65,44,84,213                    // vandps        %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,108,135,2,0        // vbroadcastss  0x2876c(%rip),%ymm13        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,45,28,158,2,0         // vbroadcastss  0x29e1c(%rip),%ymm13        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,86,213                    // vorps         %ymm13,%ymm10,%ymm10
-  .byte  196,98,125,24,45,2,136,2,0          // vbroadcastss  0x28802(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,178,158,2,0        // vbroadcastss  0x29eb2(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,37,184,236                   // vfmadd231ps   %ymm12,%ymm11,%ymm13
-  .byte  196,98,125,24,29,248,135,2,0        // vbroadcastss  0x287f8(%rip),%ymm11        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,29,168,158,2,0        // vbroadcastss  0x29ea8(%rip),%ymm11        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,45,172,221                   // vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  .byte  196,98,125,24,37,238,135,2,0        // vbroadcastss  0x287ee(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,158,158,2,0        // vbroadcastss  0x29e9e(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,228,135,2,0        // vbroadcastss  0x287e4(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,148,158,2,0        // vbroadcastss  0x29e94(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,28,94,210                    // vdivps        %ymm10,%ymm12,%ymm10
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
   .byte  196,193,100,89,218                  // vmulps        %ymm10,%ymm3,%ymm3
   .byte  196,99,125,8,211,1                  // vroundps      $0x1,%ymm3,%ymm10
   .byte  196,65,100,92,210                   // vsubps        %ymm10,%ymm3,%ymm10
-  .byte  196,98,125,24,29,197,135,2,0        // vbroadcastss  0x287c5(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,117,158,2,0        // vbroadcastss  0x29e75(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,187,135,2,0        // vbroadcastss  0x287bb(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,107,158,2,0        // vbroadcastss  0x29e6b(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,98,45,172,219                   // vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  .byte  196,226,125,24,29,177,135,2,0       // vbroadcastss  0x287b1(%rip),%ymm3        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,29,97,158,2,0        // vbroadcastss  0x29e61(%rip),%ymm3        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,193,100,92,218                  // vsubps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,167,135,2,0        // vbroadcastss  0x287a7(%rip),%ymm10        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,21,87,158,2,0         // vbroadcastss  0x29e57(%rip),%ymm10        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  197,172,94,219                      // vdivps        %ymm3,%ymm10,%ymm3
   .byte  197,164,88,219                      // vaddps        %ymm3,%ymm11,%ymm3
-  .byte  196,98,125,24,21,154,135,2,0        // vbroadcastss  0x2879a(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,74,158,2,0         // vbroadcastss  0x29e4a(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,100,89,218                  // vmulps        %ymm10,%ymm3,%ymm3
   .byte  197,253,91,219                      // vcvtps2dq     %ymm3,%ymm3
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -19443,7 +19443,7 @@
   .byte  196,195,101,74,217,128              // vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,100,95,216                  // vmaxps        %ymm8,%ymm3,%ymm3
-  .byte  196,98,125,24,5,174,134,2,0         // vbroadcastss  0x286ae(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,94,157,2,0          // vbroadcastss  0x29d5e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -19462,35 +19462,35 @@
   .byte  197,252,40,241                      // vmovaps       %ymm1,%ymm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,124,91,208                      // vcvtdq2ps     %ymm0,%ymm10
-  .byte  196,98,125,24,29,5,135,2,0          // vbroadcastss  0x28705(%rip),%ymm11        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  .byte  196,226,125,24,45,0,135,2,0         // vbroadcastss  0x28700(%rip),%ymm5        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,29,181,157,2,0        // vbroadcastss  0x29db5(%rip),%ymm11        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,226,125,24,45,176,157,2,0       // vbroadcastss  0x29db0(%rip),%ymm5        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  197,124,84,205                      // vandps        %ymm5,%ymm0,%ymm9
-  .byte  196,226,125,24,37,83,134,2,0        // vbroadcastss  0x28653(%rip),%ymm4        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,37,3,157,2,0         // vbroadcastss  0x29d03(%rip),%ymm4        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,52,86,228                       // vorps         %ymm4,%ymm9,%ymm12
-  .byte  196,98,125,24,45,234,134,2,0        // vbroadcastss  0x286ea(%rip),%ymm13        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,45,154,157,2,0        // vbroadcastss  0x29d9a(%rip),%ymm13        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,66,37,168,213                   // vfmadd213ps   %ymm13,%ymm11,%ymm10
-  .byte  196,98,125,24,13,224,134,2,0        // vbroadcastss  0x286e0(%rip),%ymm9        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,13,144,157,2,0        // vbroadcastss  0x29d90(%rip),%ymm9        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,66,29,188,209                   // vfnmadd231ps  %ymm9,%ymm12,%ymm10
   .byte  197,124,91,246                      // vcvtdq2ps     %ymm6,%ymm14
   .byte  196,66,37,168,245                   // vfmadd213ps   %ymm13,%ymm11,%ymm14
   .byte  197,124,91,255                      // vcvtdq2ps     %ymm7,%ymm15
   .byte  196,66,37,168,253                   // vfmadd213ps   %ymm13,%ymm11,%ymm15
-  .byte  196,98,125,24,29,196,134,2,0        // vbroadcastss  0x286c4(%rip),%ymm11        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,29,116,157,2,0        // vbroadcastss  0x29d74(%rip),%ymm11        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,65,28,88,227                    // vaddps        %ymm11,%ymm12,%ymm12
-  .byte  196,98,125,24,45,186,134,2,0        // vbroadcastss  0x286ba(%rip),%ymm13        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,45,106,157,2,0        // vbroadcastss  0x29d6a(%rip),%ymm13        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  196,65,20,94,228                    // vdivps        %ymm12,%ymm13,%ymm12
   .byte  196,65,44,92,212                    // vsubps        %ymm12,%ymm10,%ymm10
   .byte  196,98,125,24,32                    // vbroadcastss  (%rax),%ymm12
   .byte  196,65,44,89,212                    // vmulps        %ymm12,%ymm10,%ymm10
   .byte  196,67,125,8,194,1                  // vroundps      $0x1,%ymm10,%ymm8
   .byte  196,65,44,92,192                    // vsubps        %ymm8,%ymm10,%ymm8
-  .byte  196,226,125,24,21,150,134,2,0       // vbroadcastss  0x28696(%rip),%ymm2        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,226,125,24,21,70,157,2,0        // vbroadcastss  0x29d46(%rip),%ymm2        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  197,44,88,210                       // vaddps        %ymm2,%ymm10,%ymm10
-  .byte  196,226,125,24,29,141,134,2,0       // vbroadcastss  0x2868d(%rip),%ymm3        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,226,125,24,29,61,157,2,0        // vbroadcastss  0x29d3d(%rip),%ymm3        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,98,61,188,211                   // vfnmadd231ps  %ymm3,%ymm8,%ymm10
-  .byte  196,226,125,24,13,131,134,2,0       // vbroadcastss  0x28683(%rip),%ymm1        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,13,51,157,2,0        // vbroadcastss  0x29d33(%rip),%ymm1        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,116,92,192                   // vsubps        %ymm8,%ymm1,%ymm8
-  .byte  196,226,125,24,5,121,134,2,0        // vbroadcastss  0x28679(%rip),%ymm0        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,226,125,24,5,41,157,2,0         // vbroadcastss  0x29d29(%rip),%ymm0        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,124,94,192                   // vdivps        %ymm8,%ymm0,%ymm8
   .byte  196,65,44,88,192                    // vaddps        %ymm8,%ymm10,%ymm8
   .byte  197,204,84,245                      // vandps        %ymm5,%ymm6,%ymm6
@@ -19521,7 +19521,7 @@
   .byte  197,244,92,205                      // vsubps        %ymm5,%ymm1,%ymm1
   .byte  197,252,94,193                      // vdivps        %ymm1,%ymm0,%ymm0
   .byte  197,236,88,192                      // vaddps        %ymm0,%ymm2,%ymm0
-  .byte  196,226,125,24,13,234,133,2,0       // vbroadcastss  0x285ea(%rip),%ymm1        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,226,125,24,13,154,156,2,0       // vbroadcastss  0x29c9a(%rip),%ymm1        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,188,89,209                      // vmulps        %ymm1,%ymm8,%ymm2
   .byte  197,204,89,217                      // vmulps        %ymm1,%ymm6,%ymm3
   .byte  197,252,89,225                      // vmulps        %ymm1,%ymm0,%ymm4
@@ -19541,26 +19541,26 @@
 .globl _sk_lab_to_xyz_hsw
 FUNCTION(_sk_lab_to_xyz_hsw)
 _sk_lab_to_xyz_hsw:
-  .byte  196,98,125,24,5,168,133,2,0         // vbroadcastss  0x285a8(%rip),%ymm8        # 32cf4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
-  .byte  196,98,125,24,13,7,133,2,0          // vbroadcastss  0x28507(%rip),%ymm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  196,98,125,24,21,154,133,2,0        // vbroadcastss  0x2859a(%rip),%ymm10        # 32cf8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
+  .byte  196,98,125,24,5,88,156,2,0          // vbroadcastss  0x29c58(%rip),%ymm8        # 343a4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
+  .byte  196,98,125,24,13,183,155,2,0        // vbroadcastss  0x29bb7(%rip),%ymm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,21,74,156,2,0         // vbroadcastss  0x29c4a(%rip),%ymm10        # 343a8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
   .byte  196,194,53,168,202                  // vfmadd213ps   %ymm10,%ymm9,%ymm1
   .byte  196,194,53,168,210                  // vfmadd213ps   %ymm10,%ymm9,%ymm2
-  .byte  196,98,125,24,13,139,133,2,0        // vbroadcastss  0x2858b(%rip),%ymm9        # 32cfc <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
+  .byte  196,98,125,24,13,59,156,2,0         // vbroadcastss  0x29c3b(%rip),%ymm9        # 343ac <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
   .byte  196,66,125,184,200                  // vfmadd231ps   %ymm8,%ymm0,%ymm9
-  .byte  196,226,125,24,5,129,133,2,0        // vbroadcastss  0x28581(%rip),%ymm0        # 32d00 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
+  .byte  196,226,125,24,5,49,156,2,0         // vbroadcastss  0x29c31(%rip),%ymm0        # 343b0 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
-  .byte  196,98,125,24,5,120,133,2,0         // vbroadcastss  0x28578(%rip),%ymm8        # 32d04 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
+  .byte  196,98,125,24,5,40,156,2,0          // vbroadcastss  0x29c28(%rip),%ymm8        # 343b4 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
   .byte  196,98,117,168,192                  // vfmadd213ps   %ymm0,%ymm1,%ymm8
-  .byte  196,98,125,24,13,110,133,2,0        // vbroadcastss  0x2856e(%rip),%ymm9        # 32d08 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
+  .byte  196,98,125,24,13,30,156,2,0         // vbroadcastss  0x29c1e(%rip),%ymm9        # 343b8 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
   .byte  196,98,109,172,200                  // vfnmadd213ps  %ymm0,%ymm2,%ymm9
   .byte  196,193,60,89,200                   // vmulps        %ymm8,%ymm8,%ymm1
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
-  .byte  196,226,125,24,21,91,133,2,0        // vbroadcastss  0x2855b(%rip),%ymm2        # 32d0c <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
+  .byte  196,226,125,24,21,11,156,2,0        // vbroadcastss  0x29c0b(%rip),%ymm2        # 343bc <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
   .byte  197,108,194,209,1                   // vcmpltps      %ymm1,%ymm2,%ymm10
-  .byte  196,98,125,24,29,81,133,2,0         // vbroadcastss  0x28551(%rip),%ymm11        # 32d10 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
+  .byte  196,98,125,24,29,1,156,2,0          // vbroadcastss  0x29c01(%rip),%ymm11        # 343c0 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
   .byte  196,65,60,88,195                    // vaddps        %ymm11,%ymm8,%ymm8
-  .byte  196,98,125,24,37,71,133,2,0         // vbroadcastss  0x28547(%rip),%ymm12        # 32d14 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
+  .byte  196,98,125,24,37,247,155,2,0        // vbroadcastss  0x29bf7(%rip),%ymm12        # 343c4 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
   .byte  196,65,60,89,196                    // vmulps        %ymm12,%ymm8,%ymm8
   .byte  196,99,61,74,193,160                // vblendvps     %ymm10,%ymm1,%ymm8,%ymm8
   .byte  197,252,89,200                      // vmulps        %ymm0,%ymm0,%ymm1
@@ -19575,9 +19575,9 @@
   .byte  196,65,52,88,203                    // vaddps        %ymm11,%ymm9,%ymm9
   .byte  196,65,52,89,204                    // vmulps        %ymm12,%ymm9,%ymm9
   .byte  196,227,53,74,208,32                // vblendvps     %ymm2,%ymm0,%ymm9,%ymm2
-  .byte  196,226,125,24,5,252,132,2,0        // vbroadcastss  0x284fc(%rip),%ymm0        # 32d18 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
+  .byte  196,226,125,24,5,172,155,2,0        // vbroadcastss  0x29bac(%rip),%ymm0        # 343c8 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,243,132,2,0         // vbroadcastss  0x284f3(%rip),%ymm8        # 32d1c <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
+  .byte  196,98,125,24,5,163,155,2,0         // vbroadcastss  0x29ba3(%rip),%ymm8        # 343cc <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -19595,10 +19595,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,52                              // jne           a87e <_sk_load_a8_hsw+0x4c>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,24,139,2,0            // vpand         0x28b18(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,248,161,2,0           // vpand         0x2a1f8(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,62,132,2,0        // vbroadcastss  0x2843e(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,238,154,2,0       // vbroadcastss  0x29aee(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,217                      // vmulps        %ymm1,%ymm0,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
@@ -19671,10 +19671,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,52                              // jne           a97c <_sk_load_a8_dst_hsw+0x4c>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,26,138,2,0           // vpand         0x28a1a(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,250,160,2,0          // vpand         0x2a0fa(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,228                  // vpmovzxwd     %xmm4,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,64,131,2,0        // vbroadcastss  0x28340(%rip),%ymm5        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,45,240,153,2,0       // vbroadcastss  0x299f0(%rip),%ymm5        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,220,89,253                      // vmulps        %ymm5,%ymm4,%ymm7
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
@@ -19722,7 +19722,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,221                 // callq         ffffffffde00aa24 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd80a4>
+  .byte  232,255,255,255,221                 // callq         ffffffffde00aa24 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd69f4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,210                             // callq         *%rdx
@@ -19789,7 +19789,7 @@
   .byte  196,227,121,32,192,7                // vpinsrb       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,135,129,2,0       // vbroadcastss  0x28187(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,55,152,2,0        // vbroadcastss  0x29837(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,217                      // vmulps        %ymm1,%ymm0,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
@@ -19811,7 +19811,7 @@
   .byte  77,15,175,193                       // imul          %r9,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,4,129,2,0           // vbroadcastss  0x28104(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,180,151,2,0         // vbroadcastss  0x297b4(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,100,89,192                   // vmulps        %ymm8,%ymm3,%ymm8
   .byte  196,65,125,91,192                   // vcvtps2dq     %ymm8,%ymm8
   .byte  196,67,125,25,193,1                 // vextractf128  $0x1,%ymm8,%xmm9
@@ -19836,13 +19836,13 @@
   .byte  196,3,121,20,4,24,0                 // vpextrb       $0x0,%xmm8,(%r8,%r11,1)
   .byte  235,202                             // jmp           ab7c <_sk_store_a8_hsw+0x40>
   .byte  196,3,121,20,68,24,2,4              // vpextrb       $0x4,%xmm8,0x2(%r8,%r11,1)
-  .byte  196,98,57,0,5,189,135,2,0           // vpshufb       0x287bd(%rip),%xmm8,%xmm8        # 33380 <_sk_srcover_rgba_8888_sse2_lowp+0xa00>
+  .byte  196,98,57,0,5,157,158,2,0           // vpshufb       0x29e9d(%rip),%xmm8,%xmm8        # 34a60 <_sk_srcover_rgba_8888_sse2_lowp+0xa30>
   .byte  196,3,121,21,4,24,0                 // vpextrw       $0x0,%xmm8,(%r8,%r11,1)
   .byte  235,176                             // jmp           ab7c <_sk_store_a8_hsw+0x40>
   .byte  196,3,121,20,68,24,6,12             // vpextrb       $0xc,%xmm8,0x6(%r8,%r11,1)
   .byte  196,3,121,20,68,24,5,10             // vpextrb       $0xa,%xmm8,0x5(%r8,%r11,1)
   .byte  196,3,121,20,68,24,4,8              // vpextrb       $0x8,%xmm8,0x4(%r8,%r11,1)
-  .byte  196,98,57,0,5,163,135,2,0           // vpshufb       0x287a3(%rip),%xmm8,%xmm8        # 33390 <_sk_srcover_rgba_8888_sse2_lowp+0xa10>
+  .byte  196,98,57,0,5,131,158,2,0           // vpshufb       0x29e83(%rip),%xmm8,%xmm8        # 34a70 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
   .byte  196,1,121,126,4,24                  // vmovd         %xmm8,(%r8,%r11,1)
   .byte  235,135                             // jmp           ab7c <_sk_store_a8_hsw+0x40>
   .byte  15,31,0                             // nopl          (%rax)
@@ -19879,13 +19879,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,57                              // jne           ac65 <_sk_load_g8_hsw+0x51>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,54,135,2,0            // vpand         0x28736(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,22,158,2,0            // vpand         0x29e16(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,92,128,2,0        // vbroadcastss  0x2805c(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,12,151,2,0        // vbroadcastss  0x2970c(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,213,127,2,0       // vbroadcastss  0x27fd5(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,133,150,2,0       // vbroadcastss  0x29685(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  255,224                             // jmpq          *%rax
@@ -19930,7 +19930,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,222                 // jmpq          ffffffffdf00ad0c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd838c>
+  .byte  233,255,255,255,222                 // jmpq          ffffffffdf00ad0c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd6cdc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,211                             // callq         *%rbx
@@ -19954,13 +19954,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,57                              // jne           ad69 <_sk_load_g8_dst_hsw+0x51>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,50,134,2,0           // vpand         0x28632(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,18,157,2,0           // vpand         0x29d12(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,125,51,228                  // vpmovzxwd     %xmm4,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,88,127,2,0        // vbroadcastss  0x27f58(%rip),%ymm5        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,45,8,150,2,0         // vbroadcastss  0x29608(%rip),%ymm5        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,209,126,2,0       // vbroadcastss  0x27ed1(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,129,149,2,0       // vbroadcastss  0x29581(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,236                      // vmovaps       %ymm4,%ymm5
   .byte  197,252,40,244                      // vmovaps       %ymm4,%ymm6
   .byte  255,224                             // jmpq          *%rax
@@ -20005,7 +20005,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,222                 // jmpq          ffffffffdf00ae10 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd8490>
+  .byte  233,255,255,255,222                 // jmpq          ffffffffdf00ae10 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd6de0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,211                             // callq         *%rbx
@@ -20072,10 +20072,10 @@
   .byte  196,227,121,32,192,7                // vpinsrb       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,49,192                  // vpmovzxbd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,155,125,2,0       // vbroadcastss  0x27d9b(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,75,148,2,0        // vbroadcastss  0x2944b(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,20,125,2,0        // vbroadcastss  0x27d14(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,196,147,2,0       // vbroadcastss  0x293c4(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -20099,23 +20099,23 @@
   .byte  117,114                             // jne           afba <_sk_load_565_hsw+0x8d>
   .byte  196,129,122,111,4,88                // vmovdqu       (%r8,%r11,2),%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  196,226,125,88,5,80,125,2,0         // vpbroadcastd  0x27d50(%rip),%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,88,5,0,148,2,0          // vpbroadcastd  0x29400(%rip),%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,67,125,2,0        // vbroadcastss  0x27d43(%rip),%ymm1        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,13,243,147,2,0       // vbroadcastss  0x293f3(%rip),%ymm1        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,58,125,2,0        // vpbroadcastd  0x27d3a(%rip),%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,88,13,234,147,2,0       // vpbroadcastd  0x293ea(%rip),%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,45,125,2,0        // vbroadcastss  0x27d2d(%rip),%ymm3        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,29,221,147,2,0       // vbroadcastss  0x293dd(%rip),%ymm3        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,36,125,2,0        // vpbroadcastd  0x27d24(%rip),%ymm3        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,88,29,212,147,2,0       // vpbroadcastd  0x293d4(%rip),%ymm3        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,23,125,2,0        // vbroadcastss  0x27d17(%rip),%ymm3        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,29,199,147,2,0       // vbroadcastss  0x293c7(%rip),%ymm3        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,120,124,2,0       // vbroadcastss  0x27c78(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,40,147,2,0        // vbroadcastss  0x29328(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -20178,23 +20178,23 @@
   .byte  117,114                             // jne           b0e5 <_sk_load_565_dst_hsw+0x8d>
   .byte  196,129,122,111,36,88               // vmovdqu       (%r8,%r11,2),%xmm4
   .byte  196,226,125,51,244                  // vpmovzxwd     %xmm4,%ymm6
-  .byte  196,226,125,88,37,37,124,2,0        // vpbroadcastd  0x27c25(%rip),%ymm4        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,88,37,213,146,2,0       // vpbroadcastd  0x292d5(%rip),%ymm4        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,205,219,228                     // vpand         %ymm4,%ymm6,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,24,124,2,0        // vbroadcastss  0x27c18(%rip),%ymm5        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,45,200,146,2,0       // vbroadcastss  0x292c8(%rip),%ymm5        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
-  .byte  196,226,125,88,45,15,124,2,0        // vpbroadcastd  0x27c0f(%rip),%ymm5        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,88,45,191,146,2,0       // vpbroadcastd  0x292bf(%rip),%ymm5        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,205,219,237                     // vpand         %ymm5,%ymm6,%ymm5
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,226,125,24,61,2,124,2,0         // vbroadcastss  0x27c02(%rip),%ymm7        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,61,178,146,2,0       // vbroadcastss  0x292b2(%rip),%ymm7        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,212,89,239                      // vmulps        %ymm7,%ymm5,%ymm5
-  .byte  196,226,125,88,61,249,123,2,0       // vpbroadcastd  0x27bf9(%rip),%ymm7        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,88,61,169,146,2,0       // vpbroadcastd  0x292a9(%rip),%ymm7        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,205,219,247                     // vpand         %ymm7,%ymm6,%ymm6
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  196,226,125,24,61,236,123,2,0       // vbroadcastss  0x27bec(%rip),%ymm7        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,61,156,146,2,0       // vbroadcastss  0x2929c(%rip),%ymm7        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,204,89,247                      // vmulps        %ymm7,%ymm6,%ymm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,77,123,2,0        // vbroadcastss  0x27b4d(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,253,145,2,0       // vbroadcastss  0x291fd(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -20300,23 +20300,23 @@
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,208                  // vpmovzxwd     %xmm0,%ymm2
-  .byte  196,226,125,88,5,64,122,2,0         // vpbroadcastd  0x27a40(%rip),%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,88,5,240,144,2,0        // vpbroadcastd  0x290f0(%rip),%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,237,219,192                     // vpand         %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,51,122,2,0        // vbroadcastss  0x27a33(%rip),%ymm1        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,13,227,144,2,0       // vbroadcastss  0x290e3(%rip),%ymm1        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,42,122,2,0        // vpbroadcastd  0x27a2a(%rip),%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,88,13,218,144,2,0       // vpbroadcastd  0x290da(%rip),%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,29,122,2,0        // vbroadcastss  0x27a1d(%rip),%ymm3        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,29,205,144,2,0       // vbroadcastss  0x290cd(%rip),%ymm3        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,88,29,20,122,2,0        // vpbroadcastd  0x27a14(%rip),%ymm3        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,88,29,196,144,2,0       // vpbroadcastd  0x290c4(%rip),%ymm3        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,7,122,2,0         // vbroadcastss  0x27a07(%rip),%ymm3        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,29,183,144,2,0       // vbroadcastss  0x290b7(%rip),%ymm3        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,104,121,2,0       // vbroadcastss  0x27968(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,24,144,2,0        // vbroadcastss  0x29018(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -20334,11 +20334,11 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,48,122,2,0          // vbroadcastss  0x27a30(%rip),%ymm8        # 32d20 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
+  .byte  196,98,125,24,5,224,144,2,0         // vbroadcastss  0x290e0(%rip),%ymm8        # 343d0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,53,114,241,11               // vpslld        $0xb,%ymm9,%ymm9
-  .byte  196,98,125,24,21,27,122,2,0         // vbroadcastss  0x27a1b(%rip),%ymm10        # 32d24 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
+  .byte  196,98,125,24,21,203,144,2,0        // vbroadcastss  0x290cb(%rip),%ymm10        # 343d4 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
   .byte  196,65,116,89,210                   // vmulps        %ymm10,%ymm1,%ymm10
   .byte  196,65,125,91,210                   // vcvtps2dq     %ymm10,%ymm10
   .byte  196,193,45,114,242,5                // vpslld        $0x5,%ymm10,%ymm10
@@ -20390,7 +20390,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,224                 // callq         ffffffffe100b3bc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd8a3c>
+  .byte  232,255,255,255,224                 // callq         ffffffffe100b3bc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd738c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -20410,25 +20410,25 @@
   .byte  15,133,138,0,0,0                    // jne           b469 <_sk_load_4444_hsw+0xa9>
   .byte  196,129,122,111,4,88                // vmovdqu       (%r8,%r11,2),%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  196,226,125,88,5,53,121,2,0         // vpbroadcastd  0x27935(%rip),%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,88,5,229,143,2,0        // vpbroadcastd  0x28fe5(%rip),%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,40,121,2,0        // vbroadcastss  0x27928(%rip),%ymm1        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,13,216,143,2,0       // vbroadcastss  0x28fd8(%rip),%ymm1        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,31,121,2,0        // vpbroadcastd  0x2791f(%rip),%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,88,13,207,143,2,0       // vpbroadcastd  0x28fcf(%rip),%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,18,121,2,0        // vbroadcastss  0x27912(%rip),%ymm2        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,21,194,143,2,0       // vbroadcastss  0x28fc2(%rip),%ymm2        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,9,121,2,0         // vpbroadcastd  0x27909(%rip),%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,88,21,185,143,2,0       // vpbroadcastd  0x28fb9(%rip),%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,252,120,2,0         // vbroadcastss  0x278fc(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,172,143,2,0         // vbroadcastss  0x28fac(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,242,120,2,0         // vpbroadcastd  0x278f2(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,88,5,162,143,2,0         // vpbroadcastd  0x28fa2(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,228,120,2,0         // vbroadcastss  0x278e4(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,148,143,2,0         // vbroadcastss  0x28f94(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -20496,25 +20496,25 @@
   .byte  15,133,138,0,0,0                    // jne           b5b5 <_sk_load_4444_dst_hsw+0xa9>
   .byte  196,129,122,111,36,88               // vmovdqu       (%r8,%r11,2),%xmm4
   .byte  196,226,125,51,252                  // vpmovzxwd     %xmm4,%ymm7
-  .byte  196,226,125,88,37,233,119,2,0       // vpbroadcastd  0x277e9(%rip),%ymm4        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,88,37,153,142,2,0       // vpbroadcastd  0x28e99(%rip),%ymm4        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,197,219,228                     // vpand         %ymm4,%ymm7,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,220,119,2,0       // vbroadcastss  0x277dc(%rip),%ymm5        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,45,140,142,2,0       // vbroadcastss  0x28e8c(%rip),%ymm5        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
-  .byte  196,226,125,88,45,211,119,2,0       // vpbroadcastd  0x277d3(%rip),%ymm5        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,88,45,131,142,2,0       // vpbroadcastd  0x28e83(%rip),%ymm5        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,197,219,237                     // vpand         %ymm5,%ymm7,%ymm5
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,226,125,24,53,198,119,2,0       // vbroadcastss  0x277c6(%rip),%ymm6        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,53,118,142,2,0       // vbroadcastss  0x28e76(%rip),%ymm6        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,212,89,238                      // vmulps        %ymm6,%ymm5,%ymm5
-  .byte  196,226,125,88,53,189,119,2,0       // vpbroadcastd  0x277bd(%rip),%ymm6        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,88,53,109,142,2,0       // vpbroadcastd  0x28e6d(%rip),%ymm6        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,197,219,246                     // vpand         %ymm6,%ymm7,%ymm6
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  196,98,125,24,5,176,119,2,0         // vbroadcastss  0x277b0(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,96,142,2,0          // vbroadcastss  0x28e60(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,76,89,240                   // vmulps        %ymm8,%ymm6,%ymm6
-  .byte  196,98,125,88,5,166,119,2,0         // vpbroadcastd  0x277a6(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,88,5,86,142,2,0          // vpbroadcastd  0x28e56(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,69,219,248                  // vpand         %ymm8,%ymm7,%ymm7
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  196,98,125,24,5,152,119,2,0         // vbroadcastss  0x27798(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,72,142,2,0          // vbroadcastss  0x28e48(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,68,89,248                   // vmulps        %ymm8,%ymm7,%ymm7
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -20622,25 +20622,25 @@
   .byte  65,15,183,4,88                      // movzwl        (%r8,%rbx,2),%eax
   .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
   .byte  196,226,125,51,216                  // vpmovzxwd     %xmm0,%ymm3
-  .byte  196,226,125,88,5,232,117,2,0        // vpbroadcastd  0x275e8(%rip),%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,88,5,152,140,2,0        // vpbroadcastd  0x28c98(%rip),%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,229,219,192                     // vpand         %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,219,117,2,0       // vbroadcastss  0x275db(%rip),%ymm1        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,13,139,140,2,0       // vbroadcastss  0x28c8b(%rip),%ymm1        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,88,13,210,117,2,0       // vpbroadcastd  0x275d2(%rip),%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,88,13,130,140,2,0       // vpbroadcastd  0x28c82(%rip),%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,229,219,201                     // vpand         %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,197,117,2,0       // vbroadcastss  0x275c5(%rip),%ymm2        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,21,117,140,2,0       // vbroadcastss  0x28c75(%rip),%ymm2        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,88,21,188,117,2,0       // vpbroadcastd  0x275bc(%rip),%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,88,21,108,140,2,0       // vpbroadcastd  0x28c6c(%rip),%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,229,219,210                     // vpand         %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,175,117,2,0         // vbroadcastss  0x275af(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,95,140,2,0          // vbroadcastss  0x28c5f(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,88,5,165,117,2,0         // vpbroadcastd  0x275a5(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,88,5,85,140,2,0          // vpbroadcastd  0x28c55(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,101,219,216                 // vpand         %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,151,117,2,0         // vbroadcastss  0x27597(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,71,140,2,0          // vbroadcastss  0x28c47(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -20660,7 +20660,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,108,117,2,0         // vbroadcastss  0x2756c(%rip),%ymm8        # 32d48 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
+  .byte  196,98,125,24,5,28,140,2,0          // vbroadcastss  0x28c1c(%rip),%ymm8        # 343f8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,53,114,241,12               // vpslld        $0xc,%ymm9,%ymm9
@@ -20719,7 +20719,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,224                 // callq         ffffffffe100b8b4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd8f34>
+  .byte  232,255,255,255,224                 // callq         ffffffffe100b8b4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd7884>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -20738,14 +20738,14 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,86                              // jne           b92a <_sk_load_8888_hsw+0x72>
   .byte  196,129,126,111,28,152              // vmovdqu       (%r8,%r11,4),%ymm3
-  .byte  197,229,219,5,254,117,2,0           // vpand         0x275fe(%rip),%ymm3,%ymm0        # 32ee0 <_sk_srcover_rgba_8888_sse2_lowp+0x560>
+  .byte  197,229,219,5,158,140,2,0           // vpand         0x28c9e(%rip),%ymm3,%ymm0        # 34580 <_sk_srcover_rgba_8888_sse2_lowp+0x550>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,185,115,2,0         // vbroadcastss  0x273b9(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,105,138,2,0         // vbroadcastss  0x28a69(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,226,101,0,13,3,118,2,0          // vpshufb       0x27603(%rip),%ymm3,%ymm1        # 32f00 <_sk_srcover_rgba_8888_sse2_lowp+0x580>
+  .byte  196,226,101,0,13,163,140,2,0        // vpshufb       0x28ca3(%rip),%ymm3,%ymm1        # 345a0 <_sk_srcover_rgba_8888_sse2_lowp+0x570>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,226,101,0,21,17,118,2,0         // vpshufb       0x27611(%rip),%ymm3,%ymm2        # 32f20 <_sk_srcover_rgba_8888_sse2_lowp+0x5a0>
+  .byte  196,226,101,0,21,177,140,2,0        // vpshufb       0x28cb1(%rip),%ymm3,%ymm2        # 345c0 <_sk_srcover_rgba_8888_sse2_lowp+0x590>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  197,229,114,211,24                  // vpsrld        $0x18,%ymm3,%ymm3
@@ -20821,14 +20821,14 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,86                              // jne           ba5e <_sk_load_8888_dst_hsw+0x72>
   .byte  196,129,126,111,60,152              // vmovdqu       (%r8,%r11,4),%ymm7
-  .byte  197,197,219,37,42,117,2,0           // vpand         0x2752a(%rip),%ymm7,%ymm4        # 32f40 <_sk_srcover_rgba_8888_sse2_lowp+0x5c0>
+  .byte  197,197,219,37,202,139,2,0          // vpand         0x28bca(%rip),%ymm7,%ymm4        # 345e0 <_sk_srcover_rgba_8888_sse2_lowp+0x5b0>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,98,125,24,5,133,114,2,0         // vbroadcastss  0x27285(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,53,137,2,0          // vbroadcastss  0x28935(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,92,89,224                   // vmulps        %ymm8,%ymm4,%ymm4
-  .byte  196,226,69,0,45,47,117,2,0          // vpshufb       0x2752f(%rip),%ymm7,%ymm5        # 32f60 <_sk_srcover_rgba_8888_sse2_lowp+0x5e0>
+  .byte  196,226,69,0,45,207,139,2,0         // vpshufb       0x28bcf(%rip),%ymm7,%ymm5        # 34600 <_sk_srcover_rgba_8888_sse2_lowp+0x5d0>
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
   .byte  196,193,84,89,232                   // vmulps        %ymm8,%ymm5,%ymm5
-  .byte  196,226,69,0,53,61,117,2,0          // vpshufb       0x2753d(%rip),%ymm7,%ymm6        # 32f80 <_sk_srcover_rgba_8888_sse2_lowp+0x600>
+  .byte  196,226,69,0,53,221,139,2,0         // vpshufb       0x28bdd(%rip),%ymm7,%ymm6        # 34620 <_sk_srcover_rgba_8888_sse2_lowp+0x5f0>
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
   .byte  196,193,76,89,240                   // vmulps        %ymm8,%ymm6,%ymm6
   .byte  197,197,114,215,24                  // vpsrld        $0x18,%ymm7,%ymm7
@@ -20912,14 +20912,14 @@
   .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
   .byte  197,245,254,192                     // vpaddd        %ymm0,%ymm1,%ymm0
   .byte  196,66,101,144,4,128                // vpgatherdd    %ymm3,(%r8,%ymm0,4),%ymm8
-  .byte  197,189,219,5,41,116,2,0            // vpand         0x27429(%rip),%ymm8,%ymm0        # 32fa0 <_sk_srcover_rgba_8888_sse2_lowp+0x620>
+  .byte  197,189,219,5,201,138,2,0           // vpand         0x28ac9(%rip),%ymm8,%ymm0        # 34640 <_sk_srcover_rgba_8888_sse2_lowp+0x610>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,29,36,113,2,0        // vbroadcastss  0x27124(%rip),%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,29,212,135,2,0       // vbroadcastss  0x287d4(%rip),%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,61,0,13,47,116,2,0          // vpshufb       0x2742f(%rip),%ymm8,%ymm1        # 32fc0 <_sk_srcover_rgba_8888_sse2_lowp+0x640>
+  .byte  196,226,61,0,13,207,138,2,0         // vpshufb       0x28acf(%rip),%ymm8,%ymm1        # 34660 <_sk_srcover_rgba_8888_sse2_lowp+0x630>
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,61,0,21,62,116,2,0          // vpshufb       0x2743e(%rip),%ymm8,%ymm2        # 32fe0 <_sk_srcover_rgba_8888_sse2_lowp+0x660>
+  .byte  196,226,61,0,21,222,138,2,0         // vpshufb       0x28ade(%rip),%ymm8,%ymm2        # 34680 <_sk_srcover_rgba_8888_sse2_lowp+0x650>
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  196,193,61,114,208,24               // vpsrld        $0x18,%ymm8,%ymm8
@@ -20939,7 +20939,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,127,112,2,0         // vbroadcastss  0x2707f(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,47,135,2,0          // vbroadcastss  0x2872f(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -20993,7 +20993,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,219                 // jmpq          ffffffffdc00bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd9338>
+  .byte  233,255,255,255,219                 // jmpq          ffffffffdc00bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd7c88>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,205                             // dec           %ebp
@@ -21015,14 +21015,14 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,86                              // jne           bd32 <_sk_load_bgra_hsw+0x72>
   .byte  196,129,126,111,28,152              // vmovdqu       (%r8,%r11,4),%ymm3
-  .byte  197,229,219,5,22,115,2,0            // vpand         0x27316(%rip),%ymm3,%ymm0        # 33000 <_sk_srcover_rgba_8888_sse2_lowp+0x680>
+  .byte  197,229,219,5,182,137,2,0           // vpand         0x289b6(%rip),%ymm3,%ymm0        # 346a0 <_sk_srcover_rgba_8888_sse2_lowp+0x670>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,177,111,2,0         // vbroadcastss  0x26fb1(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,97,134,2,0          // vbroadcastss  0x28661(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,208                  // vmulps        %ymm8,%ymm0,%ymm2
-  .byte  196,226,101,0,5,27,115,2,0          // vpshufb       0x2731b(%rip),%ymm3,%ymm0        # 33020 <_sk_srcover_rgba_8888_sse2_lowp+0x6a0>
+  .byte  196,226,101,0,5,187,137,2,0         // vpshufb       0x289bb(%rip),%ymm3,%ymm0        # 346c0 <_sk_srcover_rgba_8888_sse2_lowp+0x690>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
   .byte  196,193,124,89,200                  // vmulps        %ymm8,%ymm0,%ymm1
-  .byte  196,226,101,0,5,41,115,2,0          // vpshufb       0x27329(%rip),%ymm3,%ymm0        # 33040 <_sk_srcover_rgba_8888_sse2_lowp+0x6c0>
+  .byte  196,226,101,0,5,201,137,2,0         // vpshufb       0x289c9(%rip),%ymm3,%ymm0        # 346e0 <_sk_srcover_rgba_8888_sse2_lowp+0x6b0>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,229,114,211,24                  // vpsrld        $0x18,%ymm3,%ymm3
@@ -21098,14 +21098,14 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,86                              // jne           be66 <_sk_load_bgra_dst_hsw+0x72>
   .byte  196,129,126,111,60,152              // vmovdqu       (%r8,%r11,4),%ymm7
-  .byte  197,197,219,37,66,114,2,0           // vpand         0x27242(%rip),%ymm7,%ymm4        # 33060 <_sk_srcover_rgba_8888_sse2_lowp+0x6e0>
+  .byte  197,197,219,37,226,136,2,0          // vpand         0x288e2(%rip),%ymm7,%ymm4        # 34700 <_sk_srcover_rgba_8888_sse2_lowp+0x6d0>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,98,125,24,5,125,110,2,0         // vbroadcastss  0x26e7d(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,45,133,2,0          // vbroadcastss  0x2852d(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,92,89,240                   // vmulps        %ymm8,%ymm4,%ymm6
-  .byte  196,226,69,0,37,71,114,2,0          // vpshufb       0x27247(%rip),%ymm7,%ymm4        # 33080 <_sk_srcover_rgba_8888_sse2_lowp+0x700>
+  .byte  196,226,69,0,37,231,136,2,0         // vpshufb       0x288e7(%rip),%ymm7,%ymm4        # 34720 <_sk_srcover_rgba_8888_sse2_lowp+0x6f0>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
   .byte  196,193,92,89,232                   // vmulps        %ymm8,%ymm4,%ymm5
-  .byte  196,226,69,0,37,85,114,2,0          // vpshufb       0x27255(%rip),%ymm7,%ymm4        # 330a0 <_sk_srcover_rgba_8888_sse2_lowp+0x720>
+  .byte  196,226,69,0,37,245,136,2,0         // vpshufb       0x288f5(%rip),%ymm7,%ymm4        # 34740 <_sk_srcover_rgba_8888_sse2_lowp+0x710>
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
   .byte  196,193,92,89,224                   // vmulps        %ymm8,%ymm4,%ymm4
   .byte  197,197,114,215,24                  // vpsrld        $0x18,%ymm7,%ymm7
@@ -21189,14 +21189,14 @@
   .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
   .byte  197,245,254,192                     // vpaddd        %ymm0,%ymm1,%ymm0
   .byte  196,66,101,144,4,128                // vpgatherdd    %ymm3,(%r8,%ymm0,4),%ymm8
-  .byte  197,189,219,5,65,113,2,0            // vpand         0x27141(%rip),%ymm8,%ymm0        # 330c0 <_sk_srcover_rgba_8888_sse2_lowp+0x740>
+  .byte  197,189,219,5,225,135,2,0           // vpand         0x287e1(%rip),%ymm8,%ymm0        # 34760 <_sk_srcover_rgba_8888_sse2_lowp+0x730>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,29,28,109,2,0        // vbroadcastss  0x26d1c(%rip),%ymm3        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,29,204,131,2,0       // vbroadcastss  0x283cc(%rip),%ymm3        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,211                      // vmulps        %ymm3,%ymm0,%ymm2
-  .byte  196,226,61,0,5,71,113,2,0           // vpshufb       0x27147(%rip),%ymm8,%ymm0        # 330e0 <_sk_srcover_rgba_8888_sse2_lowp+0x760>
+  .byte  196,226,61,0,5,231,135,2,0          // vpshufb       0x287e7(%rip),%ymm8,%ymm0        # 34780 <_sk_srcover_rgba_8888_sse2_lowp+0x750>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
   .byte  197,252,89,203                      // vmulps        %ymm3,%ymm0,%ymm1
-  .byte  196,226,61,0,5,86,113,2,0           // vpshufb       0x27156(%rip),%ymm8,%ymm0        # 33100 <_sk_srcover_rgba_8888_sse2_lowp+0x780>
+  .byte  196,226,61,0,5,246,135,2,0          // vpshufb       0x287f6(%rip),%ymm8,%ymm0        # 347a0 <_sk_srcover_rgba_8888_sse2_lowp+0x770>
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
   .byte  196,193,61,114,208,24               // vpsrld        $0x18,%ymm8,%ymm8
@@ -21216,7 +21216,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,119,108,2,0         // vbroadcastss  0x26c77(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,39,131,2,0          // vbroadcastss  0x28327(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,108,89,200                   // vmulps        %ymm8,%ymm2,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -21270,7 +21270,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,219                 // jmpq          ffffffffdc00c0c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd9740>
+  .byte  233,255,255,255,219                 // jmpq          ffffffffdc00c0c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd8090>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,205                             // dec           %ebp
@@ -21531,7 +21531,7 @@
   .byte  197,241,235,192                     // vpor          %xmm0,%xmm1,%xmm0
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,21,248,103,2,0        // vbroadcastss  0x267f8(%rip),%ymm10        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,21,168,126,2,0        // vbroadcastss  0x27ea8(%rip),%ymm10        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  197,185,109,202                     // vpunpckhqdq   %xmm2,%xmm8,%xmm1
   .byte  197,233,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm2
@@ -21623,7 +21623,7 @@
   .byte  197,241,235,192                     // vpor          %xmm0,%xmm1,%xmm0
   .byte  196,226,125,51,192                  // vpmovzxwd     %xmm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,103,102,2,0        // vbroadcastss  0x26667(%rip),%ymm9        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,13,23,125,2,0         // vbroadcastss  0x27d17(%rip),%ymm9        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,185,109,203                     // vpunpckhqdq   %xmm3,%xmm8,%xmm1
   .byte  197,225,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm3
@@ -21639,7 +21639,7 @@
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  196,193,108,89,209                  // vmulps        %ymm9,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,135,101,2,0       // vbroadcastss  0x26587(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,55,124,2,0        // vbroadcastss  0x27c37(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,4,64                // vmovd         (%r8,%rax,2),%xmm0
   .byte  196,65,49,239,201                   // vpxor         %xmm9,%xmm9,%xmm9
@@ -21715,7 +21715,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  73,99,193                           // movslq        %r9d,%rax
-  .byte  196,98,125,24,5,80,101,2,0          // vbroadcastss  0x26550(%rip),%ymm8        # 32d4c <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
+  .byte  196,98,125,24,5,0,124,2,0           // vbroadcastss  0x27c00(%rip),%ymm8        # 343fc <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,67,125,25,202,1                 // vextractf128  $0x1,%ymm9,%xmm10
@@ -21874,7 +21874,7 @@
   .byte  255,92,255,255                      // lcall         *-0x1(%rdi,%rdi,8)
   .byte  255,70,255                          // incl          -0x1(%rsi)
   .byte  255                                 // (bad)
-  .byte  255,53,255,255,255,40               // pushq         0x28ffffff(%rip)        # 2900cacc <_sk_srcover_rgba_8888_sse2_lowp+0x28fda14c>
+  .byte  255,53,255,255,255,40               // pushq         0x28ffffff(%rip)        # 2900cacc <_sk_srcover_rgba_8888_sse2_lowp+0x28fd8a9c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -22069,7 +22069,7 @@
   .byte  196,65,124,92,218                   // vsubps        %ymm10,%ymm0,%ymm11
   .byte  196,193,58,88,192                   // vaddss        %xmm8,%xmm8,%xmm0
   .byte  196,98,125,24,192                   // vbroadcastss  %xmm0,%ymm8
-  .byte  197,178,89,5,133,94,2,0             // vmulss        0x25e85(%rip),%xmm9,%xmm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,5,53,117,2,0             // vmulss        0x27535(%rip),%xmm9,%xmm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,89,192                      // vmulps        %ymm0,%ymm11,%ymm0
   .byte  196,227,125,8,192,1                 // vroundps      $0x1,%ymm0,%ymm0
@@ -22092,7 +22092,7 @@
   .byte  196,65,116,92,218                   // vsubps        %ymm10,%ymm1,%ymm11
   .byte  196,193,58,88,200                   // vaddss        %xmm8,%xmm8,%xmm1
   .byte  196,98,125,24,193                   // vbroadcastss  %xmm1,%ymm8
-  .byte  197,178,89,13,52,94,2,0             // vmulss        0x25e34(%rip),%xmm9,%xmm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,13,228,116,2,0           // vmulss        0x274e4(%rip),%xmm9,%xmm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,226,125,24,201                  // vbroadcastss  %xmm1,%ymm1
   .byte  197,164,89,201                      // vmulps        %ymm1,%ymm11,%ymm1
   .byte  196,227,125,8,201,1                 // vroundps      $0x1,%ymm1,%ymm1
@@ -22110,7 +22110,7 @@
 _sk_clamp_x_1_hsw:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,252,93,2,0          // vbroadcastss  0x25dfc(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,172,116,2,0         // vbroadcastss  0x274ac(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -22123,7 +22123,7 @@
   .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,214,93,2,0          // vbroadcastss  0x25dd6(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,134,116,2,0         // vbroadcastss  0x27486(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -22132,9 +22132,9 @@
 .globl _sk_mirror_x_1_hsw
 FUNCTION(_sk_mirror_x_1_hsw)
 _sk_mirror_x_1_hsw:
-  .byte  196,98,125,24,5,220,93,2,0          // vbroadcastss  0x25ddc(%rip),%ymm8        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  196,98,125,24,5,140,116,2,0         // vbroadcastss  0x2748c(%rip),%ymm8        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,13,178,93,2,0         // vbroadcastss  0x25db2(%rip),%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,13,98,116,2,0         // vbroadcastss  0x27462(%rip),%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,124,89,201                   // vmulps        %ymm9,%ymm0,%ymm9
   .byte  196,67,125,8,201,1                  // vroundps      $0x1,%ymm9,%ymm9
   .byte  196,65,52,88,201                    // vaddps        %ymm9,%ymm9,%ymm9
@@ -22144,7 +22144,7 @@
   .byte  197,60,92,200                       // vsubps        %ymm0,%ymm8,%ymm9
   .byte  197,180,84,192                      // vandps        %ymm0,%ymm9,%ymm0
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,130,93,2,0          // vbroadcastss  0x25d82(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,50,116,2,0          // vbroadcastss  0x27432(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -22153,11 +22153,11 @@
 .globl _sk_luminance_to_alpha_hsw
 FUNCTION(_sk_luminance_to_alpha_hsw)
 _sk_luminance_to_alpha_hsw:
-  .byte  196,226,125,24,29,148,94,2,0        // vbroadcastss  0x25e94(%rip),%ymm3        # 32d54 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
-  .byte  196,98,125,24,5,135,94,2,0          // vbroadcastss  0x25e87(%rip),%ymm8        # 32d50 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
+  .byte  196,226,125,24,29,68,117,2,0        // vbroadcastss  0x27544(%rip),%ymm3        # 34404 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
+  .byte  196,98,125,24,5,55,117,2,0          // vbroadcastss  0x27537(%rip),%ymm8        # 34400 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  196,226,125,184,203                 // vfmadd231ps   %ymm3,%ymm0,%ymm1
-  .byte  196,226,125,24,29,124,94,2,0        // vbroadcastss  0x25e7c(%rip),%ymm3        # 32d58 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
+  .byte  196,226,125,24,29,44,117,2,0        // vbroadcastss  0x2752c(%rip),%ymm3        # 34408 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
   .byte  196,226,109,168,217                 // vfmadd213ps   %ymm1,%ymm2,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
@@ -22429,14 +22429,14 @@
 .globl _sk_gauss_a_to_rgba_hsw
 FUNCTION(_sk_gauss_a_to_rgba_hsw)
 _sk_gauss_a_to_rgba_hsw:
-  .byte  196,226,125,24,5,98,90,2,0          // vbroadcastss  0x25a62(%rip),%ymm0        # 32d60 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
-  .byte  196,226,125,24,13,85,90,2,0         // vbroadcastss  0x25a55(%rip),%ymm1        # 32d5c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
+  .byte  196,226,125,24,5,18,113,2,0         // vbroadcastss  0x27112(%rip),%ymm0        # 34410 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
+  .byte  196,226,125,24,13,5,113,2,0         // vbroadcastss  0x27105(%rip),%ymm1        # 3440c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
   .byte  196,226,101,168,200                 // vfmadd213ps   %ymm0,%ymm3,%ymm1
-  .byte  196,226,125,24,5,79,90,2,0          // vbroadcastss  0x25a4f(%rip),%ymm0        # 32d64 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
+  .byte  196,226,125,24,5,255,112,2,0        // vbroadcastss  0x270ff(%rip),%ymm0        # 34414 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
   .byte  196,226,101,184,193                 // vfmadd231ps   %ymm1,%ymm3,%ymm0
-  .byte  196,226,125,24,13,69,90,2,0         // vbroadcastss  0x25a45(%rip),%ymm1        # 32d68 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
+  .byte  196,226,125,24,13,245,112,2,0       // vbroadcastss  0x270f5(%rip),%ymm1        # 34418 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
   .byte  196,226,101,184,200                 // vfmadd231ps   %ymm0,%ymm3,%ymm1
-  .byte  196,226,125,24,5,59,90,2,0          // vbroadcastss  0x25a3b(%rip),%ymm0        # 32d6c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
+  .byte  196,226,125,24,5,235,112,2,0        // vbroadcastss  0x270eb(%rip),%ymm0        # 3441c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
   .byte  196,226,101,184,193                 // vfmadd231ps   %ymm1,%ymm3,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
@@ -22455,7 +22455,7 @@
   .byte  76,139,72,72                        // mov           0x48(%rax),%r9
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  65,186,1,0,0,0                      // mov           $0x1,%r10d
-  .byte  196,226,125,24,21,200,88,2,0        // vbroadcastss  0x258c8(%rip),%ymm2        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,226,125,24,21,120,111,2,0       // vbroadcastss  0x26f78(%rip),%ymm2        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  196,65,53,239,201                   // vpxor         %ymm9,%ymm9,%ymm9
   .byte  196,130,125,24,28,145               // vbroadcastss  (%r9,%r10,4),%ymm3
   .byte  197,228,194,216,2                   // vcmpleps      %ymm0,%ymm3,%ymm3
@@ -22558,24 +22558,24 @@
   .byte  196,65,52,95,226                    // vmaxps        %ymm10,%ymm9,%ymm12
   .byte  196,65,36,94,220                    // vdivps        %ymm12,%ymm11,%ymm11
   .byte  196,65,36,89,227                    // vmulps        %ymm11,%ymm11,%ymm12
-  .byte  196,98,125,24,45,87,88,2,0          // vbroadcastss  0x25857(%rip),%ymm13        # 32d70 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
-  .byte  196,98,125,24,53,82,88,2,0          // vbroadcastss  0x25852(%rip),%ymm14        # 32d74 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
+  .byte  196,98,125,24,45,7,111,2,0          // vbroadcastss  0x26f07(%rip),%ymm13        # 34420 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
+  .byte  196,98,125,24,53,2,111,2,0          // vbroadcastss  0x26f02(%rip),%ymm14        # 34424 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
   .byte  196,66,29,184,245                   // vfmadd231ps   %ymm13,%ymm12,%ymm14
-  .byte  196,98,125,24,45,72,88,2,0          // vbroadcastss  0x25848(%rip),%ymm13        # 32d78 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
+  .byte  196,98,125,24,45,248,110,2,0        // vbroadcastss  0x26ef8(%rip),%ymm13        # 34428 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
   .byte  196,66,29,184,238                   // vfmadd231ps   %ymm14,%ymm12,%ymm13
-  .byte  196,98,125,24,53,62,88,2,0          // vbroadcastss  0x2583e(%rip),%ymm14        # 32d7c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
+  .byte  196,98,125,24,53,238,110,2,0        // vbroadcastss  0x26eee(%rip),%ymm14        # 3442c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
   .byte  196,66,29,184,245                   // vfmadd231ps   %ymm13,%ymm12,%ymm14
   .byte  196,65,36,89,222                    // vmulps        %ymm14,%ymm11,%ymm11
   .byte  196,65,52,194,202,1                 // vcmpltps      %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,41,88,2,0          // vbroadcastss  0x25829(%rip),%ymm10        # 32d80 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
+  .byte  196,98,125,24,21,217,110,2,0        // vbroadcastss  0x26ed9(%rip),%ymm10        # 34430 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
   .byte  196,65,44,92,211                    // vsubps        %ymm11,%ymm10,%ymm10
   .byte  196,67,37,74,202,144                // vblendvps     %ymm9,%ymm10,%ymm11,%ymm9
   .byte  196,193,124,194,192,1               // vcmpltps      %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,21,187,86,2,0         // vbroadcastss  0x256bb(%rip),%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,21,107,109,2,0        // vbroadcastss  0x26d6b(%rip),%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,92,209                    // vsubps        %ymm9,%ymm10,%ymm10
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  196,65,116,194,200,1                // vcmpltps      %ymm8,%ymm1,%ymm9
-  .byte  196,98,125,24,21,165,86,2,0         // vbroadcastss  0x256a5(%rip),%ymm10        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,21,85,109,2,0         // vbroadcastss  0x26d55(%rip),%ymm10        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,44,92,208                       // vsubps        %ymm0,%ymm10,%ymm10
   .byte  196,195,125,74,194,144              // vblendvps     %ymm9,%ymm10,%ymm0,%ymm0
   .byte  196,65,124,194,200,3                // vcmpunordps   %ymm8,%ymm0,%ymm9
@@ -22603,23 +22603,23 @@
   .byte  197,50,89,80,76                     // vmulss        0x4c(%rax),%xmm9,%xmm10
   .byte  196,66,125,24,210                   // vbroadcastss  %xmm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  196,98,125,24,29,169,87,2,0         // vbroadcastss  0x257a9(%rip),%ymm11        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,29,89,110,2,0         // vbroadcastss  0x26e59(%rip),%ymm11        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
   .byte  196,98,125,184,216                  // vfmadd231ps   %ymm0,%ymm0,%ymm11
   .byte  196,193,50,89,193                   // vmulss        %xmm9,%xmm9,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,13,136,87,2,0         // vbroadcastss  0x25788(%rip),%ymm9        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  196,98,125,24,13,56,110,2,0         // vbroadcastss  0x26e38(%rip),%ymm9        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,194,45,184,194                  // vfmadd231ps   %ymm10,%ymm10,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  196,98,125,24,13,107,87,2,0         // vbroadcastss  0x2576b(%rip),%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,13,27,110,2,0         // vbroadcastss  0x26e1b(%rip),%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,44,87,201                    // vxorps        %ymm9,%ymm10,%ymm9
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,248,85,2,0         // vbroadcastss  0x255f8(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,168,108,2,0        // vbroadcastss  0x26ca8(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  196,65,44,89,208                    // vmulps        %ymm8,%ymm10,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
@@ -22638,23 +22638,23 @@
   .byte  197,50,89,80,76                     // vmulss        0x4c(%rax),%xmm9,%xmm10
   .byte  196,66,125,24,210                   // vbroadcastss  %xmm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  196,98,125,24,29,17,87,2,0          // vbroadcastss  0x25711(%rip),%ymm11        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,29,193,109,2,0        // vbroadcastss  0x26dc1(%rip),%ymm11        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
   .byte  196,98,125,184,216                  // vfmadd231ps   %ymm0,%ymm0,%ymm11
   .byte  196,193,50,89,193                   // vmulss        %xmm9,%xmm9,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,13,240,86,2,0         // vbroadcastss  0x256f0(%rip),%ymm9        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  196,98,125,24,13,160,109,2,0        // vbroadcastss  0x26da0(%rip),%ymm9        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,194,45,184,194                  // vfmadd231ps   %ymm10,%ymm10,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  196,98,125,24,13,211,86,2,0         // vbroadcastss  0x256d3(%rip),%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,13,131,109,2,0        // vbroadcastss  0x26d83(%rip),%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,44,87,201                    // vxorps        %ymm9,%ymm10,%ymm9
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,96,85,2,0          // vbroadcastss  0x25560(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,16,108,2,0         // vbroadcastss  0x26c10(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  196,65,44,89,208                    // vmulps        %ymm8,%ymm10,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
@@ -22672,14 +22672,14 @@
   .byte  197,58,89,72,76                     // vmulss        0x4c(%rax),%xmm8,%xmm9
   .byte  196,66,125,24,201                   // vbroadcastss  %xmm9,%ymm9
   .byte  197,52,88,200                       // vaddps        %ymm0,%ymm9,%ymm9
-  .byte  196,98,125,24,21,127,86,2,0         // vbroadcastss  0x2567f(%rip),%ymm10        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,21,47,109,2,0         // vbroadcastss  0x26d2f(%rip),%ymm10        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
   .byte  197,116,89,209                      // vmulps        %ymm1,%ymm1,%ymm10
   .byte  196,98,125,184,208                  // vfmadd231ps   %ymm0,%ymm0,%ymm10
   .byte  196,193,58,89,192                   // vmulss        %xmm8,%xmm8,%xmm0
   .byte  196,226,125,24,192                  // vbroadcastss  %xmm0,%ymm0
   .byte  197,172,92,192                      // vsubps        %ymm0,%ymm10,%ymm0
-  .byte  196,98,125,24,5,98,86,2,0           // vbroadcastss  0x25662(%rip),%ymm8        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,5,18,109,2,0          // vbroadcastss  0x26d12(%rip),%ymm8        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,193,124,87,192                  // vxorps        %ymm8,%ymm0,%ymm0
   .byte  196,193,124,94,193                  // vdivps        %ymm9,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22722,7 +22722,7 @@
 FUNCTION(_sk_save_xy_hsw)
 _sk_save_xy_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,141,84,2,0          // vbroadcastss  0x2548d(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,61,107,2,0          // vbroadcastss  0x26b3d(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,124,88,200                   // vaddps        %ymm8,%ymm0,%ymm9
   .byte  196,67,125,8,209,1                  // vroundps      $0x1,%ymm9,%ymm10
   .byte  196,65,52,92,202                    // vsubps        %ymm10,%ymm9,%ymm9
@@ -22756,9 +22756,9 @@
 FUNCTION(_sk_bilinear_nx_hsw)
 _sk_bilinear_nx_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,123,85,2,0         // vbroadcastss  0x2557b(%rip),%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,5,43,108,2,0         // vbroadcastss  0x26c2b(%rip),%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,14,84,2,0           // vbroadcastss  0x2540e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,190,106,2,0         // vbroadcastss  0x26abe(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22769,7 +22769,7 @@
 FUNCTION(_sk_bilinear_px_hsw)
 _sk_bilinear_px_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,235,83,2,0         // vbroadcastss  0x253eb(%rip),%ymm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,5,155,106,2,0        // vbroadcastss  0x26a9b(%rip),%ymm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -22781,9 +22781,9 @@
 FUNCTION(_sk_bilinear_ny_hsw)
 _sk_bilinear_ny_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,44,85,2,0         // vbroadcastss  0x2552c(%rip),%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,13,220,107,2,0       // vbroadcastss  0x26bdc(%rip),%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,190,83,2,0          // vbroadcastss  0x253be(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,110,106,2,0         // vbroadcastss  0x26a6e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22794,7 +22794,7 @@
 FUNCTION(_sk_bilinear_py_hsw)
 _sk_bilinear_py_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,155,83,2,0        // vbroadcastss  0x2539b(%rip),%ymm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,13,75,106,2,0        // vbroadcastss  0x26a4b(%rip),%ymm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -22806,13 +22806,13 @@
 FUNCTION(_sk_bicubic_n3x_hsw)
 _sk_bicubic_n3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,223,84,2,0         // vbroadcastss  0x254df(%rip),%ymm0        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,226,125,24,5,143,107,2,0        // vbroadcastss  0x26b8f(%rip),%ymm0        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,110,83,2,0          // vbroadcastss  0x2536e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,30,106,2,0          // vbroadcastss  0x26a1e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,204,83,2,0         // vbroadcastss  0x253cc(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  .byte  196,98,125,24,29,183,84,2,0         // vbroadcastss  0x254b7(%rip),%ymm11        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,124,106,2,0        // vbroadcastss  0x26a7c(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,29,103,107,2,0        // vbroadcastss  0x26b67(%rip),%ymm11        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,66,61,168,218                   // vfmadd213ps   %ymm10,%ymm8,%ymm11
   .byte  196,65,36,89,193                    // vmulps        %ymm9,%ymm11,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -22824,16 +22824,16 @@
 FUNCTION(_sk_bicubic_n1x_hsw)
 _sk_bicubic_n1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,142,84,2,0         // vbroadcastss  0x2548e(%rip),%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,5,62,107,2,0         // vbroadcastss  0x26b3e(%rip),%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,33,83,2,0           // vbroadcastss  0x25321(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,209,105,2,0         // vbroadcastss  0x269d1(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,128,84,2,0         // vbroadcastss  0x25480(%rip),%ymm9        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  196,98,125,24,21,115,84,2,0         // vbroadcastss  0x25473(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,13,48,107,2,0         // vbroadcastss  0x26b30(%rip),%ymm9        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,35,107,2,0         // vbroadcastss  0x26b23(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,245,82,2,0         // vbroadcastss  0x252f5(%rip),%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,13,165,105,2,0        // vbroadcastss  0x269a5(%rip),%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,95,84,2,0          // vbroadcastss  0x2545f(%rip),%ymm10        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,21,15,107,2,0         // vbroadcastss  0x26b0f(%rip),%ymm10        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,66,61,184,209                   // vfmadd231ps   %ymm9,%ymm8,%ymm10
   .byte  197,124,17,144,0,1,0,0              // vmovups       %ymm10,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22844,14 +22844,14 @@
 FUNCTION(_sk_bicubic_p1x_hsw)
 _sk_bicubic_p1x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,203,82,2,0          // vbroadcastss  0x252cb(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,123,105,2,0         // vbroadcastss  0x2697b(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,136,128,0,0,0            // vmovups       0x80(%rax),%ymm9
-  .byte  196,98,125,24,21,42,84,2,0          // vbroadcastss  0x2542a(%rip),%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  196,98,125,24,29,29,84,2,0          // vbroadcastss  0x2541d(%rip),%ymm11        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,21,218,106,2,0        // vbroadcastss  0x26ada(%rip),%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,29,205,106,2,0        // vbroadcastss  0x26acd(%rip),%ymm11        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,66,53,168,218                   // vfmadd213ps   %ymm10,%ymm9,%ymm11
   .byte  196,66,53,168,216                   // vfmadd213ps   %ymm8,%ymm9,%ymm11
-  .byte  196,98,125,24,5,18,84,2,0           // vbroadcastss  0x25412(%rip),%ymm8        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,5,194,106,2,0         // vbroadcastss  0x26ac2(%rip),%ymm8        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,66,53,184,195                   // vfmadd231ps   %ymm11,%ymm9,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22862,12 +22862,12 @@
 FUNCTION(_sk_bicubic_p3x_hsw)
 _sk_bicubic_p3x_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,242,83,2,0         // vbroadcastss  0x253f2(%rip),%ymm0        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,226,125,24,5,162,106,2,0        // vbroadcastss  0x26aa2(%rip),%ymm0        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,220,82,2,0         // vbroadcastss  0x252dc(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  .byte  196,98,125,24,29,199,83,2,0         // vbroadcastss  0x253c7(%rip),%ymm11        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,140,105,2,0        // vbroadcastss  0x2698c(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,29,119,106,2,0        // vbroadcastss  0x26a77(%rip),%ymm11        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,66,61,168,218                   // vfmadd213ps   %ymm10,%ymm8,%ymm11
   .byte  196,65,52,89,195                    // vmulps        %ymm11,%ymm9,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -22879,13 +22879,13 @@
 FUNCTION(_sk_bicubic_n3y_hsw)
 _sk_bicubic_n3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,162,83,2,0        // vbroadcastss  0x253a2(%rip),%ymm1        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,226,125,24,13,82,106,2,0        // vbroadcastss  0x26a52(%rip),%ymm1        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,48,82,2,0           // vbroadcastss  0x25230(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,224,104,2,0         // vbroadcastss  0x268e0(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,142,82,2,0         // vbroadcastss  0x2528e(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  .byte  196,98,125,24,29,121,83,2,0         // vbroadcastss  0x25379(%rip),%ymm11        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,62,105,2,0         // vbroadcastss  0x2693e(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,29,41,106,2,0         // vbroadcastss  0x26a29(%rip),%ymm11        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,66,61,168,218                   // vfmadd213ps   %ymm10,%ymm8,%ymm11
   .byte  196,65,36,89,193                    // vmulps        %ymm9,%ymm11,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -22897,16 +22897,16 @@
 FUNCTION(_sk_bicubic_n1y_hsw)
 _sk_bicubic_n1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,80,83,2,0         // vbroadcastss  0x25350(%rip),%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,13,0,106,2,0         // vbroadcastss  0x26a00(%rip),%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,226,81,2,0          // vbroadcastss  0x251e2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,146,104,2,0         // vbroadcastss  0x26892(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,65,83,2,0          // vbroadcastss  0x25341(%rip),%ymm9        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  196,98,125,24,21,52,83,2,0          // vbroadcastss  0x25334(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,13,241,105,2,0        // vbroadcastss  0x269f1(%rip),%ymm9        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,228,105,2,0        // vbroadcastss  0x269e4(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,66,61,168,209                   // vfmadd213ps   %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,13,182,81,2,0         // vbroadcastss  0x251b6(%rip),%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,13,102,104,2,0        // vbroadcastss  0x26866(%rip),%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,66,61,184,202                   // vfmadd231ps   %ymm10,%ymm8,%ymm9
-  .byte  196,98,125,24,21,32,83,2,0          // vbroadcastss  0x25320(%rip),%ymm10        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,21,208,105,2,0        // vbroadcastss  0x269d0(%rip),%ymm10        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,66,61,184,209                   // vfmadd231ps   %ymm9,%ymm8,%ymm10
   .byte  197,124,17,144,64,1,0,0             // vmovups       %ymm10,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22917,14 +22917,14 @@
 FUNCTION(_sk_bicubic_p1y_hsw)
 _sk_bicubic_p1y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,140,81,2,0          // vbroadcastss  0x2518c(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,60,104,2,0          // vbroadcastss  0x2683c(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,72,64                    // vaddps        0x40(%rax),%ymm8,%ymm1
   .byte  197,124,16,136,192,0,0,0            // vmovups       0xc0(%rax),%ymm9
-  .byte  196,98,125,24,21,234,82,2,0         // vbroadcastss  0x252ea(%rip),%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
-  .byte  196,98,125,24,29,221,82,2,0         // vbroadcastss  0x252dd(%rip),%ymm11        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,21,154,105,2,0        // vbroadcastss  0x2699a(%rip),%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,29,141,105,2,0        // vbroadcastss  0x2698d(%rip),%ymm11        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,66,53,168,218                   // vfmadd213ps   %ymm10,%ymm9,%ymm11
   .byte  196,66,53,168,216                   // vfmadd213ps   %ymm8,%ymm9,%ymm11
-  .byte  196,98,125,24,5,210,82,2,0          // vbroadcastss  0x252d2(%rip),%ymm8        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,5,130,105,2,0         // vbroadcastss  0x26982(%rip),%ymm8        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,66,53,184,195                   // vfmadd231ps   %ymm11,%ymm9,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -22935,12 +22935,12 @@
 FUNCTION(_sk_bicubic_p3y_hsw)
 _sk_bicubic_p3y_hsw:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,178,82,2,0        // vbroadcastss  0x252b2(%rip),%ymm1        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,226,125,24,13,98,105,2,0        // vbroadcastss  0x26962(%rip),%ymm1        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,155,81,2,0         // vbroadcastss  0x2519b(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  .byte  196,98,125,24,29,134,82,2,0         // vbroadcastss  0x25286(%rip),%ymm11        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,75,104,2,0         // vbroadcastss  0x2684b(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,29,54,105,2,0         // vbroadcastss  0x26936(%rip),%ymm11        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,66,61,168,218                   // vfmadd213ps   %ymm10,%ymm8,%ymm11
   .byte  196,65,52,89,195                    // vmulps        %ymm11,%ymm9,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -23080,7 +23080,7 @@
   .byte  196,98,93,64,218                    // vpmulld       %ymm2,%ymm4,%ymm11
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  197,165,254,193                     // vpaddd        %ymm1,%ymm11,%ymm0
-  .byte  196,98,125,88,61,72,80,2,0          // vpbroadcastd  0x25048(%rip),%ymm15        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  196,98,125,88,61,248,102,2,0        // vpbroadcastd  0x266f8(%rip),%ymm15        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  196,194,125,64,199                  // vpmulld       %ymm15,%ymm0,%ymm0
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
   .byte  197,237,118,210                     // vpcmpeqd      %ymm2,%ymm2,%ymm2
@@ -23092,13 +23092,13 @@
   .byte  197,213,118,237                     // vpcmpeqd      %ymm5,%ymm5,%ymm5
   .byte  196,226,85,146,60,144               // vgatherdps    %ymm5,(%rax,%ymm2,4),%ymm7
   .byte  197,252,17,188,36,0,2,0,0           // vmovups       %ymm7,0x200(%rsp)
-  .byte  196,226,125,88,61,146,78,2,0        // vpbroadcastd  0x24e92(%rip),%ymm7        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,226,125,88,61,66,101,2,0        // vpbroadcastd  0x26542(%rip),%ymm7        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  197,253,254,199                     // vpaddd        %ymm7,%ymm0,%ymm0
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
   .byte  197,213,118,237                     // vpcmpeqd      %ymm5,%ymm5,%ymm5
   .byte  196,226,85,146,20,128               // vgatherdps    %ymm5,(%rax,%ymm0,4),%ymm2
   .byte  197,252,17,148,36,32,1,0,0          // vmovups       %ymm2,0x120(%rsp)
-  .byte  196,226,125,24,5,218,79,2,0         // vbroadcastss  0x24fda(%rip),%ymm0        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,226,125,24,5,138,102,2,0        // vbroadcastss  0x2668a(%rip),%ymm0        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  197,172,88,232                      // vaddps        %ymm0,%ymm10,%ymm5
   .byte  197,254,91,237                      // vcvttps2dq    %ymm5,%ymm5
   .byte  196,226,93,64,213                   // vpmulld       %ymm5,%ymm4,%ymm2
@@ -23360,7 +23360,7 @@
   .byte  197,254,127,132,36,0,1,0,0          // vmovdqu       %ymm0,0x100(%rsp)
   .byte  196,98,109,64,200                   // vpmulld       %ymm0,%ymm2,%ymm9
   .byte  197,181,254,199                     // vpaddd        %ymm7,%ymm9,%ymm0
-  .byte  196,98,125,88,21,115,74,2,0         // vpbroadcastd  0x24a73(%rip),%ymm10        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  196,98,125,88,21,35,97,2,0          // vpbroadcastd  0x26123(%rip),%ymm10        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  196,194,125,64,194                  // vpmulld       %ymm10,%ymm0,%ymm0
   .byte  197,213,118,237                     // vpcmpeqd      %ymm5,%ymm5,%ymm5
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
@@ -23372,13 +23372,13 @@
   .byte  196,65,36,87,219                    // vxorps        %ymm11,%ymm11,%ymm11
   .byte  196,98,61,146,28,168                // vgatherdps    %ymm8,(%rax,%ymm5,4),%ymm11
   .byte  197,124,17,156,36,192,0,0,0         // vmovups       %ymm11,0xc0(%rsp)
-  .byte  196,98,125,88,29,186,72,2,0         // vpbroadcastd  0x248ba(%rip),%ymm11        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,125,88,29,106,95,2,0         // vpbroadcastd  0x25f6a(%rip),%ymm11        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,193,125,254,195                 // vpaddd        %ymm11,%ymm0,%ymm0
   .byte  196,65,61,118,192                   // vpcmpeqd      %ymm8,%ymm8,%ymm8
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
   .byte  196,226,61,146,44,128               // vgatherdps    %ymm8,(%rax,%ymm0,4),%ymm5
   .byte  197,252,17,108,36,32                // vmovups       %ymm5,0x20(%rsp)
-  .byte  196,226,125,24,5,3,74,2,0           // vbroadcastss  0x24a03(%rip),%ymm0        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,226,125,24,5,179,96,2,0         // vbroadcastss  0x260b3(%rip),%ymm0        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  197,116,88,192                      // vaddps        %ymm0,%ymm1,%ymm8
   .byte  196,65,126,91,192                   // vcvttps2dq    %ymm8,%ymm8
   .byte  196,194,109,64,232                  // vpmulld       %ymm8,%ymm2,%ymm5
@@ -23811,7 +23811,7 @@
   .byte  196,193,100,92,210                  // vsubps        %ymm10,%ymm3,%ymm2
   .byte  196,194,77,168,210                  // vfmadd213ps   %ymm10,%ymm6,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,221,62,2,0        // vbroadcastss  0x23edd(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,141,85,2,0        // vbroadcastss  0x2558d(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,16,164,36,96,6,0,0          // vmovups       0x660(%rsp),%ymm4
   .byte  197,252,16,172,36,128,6,0,0         // vmovups       0x680(%rsp),%ymm5
   .byte  197,252,16,180,36,160,6,0,0         // vmovups       0x6a0(%rsp),%ymm6
@@ -23913,10 +23913,10 @@
   .byte  197,249,112,201,0                   // vpshufd       $0x0,%xmm1,%xmm1
   .byte  196,227,117,24,201,1                // vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,154,61,2,0        // vbroadcastss  0x23d9a(%rip),%ymm2        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,21,74,84,2,0         // vbroadcastss  0x2544a(%rip),%ymm2        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,244,88,202                      // vaddps        %ymm2,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,21,143,61,2,0        // vbroadcastss  0x23d8f(%rip),%ymm2        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,21,63,84,2,0         // vbroadcastss  0x2543f(%rip),%ymm2        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,228,87,219                      // vxorps        %ymm3,%ymm3,%ymm3
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
   .byte  197,212,87,237                      // vxorps        %ymm5,%ymm5,%ymm5
@@ -23930,46 +23930,46 @@
 _sk_dither_avx:
   .byte  197,121,110,194                     // vmovd         %edx,%xmm8
   .byte  196,65,121,112,192,0                // vpshufd       $0x0,%xmm8,%xmm8
-  .byte  197,57,254,13,215,68,2,0            // vpaddd        0x244d7(%rip),%xmm8,%xmm9        # 333a0 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
-  .byte  197,57,254,5,223,68,2,0             // vpaddd        0x244df(%rip),%xmm8,%xmm8        # 333b0 <_sk_srcover_rgba_8888_sse2_lowp+0xa30>
+  .byte  197,57,254,13,183,91,2,0            // vpaddd        0x25bb7(%rip),%xmm8,%xmm9        # 34a80 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  197,57,254,5,191,91,2,0             // vpaddd        0x25bbf(%rip),%xmm8,%xmm8        # 34a90 <_sk_srcover_rgba_8888_sse2_lowp+0xa60>
   .byte  196,67,53,24,208,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm10
   .byte  197,121,110,217                     // vmovd         %ecx,%xmm11
   .byte  196,65,121,112,219,0                // vpshufd       $0x0,%xmm11,%xmm11
   .byte  196,67,37,24,219,1                  // vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
   .byte  196,65,36,87,218                    // vxorps        %ymm10,%ymm11,%ymm11
-  .byte  196,98,125,24,21,63,61,2,0          // vbroadcastss  0x23d3f(%rip),%ymm10        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,98,125,24,21,239,83,2,0         // vbroadcastss  0x253ef(%rip),%ymm10        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  196,65,36,84,210                    // vandps        %ymm10,%ymm11,%ymm10
   .byte  196,193,25,114,242,5                // vpslld        $0x5,%xmm10,%xmm12
   .byte  196,67,125,25,210,1                 // vextractf128  $0x1,%ymm10,%xmm10
   .byte  196,193,41,114,242,5                // vpslld        $0x5,%xmm10,%xmm10
   .byte  196,67,29,24,210,1                  // vinsertf128   $0x1,%xmm10,%ymm12,%ymm10
-  .byte  197,123,18,37,150,76,2,0            // vmovddup      0x24c96(%rip),%xmm12        # 33bb0 <_sk_srcover_rgba_8888_sse2_lowp+0x1230>
+  .byte  197,123,18,37,118,99,2,0            // vmovddup      0x26376(%rip),%xmm12        # 35290 <_sk_srcover_rgba_8888_sse2_lowp+0x1260>
   .byte  196,65,49,219,236                   // vpand         %xmm12,%xmm9,%xmm13
   .byte  196,193,17,114,245,4                // vpslld        $0x4,%xmm13,%xmm13
   .byte  196,65,57,219,228                   // vpand         %xmm12,%xmm8,%xmm12
   .byte  196,193,25,114,244,4                // vpslld        $0x4,%xmm12,%xmm12
   .byte  196,67,21,24,228,1                  // vinsertf128   $0x1,%xmm12,%ymm13,%ymm12
-  .byte  196,98,125,24,45,253,60,2,0         // vbroadcastss  0x23cfd(%rip),%ymm13        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,125,24,45,173,83,2,0         // vbroadcastss  0x253ad(%rip),%ymm13        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,65,36,84,237                    // vandps        %ymm13,%ymm11,%ymm13
   .byte  196,193,9,114,245,2                 // vpslld        $0x2,%xmm13,%xmm14
   .byte  196,67,125,25,237,1                 // vextractf128  $0x1,%ymm13,%xmm13
   .byte  196,193,17,114,245,2                // vpslld        $0x2,%xmm13,%xmm13
   .byte  196,67,13,24,237,1                  // vinsertf128   $0x1,%xmm13,%ymm14,%ymm13
-  .byte  197,123,18,53,84,76,2,0             // vmovddup      0x24c54(%rip),%xmm14        # 33bb8 <_sk_srcover_rgba_8888_sse2_lowp+0x1238>
+  .byte  197,123,18,53,52,99,2,0             // vmovddup      0x26334(%rip),%xmm14        # 35298 <_sk_srcover_rgba_8888_sse2_lowp+0x1268>
   .byte  196,65,49,219,254                   // vpand         %xmm14,%xmm9,%xmm15
   .byte  196,65,1,254,255                    // vpaddd        %xmm15,%xmm15,%xmm15
   .byte  196,65,57,219,246                   // vpand         %xmm14,%xmm8,%xmm14
   .byte  196,65,9,254,246                    // vpaddd        %xmm14,%xmm14,%xmm14
   .byte  196,67,5,24,246,1                   // vinsertf128   $0x1,%xmm14,%ymm15,%ymm14
   .byte  196,65,12,86,228                    // vorps         %ymm12,%ymm14,%ymm12
-  .byte  196,98,125,24,53,172,60,2,0         // vbroadcastss  0x23cac(%rip),%ymm14        # 32c38 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
+  .byte  196,98,125,24,53,92,83,2,0          // vbroadcastss  0x2535c(%rip),%ymm14        # 342e8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
   .byte  196,65,36,84,222                    // vandps        %ymm14,%ymm11,%ymm11
   .byte  196,193,9,114,211,1                 // vpsrld        $0x1,%xmm11,%xmm14
   .byte  196,67,125,25,219,1                 // vextractf128  $0x1,%ymm11,%xmm11
   .byte  196,193,33,114,211,1                // vpsrld        $0x1,%xmm11,%xmm11
   .byte  196,67,13,24,219,1                  // vinsertf128   $0x1,%xmm11,%ymm14,%ymm11
   .byte  196,65,20,86,219                    // vorps         %ymm11,%ymm13,%ymm11
-  .byte  197,123,18,45,10,76,2,0             // vmovddup      0x24c0a(%rip),%xmm13        # 33bc0 <_sk_srcover_rgba_8888_sse2_lowp+0x1240>
+  .byte  197,123,18,45,234,98,2,0            // vmovddup      0x262ea(%rip),%xmm13        # 352a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1270>
   .byte  196,65,49,219,205                   // vpand         %xmm13,%xmm9,%xmm9
   .byte  196,65,57,219,197                   // vpand         %xmm13,%xmm8,%xmm8
   .byte  196,193,49,114,209,2                // vpsrld        $0x2,%xmm9,%xmm9
@@ -23980,9 +23980,9 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  196,65,60,86,195                    // vorps         %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,79,60,2,0          // vbroadcastss  0x23c4f(%rip),%ymm9        # 32c40 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
+  .byte  196,98,125,24,13,255,82,2,0         // vbroadcastss  0x252ff(%rip),%ymm9        # 342f0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,69,60,2,0          // vbroadcastss  0x23c45(%rip),%ymm9        # 32c44 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
+  .byte  196,98,125,24,13,245,82,2,0         // vbroadcastss  0x252f5(%rip),%ymm9        # 342f4 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  196,98,125,24,8                     // vbroadcastss  (%rax),%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
@@ -24016,7 +24016,7 @@
 FUNCTION(_sk_black_color_avx)
 _sk_black_color_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,205,59,2,0        // vbroadcastss  0x23bcd(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,125,82,2,0        // vbroadcastss  0x2527d(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
@@ -24027,7 +24027,7 @@
 FUNCTION(_sk_white_color_avx)
 _sk_white_color_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,180,59,2,0         // vbroadcastss  0x23bb4(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,100,82,2,0         // vbroadcastss  0x25264(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  197,252,40,216                      // vmovaps       %ymm0,%ymm3
@@ -24073,7 +24073,7 @@
 FUNCTION(_sk_srcatop_avx)
 _sk_srcatop_avx:
   .byte  197,252,89,199                      // vmulps        %ymm7,%ymm0,%ymm0
-  .byte  196,98,125,24,5,83,59,2,0           // vbroadcastss  0x23b53(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,3,82,2,0            // vbroadcastss  0x25203(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,204                       // vmulps        %ymm4,%ymm8,%ymm9
   .byte  197,180,88,192                      // vaddps        %ymm0,%ymm9,%ymm0
@@ -24094,7 +24094,7 @@
 FUNCTION(_sk_dstatop_avx)
 _sk_dstatop_avx:
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
-  .byte  196,98,125,24,13,17,59,2,0          // vbroadcastss  0x23b11(%rip),%ymm9        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,13,193,81,2,0         // vbroadcastss  0x251c1(%rip),%ymm9        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,52,92,207                       // vsubps        %ymm7,%ymm9,%ymm9
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
   .byte  197,188,88,192                      // vaddps        %ymm0,%ymm8,%ymm0
@@ -24136,7 +24136,7 @@
 .globl _sk_srcout_avx
 FUNCTION(_sk_srcout_avx)
 _sk_srcout_avx:
-  .byte  196,98,125,24,5,172,58,2,0          // vbroadcastss  0x23aac(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,92,81,2,0           // vbroadcastss  0x2515c(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
@@ -24149,7 +24149,7 @@
 .globl _sk_dstout_avx
 FUNCTION(_sk_dstout_avx)
 _sk_dstout_avx:
-  .byte  196,226,125,24,5,139,58,2,0         // vbroadcastss  0x23a8b(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,59,81,2,0          // vbroadcastss  0x2513b(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,92,219                      // vsubps        %ymm3,%ymm0,%ymm3
   .byte  197,228,89,196                      // vmulps        %ymm4,%ymm3,%ymm0
   .byte  197,228,89,205                      // vmulps        %ymm5,%ymm3,%ymm1
@@ -24162,7 +24162,7 @@
 .globl _sk_srcover_avx
 FUNCTION(_sk_srcover_avx)
 _sk_srcover_avx:
-  .byte  196,98,125,24,5,106,58,2,0          // vbroadcastss  0x23a6a(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,26,81,2,0           // vbroadcastss  0x2511a(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,204                       // vmulps        %ymm4,%ymm8,%ymm9
   .byte  197,180,88,192                      // vaddps        %ymm0,%ymm9,%ymm0
@@ -24179,7 +24179,7 @@
 .globl _sk_dstover_avx
 FUNCTION(_sk_dstover_avx)
 _sk_dstover_avx:
-  .byte  196,98,125,24,5,57,58,2,0           // vbroadcastss  0x23a39(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,233,80,2,0          // vbroadcastss  0x250e9(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,199                       // vsubps        %ymm7,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,252,88,196                      // vaddps        %ymm4,%ymm0,%ymm0
@@ -24207,7 +24207,7 @@
 .globl _sk_multiply_avx
 FUNCTION(_sk_multiply_avx)
 _sk_multiply_avx:
-  .byte  196,98,125,24,5,244,57,2,0          // vbroadcastss  0x239f4(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,164,80,2,0          // vbroadcastss  0x250a4(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,208                       // vmulps        %ymm0,%ymm9,%ymm10
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24238,7 +24238,7 @@
 FUNCTION(_sk_plus__avx)
 _sk_plus__avx:
   .byte  197,252,88,196                      // vaddps        %ymm4,%ymm0,%ymm0
-  .byte  196,98,125,24,5,131,57,2,0          // vbroadcastss  0x23983(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,51,80,2,0           // vbroadcastss  0x25033(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  197,244,88,205                      // vaddps        %ymm5,%ymm1,%ymm1
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
@@ -24272,7 +24272,7 @@
 .globl _sk_xor__avx
 FUNCTION(_sk_xor__avx)
 _sk_xor__avx:
-  .byte  196,98,125,24,5,34,57,2,0           // vbroadcastss  0x23922(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,210,79,2,0          // vbroadcastss  0x24fd2(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,180,89,192                      // vmulps        %ymm0,%ymm9,%ymm0
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24309,7 +24309,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,95,209                  // vmaxps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,158,56,2,0          // vbroadcastss  0x2389e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,78,79,2,0           // vbroadcastss  0x24f4e(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,199                       // vmulps        %ymm7,%ymm8,%ymm8
   .byte  197,188,88,219                      // vaddps        %ymm3,%ymm8,%ymm3
@@ -24335,7 +24335,7 @@
   .byte  197,100,89,206                      // vmulps        %ymm6,%ymm3,%ymm9
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,70,56,2,0           // vbroadcastss  0x23846(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,246,78,2,0          // vbroadcastss  0x24ef6(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,199                       // vmulps        %ymm7,%ymm8,%ymm8
   .byte  197,188,88,219                      // vaddps        %ymm3,%ymm8,%ymm3
@@ -24364,7 +24364,7 @@
   .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,226,55,2,0          // vbroadcastss  0x237e2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,146,78,2,0          // vbroadcastss  0x24e92(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,199                       // vmulps        %ymm7,%ymm8,%ymm8
   .byte  197,188,88,219                      // vaddps        %ymm3,%ymm8,%ymm3
@@ -24387,7 +24387,7 @@
   .byte  197,236,89,214                      // vmulps        %ymm6,%ymm2,%ymm2
   .byte  197,236,88,210                      // vaddps        %ymm2,%ymm2,%ymm2
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
-  .byte  196,98,125,24,5,153,55,2,0          // vbroadcastss  0x23799(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,73,78,2,0           // vbroadcastss  0x24e49(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
   .byte  197,60,89,199                       // vmulps        %ymm7,%ymm8,%ymm8
   .byte  197,188,88,219                      // vaddps        %ymm3,%ymm8,%ymm3
@@ -24398,7 +24398,7 @@
 .globl _sk_colorburn_avx
 FUNCTION(_sk_colorburn_avx)
 _sk_colorburn_avx:
-  .byte  196,98,125,24,5,128,55,2,0          // vbroadcastss  0x23780(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,48,78,2,0           // vbroadcastss  0x24e30(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,52,89,216                       // vmulps        %ymm0,%ymm9,%ymm11
   .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
@@ -24460,7 +24460,7 @@
 FUNCTION(_sk_colordodge_avx)
 _sk_colordodge_avx:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
-  .byte  196,98,125,24,13,132,54,2,0         // vbroadcastss  0x23684(%rip),%ymm9        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,13,52,77,2,0          // vbroadcastss  0x24d34(%rip),%ymm9        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,52,92,215                       // vsubps        %ymm7,%ymm9,%ymm10
   .byte  197,44,89,216                       // vmulps        %ymm0,%ymm10,%ymm11
   .byte  197,52,92,203                       // vsubps        %ymm3,%ymm9,%ymm9
@@ -24517,7 +24517,7 @@
 .globl _sk_hardlight_avx
 FUNCTION(_sk_hardlight_avx)
 _sk_hardlight_avx:
-  .byte  196,98,125,24,5,143,53,2,0          // vbroadcastss  0x2358f(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,63,76,2,0           // vbroadcastss  0x24c3f(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,215                       // vsubps        %ymm7,%ymm8,%ymm10
   .byte  197,44,89,200                       // vmulps        %ymm0,%ymm10,%ymm9
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24572,7 +24572,7 @@
 .globl _sk_overlay_avx
 FUNCTION(_sk_overlay_avx)
 _sk_overlay_avx:
-  .byte  196,98,125,24,5,180,52,2,0          // vbroadcastss  0x234b4(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,100,75,2,0          // vbroadcastss  0x24b64(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,215                       // vsubps        %ymm7,%ymm8,%ymm10
   .byte  197,44,89,200                       // vmulps        %ymm0,%ymm10,%ymm9
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24638,10 +24638,10 @@
   .byte  196,65,60,88,192                    // vaddps        %ymm8,%ymm8,%ymm8
   .byte  196,65,60,89,216                    // vmulps        %ymm8,%ymm8,%ymm11
   .byte  196,65,60,88,195                    // vaddps        %ymm11,%ymm8,%ymm8
-  .byte  196,98,125,24,29,187,51,2,0         // vbroadcastss  0x233bb(%rip),%ymm11        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  196,98,125,24,29,107,74,2,0         // vbroadcastss  0x24a6b(%rip),%ymm11        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  196,65,28,88,235                    // vaddps        %ymm11,%ymm12,%ymm13
   .byte  196,65,20,89,192                    // vmulps        %ymm8,%ymm13,%ymm8
-  .byte  196,98,125,24,45,172,51,2,0         // vbroadcastss  0x233ac(%rip),%ymm13        # 32c4c <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
+  .byte  196,98,125,24,45,92,74,2,0          // vbroadcastss  0x24a5c(%rip),%ymm13        # 342fc <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
   .byte  196,65,28,89,245                    // vmulps        %ymm13,%ymm12,%ymm14
   .byte  196,65,12,88,192                    // vaddps        %ymm8,%ymm14,%ymm8
   .byte  196,65,124,82,244                   // vrsqrtps      %ymm12,%ymm14
@@ -24652,7 +24652,7 @@
   .byte  197,4,194,255,2                     // vcmpleps      %ymm7,%ymm15,%ymm15
   .byte  196,67,13,74,240,240                // vblendvps     %ymm15,%ymm8,%ymm14,%ymm14
   .byte  197,116,88,249                      // vaddps        %ymm1,%ymm1,%ymm15
-  .byte  196,98,125,24,5,86,51,2,0           // vbroadcastss  0x23356(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,6,74,2,0            // vbroadcastss  0x24a06(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,65,60,92,228                    // vsubps        %ymm12,%ymm8,%ymm12
   .byte  197,132,92,195                      // vsubps        %ymm3,%ymm15,%ymm0
   .byte  196,65,124,89,228                   // vmulps        %ymm12,%ymm0,%ymm12
@@ -24779,12 +24779,12 @@
   .byte  196,65,28,89,219                    // vmulps        %ymm11,%ymm12,%ymm11
   .byte  196,65,36,94,222                    // vdivps        %ymm14,%ymm11,%ymm11
   .byte  196,67,37,74,224,240                // vblendvps     %ymm15,%ymm8,%ymm11,%ymm12
-  .byte  196,98,125,24,53,57,49,2,0          // vbroadcastss  0x23139(%rip),%ymm14        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,233,71,2,0         // vbroadcastss  0x247e9(%rip),%ymm14        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,47,49,2,0          // vbroadcastss  0x2312f(%rip),%ymm15        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,61,223,71,2,0         // vbroadcastss  0x247df(%rip),%ymm15        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,32,49,2,0          // vbroadcastss  0x23120(%rip),%ymm0        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,226,125,24,5,208,71,2,0         // vbroadcastss  0x247d0(%rip),%ymm0        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  197,76,89,232                       // vmulps        %ymm0,%ymm6,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
   .byte  196,65,52,89,238                    // vmulps        %ymm14,%ymm9,%ymm13
@@ -24845,7 +24845,7 @@
   .byte  196,65,36,95,208                    // vmaxps        %ymm8,%ymm11,%ymm10
   .byte  196,195,109,74,209,240              // vblendvps     %ymm15,%ymm9,%ymm2,%ymm2
   .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,205,47,2,0          // vbroadcastss  0x22fcd(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,125,70,2,0          // vbroadcastss  0x2467d(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,180,89,201                      // vmulps        %ymm1,%ymm9,%ymm1
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24902,12 +24902,12 @@
   .byte  196,65,28,89,219                    // vmulps        %ymm11,%ymm12,%ymm11
   .byte  196,65,36,94,222                    // vdivps        %ymm14,%ymm11,%ymm11
   .byte  196,67,37,74,224,240                // vblendvps     %ymm15,%ymm8,%ymm11,%ymm12
-  .byte  196,98,125,24,53,247,46,2,0         // vbroadcastss  0x22ef7(%rip),%ymm14        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,53,167,69,2,0         // vbroadcastss  0x245a7(%rip),%ymm14        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,92,89,222                    // vmulps        %ymm14,%ymm4,%ymm11
-  .byte  196,98,125,24,61,237,46,2,0         // vbroadcastss  0x22eed(%rip),%ymm15        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,61,157,69,2,0         // vbroadcastss  0x2459d(%rip),%ymm15        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,239                    // vmulps        %ymm15,%ymm5,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
-  .byte  196,226,125,24,5,222,46,2,0         // vbroadcastss  0x22ede(%rip),%ymm0        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,226,125,24,5,142,69,2,0         // vbroadcastss  0x2458e(%rip),%ymm0        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  197,76,89,232                       // vmulps        %ymm0,%ymm6,%ymm13
   .byte  196,65,36,88,221                    // vaddps        %ymm13,%ymm11,%ymm11
   .byte  196,65,52,89,238                    // vmulps        %ymm14,%ymm9,%ymm13
@@ -24968,7 +24968,7 @@
   .byte  196,65,36,95,208                    // vmaxps        %ymm8,%ymm11,%ymm10
   .byte  196,195,109,74,209,240              // vblendvps     %ymm15,%ymm9,%ymm2,%ymm2
   .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,139,45,2,0          // vbroadcastss  0x22d8b(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,59,68,2,0           // vbroadcastss  0x2443b(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,207                       // vsubps        %ymm7,%ymm8,%ymm9
   .byte  197,180,89,201                      // vmulps        %ymm1,%ymm9,%ymm1
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
@@ -24997,12 +24997,12 @@
   .byte  197,252,17,68,36,168                // vmovups       %ymm0,-0x58(%rsp)
   .byte  197,124,89,199                      // vmulps        %ymm7,%ymm0,%ymm8
   .byte  197,116,89,207                      // vmulps        %ymm7,%ymm1,%ymm9
-  .byte  196,98,125,24,45,61,45,2,0          // vbroadcastss  0x22d3d(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,45,237,67,2,0         // vbroadcastss  0x243ed(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,92,89,213                    // vmulps        %ymm13,%ymm4,%ymm10
-  .byte  196,98,125,24,53,51,45,2,0          // vbroadcastss  0x22d33(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,53,227,67,2,0         // vbroadcastss  0x243e3(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,84,89,222                    // vmulps        %ymm14,%ymm5,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,36,45,2,0          // vbroadcastss  0x22d24(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,212,67,2,0         // vbroadcastss  0x243d4(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,65,76,89,223                    // vmulps        %ymm15,%ymm6,%ymm11
   .byte  196,193,44,88,195                   // vaddps        %ymm11,%ymm10,%ymm0
   .byte  196,65,60,89,221                    // vmulps        %ymm13,%ymm8,%ymm11
@@ -25065,7 +25065,7 @@
   .byte  196,65,44,95,207                    // vmaxps        %ymm15,%ymm10,%ymm9
   .byte  196,195,37,74,192,0                 // vblendvps     %ymm0,%ymm8,%ymm11,%ymm0
   .byte  196,65,124,95,199                   // vmaxps        %ymm15,%ymm0,%ymm8
-  .byte  196,226,125,24,5,191,43,2,0         // vbroadcastss  0x22bbf(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,111,66,2,0         // vbroadcastss  0x2426f(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,124,92,215                      // vsubps        %ymm7,%ymm0,%ymm10
   .byte  197,172,89,84,36,168                // vmulps        -0x58(%rsp),%ymm10,%ymm2
   .byte  197,124,92,219                      // vsubps        %ymm3,%ymm0,%ymm11
@@ -25095,12 +25095,12 @@
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  197,100,89,196                      // vmulps        %ymm4,%ymm3,%ymm8
   .byte  197,100,89,205                      // vmulps        %ymm5,%ymm3,%ymm9
-  .byte  196,98,125,24,45,109,43,2,0         // vbroadcastss  0x22b6d(%rip),%ymm13        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,45,29,66,2,0          // vbroadcastss  0x2421d(%rip),%ymm13        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,108,89,213                   // vmulps        %ymm13,%ymm2,%ymm10
-  .byte  196,98,125,24,53,99,43,2,0          // vbroadcastss  0x22b63(%rip),%ymm14        # 32c54 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  .byte  196,98,125,24,53,19,66,2,0          // vbroadcastss  0x24213(%rip),%ymm14        # 34304 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   .byte  196,65,116,89,222                   // vmulps        %ymm14,%ymm1,%ymm11
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,61,84,43,2,0          // vbroadcastss  0x22b54(%rip),%ymm15        # 32c58 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  .byte  196,98,125,24,61,4,66,2,0           // vbroadcastss  0x24204(%rip),%ymm15        # 34308 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   .byte  196,65,28,89,223                    // vmulps        %ymm15,%ymm12,%ymm11
   .byte  196,193,44,88,195                   // vaddps        %ymm11,%ymm10,%ymm0
   .byte  196,65,60,89,221                    // vmulps        %ymm13,%ymm8,%ymm11
@@ -25163,7 +25163,7 @@
   .byte  196,65,44,95,207                    // vmaxps        %ymm15,%ymm10,%ymm9
   .byte  196,195,37,74,192,0                 // vblendvps     %ymm0,%ymm8,%ymm11,%ymm0
   .byte  196,65,124,95,199                   // vmaxps        %ymm15,%ymm0,%ymm8
-  .byte  196,226,125,24,5,239,41,2,0         // vbroadcastss  0x229ef(%rip),%ymm0        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,5,159,64,2,0         // vbroadcastss  0x2409f(%rip),%ymm0        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,124,92,215                      // vsubps        %ymm7,%ymm0,%ymm10
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  197,124,92,219                      // vsubps        %ymm3,%ymm0,%ymm11
@@ -25197,7 +25197,7 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,42,1,0,0                     // jne           103d6 <_sk_srcover_rgba_8888_avx+0x14a>
   .byte  196,1,125,16,4,152                  // vmovupd       (%r8,%r11,4),%ymm8
-  .byte  197,125,40,13,102,46,2,0            // vmovapd       0x22e66(%rip),%ymm9        # 33120 <_sk_srcover_rgba_8888_sse2_lowp+0x7a0>
+  .byte  197,125,40,13,6,69,2,0              // vmovapd       0x24506(%rip),%ymm9        # 347c0 <_sk_srcover_rgba_8888_sse2_lowp+0x790>
   .byte  196,193,61,84,225                   // vandpd        %ymm9,%ymm8,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
   .byte  196,193,81,114,208,8                // vpsrld        $0x8,%xmm8,%xmm5
@@ -25215,9 +25215,9 @@
   .byte  197,193,114,215,24                  // vpsrld        $0x18,%xmm7,%xmm7
   .byte  196,227,61,24,255,1                 // vinsertf128   $0x1,%xmm7,%ymm8,%ymm7
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  196,98,125,24,5,21,41,2,0           // vbroadcastss  0x22915(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,197,63,2,0          // vbroadcastss  0x23fc5(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,195                       // vsubps        %ymm3,%ymm8,%ymm8
-  .byte  196,98,125,24,13,52,41,2,0          // vbroadcastss  0x22934(%rip),%ymm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,13,228,63,2,0         // vbroadcastss  0x23fe4(%rip),%ymm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,60,89,212                       // vmulps        %ymm4,%ymm8,%ymm10
   .byte  196,193,124,88,194                  // vaddps        %ymm10,%ymm0,%ymm0
@@ -25361,7 +25361,7 @@
 .globl _sk_clamp_1_avx
 FUNCTION(_sk_clamp_1_avx)
 _sk_clamp_1_avx:
-  .byte  196,98,125,24,5,210,38,2,0          // vbroadcastss  0x226d2(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,130,61,2,0          // vbroadcastss  0x23d82(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
@@ -25373,7 +25373,7 @@
 .globl _sk_clamp_a_avx
 FUNCTION(_sk_clamp_a_avx)
 _sk_clamp_a_avx:
-  .byte  196,98,125,24,5,177,38,2,0          // vbroadcastss  0x226b1(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,97,61,2,0           // vbroadcastss  0x23d61(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  197,252,93,195                      // vminps        %ymm3,%ymm0,%ymm0
   .byte  197,244,93,203                      // vminps        %ymm3,%ymm1,%ymm1
@@ -25385,7 +25385,7 @@
 .globl _sk_clamp_a_dst_avx
 FUNCTION(_sk_clamp_a_dst_avx)
 _sk_clamp_a_dst_avx:
-  .byte  196,98,125,24,5,147,38,2,0          // vbroadcastss  0x22693(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,67,61,2,0           // vbroadcastss  0x23d43(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,68,93,248                   // vminps        %ymm8,%ymm7,%ymm7
   .byte  197,220,93,231                      // vminps        %ymm7,%ymm4,%ymm4
   .byte  197,212,93,239                      // vminps        %ymm7,%ymm5,%ymm5
@@ -25418,7 +25418,7 @@
 .globl _sk_invert_avx
 FUNCTION(_sk_invert_avx)
 _sk_invert_avx:
-  .byte  196,98,125,24,5,78,38,2,0           // vbroadcastss  0x2264e(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,254,60,2,0          // vbroadcastss  0x23cfe(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,188,92,192                      // vsubps        %ymm0,%ymm8,%ymm0
   .byte  197,188,92,201                      // vsubps        %ymm1,%ymm8,%ymm1
   .byte  197,188,92,210                      // vsubps        %ymm2,%ymm8,%ymm2
@@ -25472,9 +25472,9 @@
 .globl _sk_unpremul_avx
 FUNCTION(_sk_unpremul_avx)
 _sk_unpremul_avx:
-  .byte  196,98,125,24,5,233,37,2,0          // vbroadcastss  0x225e9(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,153,60,2,0          // vbroadcastss  0x23c99(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,94,195                       // vdivps        %ymm3,%ymm8,%ymm8
-  .byte  196,98,125,24,13,12,38,2,0          // vbroadcastss  0x2260c(%rip),%ymm9        # 32c60 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
+  .byte  196,98,125,24,13,188,60,2,0         // vbroadcastss  0x23cbc(%rip),%ymm9        # 34310 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
   .byte  196,65,60,194,201,1                 // vcmpltps      %ymm9,%ymm8,%ymm9
   .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
   .byte  196,67,45,74,192,144                // vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
@@ -25488,17 +25488,17 @@
 .globl _sk_from_srgb_avx
 FUNCTION(_sk_from_srgb_avx)
 _sk_from_srgb_avx:
-  .byte  196,98,125,24,5,230,37,2,0          // vbroadcastss  0x225e6(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,150,60,2,0          // vbroadcastss  0x23c96(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  197,124,89,208                      // vmulps        %ymm0,%ymm0,%ymm10
-  .byte  196,98,125,24,29,192,37,2,0         // vbroadcastss  0x225c0(%rip),%ymm11        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,112,60,2,0         // vbroadcastss  0x23c70(%rip),%ymm11        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,124,89,227                   // vmulps        %ymm11,%ymm0,%ymm12
-  .byte  196,98,125,24,45,202,37,2,0         // vbroadcastss  0x225ca(%rip),%ymm13        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,45,122,60,2,0         // vbroadcastss  0x23c7a(%rip),%ymm13        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,44,89,212                    // vmulps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,187,37,2,0         // vbroadcastss  0x225bb(%rip),%ymm12        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,37,107,60,2,0         // vbroadcastss  0x23c6b(%rip),%ymm12        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,53,177,37,2,0         // vbroadcastss  0x225b1(%rip),%ymm14        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,53,97,60,2,0          // vbroadcastss  0x23c61(%rip),%ymm14        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  196,193,124,194,198,1               // vcmpltps      %ymm14,%ymm0,%ymm0
   .byte  196,195,45,74,193,0                 // vblendvps     %ymm0,%ymm9,%ymm10,%ymm0
   .byte  196,65,116,89,200                   // vmulps        %ymm8,%ymm1,%ymm9
@@ -25524,17 +25524,17 @@
 .globl _sk_from_srgb_dst_avx
 FUNCTION(_sk_from_srgb_dst_avx)
 _sk_from_srgb_dst_avx:
-  .byte  196,98,125,24,5,58,37,2,0           // vbroadcastss  0x2253a(%rip),%ymm8        # 32c64 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  .byte  196,98,125,24,5,234,59,2,0          // vbroadcastss  0x23bea(%rip),%ymm8        # 34314 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   .byte  196,65,92,89,200                    // vmulps        %ymm8,%ymm4,%ymm9
   .byte  197,92,89,212                       // vmulps        %ymm4,%ymm4,%ymm10
-  .byte  196,98,125,24,29,20,37,2,0          // vbroadcastss  0x22514(%rip),%ymm11        # 32c50 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  .byte  196,98,125,24,29,196,59,2,0         // vbroadcastss  0x23bc4(%rip),%ymm11        # 34300 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   .byte  196,65,92,89,227                    // vmulps        %ymm11,%ymm4,%ymm12
-  .byte  196,98,125,24,45,30,37,2,0          // vbroadcastss  0x2251e(%rip),%ymm13        # 32c68 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  .byte  196,98,125,24,45,206,59,2,0         // vbroadcastss  0x23bce(%rip),%ymm13        # 34318 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,44,89,212                    // vmulps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,37,15,37,2,0          // vbroadcastss  0x2250f(%rip),%ymm12        # 32c6c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  .byte  196,98,125,24,37,191,59,2,0         // vbroadcastss  0x23bbf(%rip),%ymm12        # 3431c <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   .byte  196,65,44,88,212                    // vaddps        %ymm12,%ymm10,%ymm10
-  .byte  196,98,125,24,53,5,37,2,0           // vbroadcastss  0x22505(%rip),%ymm14        # 32c70 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  .byte  196,98,125,24,53,181,59,2,0         // vbroadcastss  0x23bb5(%rip),%ymm14        # 34320 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   .byte  196,193,92,194,230,1                // vcmpltps      %ymm14,%ymm4,%ymm4
   .byte  196,195,45,74,225,64                // vblendvps     %ymm4,%ymm9,%ymm10,%ymm4
   .byte  196,65,84,89,200                    // vmulps        %ymm8,%ymm5,%ymm9
@@ -25561,20 +25561,20 @@
 FUNCTION(_sk_to_srgb_avx)
 _sk_to_srgb_avx:
   .byte  197,124,82,200                      // vrsqrtps      %ymm0,%ymm9
-  .byte  196,98,125,24,5,154,36,2,0          // vbroadcastss  0x2249a(%rip),%ymm8        # 32c74 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
+  .byte  196,98,125,24,5,74,59,2,0           // vbroadcastss  0x23b4a(%rip),%ymm8        # 34324 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
   .byte  196,65,124,89,208                   // vmulps        %ymm8,%ymm0,%ymm10
-  .byte  196,98,125,24,29,148,36,2,0         // vbroadcastss  0x22494(%rip),%ymm11        # 32c7c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
+  .byte  196,98,125,24,29,68,59,2,0          // vbroadcastss  0x23b44(%rip),%ymm11        # 3432c <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
   .byte  196,65,52,89,227                    // vmulps        %ymm11,%ymm9,%ymm12
-  .byte  196,98,125,24,45,130,36,2,0         // vbroadcastss  0x22482(%rip),%ymm13        # 32c78 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
+  .byte  196,98,125,24,45,50,59,2,0          // vbroadcastss  0x23b32(%rip),%ymm13        # 34328 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,52,89,228                    // vmulps        %ymm12,%ymm9,%ymm12
-  .byte  196,98,125,24,53,167,37,2,0         // vbroadcastss  0x225a7(%rip),%ymm14        # 32db0 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
+  .byte  196,98,125,24,53,87,60,2,0          // vbroadcastss  0x23c57(%rip),%ymm14        # 34460 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
   .byte  196,65,28,88,230                    // vaddps        %ymm14,%ymm12,%ymm12
-  .byte  196,98,125,24,61,157,37,2,0         // vbroadcastss  0x2259d(%rip),%ymm15        # 32db4 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
+  .byte  196,98,125,24,61,77,60,2,0          // vbroadcastss  0x23c4d(%rip),%ymm15        # 34464 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
   .byte  196,65,52,88,207                    // vaddps        %ymm15,%ymm9,%ymm9
   .byte  196,65,124,83,201                   // vrcpps        %ymm9,%ymm9
   .byte  196,65,52,89,204                    // vmulps        %ymm12,%ymm9,%ymm9
-  .byte  196,98,125,24,37,89,36,2,0          // vbroadcastss  0x22459(%rip),%ymm12        # 32c88 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
+  .byte  196,98,125,24,37,9,59,2,0           // vbroadcastss  0x23b09(%rip),%ymm12        # 34338 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
   .byte  196,193,124,194,196,1               // vcmpltps      %ymm12,%ymm0,%ymm0
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  197,124,82,201                      // vrsqrtps      %ymm1,%ymm9
@@ -25611,7 +25611,7 @@
   .byte  197,116,93,202                      // vminps        %ymm2,%ymm1,%ymm9
   .byte  196,65,124,93,201                   // vminps        %ymm9,%ymm0,%ymm9
   .byte  196,65,60,92,209                    // vsubps        %ymm9,%ymm8,%ymm10
-  .byte  196,98,125,24,29,97,35,2,0          // vbroadcastss  0x22361(%rip),%ymm11        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,29,17,58,2,0          // vbroadcastss  0x23a11(%rip),%ymm11        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,65,36,94,218                    // vdivps        %ymm10,%ymm11,%ymm11
   .byte  197,116,92,226                      // vsubps        %ymm2,%ymm1,%ymm12
   .byte  196,65,28,89,227                    // vmulps        %ymm11,%ymm12,%ymm12
@@ -25621,19 +25621,19 @@
   .byte  196,193,108,89,211                  // vmulps        %ymm11,%ymm2,%ymm2
   .byte  197,252,92,201                      // vsubps        %ymm1,%ymm0,%ymm1
   .byte  196,193,116,89,203                  // vmulps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,146,35,2,0         // vbroadcastss  0x22392(%rip),%ymm11        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  196,98,125,24,29,66,58,2,0          // vbroadcastss  0x23a42(%rip),%ymm11        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,128,35,2,0         // vbroadcastss  0x22380(%rip),%ymm11        # 32c90 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  .byte  196,98,125,24,29,48,58,2,0          // vbroadcastss  0x23a30(%rip),%ymm11        # 34340 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,227,117,74,202,224              // vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,104,35,2,0        // vbroadcastss  0x22368(%rip),%ymm2        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,226,125,24,21,24,58,2,0         // vbroadcastss  0x23a18(%rip),%ymm2        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  196,65,12,87,246                    // vxorps        %ymm14,%ymm14,%ymm14
   .byte  196,227,13,74,210,208               // vblendvps     %ymm13,%ymm2,%ymm14,%ymm2
   .byte  197,188,194,192,0                   // vcmpeqps      %ymm0,%ymm8,%ymm0
   .byte  196,193,108,88,212                  // vaddps        %ymm12,%ymm2,%ymm2
   .byte  196,227,117,74,194,0                // vblendvps     %ymm0,%ymm2,%ymm1,%ymm0
   .byte  196,193,60,88,201                   // vaddps        %ymm9,%ymm8,%ymm1
-  .byte  196,98,125,24,37,223,34,2,0         // vbroadcastss  0x222df(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,143,57,2,0         // vbroadcastss  0x2398f(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,116,89,212                  // vmulps        %ymm12,%ymm1,%ymm2
   .byte  197,28,194,226,1                    // vcmpltps      %ymm2,%ymm12,%ymm12
   .byte  196,65,36,92,216                    // vsubps        %ymm8,%ymm11,%ymm11
@@ -25643,7 +25643,7 @@
   .byte  197,172,94,201                      // vdivps        %ymm1,%ymm10,%ymm1
   .byte  196,195,125,74,198,128              // vblendvps     %ymm8,%ymm14,%ymm0,%ymm0
   .byte  196,195,117,74,206,128              // vblendvps     %ymm8,%ymm14,%ymm1,%ymm1
-  .byte  196,98,125,24,5,18,35,2,0           // vbroadcastss  0x22312(%rip),%ymm8        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  196,98,125,24,5,194,57,2,0          // vbroadcastss  0x239c2(%rip),%ymm8        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -25660,7 +25660,7 @@
   .byte  197,252,17,92,36,128                // vmovups       %ymm3,-0x80(%rsp)
   .byte  197,252,40,225                      // vmovaps       %ymm1,%ymm4
   .byte  197,252,40,216                      // vmovaps       %ymm0,%ymm3
-  .byte  196,98,125,24,5,107,34,2,0          // vbroadcastss  0x2226b(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,27,57,2,0           // vbroadcastss  0x2391b(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,60,194,202,2                    // vcmpleps      %ymm2,%ymm8,%ymm9
   .byte  197,92,89,210                       // vmulps        %ymm2,%ymm4,%ymm10
   .byte  196,65,92,92,218                    // vsubps        %ymm10,%ymm4,%ymm11
@@ -25668,23 +25668,23 @@
   .byte  197,52,88,210                       // vaddps        %ymm2,%ymm9,%ymm10
   .byte  197,108,88,202                      // vaddps        %ymm2,%ymm2,%ymm9
   .byte  196,65,52,92,202                    // vsubps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,29,177,34,2,0         // vbroadcastss  0x222b1(%rip),%ymm11        # 32c9c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
+  .byte  196,98,125,24,29,97,57,2,0          // vbroadcastss  0x23961(%rip),%ymm11        # 3434c <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
   .byte  196,65,100,88,219                   // vaddps        %ymm11,%ymm3,%ymm11
   .byte  196,67,125,8,227,1                  // vroundps      $0x1,%ymm11,%ymm12
   .byte  196,65,36,92,252                    // vsubps        %ymm12,%ymm11,%ymm15
   .byte  196,65,44,92,217                    // vsubps        %ymm9,%ymm10,%ymm11
-  .byte  196,98,125,24,37,131,34,2,0         // vbroadcastss  0x22283(%rip),%ymm12        # 32c8c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  .byte  196,98,125,24,37,51,57,2,0          // vbroadcastss  0x23933(%rip),%ymm12        # 3433c <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   .byte  196,193,4,89,196                    // vmulps        %ymm12,%ymm15,%ymm0
-  .byte  196,98,125,24,45,125,34,2,0         // vbroadcastss  0x2227d(%rip),%ymm13        # 32c94 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  .byte  196,98,125,24,45,45,57,2,0          // vbroadcastss  0x2392d(%rip),%ymm13        # 34344 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   .byte  197,20,92,240                       // vsubps        %ymm0,%ymm13,%ymm14
   .byte  196,65,36,89,246                    // vmulps        %ymm14,%ymm11,%ymm14
   .byte  196,65,52,88,246                    // vaddps        %ymm14,%ymm9,%ymm14
-  .byte  196,226,125,24,13,114,34,2,0        // vbroadcastss  0x22272(%rip),%ymm1        # 32ca0 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
+  .byte  196,226,125,24,13,34,57,2,0         // vbroadcastss  0x23922(%rip),%ymm1        # 34350 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
   .byte  196,193,116,194,255,2               // vcmpleps      %ymm15,%ymm1,%ymm7
   .byte  196,195,13,74,249,112               // vblendvps     %ymm7,%ymm9,%ymm14,%ymm7
   .byte  196,65,60,194,247,2                 // vcmpleps      %ymm15,%ymm8,%ymm14
   .byte  196,227,45,74,255,224               // vblendvps     %ymm14,%ymm7,%ymm10,%ymm7
-  .byte  196,98,125,24,53,73,34,2,0          // vbroadcastss  0x22249(%rip),%ymm14        # 32c98 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  .byte  196,98,125,24,53,249,56,2,0         // vbroadcastss  0x238f9(%rip),%ymm14        # 34348 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   .byte  196,65,12,194,255,2                 // vcmpleps      %ymm15,%ymm14,%ymm15
   .byte  196,193,124,89,195                  // vmulps        %ymm11,%ymm0,%ymm0
   .byte  197,180,88,192                      // vaddps        %ymm0,%ymm9,%ymm0
@@ -25703,7 +25703,7 @@
   .byte  197,164,89,247                      // vmulps        %ymm7,%ymm11,%ymm6
   .byte  197,180,88,246                      // vaddps        %ymm6,%ymm9,%ymm6
   .byte  196,227,77,74,237,0                 // vblendvps     %ymm0,%ymm5,%ymm6,%ymm5
-  .byte  196,226,125,24,5,243,33,2,0         // vbroadcastss  0x221f3(%rip),%ymm0        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,226,125,24,5,163,56,2,0         // vbroadcastss  0x238a3(%rip),%ymm0        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  197,228,88,192                      // vaddps        %ymm0,%ymm3,%ymm0
   .byte  196,227,125,8,216,1                 // vroundps      $0x1,%ymm0,%ymm3
   .byte  197,252,92,195                      // vsubps        %ymm3,%ymm0,%ymm0
@@ -25759,13 +25759,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,75                              // jne           10bb7 <_sk_scale_u8_avx+0x63>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,246,39,2,0             // vpand         0x227f6(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,214,62,2,0             // vpand         0x23ed6(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,121,51,200                   // vpmovzxwd     %xmm8,%xmm9
   .byte  196,65,121,112,192,78               // vpshufd       $0x4e,%xmm8,%xmm8
   .byte  196,66,121,51,192                   // vpmovzxwd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,10,33,2,0          // vbroadcastss  0x2210a(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,186,55,2,0         // vbroadcastss  0x237ba(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  197,188,89,201                      // vmulps        %ymm1,%ymm8,%ymm1
@@ -25845,20 +25845,20 @@
   .byte  196,65,121,112,192,78               // vpshufd       $0x4e,%xmm8,%xmm8
   .byte  196,66,121,51,192                   // vpmovzxwd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
-  .byte  196,98,125,24,13,252,31,2,0         // vbroadcastss  0x21ffc(%rip),%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,98,125,24,13,172,54,2,0         // vbroadcastss  0x236ac(%rip),%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,60,84,201                    // vandps        %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,237,31,2,0         // vbroadcastss  0x21fed(%rip),%ymm10        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,98,125,24,21,157,54,2,0         // vbroadcastss  0x2369d(%rip),%ymm10        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,227,31,2,0         // vbroadcastss  0x21fe3(%rip),%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,98,125,24,21,147,54,2,0         // vbroadcastss  0x23693(%rip),%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,60,84,210                    // vandps        %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,212,31,2,0         // vbroadcastss  0x21fd4(%rip),%ymm11        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,98,125,24,29,132,54,2,0         // vbroadcastss  0x23684(%rip),%ymm11        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,29,202,31,2,0         // vbroadcastss  0x21fca(%rip),%ymm11        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,98,125,24,29,122,54,2,0         // vbroadcastss  0x2367a(%rip),%ymm11        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,60,84,195                    // vandps        %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,187,31,2,0         // vbroadcastss  0x21fbb(%rip),%ymm11        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,98,125,24,29,107,54,2,0         // vbroadcastss  0x2366b(%rip),%ymm11        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  197,100,194,223,1                   // vcmpltps      %ymm7,%ymm3,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -25951,13 +25951,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,111                             // jne           10ea6 <_sk_lerp_u8_avx+0x87>
   .byte  196,2,121,48,4,24                   // vpmovzxbw     (%r8,%r11,1),%xmm8
-  .byte  197,57,219,5,43,37,2,0              // vpand         0x2252b(%rip),%xmm8,%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,57,219,5,11,60,2,0              // vpand         0x23c0b(%rip),%xmm8,%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,66,121,51,200                   // vpmovzxwd     %xmm8,%xmm9
   .byte  196,65,121,112,192,78               // vpshufd       $0x4e,%xmm8,%xmm8
   .byte  196,66,121,51,192                   // vpmovzxwd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,13,63,30,2,0          // vbroadcastss  0x21e3f(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,239,52,2,0         // vbroadcastss  0x234ef(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,252,92,196                      // vsubps        %ymm4,%ymm0,%ymm0
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
@@ -26013,7 +26013,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,221                 // callq         ffffffffde010f54 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfde5d4>
+  .byte  232,255,255,255,221                 // callq         ffffffffde010f54 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfdcf24>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,210                             // callq         *%rdx
@@ -26042,20 +26042,20 @@
   .byte  196,65,121,112,192,78               // vpshufd       $0x4e,%xmm8,%xmm8
   .byte  196,66,121,51,192                   // vpmovzxwd     %xmm8,%xmm8
   .byte  196,67,53,24,192,1                  // vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
-  .byte  196,98,125,24,13,8,29,2,0           // vbroadcastss  0x21d08(%rip),%ymm9        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,98,125,24,13,184,51,2,0         // vbroadcastss  0x233b8(%rip),%ymm9        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  196,65,60,84,201                    // vandps        %ymm9,%ymm8,%ymm9
   .byte  196,65,124,91,201                   // vcvtdq2ps     %ymm9,%ymm9
-  .byte  196,98,125,24,21,249,28,2,0         // vbroadcastss  0x21cf9(%rip),%ymm10        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,98,125,24,21,169,51,2,0         // vbroadcastss  0x233a9(%rip),%ymm10        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,239,28,2,0         // vbroadcastss  0x21cef(%rip),%ymm10        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,98,125,24,21,159,51,2,0         // vbroadcastss  0x2339f(%rip),%ymm10        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  196,65,60,84,210                    // vandps        %ymm10,%ymm8,%ymm10
   .byte  196,65,124,91,210                   // vcvtdq2ps     %ymm10,%ymm10
-  .byte  196,98,125,24,29,224,28,2,0         // vbroadcastss  0x21ce0(%rip),%ymm11        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,98,125,24,29,144,51,2,0         // vbroadcastss  0x23390(%rip),%ymm11        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,24,29,214,28,2,0         // vbroadcastss  0x21cd6(%rip),%ymm11        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,98,125,24,29,134,51,2,0         // vbroadcastss  0x23386(%rip),%ymm11        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  196,65,60,84,195                    // vandps        %ymm11,%ymm8,%ymm8
   .byte  196,65,124,91,192                   // vcvtdq2ps     %ymm8,%ymm8
-  .byte  196,98,125,24,29,199,28,2,0         // vbroadcastss  0x21cc7(%rip),%ymm11        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,98,125,24,29,119,51,2,0         // vbroadcastss  0x23377(%rip),%ymm11        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  197,100,194,223,1                   // vcmpltps      %ymm7,%ymm3,%ymm11
   .byte  196,65,44,93,224                    // vminps        %ymm8,%ymm10,%ymm12
@@ -26132,7 +26132,7 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,4,2,0,0                      // jne           11310 <_sk_load_tables_avx+0x218>
   .byte  196,65,125,16,20,144                // vmovupd       (%r8,%rdx,4),%ymm10
-  .byte  197,125,40,13,38,32,2,0             // vmovapd       0x22026(%rip),%ymm9        # 33140 <_sk_srcover_rgba_8888_sse2_lowp+0x7c0>
+  .byte  197,125,40,13,198,54,2,0            // vmovapd       0x236c6(%rip),%ymm9        # 347e0 <_sk_srcover_rgba_8888_sse2_lowp+0x7b0>
   .byte  196,193,45,84,201                   // vandpd        %ymm9,%ymm10,%ymm1
   .byte  196,227,125,25,200,1                // vextractf128  $0x1,%ymm1,%xmm0
   .byte  196,193,249,126,192                 // vmovq         %xmm0,%r8
@@ -26224,7 +26224,7 @@
   .byte  196,193,65,114,213,24               // vpsrld        $0x18,%xmm13,%xmm7
   .byte  196,227,101,24,223,1                // vinsertf128   $0x1,%xmm7,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,226,125,24,61,166,25,2,0        // vbroadcastss  0x219a6(%rip),%ymm7        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,61,86,48,2,0         // vbroadcastss  0x23056(%rip),%ymm7        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,228,89,223                      // vmulps        %ymm7,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,124,36,200               // vmovups       -0x38(%rsp),%ymm7
@@ -26310,7 +26310,7 @@
   .byte  197,177,108,208                     // vpunpcklqdq   %xmm0,%xmm9,%xmm2
   .byte  197,49,109,232                      // vpunpckhqdq   %xmm0,%xmm9,%xmm13
   .byte  196,65,57,108,212                   // vpunpcklqdq   %xmm12,%xmm8,%xmm10
-  .byte  197,121,111,29,24,31,2,0            // vmovdqa       0x21f18(%rip),%xmm11        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,121,111,29,248,53,2,0           // vmovdqa       0x235f8(%rip),%xmm11        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,193,105,219,195                 // vpand         %xmm11,%xmm2,%xmm0
   .byte  197,249,112,208,78                  // vpshufd       $0x4e,%xmm0,%xmm2
   .byte  196,226,121,51,210                  // vpmovzxwd     %xmm2,%xmm2
@@ -26404,7 +26404,7 @@
   .byte  196,226,121,51,219                  // vpmovzxwd     %xmm3,%xmm3
   .byte  196,227,77,24,219,1                 // vinsertf128   $0x1,%xmm3,%ymm6,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,226,125,24,53,135,22,2,0        // vbroadcastss  0x21687(%rip),%ymm6        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,226,125,24,53,55,45,2,0         // vbroadcastss  0x22d37(%rip),%ymm6        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  197,228,89,222                      // vmulps        %ymm6,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,16,116,36,168               // vmovups       -0x58(%rsp),%ymm6
@@ -26468,7 +26468,7 @@
   .byte  197,105,108,216                     // vpunpcklqdq   %xmm0,%xmm2,%xmm11
   .byte  197,241,108,211                     // vpunpcklqdq   %xmm3,%xmm1,%xmm2
   .byte  197,241,109,203                     // vpunpckhqdq   %xmm3,%xmm1,%xmm1
-  .byte  197,121,111,53,24,28,2,0            // vmovdqa       0x21c18(%rip),%xmm14        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,121,111,53,248,50,2,0           // vmovdqa       0x232f8(%rip),%xmm14        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,193,105,219,214                 // vpand         %xmm14,%xmm2,%xmm2
   .byte  197,249,112,218,78                  // vpshufd       $0x4e,%xmm2,%xmm3
   .byte  196,226,121,51,219                  // vpmovzxwd     %xmm3,%xmm3
@@ -26554,7 +26554,7 @@
   .byte  196,227,105,33,215,48               // vinsertps     $0x30,%xmm7,%xmm2,%xmm2
   .byte  196,227,77,24,210,1                 // vinsertf128   $0x1,%xmm2,%ymm6,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,29,19,2,0         // vbroadcastss  0x2131d(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,205,41,2,0        // vbroadcastss  0x229cd(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,16,116,36,168               // vmovups       -0x58(%rsp),%ymm6
   .byte  197,252,16,124,36,200               // vmovups       -0x38(%rsp),%ymm7
   .byte  255,224                             // jmpq          *%rax
@@ -26625,7 +26625,7 @@
 FUNCTION(_sk_byte_tables_avx)
 _sk_byte_tables_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,6,18,2,0            // vbroadcastss  0x21206(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,182,40,2,0          // vbroadcastss  0x228b6(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  197,125,91,200                      // vcvtps2dq     %ymm0,%ymm9
   .byte  196,65,249,126,200                  // vmovq         %xmm9,%r8
@@ -26760,7 +26760,7 @@
   .byte  196,194,121,49,204                  // vpmovzxbd     %xmm12,%xmm1
   .byte  196,194,121,49,213                  // vpmovzxbd     %xmm13,%xmm2
   .byte  196,227,117,24,202,1                // vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
-  .byte  196,98,125,24,13,194,15,2,0         // vbroadcastss  0x20fc2(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,114,38,2,0         // vbroadcastss  0x22672(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
   .byte  196,193,116,89,201                  // vmulps        %ymm9,%ymm1,%ymm1
@@ -26888,7 +26888,7 @@
   .byte  196,194,121,49,203                  // vpmovzxbd     %xmm11,%xmm1
   .byte  196,227,125,24,193,1                // vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,119,13,2,0         // vbroadcastss  0x20d77(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,39,36,2,0          // vbroadcastss  0x22427(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,194,121,49,202                  // vpmovzxbd     %xmm10,%xmm1
   .byte  196,194,121,49,212                  // vpmovzxbd     %xmm12,%xmm2
@@ -27125,36 +27125,36 @@
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
   .byte  196,98,125,24,16                    // vbroadcastss  (%rax),%ymm10
   .byte  197,124,91,216                      // vcvtdq2ps     %ymm0,%ymm11
-  .byte  196,98,125,24,37,8,10,2,0           // vbroadcastss  0x20a08(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,37,184,32,2,0         // vbroadcastss  0x220b8(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,254,9,2,0          // vbroadcastss  0x209fe(%rip),%ymm12        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,174,32,2,0         // vbroadcastss  0x220ae(%rip),%ymm12        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,124,84,196                  // vandps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,80,9,2,0           // vbroadcastss  0x20950(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,0,32,2,0           // vbroadcastss  0x22000(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,124,86,196                  // vorps         %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,230,9,2,0          // vbroadcastss  0x209e6(%rip),%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,37,150,32,2,0         // vbroadcastss  0x22096(%rip),%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,220,9,2,0          // vbroadcastss  0x209dc(%rip),%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,37,140,32,2,0         // vbroadcastss  0x2208c(%rip),%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,65,124,89,228                   // vmulps        %ymm12,%ymm0,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,205,9,2,0          // vbroadcastss  0x209cd(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,125,32,2,0         // vbroadcastss  0x2207d(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,124,88,196                  // vaddps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,37,195,9,2,0          // vbroadcastss  0x209c3(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,115,32,2,0         // vbroadcastss  0x22073(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  197,156,94,192                      // vdivps        %ymm0,%ymm12,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
   .byte  197,172,89,192                      // vmulps        %ymm0,%ymm10,%ymm0
   .byte  196,99,125,8,208,1                  // vroundps      $0x1,%ymm0,%ymm10
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,167,9,2,0          // vbroadcastss  0x209a7(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,87,32,2,0          // vbroadcastss  0x22057(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,124,88,195                  // vaddps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,157,9,2,0          // vbroadcastss  0x2099d(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,77,32,2,0          // vbroadcastss  0x2204d(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,124,92,195                  // vsubps        %ymm11,%ymm0,%ymm0
-  .byte  196,98,125,24,29,142,9,2,0          // vbroadcastss  0x2098e(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,98,125,24,29,62,32,2,0          // vbroadcastss  0x2203e(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,132,9,2,0          // vbroadcastss  0x20984(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,52,32,2,0          // vbroadcastss  0x22034(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,124,88,194                  // vaddps        %ymm10,%ymm0,%ymm0
-  .byte  196,98,125,24,21,117,9,2,0          // vbroadcastss  0x20975(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,37,32,2,0          // vbroadcastss  0x22025(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,124,89,194                  // vmulps        %ymm10,%ymm0,%ymm0
   .byte  197,253,91,192                      // vcvtps2dq     %ymm0,%ymm0
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -27162,7 +27162,7 @@
   .byte  196,195,125,74,193,128              // vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,137,8,2,0           // vbroadcastss  0x20889(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,57,31,2,0           // vbroadcastss  0x21f39(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -27184,36 +27184,36 @@
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
   .byte  196,98,125,24,16                    // vbroadcastss  (%rax),%ymm10
   .byte  197,124,91,217                      // vcvtdq2ps     %ymm1,%ymm11
-  .byte  196,98,125,24,37,206,8,2,0          // vbroadcastss  0x208ce(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,37,126,31,2,0         // vbroadcastss  0x21f7e(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,196,8,2,0          // vbroadcastss  0x208c4(%rip),%ymm12        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,116,31,2,0         // vbroadcastss  0x21f74(%rip),%ymm12        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,116,84,204                  // vandps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,22,8,2,0           // vbroadcastss  0x20816(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,198,30,2,0         // vbroadcastss  0x21ec6(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,116,86,204                  // vorps         %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,172,8,2,0          // vbroadcastss  0x208ac(%rip),%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,37,92,31,2,0          // vbroadcastss  0x21f5c(%rip),%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,162,8,2,0          // vbroadcastss  0x208a2(%rip),%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,37,82,31,2,0          // vbroadcastss  0x21f52(%rip),%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,65,116,89,228                   // vmulps        %ymm12,%ymm1,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,147,8,2,0          // vbroadcastss  0x20893(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,67,31,2,0          // vbroadcastss  0x21f43(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,116,88,204                  // vaddps        %ymm12,%ymm1,%ymm1
-  .byte  196,98,125,24,37,137,8,2,0          // vbroadcastss  0x20889(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,57,31,2,0          // vbroadcastss  0x21f39(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  197,156,94,201                      // vdivps        %ymm1,%ymm12,%ymm1
   .byte  197,164,92,201                      // vsubps        %ymm1,%ymm11,%ymm1
   .byte  197,172,89,201                      // vmulps        %ymm1,%ymm10,%ymm1
   .byte  196,99,125,8,209,1                  // vroundps      $0x1,%ymm1,%ymm10
   .byte  196,65,116,92,210                   // vsubps        %ymm10,%ymm1,%ymm10
-  .byte  196,98,125,24,29,109,8,2,0          // vbroadcastss  0x2086d(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,29,31,2,0          // vbroadcastss  0x21f1d(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,99,8,2,0           // vbroadcastss  0x20863(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,19,31,2,0          // vbroadcastss  0x21f13(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,116,92,203                  // vsubps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,29,84,8,2,0           // vbroadcastss  0x20854(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,98,125,24,29,4,31,2,0           // vbroadcastss  0x21f04(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,74,8,2,0           // vbroadcastss  0x2084a(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,250,30,2,0         // vbroadcastss  0x21efa(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,116,88,202                  // vaddps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,21,59,8,2,0           // vbroadcastss  0x2083b(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,235,30,2,0         // vbroadcastss  0x21eeb(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,116,89,202                  // vmulps        %ymm10,%ymm1,%ymm1
   .byte  197,253,91,201                      // vcvtps2dq     %ymm1,%ymm1
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -27221,7 +27221,7 @@
   .byte  196,195,117,74,201,128              // vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,116,95,200                  // vmaxps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,5,79,7,2,0            // vbroadcastss  0x2074f(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,255,29,2,0          // vbroadcastss  0x21dff(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -27243,36 +27243,36 @@
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,98,125,24,16                    // vbroadcastss  (%rax),%ymm10
   .byte  197,124,91,218                      // vcvtdq2ps     %ymm2,%ymm11
-  .byte  196,98,125,24,37,148,7,2,0          // vbroadcastss  0x20794(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,37,68,30,2,0          // vbroadcastss  0x21e44(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,138,7,2,0          // vbroadcastss  0x2078a(%rip),%ymm12        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,58,30,2,0          // vbroadcastss  0x21e3a(%rip),%ymm12        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,108,84,212                  // vandps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,220,6,2,0          // vbroadcastss  0x206dc(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,140,29,2,0         // vbroadcastss  0x21d8c(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,108,86,212                  // vorps         %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,114,7,2,0          // vbroadcastss  0x20772(%rip),%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,37,34,30,2,0          // vbroadcastss  0x21e22(%rip),%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,104,7,2,0          // vbroadcastss  0x20768(%rip),%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,37,24,30,2,0          // vbroadcastss  0x21e18(%rip),%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,65,108,89,228                   // vmulps        %ymm12,%ymm2,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,89,7,2,0           // vbroadcastss  0x20759(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,9,30,2,0           // vbroadcastss  0x21e09(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,108,88,212                  // vaddps        %ymm12,%ymm2,%ymm2
-  .byte  196,98,125,24,37,79,7,2,0           // vbroadcastss  0x2074f(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,255,29,2,0         // vbroadcastss  0x21dff(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  197,156,94,210                      // vdivps        %ymm2,%ymm12,%ymm2
   .byte  197,164,92,210                      // vsubps        %ymm2,%ymm11,%ymm2
   .byte  197,172,89,210                      // vmulps        %ymm2,%ymm10,%ymm2
   .byte  196,99,125,8,210,1                  // vroundps      $0x1,%ymm2,%ymm10
   .byte  196,65,108,92,210                   // vsubps        %ymm10,%ymm2,%ymm10
-  .byte  196,98,125,24,29,51,7,2,0           // vbroadcastss  0x20733(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,227,29,2,0         // vbroadcastss  0x21de3(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,41,7,2,0           // vbroadcastss  0x20729(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,217,29,2,0         // vbroadcastss  0x21dd9(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,108,92,211                  // vsubps        %ymm11,%ymm2,%ymm2
-  .byte  196,98,125,24,29,26,7,2,0           // vbroadcastss  0x2071a(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,98,125,24,29,202,29,2,0         // vbroadcastss  0x21dca(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,16,7,2,0           // vbroadcastss  0x20710(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,192,29,2,0         // vbroadcastss  0x21dc0(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,108,88,210                  // vaddps        %ymm10,%ymm2,%ymm2
-  .byte  196,98,125,24,21,1,7,2,0            // vbroadcastss  0x20701(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,177,29,2,0         // vbroadcastss  0x21db1(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,108,89,210                  // vmulps        %ymm10,%ymm2,%ymm2
   .byte  197,253,91,210                      // vcvtps2dq     %ymm2,%ymm2
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -27280,7 +27280,7 @@
   .byte  196,195,109,74,209,128              // vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,21,6,2,0            // vbroadcastss  0x20615(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,197,28,2,0          // vbroadcastss  0x21cc5(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -27302,36 +27302,36 @@
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
   .byte  196,98,125,24,16                    // vbroadcastss  (%rax),%ymm10
   .byte  197,124,91,219                      // vcvtdq2ps     %ymm3,%ymm11
-  .byte  196,98,125,24,37,90,6,2,0           // vbroadcastss  0x2065a(%rip),%ymm12        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,37,10,29,2,0          // vbroadcastss  0x21d0a(%rip),%ymm12        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,80,6,2,0           // vbroadcastss  0x20650(%rip),%ymm12        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,37,0,29,2,0           // vbroadcastss  0x21d00(%rip),%ymm12        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,100,84,220                  // vandps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,162,5,2,0          // vbroadcastss  0x205a2(%rip),%ymm12        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,37,82,28,2,0          // vbroadcastss  0x21c52(%rip),%ymm12        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,100,86,220                  // vorps         %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,56,6,2,0           // vbroadcastss  0x20638(%rip),%ymm12        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,37,232,28,2,0         // vbroadcastss  0x21ce8(%rip),%ymm12        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,65,36,88,220                    // vaddps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,46,6,2,0           // vbroadcastss  0x2062e(%rip),%ymm12        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,37,222,28,2,0         // vbroadcastss  0x21cde(%rip),%ymm12        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,65,100,89,228                   // vmulps        %ymm12,%ymm3,%ymm12
   .byte  196,65,36,92,220                    // vsubps        %ymm12,%ymm11,%ymm11
-  .byte  196,98,125,24,37,31,6,2,0           // vbroadcastss  0x2061f(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,207,28,2,0         // vbroadcastss  0x21ccf(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,100,88,220                  // vaddps        %ymm12,%ymm3,%ymm3
-  .byte  196,98,125,24,37,21,6,2,0           // vbroadcastss  0x20615(%rip),%ymm12        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,37,197,28,2,0         // vbroadcastss  0x21cc5(%rip),%ymm12        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  197,156,94,219                      // vdivps        %ymm3,%ymm12,%ymm3
   .byte  197,164,92,219                      // vsubps        %ymm3,%ymm11,%ymm3
   .byte  197,172,89,219                      // vmulps        %ymm3,%ymm10,%ymm3
   .byte  196,99,125,8,211,1                  // vroundps      $0x1,%ymm3,%ymm10
   .byte  196,65,100,92,210                   // vsubps        %ymm10,%ymm3,%ymm10
-  .byte  196,98,125,24,29,249,5,2,0          // vbroadcastss  0x205f9(%rip),%ymm11        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,29,169,28,2,0         // vbroadcastss  0x21ca9(%rip),%ymm11        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,100,88,219                  // vaddps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,239,5,2,0          // vbroadcastss  0x205ef(%rip),%ymm11        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,98,125,24,29,159,28,2,0         // vbroadcastss  0x21c9f(%rip),%ymm11        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  196,65,44,89,219                    // vmulps        %ymm11,%ymm10,%ymm11
   .byte  196,193,100,92,219                  // vsubps        %ymm11,%ymm3,%ymm3
-  .byte  196,98,125,24,29,224,5,2,0          // vbroadcastss  0x205e0(%rip),%ymm11        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,98,125,24,29,144,28,2,0         // vbroadcastss  0x21c90(%rip),%ymm11        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  196,65,36,92,210                    // vsubps        %ymm10,%ymm11,%ymm10
-  .byte  196,98,125,24,29,214,5,2,0          // vbroadcastss  0x205d6(%rip),%ymm11        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,98,125,24,29,134,28,2,0         // vbroadcastss  0x21c86(%rip),%ymm11        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  196,65,36,94,210                    // vdivps        %ymm10,%ymm11,%ymm10
   .byte  196,193,100,88,218                  // vaddps        %ymm10,%ymm3,%ymm3
-  .byte  196,98,125,24,21,199,5,2,0          // vbroadcastss  0x205c7(%rip),%ymm10        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,98,125,24,21,119,28,2,0         // vbroadcastss  0x21c77(%rip),%ymm10        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  196,193,100,89,218                  // vmulps        %ymm10,%ymm3,%ymm3
   .byte  197,253,91,219                      // vcvtps2dq     %ymm3,%ymm3
   .byte  196,98,125,24,80,20                 // vbroadcastss  0x14(%rax),%ymm10
@@ -27339,7 +27339,7 @@
   .byte  196,195,101,74,217,128              // vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  196,193,100,95,216                  // vmaxps        %ymm8,%ymm3,%ymm3
-  .byte  196,98,125,24,5,219,4,2,0           // vbroadcastss  0x204db(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,139,27,2,0          // vbroadcastss  0x21b8b(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -27358,34 +27358,34 @@
   .byte  197,252,40,233                      // vmovaps       %ymm1,%ymm5
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,91,200                      // vcvtdq2ps     %ymm0,%ymm1
-  .byte  196,98,125,24,5,50,5,2,0            // vbroadcastss  0x20532(%rip),%ymm8        # 32cc8 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  .byte  196,98,125,24,5,226,27,2,0          // vbroadcastss  0x21be2(%rip),%ymm8        # 34378 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,13,40,5,2,0           // vbroadcastss  0x20528(%rip),%ymm9        # 32ccc <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  .byte  196,98,125,24,13,216,27,2,0         // vbroadcastss  0x21bd8(%rip),%ymm9        # 3437c <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   .byte  196,193,124,84,193                  // vandps        %ymm9,%ymm0,%ymm0
-  .byte  196,226,125,24,37,122,4,2,0         // vbroadcastss  0x2047a(%rip),%ymm4        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,37,42,27,2,0         // vbroadcastss  0x21b2a(%rip),%ymm4        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,252,86,196                      // vorps         %ymm4,%ymm0,%ymm0
-  .byte  196,98,125,24,21,17,5,2,0           // vbroadcastss  0x20511(%rip),%ymm10        # 32cd0 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  .byte  196,98,125,24,21,193,27,2,0         // vbroadcastss  0x21bc1(%rip),%ymm10        # 34380 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   .byte  196,193,116,88,202                  // vaddps        %ymm10,%ymm1,%ymm1
-  .byte  196,98,125,24,29,7,5,2,0            // vbroadcastss  0x20507(%rip),%ymm11        # 32cd4 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  .byte  196,98,125,24,29,183,27,2,0         // vbroadcastss  0x21bb7(%rip),%ymm11        # 34384 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   .byte  196,193,124,89,211                  // vmulps        %ymm11,%ymm0,%ymm2
   .byte  197,244,92,202                      // vsubps        %ymm2,%ymm1,%ymm1
-  .byte  196,98,125,24,37,249,4,2,0          // vbroadcastss  0x204f9(%rip),%ymm12        # 32cd8 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  .byte  196,98,125,24,37,169,27,2,0         // vbroadcastss  0x21ba9(%rip),%ymm12        # 34388 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   .byte  196,193,124,88,196                  // vaddps        %ymm12,%ymm0,%ymm0
-  .byte  196,98,125,24,45,239,4,2,0          // vbroadcastss  0x204ef(%rip),%ymm13        # 32cdc <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  .byte  196,98,125,24,45,159,27,2,0         // vbroadcastss  0x21b9f(%rip),%ymm13        # 3438c <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   .byte  197,148,94,192                      // vdivps        %ymm0,%ymm13,%ymm0
   .byte  197,244,92,192                      // vsubps        %ymm0,%ymm1,%ymm0
   .byte  196,98,125,24,48                    // vbroadcastss  (%rax),%ymm14
   .byte  196,193,124,89,198                  // vmulps        %ymm14,%ymm0,%ymm0
   .byte  196,227,125,8,200,1                 // vroundps      $0x1,%ymm0,%ymm1
   .byte  197,252,92,241                      // vsubps        %ymm1,%ymm0,%ymm6
-  .byte  196,98,125,24,61,206,4,2,0          // vbroadcastss  0x204ce(%rip),%ymm15        # 32ce0 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  .byte  196,98,125,24,61,126,27,2,0         // vbroadcastss  0x21b7e(%rip),%ymm15        # 34390 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   .byte  196,193,124,88,199                  // vaddps        %ymm15,%ymm0,%ymm0
-  .byte  196,226,125,24,29,196,4,2,0         // vbroadcastss  0x204c4(%rip),%ymm3        # 32ce4 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  .byte  196,226,125,24,29,116,27,2,0        // vbroadcastss  0x21b74(%rip),%ymm3        # 34394 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   .byte  197,204,89,203                      // vmulps        %ymm3,%ymm6,%ymm1
   .byte  197,252,92,201                      // vsubps        %ymm1,%ymm0,%ymm1
-  .byte  196,226,125,24,21,183,4,2,0         // vbroadcastss  0x204b7(%rip),%ymm2        # 32ce8 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  .byte  196,226,125,24,21,103,27,2,0        // vbroadcastss  0x21b67(%rip),%ymm2        # 34398 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   .byte  197,236,92,198                      // vsubps        %ymm6,%ymm2,%ymm0
-  .byte  196,226,125,24,53,174,4,2,0         // vbroadcastss  0x204ae(%rip),%ymm6        # 32cec <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  .byte  196,226,125,24,53,94,27,2,0         // vbroadcastss  0x21b5e(%rip),%ymm6        # 3439c <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   .byte  197,204,94,192                      // vdivps        %ymm0,%ymm6,%ymm0
   .byte  197,244,88,192                      // vaddps        %ymm0,%ymm1,%ymm0
   .byte  197,252,17,68,36,128                // vmovups       %ymm0,-0x80(%rsp)
@@ -27427,7 +27427,7 @@
   .byte  197,236,92,212                      // vsubps        %ymm4,%ymm2,%ymm2
   .byte  197,204,94,210                      // vdivps        %ymm2,%ymm6,%ymm2
   .byte  197,244,88,202                      // vaddps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,241,3,2,0         // vbroadcastss  0x203f1(%rip),%ymm2        # 32cf0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  .byte  196,226,125,24,21,161,26,2,0        // vbroadcastss  0x21aa1(%rip),%ymm2        # 343a0 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   .byte  197,236,89,92,36,128                // vmulps        -0x80(%rsp),%ymm2,%ymm3
   .byte  197,252,89,226                      // vmulps        %ymm2,%ymm0,%ymm4
   .byte  197,244,89,210                      // vmulps        %ymm2,%ymm1,%ymm2
@@ -27447,31 +27447,31 @@
 .globl _sk_lab_to_xyz_avx
 FUNCTION(_sk_lab_to_xyz_avx)
 _sk_lab_to_xyz_avx:
-  .byte  196,98,125,24,5,173,3,2,0           // vbroadcastss  0x203ad(%rip),%ymm8        # 32cf4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
+  .byte  196,98,125,24,5,93,26,2,0           // vbroadcastss  0x21a5d(%rip),%ymm8        # 343a4 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,7,3,2,0             // vbroadcastss  0x20307(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,183,25,2,0          // vbroadcastss  0x219b7(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
-  .byte  196,98,125,24,13,149,3,2,0          // vbroadcastss  0x20395(%rip),%ymm9        # 32cf8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
+  .byte  196,98,125,24,13,69,26,2,0          // vbroadcastss  0x21a45(%rip),%ymm9        # 343a8 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
   .byte  196,193,116,88,201                  // vaddps        %ymm9,%ymm1,%ymm1
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  196,193,108,88,209                  // vaddps        %ymm9,%ymm2,%ymm2
-  .byte  196,98,125,24,5,129,3,2,0           // vbroadcastss  0x20381(%rip),%ymm8        # 32cfc <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
+  .byte  196,98,125,24,5,49,26,2,0           // vbroadcastss  0x21a31(%rip),%ymm8        # 343ac <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,119,3,2,0           // vbroadcastss  0x20377(%rip),%ymm8        # 32d00 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
+  .byte  196,98,125,24,5,39,26,2,0           // vbroadcastss  0x21a27(%rip),%ymm8        # 343b0 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,5,109,3,2,0           // vbroadcastss  0x2036d(%rip),%ymm8        # 32d04 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
+  .byte  196,98,125,24,5,29,26,2,0           // vbroadcastss  0x21a1d(%rip),%ymm8        # 343b4 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
   .byte  196,193,116,89,200                  // vmulps        %ymm8,%ymm1,%ymm1
   .byte  197,252,88,201                      // vaddps        %ymm1,%ymm0,%ymm1
-  .byte  196,98,125,24,5,95,3,2,0            // vbroadcastss  0x2035f(%rip),%ymm8        # 32d08 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
+  .byte  196,98,125,24,5,15,26,2,0           // vbroadcastss  0x21a0f(%rip),%ymm8        # 343b8 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  197,252,92,210                      // vsubps        %ymm2,%ymm0,%ymm2
   .byte  197,116,89,193                      // vmulps        %ymm1,%ymm1,%ymm8
   .byte  196,65,116,89,192                   // vmulps        %ymm8,%ymm1,%ymm8
-  .byte  196,98,125,24,13,72,3,2,0           // vbroadcastss  0x20348(%rip),%ymm9        # 32d0c <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
+  .byte  196,98,125,24,13,248,25,2,0         // vbroadcastss  0x219f8(%rip),%ymm9        # 343bc <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
   .byte  196,65,52,194,208,1                 // vcmpltps      %ymm8,%ymm9,%ymm10
-  .byte  196,98,125,24,29,61,3,2,0           // vbroadcastss  0x2033d(%rip),%ymm11        # 32d10 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
+  .byte  196,98,125,24,29,237,25,2,0         // vbroadcastss  0x219ed(%rip),%ymm11        # 343c0 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
   .byte  196,193,116,88,203                  // vaddps        %ymm11,%ymm1,%ymm1
-  .byte  196,98,125,24,37,51,3,2,0           // vbroadcastss  0x20333(%rip),%ymm12        # 32d14 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
+  .byte  196,98,125,24,37,227,25,2,0         // vbroadcastss  0x219e3(%rip),%ymm12        # 343c4 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
   .byte  196,193,116,89,204                  // vmulps        %ymm12,%ymm1,%ymm1
   .byte  196,67,117,74,192,160               // vblendvps     %ymm10,%ymm8,%ymm1,%ymm8
   .byte  197,252,89,200                      // vmulps        %ymm0,%ymm0,%ymm1
@@ -27486,9 +27486,9 @@
   .byte  196,193,108,88,211                  // vaddps        %ymm11,%ymm2,%ymm2
   .byte  196,193,108,89,212                  // vmulps        %ymm12,%ymm2,%ymm2
   .byte  196,227,109,74,208,144              // vblendvps     %ymm9,%ymm0,%ymm2,%ymm2
-  .byte  196,226,125,24,5,233,2,2,0          // vbroadcastss  0x202e9(%rip),%ymm0        # 32d18 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
+  .byte  196,226,125,24,5,153,25,2,0         // vbroadcastss  0x21999(%rip),%ymm0        # 343c8 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,224,2,2,0           // vbroadcastss  0x202e0(%rip),%ymm8        # 32d1c <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
+  .byte  196,98,125,24,5,144,25,2,0          // vbroadcastss  0x21990(%rip),%ymm8        # 343cc <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -27506,13 +27506,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,68                              // jne           12aa1 <_sk_load_a8_avx+0x5c>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,5,9,2,0               // vpand         0x20905(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,229,31,2,0            // vpand         0x21fe5(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,121,51,200                  // vpmovzxwd     %xmm0,%xmm1
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,27,2,2,0          // vbroadcastss  0x2021b(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,203,24,2,0        // vbroadcastss  0x218cb(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,217                      // vmulps        %ymm1,%ymm0,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
@@ -27560,7 +27560,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,222                 // jmpq          ffffffffdf012b48 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefe01c8>
+  .byte  233,255,255,255,222                 // jmpq          ffffffffdf012b48 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefdeb18>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,211                             // callq         *%rbx
@@ -27584,13 +27584,13 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,68                              // jne           12bb0 <_sk_load_a8_dst_avx+0x5c>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,246,7,2,0            // vpand         0x207f6(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,214,30,2,0           // vpand         0x21ed6(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,121,51,236                  // vpmovzxwd     %xmm4,%xmm5
   .byte  197,249,112,228,78                  // vpshufd       $0x4e,%xmm4,%xmm4
   .byte  196,226,121,51,228                  // vpmovzxwd     %xmm4,%xmm4
   .byte  196,227,85,24,228,1                 // vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,12,1,2,0          // vbroadcastss  0x2010c(%rip),%ymm5        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,45,188,23,2,0        // vbroadcastss  0x217bc(%rip),%ymm5        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,220,89,253                      // vmulps        %ymm5,%ymm4,%ymm7
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,220,87,228                      // vxorps        %ymm4,%ymm4,%ymm4
@@ -27638,7 +27638,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,221                 // callq         ffffffffde012c58 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfe02d8>
+  .byte  232,255,255,255,221                 // callq         ffffffffde012c58 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfdec28>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,210                             // callq         *%rdx
@@ -27718,7 +27718,7 @@
   .byte  196,226,121,49,201                  // vpmovzxbd     %xmm1,%xmm1
   .byte  196,227,125,24,193,1                // vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,18,255,1,0        // vbroadcastss  0x1ff12(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,194,21,2,0        // vbroadcastss  0x215c2(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,217                      // vmulps        %ymm1,%ymm0,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
@@ -27741,7 +27741,7 @@
   .byte  77,15,175,193                       // imul          %r9,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,142,254,1,0         // vbroadcastss  0x1fe8e(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,62,21,2,0           // vbroadcastss  0x2153e(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,100,89,192                   // vmulps        %ymm8,%ymm3,%ymm8
   .byte  196,65,125,91,192                   // vcvtps2dq     %ymm8,%ymm8
   .byte  196,67,125,25,193,1                 // vextractf128  $0x1,%ymm8,%xmm9
@@ -27766,13 +27766,13 @@
   .byte  196,3,121,20,4,24,0                 // vpextrb       $0x0,%xmm8,(%r8,%r11,1)
   .byte  235,202                             // jmp           12df2 <_sk_store_a8_avx+0x40>
   .byte  196,3,121,20,68,24,2,4              // vpextrb       $0x4,%xmm8,0x2(%r8,%r11,1)
-  .byte  196,98,57,0,5,71,5,2,0              // vpshufb       0x20547(%rip),%xmm8,%xmm8        # 33380 <_sk_srcover_rgba_8888_sse2_lowp+0xa00>
+  .byte  196,98,57,0,5,39,28,2,0             // vpshufb       0x21c27(%rip),%xmm8,%xmm8        # 34a60 <_sk_srcover_rgba_8888_sse2_lowp+0xa30>
   .byte  196,3,121,21,4,24,0                 // vpextrw       $0x0,%xmm8,(%r8,%r11,1)
   .byte  235,176                             // jmp           12df2 <_sk_store_a8_avx+0x40>
   .byte  196,3,121,20,68,24,6,12             // vpextrb       $0xc,%xmm8,0x6(%r8,%r11,1)
   .byte  196,3,121,20,68,24,5,10             // vpextrb       $0xa,%xmm8,0x5(%r8,%r11,1)
   .byte  196,3,121,20,68,24,4,8              // vpextrb       $0x8,%xmm8,0x4(%r8,%r11,1)
-  .byte  196,98,57,0,5,45,5,2,0              // vpshufb       0x2052d(%rip),%xmm8,%xmm8        # 33390 <_sk_srcover_rgba_8888_sse2_lowp+0xa10>
+  .byte  196,98,57,0,5,13,28,2,0             // vpshufb       0x21c0d(%rip),%xmm8,%xmm8        # 34a70 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
   .byte  196,1,121,126,4,24                  // vmovd         %xmm8,(%r8,%r11,1)
   .byte  235,135                             // jmp           12df2 <_sk_store_a8_avx+0x40>
   .byte  144                                 // nop
@@ -27809,16 +27809,16 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,73                              // jne           12ee9 <_sk_load_g8_avx+0x61>
   .byte  196,130,121,48,4,24                 // vpmovzxbw     (%r8,%r11,1),%xmm0
-  .byte  197,249,219,5,194,4,2,0             // vpand         0x204c2(%rip),%xmm0,%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,249,219,5,162,27,2,0            // vpand         0x21ba2(%rip),%xmm0,%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,121,51,200                  // vpmovzxwd     %xmm0,%xmm1
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,216,253,1,0       // vbroadcastss  0x1fdd8(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,136,20,2,0        // vbroadcastss  0x21488(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,81,253,1,0        // vbroadcastss  0x1fd51(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,1,20,2,0          // vbroadcastss  0x21401(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  255,224                             // jmpq          *%rax
@@ -27863,7 +27863,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,222                 // jmpq          ffffffffdf012f90 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefe0610>
+  .byte  233,255,255,255,222                 // jmpq          ffffffffdf012f90 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefdef60>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,211                             // callq         *%rbx
@@ -27887,16 +27887,16 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,73                              // jne           12ffd <_sk_load_g8_dst_avx+0x61>
   .byte  196,130,121,48,36,24                // vpmovzxbw     (%r8,%r11,1),%xmm4
-  .byte  197,217,219,37,174,3,2,0            // vpand         0x203ae(%rip),%xmm4,%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  197,217,219,37,142,26,2,0           // vpand         0x21a8e(%rip),%xmm4,%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  196,226,121,51,236                  // vpmovzxwd     %xmm4,%xmm5
   .byte  197,249,112,228,78                  // vpshufd       $0x4e,%xmm4,%xmm4
   .byte  196,226,121,51,228                  // vpmovzxwd     %xmm4,%xmm4
   .byte  196,227,85,24,228,1                 // vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,196,252,1,0       // vbroadcastss  0x1fcc4(%rip),%ymm5        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,45,116,19,2,0        // vbroadcastss  0x21374(%rip),%ymm5        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,61,252,1,0        // vbroadcastss  0x1fc3d(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,237,18,2,0        // vbroadcastss  0x212ed(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,236                      // vmovaps       %ymm4,%ymm5
   .byte  197,252,40,244                      // vmovaps       %ymm4,%ymm6
   .byte  255,224                             // jmpq          *%rax
@@ -27941,7 +27941,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  233,255,255,255,222                 // jmpq          ffffffffdf0130a4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefe0724>
+  .byte  233,255,255,255,222                 // jmpq          ffffffffdf0130a4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefdf074>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,211                             // callq         *%rbx
@@ -28021,10 +28021,10 @@
   .byte  196,226,121,49,201                  // vpmovzxbd     %xmm1,%xmm1
   .byte  196,227,125,24,193,1                // vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,198,250,1,0       // vbroadcastss  0x1fac6(%rip),%ymm1        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,226,125,24,13,118,17,2,0        // vbroadcastss  0x21176(%rip),%ymm1        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,63,250,1,0        // vbroadcastss  0x1fa3f(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,239,16,2,0        // vbroadcastss  0x210ef(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
   .byte  197,252,40,208                      // vmovaps       %ymm0,%ymm2
   .byte  91                                  // pop           %rbx
@@ -28052,23 +28052,23 @@
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,208,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm2
-  .byte  196,226,125,24,5,102,250,1,0        // vbroadcastss  0x1fa66(%rip),%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,24,5,22,17,2,0          // vbroadcastss  0x21116(%rip),%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,89,250,1,0        // vbroadcastss  0x1fa59(%rip),%ymm1        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,13,9,17,2,0          // vbroadcastss  0x21109(%rip),%ymm1        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,80,250,1,0        // vbroadcastss  0x1fa50(%rip),%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,24,13,0,17,2,0          // vbroadcastss  0x21100(%rip),%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,67,250,1,0        // vbroadcastss  0x1fa43(%rip),%ymm3        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,29,243,16,2,0        // vbroadcastss  0x210f3(%rip),%ymm3        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,58,250,1,0        // vbroadcastss  0x1fa3a(%rip),%ymm3        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,24,29,234,16,2,0        // vbroadcastss  0x210ea(%rip),%ymm3        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,45,250,1,0        // vbroadcastss  0x1fa2d(%rip),%ymm3        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,29,221,16,2,0        // vbroadcastss  0x210dd(%rip),%ymm3        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,142,249,1,0       // vbroadcastss  0x1f98e(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,62,16,2,0         // vbroadcastss  0x2103e(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -28133,23 +28133,23 @@
   .byte  197,249,112,228,78                  // vpshufd       $0x4e,%xmm4,%xmm4
   .byte  196,226,121,51,228                  // vpmovzxwd     %xmm4,%xmm4
   .byte  196,227,85,24,244,1                 // vinsertf128   $0x1,%xmm4,%ymm5,%ymm6
-  .byte  196,226,125,24,37,37,249,1,0        // vbroadcastss  0x1f925(%rip),%ymm4        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,24,37,213,15,2,0        // vbroadcastss  0x20fd5(%rip),%ymm4        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,204,84,228                      // vandps        %ymm4,%ymm6,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,24,249,1,0        // vbroadcastss  0x1f918(%rip),%ymm5        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,45,200,15,2,0        // vbroadcastss  0x20fc8(%rip),%ymm5        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
-  .byte  196,226,125,24,45,15,249,1,0        // vbroadcastss  0x1f90f(%rip),%ymm5        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,24,45,191,15,2,0        // vbroadcastss  0x20fbf(%rip),%ymm5        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,204,84,237                      // vandps        %ymm5,%ymm6,%ymm5
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,226,125,24,61,2,249,1,0         // vbroadcastss  0x1f902(%rip),%ymm7        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,61,178,15,2,0        // vbroadcastss  0x20fb2(%rip),%ymm7        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,212,89,239                      // vmulps        %ymm7,%ymm5,%ymm5
-  .byte  196,226,125,24,61,249,248,1,0       // vbroadcastss  0x1f8f9(%rip),%ymm7        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,24,61,169,15,2,0        // vbroadcastss  0x20fa9(%rip),%ymm7        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,204,84,247                      // vandps        %ymm7,%ymm6,%ymm6
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  196,226,125,24,61,236,248,1,0       // vbroadcastss  0x1f8ec(%rip),%ymm7        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,61,156,15,2,0        // vbroadcastss  0x20f9c(%rip),%ymm7        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,204,89,247                      // vmulps        %ymm7,%ymm6,%ymm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,61,77,248,1,0        // vbroadcastss  0x1f84d(%rip),%ymm7        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,61,253,14,2,0        // vbroadcastss  0x20efd(%rip),%ymm7        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,7                        // and           $0x7,%r9b
@@ -28269,23 +28269,23 @@
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,208,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm2
-  .byte  196,226,125,24,5,245,246,1,0        // vbroadcastss  0x1f6f5(%rip),%ymm0        # 32cac <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  .byte  196,226,125,24,5,165,13,2,0         // vbroadcastss  0x20da5(%rip),%ymm0        # 3435c <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   .byte  197,236,84,192                      // vandps        %ymm0,%ymm2,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,232,246,1,0       // vbroadcastss  0x1f6e8(%rip),%ymm1        # 32cb0 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  .byte  196,226,125,24,13,152,13,2,0        // vbroadcastss  0x20d98(%rip),%ymm1        # 34360 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,223,246,1,0       // vbroadcastss  0x1f6df(%rip),%ymm1        # 32cb4 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  .byte  196,226,125,24,13,143,13,2,0        // vbroadcastss  0x20d8f(%rip),%ymm1        # 34364 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   .byte  197,236,84,201                      // vandps        %ymm1,%ymm2,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,29,210,246,1,0       // vbroadcastss  0x1f6d2(%rip),%ymm3        # 32cb8 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  .byte  196,226,125,24,29,130,13,2,0        // vbroadcastss  0x20d82(%rip),%ymm3        # 34368 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
-  .byte  196,226,125,24,29,201,246,1,0       // vbroadcastss  0x1f6c9(%rip),%ymm3        # 32cbc <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  .byte  196,226,125,24,29,121,13,2,0        // vbroadcastss  0x20d79(%rip),%ymm3        # 3436c <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   .byte  197,236,84,211                      // vandps        %ymm3,%ymm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,188,246,1,0       // vbroadcastss  0x1f6bc(%rip),%ymm3        # 32cc0 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  .byte  196,226,125,24,29,108,13,2,0        // vbroadcastss  0x20d6c(%rip),%ymm3        # 34370 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   .byte  197,236,89,211                      // vmulps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,29,246,1,0        // vbroadcastss  0x1f61d(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,205,12,2,0        // vbroadcastss  0x20ccd(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
   .byte  65,94                               // pop           %r14
@@ -28304,14 +28304,14 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,228,246,1,0         // vbroadcastss  0x1f6e4(%rip),%ymm8        # 32d20 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
+  .byte  196,98,125,24,5,148,13,2,0          // vbroadcastss  0x20d94(%rip),%ymm8        # 343d0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,41,114,241,11               // vpslld        $0xb,%xmm9,%xmm10
   .byte  196,67,125,25,201,1                 // vextractf128  $0x1,%ymm9,%xmm9
   .byte  196,193,49,114,241,11               // vpslld        $0xb,%xmm9,%xmm9
   .byte  196,67,45,24,201,1                  // vinsertf128   $0x1,%xmm9,%ymm10,%ymm9
-  .byte  196,98,125,24,21,189,246,1,0        // vbroadcastss  0x1f6bd(%rip),%ymm10        # 32d24 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
+  .byte  196,98,125,24,21,109,13,2,0         // vbroadcastss  0x20d6d(%rip),%ymm10        # 343d4 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
   .byte  196,65,116,89,210                   // vmulps        %ymm10,%ymm1,%ymm10
   .byte  196,65,125,91,210                   // vcvtps2dq     %ymm10,%ymm10
   .byte  196,193,33,114,242,5                // vpslld        $0x5,%xmm10,%xmm11
@@ -28366,7 +28366,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,224                 // callq         ffffffffe101372c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fe0dac>
+  .byte  232,255,255,255,224                 // callq         ffffffffe101372c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fdf6fc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -28389,25 +28389,25 @@
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,216,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm3
-  .byte  196,226,125,24,5,181,245,1,0        // vbroadcastss  0x1f5b5(%rip),%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,24,5,101,12,2,0         // vbroadcastss  0x20c65(%rip),%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,168,245,1,0       // vbroadcastss  0x1f5a8(%rip),%ymm1        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,13,88,12,2,0         // vbroadcastss  0x20c58(%rip),%ymm1        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,159,245,1,0       // vbroadcastss  0x1f59f(%rip),%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,24,13,79,12,2,0         // vbroadcastss  0x20c4f(%rip),%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,146,245,1,0       // vbroadcastss  0x1f592(%rip),%ymm2        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,21,66,12,2,0         // vbroadcastss  0x20c42(%rip),%ymm2        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,137,245,1,0       // vbroadcastss  0x1f589(%rip),%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,24,21,57,12,2,0         // vbroadcastss  0x20c39(%rip),%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,124,245,1,0         // vbroadcastss  0x1f57c(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,44,12,2,0           // vbroadcastss  0x20c2c(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,114,245,1,0         // vbroadcastss  0x1f572(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,24,5,34,12,2,0           // vbroadcastss  0x20c22(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,100,245,1,0         // vbroadcastss  0x1f564(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,20,12,2,0           // vbroadcastss  0x20c14(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -28478,25 +28478,25 @@
   .byte  197,249,112,228,78                  // vpshufd       $0x4e,%xmm4,%xmm4
   .byte  196,226,121,51,228                  // vpmovzxwd     %xmm4,%xmm4
   .byte  196,227,85,24,252,1                 // vinsertf128   $0x1,%xmm4,%ymm5,%ymm7
-  .byte  196,226,125,24,37,89,244,1,0        // vbroadcastss  0x1f459(%rip),%ymm4        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,24,37,9,11,2,0          // vbroadcastss  0x20b09(%rip),%ymm4        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,196,84,228                      // vandps        %ymm4,%ymm7,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,226,125,24,45,76,244,1,0        // vbroadcastss  0x1f44c(%rip),%ymm5        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,45,252,10,2,0        // vbroadcastss  0x20afc(%rip),%ymm5        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,220,89,229                      // vmulps        %ymm5,%ymm4,%ymm4
-  .byte  196,226,125,24,45,67,244,1,0        // vbroadcastss  0x1f443(%rip),%ymm5        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,24,45,243,10,2,0        // vbroadcastss  0x20af3(%rip),%ymm5        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,196,84,237                      // vandps        %ymm5,%ymm7,%ymm5
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,226,125,24,53,54,244,1,0        // vbroadcastss  0x1f436(%rip),%ymm6        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,53,230,10,2,0        // vbroadcastss  0x20ae6(%rip),%ymm6        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,212,89,238                      // vmulps        %ymm6,%ymm5,%ymm5
-  .byte  196,226,125,24,53,45,244,1,0        // vbroadcastss  0x1f42d(%rip),%ymm6        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,24,53,221,10,2,0        // vbroadcastss  0x20add(%rip),%ymm6        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,196,84,246                      // vandps        %ymm6,%ymm7,%ymm6
   .byte  197,252,91,246                      // vcvtdq2ps     %ymm6,%ymm6
-  .byte  196,98,125,24,5,32,244,1,0          // vbroadcastss  0x1f420(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,208,10,2,0          // vbroadcastss  0x20ad0(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,76,89,240                   // vmulps        %ymm8,%ymm6,%ymm6
-  .byte  196,98,125,24,5,22,244,1,0          // vbroadcastss  0x1f416(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,24,5,198,10,2,0          // vbroadcastss  0x20ac6(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,68,84,248                   // vandps        %ymm8,%ymm7,%ymm7
   .byte  197,252,91,255                      // vcvtdq2ps     %ymm7,%ymm7
-  .byte  196,98,125,24,5,8,244,1,0           // vbroadcastss  0x1f408(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,184,10,2,0          // vbroadcastss  0x20ab8(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,68,89,248                   // vmulps        %ymm8,%ymm7,%ymm7
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -28618,25 +28618,25 @@
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,216,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm3
-  .byte  196,226,125,24,5,17,242,1,0         // vbroadcastss  0x1f211(%rip),%ymm0        # 32d28 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  .byte  196,226,125,24,5,193,8,2,0          // vbroadcastss  0x208c1(%rip),%ymm0        # 343d8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   .byte  197,228,84,192                      // vandps        %ymm0,%ymm3,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,226,125,24,13,4,242,1,0         // vbroadcastss  0x1f204(%rip),%ymm1        # 32d2c <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  .byte  196,226,125,24,13,180,8,2,0         // vbroadcastss  0x208b4(%rip),%ymm1        # 343dc <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   .byte  197,252,89,193                      // vmulps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,251,241,1,0       // vbroadcastss  0x1f1fb(%rip),%ymm1        # 32d30 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  .byte  196,226,125,24,13,171,8,2,0         // vbroadcastss  0x208ab(%rip),%ymm1        # 343e0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   .byte  197,228,84,201                      // vandps        %ymm1,%ymm3,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,226,125,24,21,238,241,1,0       // vbroadcastss  0x1f1ee(%rip),%ymm2        # 32d34 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  .byte  196,226,125,24,21,158,8,2,0         // vbroadcastss  0x2089e(%rip),%ymm2        # 343e4 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   .byte  197,244,89,202                      // vmulps        %ymm2,%ymm1,%ymm1
-  .byte  196,226,125,24,21,229,241,1,0       // vbroadcastss  0x1f1e5(%rip),%ymm2        # 32d38 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  .byte  196,226,125,24,21,149,8,2,0         // vbroadcastss  0x20895(%rip),%ymm2        # 343e8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   .byte  197,228,84,210                      // vandps        %ymm2,%ymm3,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,98,125,24,5,216,241,1,0         // vbroadcastss  0x1f1d8(%rip),%ymm8        # 32d3c <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  .byte  196,98,125,24,5,136,8,2,0           // vbroadcastss  0x20888(%rip),%ymm8        # 343ec <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   .byte  196,193,108,89,208                  // vmulps        %ymm8,%ymm2,%ymm2
-  .byte  196,98,125,24,5,206,241,1,0         // vbroadcastss  0x1f1ce(%rip),%ymm8        # 32d40 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  .byte  196,98,125,24,5,126,8,2,0           // vbroadcastss  0x2087e(%rip),%ymm8        # 343f0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   .byte  196,193,100,84,216                  // vandps        %ymm8,%ymm3,%ymm3
   .byte  197,252,91,219                      // vcvtdq2ps     %ymm3,%ymm3
-  .byte  196,98,125,24,5,192,241,1,0         // vbroadcastss  0x1f1c0(%rip),%ymm8        # 32d44 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  .byte  196,98,125,24,5,112,8,2,0           // vbroadcastss  0x20870(%rip),%ymm8        # 343f4 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   .byte  196,193,100,89,216                  // vmulps        %ymm8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
@@ -28657,7 +28657,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,148,241,1,0         // vbroadcastss  0x1f194(%rip),%ymm8        # 32d48 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
+  .byte  196,98,125,24,5,68,8,2,0            // vbroadcastss  0x20844(%rip),%ymm8        # 343f8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,193,41,114,241,12               // vpslld        $0xc,%xmm9,%xmm10
@@ -28745,10 +28745,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,135,0,0,0                    // jne           13d6f <_sk_load_8888_avx+0xa7>
   .byte  196,1,125,16,12,152                 // vmovupd       (%r8,%r11,4),%ymm9
-  .byte  197,125,40,21,106,244,1,0           // vmovapd       0x1f46a(%rip),%ymm10        # 33160 <_sk_srcover_rgba_8888_sse2_lowp+0x7e0>
+  .byte  197,125,40,21,10,11,2,0             // vmovapd       0x20b0a(%rip),%ymm10        # 34800 <_sk_srcover_rgba_8888_sse2_lowp+0x7d0>
   .byte  196,193,53,84,194                   // vandpd        %ymm10,%ymm9,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,5,160,239,1,0         // vbroadcastss  0x1efa0(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,80,6,2,0            // vbroadcastss  0x20650(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,192                  // vmulps        %ymm8,%ymm0,%ymm0
   .byte  196,193,113,114,209,8               // vpsrld        $0x8,%xmm9,%xmm1
   .byte  196,99,125,25,203,1                 // vextractf128  $0x1,%ymm9,%xmm3
@@ -28840,10 +28840,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,135,0,0,0                    // jne           13ee7 <_sk_load_8888_dst_avx+0xa7>
   .byte  196,1,125,16,12,152                 // vmovupd       (%r8,%r11,4),%ymm9
-  .byte  197,125,40,21,18,243,1,0            // vmovapd       0x1f312(%rip),%ymm10        # 33180 <_sk_srcover_rgba_8888_sse2_lowp+0x800>
+  .byte  197,125,40,21,178,9,2,0             // vmovapd       0x209b2(%rip),%ymm10        # 34820 <_sk_srcover_rgba_8888_sse2_lowp+0x7f0>
   .byte  196,193,53,84,226                   // vandpd        %ymm10,%ymm9,%ymm4
   .byte  197,252,91,228                      // vcvtdq2ps     %ymm4,%ymm4
-  .byte  196,98,125,24,5,40,238,1,0          // vbroadcastss  0x1ee28(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,216,4,2,0           // vbroadcastss  0x204d8(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,92,89,224                   // vmulps        %ymm8,%ymm4,%ymm4
   .byte  196,193,81,114,209,8                // vpsrld        $0x8,%xmm9,%xmm5
   .byte  196,99,125,25,207,1                 // vextractf128  $0x1,%ymm9,%xmm7
@@ -28978,10 +28978,10 @@
   .byte  196,131,121,34,4,176,2              // vpinsrd       $0x2,(%r8,%r14,4),%xmm0,%xmm0
   .byte  196,195,121,34,28,152,3             // vpinsrd       $0x3,(%r8,%rbx,4),%xmm0,%xmm3
   .byte  196,227,61,24,195,1                 // vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  .byte  197,124,40,21,228,240,1,0           // vmovaps       0x1f0e4(%rip),%ymm10        # 331a0 <_sk_srcover_rgba_8888_sse2_lowp+0x820>
+  .byte  197,124,40,21,132,7,2,0             // vmovaps       0x20784(%rip),%ymm10        # 34840 <_sk_srcover_rgba_8888_sse2_lowp+0x810>
   .byte  196,193,124,84,194                  // vandps        %ymm10,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,218,235,1,0        // vbroadcastss  0x1ebda(%rip),%ymm9        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,13,138,2,2,0          // vbroadcastss  0x2028a(%rip),%ymm9        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  196,193,113,114,208,8               // vpsrld        $0x8,%xmm8,%xmm1
   .byte  197,233,114,211,8                   // vpsrld        $0x8,%xmm3,%xmm2
@@ -29018,7 +29018,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,6,235,1,0           // vbroadcastss  0x1eb06(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,182,1,2,0           // vbroadcastss  0x201b6(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -29081,7 +29081,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,218                 // callq         ffffffffdb014268 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe18e8>
+  .byte  232,255,255,255,218                 // callq         ffffffffdb014268 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe0238>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,204                             // dec           %esp
@@ -29103,10 +29103,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,135,0,0,0                    // jne           14317 <_sk_load_bgra_avx+0xa7>
   .byte  196,1,125,16,12,152                 // vmovupd       (%r8,%r11,4),%ymm9
-  .byte  197,125,40,21,34,239,1,0            // vmovapd       0x1ef22(%rip),%ymm10        # 331c0 <_sk_srcover_rgba_8888_sse2_lowp+0x840>
+  .byte  197,125,40,21,194,5,2,0             // vmovapd       0x205c2(%rip),%ymm10        # 34860 <_sk_srcover_rgba_8888_sse2_lowp+0x830>
   .byte  196,193,53,84,202                   // vandpd        %ymm10,%ymm9,%ymm1
   .byte  197,252,91,201                      // vcvtdq2ps     %ymm1,%ymm1
-  .byte  196,98,125,24,5,248,233,1,0         // vbroadcastss  0x1e9f8(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,168,0,2,0           // vbroadcastss  0x200a8(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,116,89,208                  // vmulps        %ymm8,%ymm1,%ymm2
   .byte  196,193,113,114,209,8               // vpsrld        $0x8,%xmm9,%xmm1
   .byte  196,99,125,25,203,1                 // vextractf128  $0x1,%ymm9,%xmm3
@@ -29198,10 +29198,10 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,135,0,0,0                    // jne           1448f <_sk_load_bgra_dst_avx+0xa7>
   .byte  196,1,125,16,12,152                 // vmovupd       (%r8,%r11,4),%ymm9
-  .byte  197,125,40,21,202,237,1,0           // vmovapd       0x1edca(%rip),%ymm10        # 331e0 <_sk_srcover_rgba_8888_sse2_lowp+0x860>
+  .byte  197,125,40,21,106,4,2,0             // vmovapd       0x2046a(%rip),%ymm10        # 34880 <_sk_srcover_rgba_8888_sse2_lowp+0x850>
   .byte  196,193,53,84,234                   // vandpd        %ymm10,%ymm9,%ymm5
   .byte  197,252,91,237                      // vcvtdq2ps     %ymm5,%ymm5
-  .byte  196,98,125,24,5,128,232,1,0         // vbroadcastss  0x1e880(%rip),%ymm8        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,5,48,255,1,0          // vbroadcastss  0x1ff30(%rip),%ymm8        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,84,89,240                   // vmulps        %ymm8,%ymm5,%ymm6
   .byte  196,193,81,114,209,8                // vpsrld        $0x8,%xmm9,%xmm5
   .byte  196,99,125,25,207,1                 // vextractf128  $0x1,%ymm9,%xmm7
@@ -29336,10 +29336,10 @@
   .byte  196,131,121,34,4,176,2              // vpinsrd       $0x2,(%r8,%r14,4),%xmm0,%xmm0
   .byte  196,195,121,34,28,152,3             // vpinsrd       $0x3,(%r8,%rbx,4),%xmm0,%xmm3
   .byte  196,227,61,24,195,1                 // vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  .byte  197,124,40,13,156,235,1,0           // vmovaps       0x1eb9c(%rip),%ymm9        # 33200 <_sk_srcover_rgba_8888_sse2_lowp+0x880>
+  .byte  197,124,40,13,60,2,2,0              // vmovaps       0x2023c(%rip),%ymm9        # 348a0 <_sk_srcover_rgba_8888_sse2_lowp+0x870>
   .byte  196,193,124,84,193                  // vandps        %ymm9,%ymm0,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,21,50,230,1,0         // vbroadcastss  0x1e632(%rip),%ymm10        # 32ca8 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  .byte  196,98,125,24,21,226,252,1,0        // vbroadcastss  0x1fce2(%rip),%ymm10        # 34358 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   .byte  196,193,124,89,210                  // vmulps        %ymm10,%ymm0,%ymm2
   .byte  196,193,121,114,208,8               // vpsrld        $0x8,%xmm8,%xmm0
   .byte  197,241,114,211,8                   // vpsrld        $0x8,%xmm3,%xmm1
@@ -29376,7 +29376,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,218                           // movslq        %edx,%r11
-  .byte  196,98,125,24,5,94,229,1,0          // vbroadcastss  0x1e55e(%rip),%ymm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,125,24,5,14,252,1,0          // vbroadcastss  0x1fc0e(%rip),%ymm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  196,65,108,89,200                   // vmulps        %ymm8,%ymm2,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,65,116,89,208                   // vmulps        %ymm8,%ymm1,%ymm10
@@ -29439,7 +29439,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,255,255,255,218                 // callq         ffffffffdb014810 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe1e90>
+  .byte  232,255,255,255,218                 // callq         ffffffffdb014810 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe07e0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,204                             // dec           %esp
@@ -29481,14 +29481,14 @@
   .byte  196,226,121,51,200                  // vpmovzxwd     %xmm0,%xmm1
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
-  .byte  197,123,18,21,27,243,1,0            // vmovddup      0x1f31b(%rip),%xmm10        # 33bc8 <_sk_srcover_rgba_8888_sse2_lowp+0x1248>
+  .byte  197,123,18,21,251,9,2,0             // vmovddup      0x209fb(%rip),%xmm10        # 352a8 <_sk_srcover_rgba_8888_sse2_lowp+0x1278>
   .byte  196,193,121,219,218                 // vpand         %xmm10,%xmm0,%xmm3
-  .byte  196,98,121,24,29,253,228,1,0        // vbroadcastss  0x1e4fd(%rip),%xmm11        # 32db8 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
+  .byte  196,98,121,24,29,173,251,1,0        // vbroadcastss  0x1fbad(%rip),%xmm11        # 34468 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
   .byte  197,33,102,227                      // vpcmpgtd      %xmm3,%xmm11,%xmm12
   .byte  196,193,113,219,210                 // vpand         %xmm10,%xmm1,%xmm2
   .byte  197,33,102,234                      // vpcmpgtd      %xmm2,%xmm11,%xmm13
   .byte  196,195,21,24,252,1                 // vinsertf128   $0x1,%xmm12,%ymm13,%ymm7
-  .byte  197,123,18,37,250,242,1,0           // vmovddup      0x1f2fa(%rip),%xmm12        # 33bd0 <_sk_srcover_rgba_8888_sse2_lowp+0x1250>
+  .byte  197,123,18,37,218,9,2,0             // vmovddup      0x209da(%rip),%xmm12        # 352b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1280>
   .byte  196,193,113,219,204                 // vpand         %xmm12,%xmm1,%xmm1
   .byte  197,241,114,241,16                  // vpslld        $0x10,%xmm1,%xmm1
   .byte  196,193,121,219,196                 // vpand         %xmm12,%xmm0,%xmm0
@@ -29497,7 +29497,7 @@
   .byte  197,241,235,202                     // vpor          %xmm2,%xmm1,%xmm1
   .byte  197,233,114,243,13                  // vpslld        $0xd,%xmm3,%xmm2
   .byte  197,249,235,194                     // vpor          %xmm2,%xmm0,%xmm0
-  .byte  196,226,121,24,29,183,228,1,0       // vbroadcastss  0x1e4b7(%rip),%xmm3        # 32dbc <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
+  .byte  196,226,121,24,29,103,251,1,0       // vbroadcastss  0x1fb67(%rip),%xmm3        # 3446c <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
   .byte  197,249,254,195                     // vpaddd        %xmm3,%xmm0,%xmm0
   .byte  197,241,254,203                     // vpaddd        %xmm3,%xmm1,%xmm1
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
@@ -29634,14 +29634,14 @@
   .byte  196,226,121,51,236                  // vpmovzxwd     %xmm4,%xmm5
   .byte  197,249,112,228,78                  // vpshufd       $0x4e,%xmm4,%xmm4
   .byte  196,226,121,51,228                  // vpmovzxwd     %xmm4,%xmm4
-  .byte  197,123,18,21,68,240,1,0            // vmovddup      0x1f044(%rip),%xmm10        # 33bc8 <_sk_srcover_rgba_8888_sse2_lowp+0x1248>
+  .byte  197,123,18,21,36,7,2,0              // vmovddup      0x20724(%rip),%xmm10        # 352a8 <_sk_srcover_rgba_8888_sse2_lowp+0x1278>
   .byte  196,193,89,219,250                  // vpand         %xmm10,%xmm4,%xmm7
-  .byte  196,98,121,24,29,38,226,1,0         // vbroadcastss  0x1e226(%rip),%xmm11        # 32db8 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
+  .byte  196,98,121,24,29,214,248,1,0        // vbroadcastss  0x1f8d6(%rip),%xmm11        # 34468 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
   .byte  197,33,102,231                      // vpcmpgtd      %xmm7,%xmm11,%xmm12
   .byte  196,193,81,219,242                  // vpand         %xmm10,%xmm5,%xmm6
   .byte  197,33,102,238                      // vpcmpgtd      %xmm6,%xmm11,%xmm13
   .byte  196,195,21,24,220,1                 // vinsertf128   $0x1,%xmm12,%ymm13,%ymm3
-  .byte  197,123,18,37,35,240,1,0            // vmovddup      0x1f023(%rip),%xmm12        # 33bd0 <_sk_srcover_rgba_8888_sse2_lowp+0x1250>
+  .byte  197,123,18,37,3,7,2,0               // vmovddup      0x20703(%rip),%xmm12        # 352b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1280>
   .byte  196,193,81,219,236                  // vpand         %xmm12,%xmm5,%xmm5
   .byte  197,209,114,245,16                  // vpslld        $0x10,%xmm5,%xmm5
   .byte  196,193,89,219,228                  // vpand         %xmm12,%xmm4,%xmm4
@@ -29650,7 +29650,7 @@
   .byte  197,209,235,238                     // vpor          %xmm6,%xmm5,%xmm5
   .byte  197,201,114,247,13                  // vpslld        $0xd,%xmm7,%xmm6
   .byte  197,217,235,230                     // vpor          %xmm6,%xmm4,%xmm4
-  .byte  196,226,121,24,61,224,225,1,0       // vbroadcastss  0x1e1e0(%rip),%xmm7        # 32dbc <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
+  .byte  196,226,121,24,61,144,248,1,0       // vbroadcastss  0x1f890(%rip),%xmm7        # 3446c <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
   .byte  197,217,254,231                     // vpaddd        %xmm7,%xmm4,%xmm4
   .byte  197,209,254,239                     // vpaddd        %xmm7,%xmm5,%xmm5
   .byte  196,227,85,24,228,1                 // vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
@@ -29830,14 +29830,14 @@
   .byte  196,226,121,51,208                  // vpmovzxwd     %xmm0,%xmm2
   .byte  197,249,112,192,78                  // vpshufd       $0x4e,%xmm0,%xmm0
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
-  .byte  197,123,18,29,170,236,1,0           // vmovddup      0x1ecaa(%rip),%xmm11        # 33bc8 <_sk_srcover_rgba_8888_sse2_lowp+0x1248>
+  .byte  197,123,18,29,138,3,2,0             // vmovddup      0x2038a(%rip),%xmm11        # 352a8 <_sk_srcover_rgba_8888_sse2_lowp+0x1278>
   .byte  196,193,121,219,219                 // vpand         %xmm11,%xmm0,%xmm3
-  .byte  196,98,121,24,37,140,222,1,0        // vbroadcastss  0x1de8c(%rip),%xmm12        # 32db8 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
+  .byte  196,98,121,24,37,60,245,1,0         // vbroadcastss  0x1f53c(%rip),%xmm12        # 34468 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
   .byte  197,25,102,235                      // vpcmpgtd      %xmm3,%xmm12,%xmm13
   .byte  196,193,105,219,203                 // vpand         %xmm11,%xmm2,%xmm1
   .byte  197,25,102,193                      // vpcmpgtd      %xmm1,%xmm12,%xmm8
   .byte  196,67,61,24,197,1                  // vinsertf128   $0x1,%xmm13,%ymm8,%ymm8
-  .byte  197,123,18,45,137,236,1,0           // vmovddup      0x1ec89(%rip),%xmm13        # 33bd0 <_sk_srcover_rgba_8888_sse2_lowp+0x1250>
+  .byte  197,123,18,45,105,3,2,0             // vmovddup      0x20369(%rip),%xmm13        # 352b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1280>
   .byte  196,193,105,219,213                 // vpand         %xmm13,%xmm2,%xmm2
   .byte  197,233,114,242,16                  // vpslld        $0x10,%xmm2,%xmm2
   .byte  197,241,114,241,13                  // vpslld        $0xd,%xmm1,%xmm1
@@ -29846,7 +29846,7 @@
   .byte  197,249,114,240,16                  // vpslld        $0x10,%xmm0,%xmm0
   .byte  197,233,114,243,13                  // vpslld        $0xd,%xmm3,%xmm2
   .byte  197,249,235,194                     // vpor          %xmm2,%xmm0,%xmm0
-  .byte  196,226,121,24,29,70,222,1,0        // vbroadcastss  0x1de46(%rip),%xmm3        # 32dbc <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
+  .byte  196,226,121,24,29,246,244,1,0       // vbroadcastss  0x1f4f6(%rip),%xmm3        # 3446c <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
   .byte  197,249,254,195                     // vpaddd        %xmm3,%xmm0,%xmm0
   .byte  197,241,254,203                     // vpaddd        %xmm3,%xmm1,%xmm1
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
@@ -29936,12 +29936,12 @@
   .byte  197,252,17,108,36,192               // vmovups       %ymm5,-0x40(%rsp)
   .byte  197,252,17,100,36,160               // vmovups       %ymm4,-0x60(%rsp)
   .byte  197,252,40,225                      // vmovaps       %ymm1,%ymm4
-  .byte  196,98,125,24,5,124,220,1,0         // vbroadcastss  0x1dc7c(%rip),%ymm8        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,5,44,243,1,0          // vbroadcastss  0x1f32c(%rip),%ymm8        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,124,84,216                   // vandps        %ymm8,%ymm0,%ymm11
-  .byte  196,98,125,24,21,162,220,1,0        // vbroadcastss  0x1dca2(%rip),%ymm10        # 32dc0 <_sk_srcover_rgba_8888_sse2_lowp+0x440>
+  .byte  196,98,125,24,21,82,243,1,0         // vbroadcastss  0x1f352(%rip),%ymm10        # 34470 <_sk_srcover_rgba_8888_sse2_lowp+0x440>
   .byte  196,65,124,84,226                   // vandps        %ymm10,%ymm0,%ymm12
   .byte  196,67,125,25,229,1                 // vextractf128  $0x1,%ymm12,%xmm13
-  .byte  196,98,121,24,13,146,220,1,0        // vbroadcastss  0x1dc92(%rip),%xmm9        # 32dc4 <_sk_srcover_rgba_8888_sse2_lowp+0x444>
+  .byte  196,98,121,24,13,66,243,1,0         // vbroadcastss  0x1f342(%rip),%xmm9        # 34474 <_sk_srcover_rgba_8888_sse2_lowp+0x444>
   .byte  196,65,49,102,245                   // vpcmpgtd      %xmm13,%xmm9,%xmm14
   .byte  196,65,49,102,252                   // vpcmpgtd      %xmm12,%xmm9,%xmm15
   .byte  196,67,5,24,246,1                   // vinsertf128   $0x1,%xmm14,%ymm15,%ymm14
@@ -29952,7 +29952,7 @@
   .byte  196,65,17,254,239                   // vpaddd        %xmm15,%xmm13,%xmm13
   .byte  196,193,25,114,212,13               // vpsrld        $0xd,%xmm12,%xmm12
   .byte  196,65,25,254,227                   // vpaddd        %xmm11,%xmm12,%xmm12
-  .byte  196,98,121,24,29,85,220,1,0         // vbroadcastss  0x1dc55(%rip),%xmm11        # 32dc8 <_sk_srcover_rgba_8888_sse2_lowp+0x448>
+  .byte  196,98,121,24,29,5,243,1,0          // vbroadcastss  0x1f305(%rip),%xmm11        # 34478 <_sk_srcover_rgba_8888_sse2_lowp+0x448>
   .byte  196,65,17,254,235                   // vpaddd        %xmm11,%xmm13,%xmm13
   .byte  196,65,25,254,227                   // vpaddd        %xmm11,%xmm12,%xmm12
   .byte  196,67,29,24,237,1                  // vinsertf128   $0x1,%xmm13,%ymm12,%ymm13
@@ -30103,7 +30103,7 @@
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,148,216,1,0        // vbroadcastss  0x1d894(%rip),%ymm9        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,13,68,239,1,0         // vbroadcastss  0x1ef44(%rip),%ymm9        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,169,109,202                     // vpunpckhqdq   %xmm2,%xmm10,%xmm1
   .byte  197,233,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm2
@@ -30207,7 +30207,7 @@
   .byte  196,226,121,51,192                  // vpmovzxwd     %xmm0,%xmm0
   .byte  196,227,117,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   .byte  197,252,91,192                      // vcvtdq2ps     %ymm0,%ymm0
-  .byte  196,98,125,24,13,195,214,1,0        // vbroadcastss  0x1d6c3(%rip),%ymm9        # 32cc4 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  .byte  196,98,125,24,13,115,237,1,0        // vbroadcastss  0x1ed73(%rip),%ymm9        # 34374 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   .byte  196,193,124,89,193                  // vmulps        %ymm9,%ymm0,%ymm0
   .byte  197,185,109,203                     // vpunpckhqdq   %xmm3,%xmm8,%xmm1
   .byte  197,225,113,241,8                   // vpsllw        $0x8,%xmm1,%xmm3
@@ -30229,7 +30229,7 @@
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
   .byte  196,193,108,89,209                  // vmulps        %ymm9,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,195,213,1,0       // vbroadcastss  0x1d5c3(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,115,236,1,0       // vbroadcastss  0x1ec73(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,4,64                // vmovd         (%r8,%rax,2),%xmm0
   .byte  196,65,49,239,201                   // vpxor         %xmm9,%xmm9,%xmm9
@@ -30305,7 +30305,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  73,99,193                           // movslq        %r9d,%rax
-  .byte  196,98,125,24,5,140,213,1,0         // vbroadcastss  0x1d58c(%rip),%ymm8        # 32d4c <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
+  .byte  196,98,125,24,5,60,236,1,0          // vbroadcastss  0x1ec3c(%rip),%ymm8        # 343fc <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
   .byte  196,65,124,89,200                   // vmulps        %ymm8,%ymm0,%ymm9
   .byte  196,65,125,91,201                   // vcvtps2dq     %ymm9,%ymm9
   .byte  196,67,125,25,202,1                 // vextractf128  $0x1,%ymm9,%xmm10
@@ -30464,7 +30464,7 @@
   .byte  255,92,255,255                      // lcall         *-0x1(%rdi,%rdi,8)
   .byte  255,70,255                          // incl          -0x1(%rsi)
   .byte  255                                 // (bad)
-  .byte  255,53,255,255,255,40               // pushq         0x28ffffff(%rip)        # 29015a90 <_sk_srcover_rgba_8888_sse2_lowp+0x28fe3110>
+  .byte  255,53,255,255,255,40               // pushq         0x28ffffff(%rip)        # 29015a90 <_sk_srcover_rgba_8888_sse2_lowp+0x28fe1a60>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -30663,7 +30663,7 @@
   .byte  196,193,58,88,192                   // vaddss        %xmm8,%xmm8,%xmm0
   .byte  196,227,121,4,192,0                 // vpermilps     $0x0,%xmm0,%xmm0
   .byte  196,99,125,24,192,1                 // vinsertf128   $0x1,%xmm0,%ymm0,%ymm8
-  .byte  197,178,89,5,169,206,1,0            // vmulss        0x1cea9(%rip),%xmm9,%xmm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,5,89,229,1,0             // vmulss        0x1e559(%rip),%xmm9,%xmm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,227,121,4,192,0                 // vpermilps     $0x0,%xmm0,%xmm0
   .byte  196,227,125,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   .byte  197,164,89,192                      // vmulps        %ymm0,%ymm11,%ymm0
@@ -30690,7 +30690,7 @@
   .byte  196,193,58,88,200                   // vaddss        %xmm8,%xmm8,%xmm1
   .byte  196,227,121,4,201,0                 // vpermilps     $0x0,%xmm1,%xmm1
   .byte  196,99,117,24,193,1                 // vinsertf128   $0x1,%xmm1,%ymm1,%ymm8
-  .byte  197,178,89,13,63,206,1,0            // vmulss        0x1ce3f(%rip),%xmm9,%xmm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,178,89,13,239,228,1,0           // vmulss        0x1e4ef(%rip),%xmm9,%xmm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,227,121,4,201,0                 // vpermilps     $0x0,%xmm1,%xmm1
   .byte  196,227,117,24,201,1                // vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
   .byte  197,164,89,201                      // vmulps        %ymm1,%ymm11,%ymm1
@@ -30710,7 +30710,7 @@
 _sk_clamp_x_1_avx:
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,252,205,1,0         // vbroadcastss  0x1cdfc(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,172,228,1,0         // vbroadcastss  0x1e4ac(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -30723,7 +30723,7 @@
   .byte  196,193,124,92,192                  // vsubps        %ymm8,%ymm0,%ymm0
   .byte  196,65,60,87,192                    // vxorps        %ymm8,%ymm8,%ymm8
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,214,205,1,0         // vbroadcastss  0x1cdd6(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,134,228,1,0         // vbroadcastss  0x1e486(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -30732,9 +30732,9 @@
 .globl _sk_mirror_x_1_avx
 FUNCTION(_sk_mirror_x_1_avx)
 _sk_mirror_x_1_avx:
-  .byte  196,98,125,24,5,220,205,1,0         // vbroadcastss  0x1cddc(%rip),%ymm8        # 32c48 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  .byte  196,98,125,24,5,140,228,1,0         // vbroadcastss  0x1e48c(%rip),%ymm8        # 342f8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,13,178,205,1,0        // vbroadcastss  0x1cdb2(%rip),%ymm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,13,98,228,1,0         // vbroadcastss  0x1e462(%rip),%ymm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,124,89,201                   // vmulps        %ymm9,%ymm0,%ymm9
   .byte  196,67,125,8,201,1                  // vroundps      $0x1,%ymm9,%ymm9
   .byte  196,65,52,88,201                    // vaddps        %ymm9,%ymm9,%ymm9
@@ -30744,7 +30744,7 @@
   .byte  197,60,92,200                       // vsubps        %ymm0,%ymm8,%ymm9
   .byte  197,180,84,192                      // vandps        %ymm0,%ymm9,%ymm0
   .byte  197,188,95,192                      // vmaxps        %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,24,5,130,205,1,0         // vbroadcastss  0x1cd82(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,50,228,1,0          // vbroadcastss  0x1e432(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -30753,12 +30753,12 @@
 .globl _sk_luminance_to_alpha_avx
 FUNCTION(_sk_luminance_to_alpha_avx)
 _sk_luminance_to_alpha_avx:
-  .byte  196,226,125,24,29,148,206,1,0       // vbroadcastss  0x1ce94(%rip),%ymm3        # 32d54 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
+  .byte  196,226,125,24,29,68,229,1,0        // vbroadcastss  0x1e544(%rip),%ymm3        # 34404 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,29,131,206,1,0       // vbroadcastss  0x1ce83(%rip),%ymm3        # 32d50 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
+  .byte  196,226,125,24,29,51,229,1,0        // vbroadcastss  0x1e533(%rip),%ymm3        # 34400 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
   .byte  197,244,89,203                      // vmulps        %ymm3,%ymm1,%ymm1
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
-  .byte  196,226,125,24,13,122,206,1,0       // vbroadcastss  0x1ce7a(%rip),%ymm1        # 32d58 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
+  .byte  196,226,125,24,13,42,229,1,0        // vbroadcastss  0x1e52a(%rip),%ymm1        # 34408 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
   .byte  197,236,89,201                      // vmulps        %ymm1,%ymm2,%ymm1
   .byte  197,252,88,217                      // vaddps        %ymm1,%ymm0,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31166,18 +31166,18 @@
 .globl _sk_gauss_a_to_rgba_avx
 FUNCTION(_sk_gauss_a_to_rgba_avx)
 _sk_gauss_a_to_rgba_avx:
-  .byte  196,226,125,24,5,145,199,1,0        // vbroadcastss  0x1c791(%rip),%ymm0        # 32d5c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
+  .byte  196,226,125,24,5,65,222,1,0         // vbroadcastss  0x1de41(%rip),%ymm0        # 3440c <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
   .byte  197,228,89,192                      // vmulps        %ymm0,%ymm3,%ymm0
-  .byte  196,226,125,24,13,136,199,1,0       // vbroadcastss  0x1c788(%rip),%ymm1        # 32d60 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
+  .byte  196,226,125,24,13,56,222,1,0        // vbroadcastss  0x1de38(%rip),%ymm1        # 34410 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,13,123,199,1,0       // vbroadcastss  0x1c77b(%rip),%ymm1        # 32d64 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
+  .byte  196,226,125,24,13,43,222,1,0        // vbroadcastss  0x1de2b(%rip),%ymm1        # 34414 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,13,110,199,1,0       // vbroadcastss  0x1c76e(%rip),%ymm1        # 32d68 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
+  .byte  196,226,125,24,13,30,222,1,0        // vbroadcastss  0x1de1e(%rip),%ymm1        # 34418 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  197,252,89,195                      // vmulps        %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,24,13,97,199,1,0        // vbroadcastss  0x1c761(%rip),%ymm1        # 32d6c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
+  .byte  196,226,125,24,13,17,222,1,0        // vbroadcastss  0x1de11(%rip),%ymm1        # 3441c <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
   .byte  197,252,88,193                      // vaddps        %ymm1,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,40,200                      // vmovaps       %ymm0,%ymm1
@@ -31204,7 +31204,7 @@
   .byte  73,255,200                          // dec           %r8
   .byte  72,131,195,4                        // add           $0x4,%rbx
   .byte  196,65,52,87,201                    // vxorps        %ymm9,%ymm9,%ymm9
-  .byte  196,98,125,24,21,222,197,1,0        // vbroadcastss  0x1c5de(%rip),%ymm10        # 32c34 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  .byte  196,98,125,24,21,142,220,1,0        // vbroadcastss  0x1dc8e(%rip),%ymm10        # 342e4 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  196,98,125,24,3                     // vbroadcastss  (%rbx),%ymm8
   .byte  197,60,194,192,2                    // vcmpleps      %ymm0,%ymm8,%ymm8
@@ -31398,27 +31398,27 @@
   .byte  196,65,52,95,226                    // vmaxps        %ymm10,%ymm9,%ymm12
   .byte  196,65,36,94,220                    // vdivps        %ymm12,%ymm11,%ymm11
   .byte  196,65,36,89,227                    // vmulps        %ymm11,%ymm11,%ymm12
-  .byte  196,98,125,24,45,12,195,1,0         // vbroadcastss  0x1c30c(%rip),%ymm13        # 32d70 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
+  .byte  196,98,125,24,45,188,217,1,0        // vbroadcastss  0x1d9bc(%rip),%ymm13        # 34420 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,2,195,1,0          // vbroadcastss  0x1c302(%rip),%ymm14        # 32d74 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
+  .byte  196,98,125,24,53,178,217,1,0        // vbroadcastss  0x1d9b2(%rip),%ymm14        # 34424 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,237                    // vmulps        %ymm13,%ymm12,%ymm13
-  .byte  196,98,125,24,53,243,194,1,0        // vbroadcastss  0x1c2f3(%rip),%ymm14        # 32d78 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
+  .byte  196,98,125,24,53,163,217,1,0        // vbroadcastss  0x1d9a3(%rip),%ymm14        # 34428 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
   .byte  196,65,20,88,238                    // vaddps        %ymm14,%ymm13,%ymm13
   .byte  196,65,28,89,229                    // vmulps        %ymm13,%ymm12,%ymm12
-  .byte  196,98,125,24,45,228,194,1,0        // vbroadcastss  0x1c2e4(%rip),%ymm13        # 32d7c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
+  .byte  196,98,125,24,45,148,217,1,0        // vbroadcastss  0x1d994(%rip),%ymm13        # 3442c <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
   .byte  196,65,28,88,229                    // vaddps        %ymm13,%ymm12,%ymm12
   .byte  196,65,36,89,220                    // vmulps        %ymm12,%ymm11,%ymm11
   .byte  196,65,52,194,202,1                 // vcmpltps      %ymm10,%ymm9,%ymm9
-  .byte  196,98,125,24,21,207,194,1,0        // vbroadcastss  0x1c2cf(%rip),%ymm10        # 32d80 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
+  .byte  196,98,125,24,21,127,217,1,0        // vbroadcastss  0x1d97f(%rip),%ymm10        # 34430 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
   .byte  196,65,44,92,211                    // vsubps        %ymm11,%ymm10,%ymm10
   .byte  196,67,37,74,202,144                // vblendvps     %ymm9,%ymm10,%ymm11,%ymm9
   .byte  196,193,124,194,192,1               // vcmpltps      %ymm8,%ymm0,%ymm0
-  .byte  196,98,125,24,21,97,193,1,0         // vbroadcastss  0x1c161(%rip),%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,21,17,216,1,0         // vbroadcastss  0x1d811(%rip),%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,44,92,209                    // vsubps        %ymm9,%ymm10,%ymm10
   .byte  196,195,53,74,194,0                 // vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   .byte  196,65,116,194,200,1                // vcmpltps      %ymm8,%ymm1,%ymm9
-  .byte  196,98,125,24,21,75,193,1,0         // vbroadcastss  0x1c14b(%rip),%ymm10        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,21,251,215,1,0        // vbroadcastss  0x1d7fb(%rip),%ymm10        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,44,92,208                       // vsubps        %ymm0,%ymm10,%ymm10
   .byte  196,195,125,74,194,144              // vblendvps     %ymm9,%ymm10,%ymm0,%ymm0
   .byte  196,65,124,194,200,3                // vcmpunordps   %ymm8,%ymm0,%ymm9
@@ -31448,7 +31448,7 @@
   .byte  196,67,121,4,210,0                  // vpermilps     $0x0,%xmm10,%xmm10
   .byte  196,67,45,24,210,1                  // vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  196,98,125,24,29,69,194,1,0         // vbroadcastss  0x1c245(%rip),%ymm11        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,29,245,216,1,0        // vbroadcastss  0x1d8f5(%rip),%ymm11        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
   .byte  197,252,89,192                      // vmulps        %ymm0,%ymm0,%ymm0
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
@@ -31457,17 +31457,17 @@
   .byte  196,227,121,4,192,0                 // vpermilps     $0x0,%xmm0,%xmm0
   .byte  196,227,125,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,13,25,194,1,0         // vbroadcastss  0x1c219(%rip),%ymm9        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  196,98,125,24,13,201,216,1,0        // vbroadcastss  0x1d8c9(%rip),%ymm9        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,65,44,89,194                    // vmulps        %ymm10,%ymm10,%ymm8
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  196,98,125,24,13,247,193,1,0        // vbroadcastss  0x1c1f7(%rip),%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,13,167,216,1,0        // vbroadcastss  0x1d8a7(%rip),%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,44,87,201                    // vxorps        %ymm9,%ymm10,%ymm9
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,132,192,1,0        // vbroadcastss  0x1c084(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,52,215,1,0         // vbroadcastss  0x1d734(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  196,65,60,89,210                    // vmulps        %ymm10,%ymm8,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
@@ -31487,7 +31487,7 @@
   .byte  196,67,121,4,210,0                  // vpermilps     $0x0,%xmm10,%xmm10
   .byte  196,67,45,24,210,1                  // vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
   .byte  197,44,88,208                       // vaddps        %ymm0,%ymm10,%ymm10
-  .byte  196,98,125,24,29,151,193,1,0        // vbroadcastss  0x1c197(%rip),%ymm11        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,29,71,216,1,0         // vbroadcastss  0x1d847(%rip),%ymm11        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,44,89,211                    // vmulps        %ymm11,%ymm10,%ymm10
   .byte  197,252,89,192                      // vmulps        %ymm0,%ymm0,%ymm0
   .byte  197,116,89,217                      // vmulps        %ymm1,%ymm1,%ymm11
@@ -31496,17 +31496,17 @@
   .byte  196,227,121,4,192,0                 // vpermilps     $0x0,%xmm0,%xmm0
   .byte  196,227,125,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   .byte  197,164,92,192                      // vsubps        %ymm0,%ymm11,%ymm0
-  .byte  196,98,125,24,13,107,193,1,0        // vbroadcastss  0x1c16b(%rip),%ymm9        # 32d88 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  .byte  196,98,125,24,13,27,216,1,0         // vbroadcastss  0x1d81b(%rip),%ymm9        # 34438 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
   .byte  197,188,89,192                      // vmulps        %ymm0,%ymm8,%ymm0
   .byte  196,65,44,89,194                    // vmulps        %ymm10,%ymm10,%ymm8
   .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
   .byte  197,252,81,192                      // vsqrtps       %ymm0,%ymm0
   .byte  196,98,125,24,64,68                 // vbroadcastss  0x44(%rax),%ymm8
-  .byte  196,98,125,24,13,73,193,1,0         // vbroadcastss  0x1c149(%rip),%ymm9        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,13,249,215,1,0        // vbroadcastss  0x1d7f9(%rip),%ymm9        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,65,44,87,201                    // vxorps        %ymm9,%ymm10,%ymm9
   .byte  196,65,124,92,210                   // vsubps        %ymm10,%ymm0,%ymm10
-  .byte  196,98,125,24,29,214,191,1,0        // vbroadcastss  0x1bfd6(%rip),%ymm11        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,29,134,214,1,0        // vbroadcastss  0x1d686(%rip),%ymm11        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,60,89,195                    // vmulps        %ymm11,%ymm8,%ymm8
   .byte  196,65,60,89,210                    // vmulps        %ymm10,%ymm8,%ymm10
   .byte  197,180,92,192                      // vsubps        %ymm0,%ymm9,%ymm0
@@ -31525,7 +31525,7 @@
   .byte  196,67,121,4,201,0                  // vpermilps     $0x0,%xmm9,%xmm9
   .byte  196,67,53,24,201,1                  // vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
   .byte  197,52,88,200                       // vaddps        %ymm0,%ymm9,%ymm9
-  .byte  196,98,125,24,21,239,192,1,0        // vbroadcastss  0x1c0ef(%rip),%ymm10        # 32d84 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  .byte  196,98,125,24,21,159,215,1,0        // vbroadcastss  0x1d79f(%rip),%ymm10        # 34434 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   .byte  196,65,52,89,202                    // vmulps        %ymm10,%ymm9,%ymm9
   .byte  197,252,89,192                      // vmulps        %ymm0,%ymm0,%ymm0
   .byte  197,116,89,209                      // vmulps        %ymm1,%ymm1,%ymm10
@@ -31534,7 +31534,7 @@
   .byte  196,227,121,4,192,0                 // vpermilps     $0x0,%xmm0,%xmm0
   .byte  196,227,125,24,192,1                // vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   .byte  197,172,92,192                      // vsubps        %ymm0,%ymm10,%ymm0
-  .byte  196,98,125,24,5,199,192,1,0         // vbroadcastss  0x1c0c7(%rip),%ymm8        # 32d8c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  .byte  196,98,125,24,5,119,215,1,0         // vbroadcastss  0x1d777(%rip),%ymm8        # 3443c <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   .byte  196,193,124,87,192                  // vxorps        %ymm8,%ymm0,%ymm0
   .byte  196,193,124,94,193                  // vdivps        %ymm9,%ymm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31578,7 +31578,7 @@
 FUNCTION(_sk_save_xy_avx)
 _sk_save_xy_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,237,190,1,0         // vbroadcastss  0x1beed(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,157,213,1,0         // vbroadcastss  0x1d59d(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,124,88,200                   // vaddps        %ymm8,%ymm0,%ymm9
   .byte  196,67,125,8,209,1                  // vroundps      $0x1,%ymm9,%ymm10
   .byte  196,65,52,92,202                    // vsubps        %ymm10,%ymm9,%ymm9
@@ -31615,9 +31615,9 @@
 FUNCTION(_sk_bilinear_nx_avx)
 _sk_bilinear_nx_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,211,191,1,0        // vbroadcastss  0x1bfd3(%rip),%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,5,131,214,1,0        // vbroadcastss  0x1d683(%rip),%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,102,190,1,0         // vbroadcastss  0x1be66(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,22,213,1,0          // vbroadcastss  0x1d516(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31628,7 +31628,7 @@
 FUNCTION(_sk_bilinear_px_avx)
 _sk_bilinear_px_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,67,190,1,0         // vbroadcastss  0x1be43(%rip),%ymm0        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,5,243,212,1,0        // vbroadcastss  0x1d4f3(%rip),%ymm0        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -31640,9 +31640,9 @@
 FUNCTION(_sk_bilinear_ny_avx)
 _sk_bilinear_ny_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,132,191,1,0       // vbroadcastss  0x1bf84(%rip),%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,13,52,214,1,0        // vbroadcastss  0x1d634(%rip),%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,22,190,1,0          // vbroadcastss  0x1be16(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,198,212,1,0         // vbroadcastss  0x1d4c6(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31653,7 +31653,7 @@
 FUNCTION(_sk_bilinear_py_avx)
 _sk_bilinear_py_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,243,189,1,0       // vbroadcastss  0x1bdf3(%rip),%ymm1        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,13,163,212,1,0       // vbroadcastss  0x1d4a3(%rip),%ymm1        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -31665,14 +31665,14 @@
 FUNCTION(_sk_bicubic_n3x_avx)
 _sk_bicubic_n3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,55,191,1,0         // vbroadcastss  0x1bf37(%rip),%ymm0        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,226,125,24,5,231,213,1,0        // vbroadcastss  0x1d5e7(%rip),%ymm0        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,198,189,1,0         // vbroadcastss  0x1bdc6(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,118,212,1,0         // vbroadcastss  0x1d476(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,24,191,1,0         // vbroadcastss  0x1bf18(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,200,213,1,0        // vbroadcastss  0x1d5c8(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,22,190,1,0         // vbroadcastss  0x1be16(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,198,212,1,0        // vbroadcastss  0x1d4c6(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,60,88,194                    // vaddps        %ymm10,%ymm8,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -31684,19 +31684,19 @@
 FUNCTION(_sk_bicubic_n1x_avx)
 _sk_bicubic_n1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,225,190,1,0        // vbroadcastss  0x1bee1(%rip),%ymm0        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,5,145,213,1,0        // vbroadcastss  0x1d591(%rip),%ymm0        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
-  .byte  196,98,125,24,5,116,189,1,0         // vbroadcastss  0x1bd74(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,36,212,1,0          // vbroadcastss  0x1d424(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,128,0,0,0             // vsubps        0x80(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,207,190,1,0        // vbroadcastss  0x1becf(%rip),%ymm9        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,13,127,213,1,0        // vbroadcastss  0x1d57f(%rip),%ymm9        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,197,190,1,0        // vbroadcastss  0x1bec5(%rip),%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,117,213,1,0        // vbroadcastss  0x1d575(%rip),%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,62,189,1,0         // vbroadcastss  0x1bd3e(%rip),%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,21,238,211,1,0        // vbroadcastss  0x1d3ee(%rip),%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,163,190,1,0        // vbroadcastss  0x1bea3(%rip),%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,83,213,1,0         // vbroadcastss  0x1d553(%rip),%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31707,17 +31707,17 @@
 FUNCTION(_sk_bicubic_p1x_avx)
 _sk_bicubic_p1x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,15,189,1,0          // vbroadcastss  0x1bd0f(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,191,211,1,0         // vbroadcastss  0x1d3bf(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,0                        // vaddps        (%rax),%ymm8,%ymm0
   .byte  197,124,16,136,128,0,0,0            // vmovups       0x80(%rax),%ymm9
-  .byte  196,98,125,24,21,106,190,1,0        // vbroadcastss  0x1be6a(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,21,26,213,1,0         // vbroadcastss  0x1d51a(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,96,190,1,0         // vbroadcastss  0x1be60(%rip),%ymm11        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,29,16,213,1,0         // vbroadcastss  0x1d510(%rip),%ymm11        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
   .byte  196,65,44,88,192                    // vaddps        %ymm8,%ymm10,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
-  .byte  196,98,125,24,13,71,190,1,0         // vbroadcastss  0x1be47(%rip),%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,247,212,1,0        // vbroadcastss  0x1d4f7(%rip),%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31728,13 +31728,13 @@
 FUNCTION(_sk_bicubic_p3x_avx)
 _sk_bicubic_p3x_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,5,39,190,1,0         // vbroadcastss  0x1be27(%rip),%ymm0        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,226,125,24,5,215,212,1,0        // vbroadcastss  0x1d4d7(%rip),%ymm0        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  197,252,88,0                        // vaddps        (%rax),%ymm0,%ymm0
   .byte  197,124,16,128,128,0,0,0            // vmovups       0x80(%rax),%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,5,190,1,0          // vbroadcastss  0x1be05(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,181,212,1,0        // vbroadcastss  0x1d4b5(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,3,189,1,0          // vbroadcastss  0x1bd03(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,179,211,1,0        // vbroadcastss  0x1d3b3(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,60,88,194                    // vaddps        %ymm10,%ymm8,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
   .byte  197,124,17,128,0,1,0,0              // vmovups       %ymm8,0x100(%rax)
@@ -31746,14 +31746,14 @@
 FUNCTION(_sk_bicubic_n3y_avx)
 _sk_bicubic_n3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,210,189,1,0       // vbroadcastss  0x1bdd2(%rip),%ymm1        # 32d94 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  .byte  196,226,125,24,13,130,212,1,0       // vbroadcastss  0x1d482(%rip),%ymm1        # 34444 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,96,188,1,0          // vbroadcastss  0x1bc60(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,16,211,1,0          // vbroadcastss  0x1d310(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,178,189,1,0        // vbroadcastss  0x1bdb2(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,98,212,1,0         // vbroadcastss  0x1d462(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,176,188,1,0        // vbroadcastss  0x1bcb0(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,96,211,1,0         // vbroadcastss  0x1d360(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,60,88,194                    // vaddps        %ymm10,%ymm8,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -31765,19 +31765,19 @@
 FUNCTION(_sk_bicubic_n1y_avx)
 _sk_bicubic_n1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,123,189,1,0       // vbroadcastss  0x1bd7b(%rip),%ymm1        # 32d90 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  .byte  196,226,125,24,13,43,212,1,0        // vbroadcastss  0x1d42b(%rip),%ymm1        # 34440 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
-  .byte  196,98,125,24,5,13,188,1,0          // vbroadcastss  0x1bc0d(%rip),%ymm8        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,98,125,24,5,189,210,1,0         // vbroadcastss  0x1d2bd(%rip),%ymm8        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,60,92,128,192,0,0,0             // vsubps        0xc0(%rax),%ymm8,%ymm8
-  .byte  196,98,125,24,13,104,189,1,0        // vbroadcastss  0x1bd68(%rip),%ymm9        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,13,24,212,1,0         // vbroadcastss  0x1d418(%rip),%ymm9        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,94,189,1,0         // vbroadcastss  0x1bd5e(%rip),%ymm10        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,21,14,212,1,0         // vbroadcastss  0x1d40e(%rip),%ymm10        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,201                    // vmulps        %ymm9,%ymm8,%ymm9
-  .byte  196,98,125,24,21,215,187,1,0        // vbroadcastss  0x1bbd7(%rip),%ymm10        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,21,135,210,1,0        // vbroadcastss  0x1d287(%rip),%ymm10        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,65,52,88,202                    // vaddps        %ymm10,%ymm9,%ymm9
   .byte  196,65,60,89,193                    // vmulps        %ymm9,%ymm8,%ymm8
-  .byte  196,98,125,24,13,60,189,1,0         // vbroadcastss  0x1bd3c(%rip),%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,236,211,1,0        // vbroadcastss  0x1d3ec(%rip),%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31788,17 +31788,17 @@
 FUNCTION(_sk_bicubic_p1y_avx)
 _sk_bicubic_p1y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,98,125,24,5,168,187,1,0         // vbroadcastss  0x1bba8(%rip),%ymm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,98,125,24,5,88,210,1,0          // vbroadcastss  0x1d258(%rip),%ymm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,188,88,72,64                    // vaddps        0x40(%rax),%ymm8,%ymm1
   .byte  197,124,16,136,192,0,0,0            // vmovups       0xc0(%rax),%ymm9
-  .byte  196,98,125,24,21,2,189,1,0          // vbroadcastss  0x1bd02(%rip),%ymm10        # 32d9c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  .byte  196,98,125,24,21,178,211,1,0        // vbroadcastss  0x1d3b2(%rip),%ymm10        # 3444c <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
-  .byte  196,98,125,24,29,248,188,1,0        // vbroadcastss  0x1bcf8(%rip),%ymm11        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,98,125,24,29,168,211,1,0        // vbroadcastss  0x1d3a8(%rip),%ymm11        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  196,65,44,88,211                    // vaddps        %ymm11,%ymm10,%ymm10
   .byte  196,65,52,89,210                    // vmulps        %ymm10,%ymm9,%ymm10
   .byte  196,65,44,88,192                    // vaddps        %ymm8,%ymm10,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
-  .byte  196,98,125,24,13,223,188,1,0        // vbroadcastss  0x1bcdf(%rip),%ymm9        # 32da4 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  .byte  196,98,125,24,13,143,211,1,0        // vbroadcastss  0x1d38f(%rip),%ymm9        # 34454 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   .byte  196,65,60,88,193                    // vaddps        %ymm9,%ymm8,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -31809,13 +31809,13 @@
 FUNCTION(_sk_bicubic_p3y_avx)
 _sk_bicubic_p3y_avx:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,13,191,188,1,0       // vbroadcastss  0x1bcbf(%rip),%ymm1        # 32da0 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  .byte  196,226,125,24,13,111,211,1,0       // vbroadcastss  0x1d36f(%rip),%ymm1        # 34450 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   .byte  197,244,88,72,64                    // vaddps        0x40(%rax),%ymm1,%ymm1
   .byte  197,124,16,128,192,0,0,0            // vmovups       0xc0(%rax),%ymm8
   .byte  196,65,60,89,200                    // vmulps        %ymm8,%ymm8,%ymm9
-  .byte  196,98,125,24,21,156,188,1,0        // vbroadcastss  0x1bc9c(%rip),%ymm10        # 32d98 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  .byte  196,98,125,24,21,76,211,1,0         // vbroadcastss  0x1d34c(%rip),%ymm10        # 34448 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   .byte  196,65,60,89,194                    // vmulps        %ymm10,%ymm8,%ymm8
-  .byte  196,98,125,24,21,154,187,1,0        // vbroadcastss  0x1bb9a(%rip),%ymm10        # 32ca4 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  .byte  196,98,125,24,21,74,210,1,0         // vbroadcastss  0x1d24a(%rip),%ymm10        # 34354 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   .byte  196,65,60,88,194                    // vaddps        %ymm10,%ymm8,%ymm8
   .byte  196,65,52,89,192                    // vmulps        %ymm8,%ymm9,%ymm8
   .byte  197,124,17,128,64,1,0,0             // vmovups       %ymm8,0x140(%rax)
@@ -31963,7 +31963,7 @@
   .byte  72,139,0                            // mov           (%rax),%rax
   .byte  197,217,254,207                     // vpaddd        %xmm7,%xmm4,%xmm1
   .byte  197,249,254,193                     // vpaddd        %xmm1,%xmm0,%xmm0
-  .byte  196,226,121,24,37,52,186,1,0        // vbroadcastss  0x1ba34(%rip),%xmm4        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  196,226,121,24,37,228,208,1,0       // vbroadcastss  0x1d0e4(%rip),%xmm4        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  196,98,121,64,228                   // vpmulld       %xmm4,%xmm0,%xmm12
   .byte  196,226,49,64,198                   // vpmulld       %xmm6,%xmm9,%xmm0
   .byte  197,249,127,132,36,128,0,0,0        // vmovdqa       %xmm0,0x80(%rsp)
@@ -32017,7 +32017,7 @@
   .byte  196,161,122,16,60,128               // vmovss        (%rax,%r8,4),%xmm7
   .byte  196,227,73,33,247,48                // vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   .byte  196,227,77,24,237,1                 // vinsertf128   $0x1,%xmm5,%ymm6,%ymm5
-  .byte  196,98,121,24,21,153,183,1,0        // vbroadcastss  0x1b799(%rip),%xmm10        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,121,24,21,73,206,1,0         // vbroadcastss  0x1ce49(%rip),%xmm10        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,193,105,254,210                 // vpaddd        %xmm10,%xmm2,%xmm2
   .byte  196,195,249,22,208,1                // vpextrq       $0x1,%xmm2,%r8
   .byte  196,193,249,126,209                 // vmovq         %xmm2,%r9
@@ -32045,7 +32045,7 @@
   .byte  196,161,122,16,60,128               // vmovss        (%rax,%r8,4),%xmm7
   .byte  196,227,73,33,247,48                // vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   .byte  196,227,77,24,210,1                 // vinsertf128   $0x1,%xmm2,%ymm6,%ymm2
-  .byte  196,226,125,24,61,113,184,1,0       // vbroadcastss  0x1b871(%rip),%ymm7        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,226,125,24,61,33,207,1,0        // vbroadcastss  0x1cf21(%rip),%ymm7        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  197,148,88,247                      // vaddps        %ymm7,%ymm13,%ymm6
   .byte  197,124,40,231                      // vmovaps       %ymm7,%ymm12
   .byte  197,124,17,36,36                    // vmovups       %ymm12,(%rsp)
@@ -32788,7 +32788,7 @@
   .byte  197,249,127,68,36,32                // vmovdqa       %xmm0,0x20(%rsp)
   .byte  197,225,254,216                     // vpaddd        %xmm0,%xmm3,%xmm3
   .byte  197,233,254,195                     // vpaddd        %xmm3,%xmm2,%xmm0
-  .byte  196,98,121,24,13,53,169,1,0         // vbroadcastss  0x1a935(%rip),%xmm9        # 32dac <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  .byte  196,98,121,24,13,229,191,1,0        // vbroadcastss  0x1bfe5(%rip),%xmm9        # 3445c <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   .byte  196,194,121,64,209                  // vpmulld       %xmm9,%xmm0,%xmm2
   .byte  196,226,81,64,199                   // vpmulld       %xmm7,%xmm5,%xmm0
   .byte  197,249,127,132,36,96,1,0,0         // vmovdqa       %xmm0,0x160(%rsp)
@@ -32845,7 +32845,7 @@
   .byte  196,161,122,16,60,128               // vmovss        (%rax,%r8,4),%xmm7
   .byte  196,227,73,33,247,48                // vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   .byte  196,227,77,24,237,1                 // vinsertf128   $0x1,%xmm5,%ymm6,%ymm5
-  .byte  196,98,121,24,21,138,166,1,0        // vbroadcastss  0x1a68a(%rip),%xmm10        # 32c3c <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  .byte  196,98,121,24,21,58,189,1,0         // vbroadcastss  0x1bd3a(%rip),%xmm10        # 342ec <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   .byte  196,193,121,254,194                 // vpaddd        %xmm10,%xmm0,%xmm0
   .byte  196,195,249,22,192,1                // vpextrq       $0x1,%xmm0,%r8
   .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
@@ -32873,7 +32873,7 @@
   .byte  196,161,122,16,52,128               // vmovss        (%rax,%r8,4),%xmm6
   .byte  196,227,105,33,214,48               // vinsertps     $0x30,%xmm6,%xmm2,%xmm2
   .byte  196,227,109,24,208,1                // vinsertf128   $0x1,%xmm0,%ymm2,%ymm2
-  .byte  196,98,125,24,37,98,167,1,0         // vbroadcastss  0x1a762(%rip),%ymm12        # 32da8 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  .byte  196,98,125,24,37,18,190,1,0         // vbroadcastss  0x1be12(%rip),%ymm12        # 34458 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   .byte  196,193,4,88,196                    // vaddps        %ymm12,%ymm15,%ymm0
   .byte  197,124,17,36,36                    // vmovups       %ymm12,(%rsp)
   .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
@@ -34271,7 +34271,7 @@
   .byte  197,228,89,210                      // vmulps        %ymm2,%ymm3,%ymm2
   .byte  197,220,88,210                      // vaddps        %ymm2,%ymm4,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,24,29,171,136,1,0       // vbroadcastss  0x188ab(%rip),%ymm3        # 32c30 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  .byte  196,226,125,24,29,91,159,1,0        // vbroadcastss  0x19f5b(%rip),%ymm3        # 342e0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   .byte  197,252,16,164,36,64,2,0,0          // vmovups       0x240(%rsp),%ymm4
   .byte  197,252,16,172,36,96,2,0,0          // vmovups       0x260(%rsp),%ymm5
   .byte  197,252,16,180,36,128,2,0,0         // vmovups       0x280(%rsp),%ymm6
@@ -34375,9 +34375,9 @@
   .byte  102,15,110,201                      // movd          %ecx,%xmm1
   .byte  102,15,112,201,0                    // pshufd        $0x0,%xmm1,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,88,13,26,143,1,0                 // addps         0x18f1a(%rip),%xmm1        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,13,250,165,1,0                // addps         0x1a5fa(%rip),%xmm1        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,21,33,143,1,0                 // movaps        0x18f21(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,1,166,1,0                  // movaps        0x1a601(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -34392,18 +34392,18 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  102,68,15,110,194                   // movd          %edx,%xmm8
   .byte  102,69,15,112,192,0                 // pshufd        $0x0,%xmm8,%xmm8
-  .byte  102,68,15,254,5,202,142,1,0         // paddd         0x18eca(%rip),%xmm8        # 333a0 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,68,15,254,5,170,165,1,0         // paddd         0x1a5aa(%rip),%xmm8        # 34a80 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
   .byte  102,68,15,110,201                   // movd          %ecx,%xmm9
   .byte  102,69,15,112,201,0                 // pshufd        $0x0,%xmm9,%xmm9
   .byte  102,69,15,239,200                   // pxor          %xmm8,%xmm9
-  .byte  102,68,15,111,21,241,142,1,0        // movdqa        0x18ef1(%rip),%xmm10        # 333e0 <_sk_srcover_rgba_8888_sse2_lowp+0xa60>
+  .byte  102,68,15,111,21,209,165,1,0        // movdqa        0x1a5d1(%rip),%xmm10        # 34ac0 <_sk_srcover_rgba_8888_sse2_lowp+0xa90>
   .byte  102,69,15,111,217                   // movdqa        %xmm9,%xmm11
   .byte  102,69,15,219,218                   // pand          %xmm10,%xmm11
   .byte  102,65,15,114,243,5                 // pslld         $0x5,%xmm11
   .byte  102,69,15,219,208                   // pand          %xmm8,%xmm10
   .byte  102,65,15,114,242,4                 // pslld         $0x4,%xmm10
-  .byte  102,68,15,111,37,221,142,1,0        // movdqa        0x18edd(%rip),%xmm12        # 333f0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
-  .byte  102,68,15,111,45,228,142,1,0        // movdqa        0x18ee4(%rip),%xmm13        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,37,189,165,1,0        // movdqa        0x1a5bd(%rip),%xmm12        # 34ad0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa0>
+  .byte  102,68,15,111,45,196,165,1,0        // movdqa        0x1a5c4(%rip),%xmm13        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,69,15,111,241                   // movdqa        %xmm9,%xmm14
   .byte  102,69,15,219,245                   // pand          %xmm13,%xmm14
   .byte  102,65,15,114,246,2                 // pslld         $0x2,%xmm14
@@ -34419,8 +34419,8 @@
   .byte  102,69,15,235,198                   // por           %xmm14,%xmm8
   .byte  102,69,15,235,193                   // por           %xmm9,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,159,142,1,0              // mulps         0x18e9f(%rip),%xmm8        # 33410 <_sk_srcover_rgba_8888_sse2_lowp+0xa90>
-  .byte  68,15,88,5,167,142,1,0              // addps         0x18ea7(%rip),%xmm8        # 33420 <_sk_srcover_rgba_8888_sse2_lowp+0xaa0>
+  .byte  68,15,89,5,127,165,1,0              // mulps         0x1a57f(%rip),%xmm8        # 34af0 <_sk_srcover_rgba_8888_sse2_lowp+0xac0>
+  .byte  68,15,88,5,135,165,1,0              // addps         0x1a587(%rip),%xmm8        # 34b00 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
   .byte  243,68,15,16,16                     // movss         (%rax),%xmm10
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
@@ -34462,7 +34462,7 @@
 FUNCTION(_sk_black_color_sse41)
 _sk_black_color_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,221,141,1,0                // movaps        0x18ddd(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,189,164,1,0                // movaps        0x1a4bd(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -34473,7 +34473,7 @@
 FUNCTION(_sk_white_color_sse41)
 _sk_white_color_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,5,201,141,1,0                 // movaps        0x18dc9(%rip),%xmm0        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,5,169,164,1,0                 // movaps        0x1a4a9(%rip),%xmm0        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -34519,7 +34519,7 @@
 FUNCTION(_sk_srcatop_sse41)
 _sk_srcatop_sse41:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,121,141,1,0              // movaps        0x18d79(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,89,164,1,0               // movaps        0x1a459(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -34544,7 +34544,7 @@
 _sk_dstatop_sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,44,141,1,0              // movaps        0x18d2c(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,12,164,1,0              // movaps        0x1a40c(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -34591,7 +34591,7 @@
 .globl _sk_srcout_sse41
 FUNCTION(_sk_srcout_sse41)
 _sk_srcout_sse41:
-  .byte  68,15,40,5,192,140,1,0              // movaps        0x18cc0(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,160,163,1,0              // movaps        0x1a3a0(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -34604,7 +34604,7 @@
 .globl _sk_dstout_sse41
 FUNCTION(_sk_dstout_sse41)
 _sk_dstout_sse41:
-  .byte  68,15,40,5,160,140,1,0              // movaps        0x18ca0(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,128,163,1,0              // movaps        0x1a380(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -34621,7 +34621,7 @@
 .globl _sk_srcover_sse41
 FUNCTION(_sk_srcover_sse41)
 _sk_srcover_sse41:
-  .byte  68,15,40,5,115,140,1,0              // movaps        0x18c73(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,83,163,1,0               // movaps        0x1a353(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -34641,7 +34641,7 @@
 .globl _sk_dstover_sse41
 FUNCTION(_sk_dstover_sse41)
 _sk_dstover_sse41:
-  .byte  68,15,40,5,55,140,1,0               // movaps        0x18c37(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,23,163,1,0               // movaps        0x1a317(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -34669,7 +34669,7 @@
 .globl _sk_multiply_sse41
 FUNCTION(_sk_multiply_sse41)
 _sk_multiply_sse41:
-  .byte  68,15,40,5,251,139,1,0              // movaps        0x18bfb(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,219,162,1,0              // movaps        0x1a2db(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -34707,7 +34707,7 @@
 FUNCTION(_sk_plus__sse41)
 _sk_plus__sse41:
   .byte  15,88,196                           // addps         %xmm4,%xmm0
-  .byte  68,15,40,5,124,139,1,0              // movaps        0x18b7c(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,92,162,1,0               // movaps        0x1a25c(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  15,88,205                           // addps         %xmm5,%xmm1
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
@@ -34749,7 +34749,7 @@
 FUNCTION(_sk_xor__sse41)
 _sk_xor__sse41:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,8,139,1,0                  // movaps        0x18b08(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,232,161,1,0                // movaps        0x1a1e8(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -34797,7 +34797,7 @@
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,95,209                        // maxps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,99,138,1,0                 // movaps        0x18a63(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,67,161,1,0                 // movaps        0x1a143(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -34831,7 +34831,7 @@
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,248,137,1,0                // movaps        0x189f8(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,216,160,1,0                // movaps        0x1a0d8(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -34868,7 +34868,7 @@
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,130,137,1,0                // movaps        0x18982(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,98,160,1,0                 // movaps        0x1a062(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -34896,7 +34896,7 @@
   .byte  15,89,214                           // mulps         %xmm6,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,50,137,1,0                 // movaps        0x18932(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,18,160,1,0                 // movaps        0x1a012(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -34908,7 +34908,7 @@
 .globl _sk_colorburn_sse41
 FUNCTION(_sk_colorburn_sse41)
 _sk_colorburn_sse41:
-  .byte  68,15,40,29,25,137,1,0              // movaps        0x18919(%rip),%xmm11        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,29,249,159,1,0             // movaps        0x19ff9(%rip),%xmm11        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
   .byte  68,15,92,231                        // subps         %xmm7,%xmm12
   .byte  69,15,40,204                        // movaps        %xmm12,%xmm9
@@ -34992,7 +34992,7 @@
 FUNCTION(_sk_colordodge_sse41)
 _sk_colordodge_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,21,227,135,1,0             // movaps        0x187e3(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,195,158,1,0             // movaps        0x19ec3(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -35074,7 +35074,7 @@
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
-  .byte  68,15,40,21,174,134,1,0             // movaps        0x186ae(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,142,157,1,0             // movaps        0x19d8e(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,40,234                        // movaps        %xmm10,%xmm5
   .byte  15,92,239                           // subps         %xmm7,%xmm5
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
@@ -35157,7 +35157,7 @@
 _sk_overlay_sse41:
   .byte  68,15,40,201                        // movaps        %xmm1,%xmm9
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
-  .byte  68,15,40,21,131,133,1,0             // movaps        0x18583(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,99,156,1,0              // movaps        0x19c63(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
@@ -35242,7 +35242,7 @@
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  65,15,84,193                        // andps         %xmm9,%xmm0
-  .byte  15,40,13,74,132,1,0                 // movaps        0x1844a(%rip),%xmm1        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,13,42,155,1,0                 // movaps        0x19b2a(%rip),%xmm1        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  68,15,92,208                        // subps         %xmm0,%xmm10
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
@@ -35255,10 +35255,10 @@
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,89,210                           // mulps         %xmm2,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
-  .byte  68,15,40,45,120,132,1,0             // movaps        0x18478(%rip),%xmm13        # 33430 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
+  .byte  68,15,40,45,88,155,1,0              // movaps        0x19b58(%rip),%xmm13        # 34b10 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
   .byte  69,15,88,245                        // addps         %xmm13,%xmm14
   .byte  68,15,89,242                        // mulps         %xmm2,%xmm14
-  .byte  68,15,40,37,120,132,1,0             // movaps        0x18478(%rip),%xmm12        # 33440 <_sk_srcover_rgba_8888_sse2_lowp+0xac0>
+  .byte  68,15,40,37,88,155,1,0              // movaps        0x19b58(%rip),%xmm12        # 34b20 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
   .byte  69,15,89,252                        // mulps         %xmm12,%xmm15
   .byte  69,15,88,254                        // addps         %xmm14,%xmm15
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
@@ -35444,12 +35444,12 @@
   .byte  68,15,84,208                        // andps         %xmm0,%xmm10
   .byte  15,84,200                           // andps         %xmm0,%xmm1
   .byte  68,15,84,232                        // andps         %xmm0,%xmm13
-  .byte  15,40,5,227,129,1,0                 // movaps        0x181e3(%rip),%xmm0        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  15,40,5,195,152,1,0                 // movaps        0x198c3(%rip),%xmm0        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,89,224                        // mulps         %xmm0,%xmm12
-  .byte  15,40,21,232,129,1,0                // movaps        0x181e8(%rip),%xmm2        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  15,40,21,200,152,1,0                // movaps        0x198c8(%rip),%xmm2        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  15,89,250                           // mulps         %xmm2,%xmm7
   .byte  65,15,88,252                        // addps         %xmm12,%xmm7
-  .byte  68,15,40,53,233,129,1,0             // movaps        0x181e9(%rip),%xmm14        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,53,201,152,1,0             // movaps        0x198c9(%rip),%xmm14        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  68,15,40,252                        // movaps        %xmm4,%xmm15
   .byte  69,15,89,254                        // mulps         %xmm14,%xmm15
   .byte  68,15,88,255                        // addps         %xmm7,%xmm15
@@ -35532,7 +35532,7 @@
   .byte  65,15,88,214                        // addps         %xmm14,%xmm2
   .byte  15,40,196                           // movaps        %xmm4,%xmm0
   .byte  102,15,56,20,202                    // blendvps      %xmm0,%xmm2,%xmm1
-  .byte  68,15,40,13,253,127,1,0             // movaps        0x17ffd(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,221,150,1,0             // movaps        0x196dd(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,40,225                        // movaps        %xmm9,%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  15,40,68,36,200                     // movaps        -0x38(%rsp),%xmm0
@@ -35626,14 +35626,14 @@
   .byte  68,15,84,215                        // andps         %xmm7,%xmm10
   .byte  68,15,84,223                        // andps         %xmm7,%xmm11
   .byte  68,15,84,199                        // andps         %xmm7,%xmm8
-  .byte  15,40,21,39,127,1,0                 // movaps        0x17f27(%rip),%xmm2        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  15,40,21,7,150,1,0                  // movaps        0x19607(%rip),%xmm2        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,218                           // mulps         %xmm2,%xmm3
-  .byte  15,40,13,42,127,1,0                 // movaps        0x17f2a(%rip),%xmm1        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  15,40,13,10,150,1,0                 // movaps        0x1960a(%rip),%xmm1        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
   .byte  15,89,249                           // mulps         %xmm1,%xmm7
   .byte  15,88,251                           // addps         %xmm3,%xmm7
-  .byte  68,15,40,45,41,127,1,0              // movaps        0x17f29(%rip),%xmm13        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,45,9,150,1,0               // movaps        0x19609(%rip),%xmm13        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  69,15,89,245                        // mulps         %xmm13,%xmm14
   .byte  68,15,88,247                        // addps         %xmm7,%xmm14
   .byte  65,15,40,218                        // movaps        %xmm10,%xmm3
@@ -35714,7 +35714,7 @@
   .byte  65,15,88,253                        // addps         %xmm13,%xmm7
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  102,68,15,56,20,223                 // blendvps      %xmm0,%xmm7,%xmm11
-  .byte  68,15,40,13,63,125,1,0              // movaps        0x17d3f(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,31,148,1,0              // movaps        0x1941f(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  68,15,92,204                        // subps         %xmm4,%xmm9
   .byte  15,40,124,36,168                    // movaps        -0x58(%rsp),%xmm7
@@ -35769,14 +35769,14 @@
   .byte  15,40,231                           // movaps        %xmm7,%xmm4
   .byte  68,15,89,244                        // mulps         %xmm4,%xmm14
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
-  .byte  68,15,40,13,250,124,1,0             // movaps        0x17cfa(%rip),%xmm9        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,13,218,147,1,0             // movaps        0x193da(%rip),%xmm9        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  65,15,40,250                        // movaps        %xmm10,%xmm7
   .byte  65,15,89,249                        // mulps         %xmm9,%xmm7
-  .byte  68,15,40,21,250,124,1,0             // movaps        0x17cfa(%rip),%xmm10        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  68,15,40,21,218,147,1,0             // movaps        0x193da(%rip),%xmm10        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  65,15,40,219                        // movaps        %xmm11,%xmm3
   .byte  65,15,89,218                        // mulps         %xmm10,%xmm3
   .byte  15,88,223                           // addps         %xmm7,%xmm3
-  .byte  68,15,40,29,247,124,1,0             // movaps        0x17cf7(%rip),%xmm11        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,29,215,147,1,0             // movaps        0x193d7(%rip),%xmm11        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  69,15,40,236                        // movaps        %xmm12,%xmm13
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
@@ -35861,7 +35861,7 @@
   .byte  65,15,88,251                        // addps         %xmm11,%xmm7
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  102,15,56,20,207                    // blendvps      %xmm0,%xmm7,%xmm1
-  .byte  68,15,40,13,3,123,1,0               // movaps        0x17b03(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,227,145,1,0             // movaps        0x191e3(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  68,15,89,192                        // mulps         %xmm0,%xmm8
@@ -35913,13 +35913,13 @@
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,89,205                        // mulps         %xmm5,%xmm9
-  .byte  68,15,40,5,203,122,1,0              // movaps        0x17acb(%rip),%xmm8        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,5,171,145,1,0              // movaps        0x191ab(%rip),%xmm8        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
-  .byte  68,15,40,21,207,122,1,0             // movaps        0x17acf(%rip),%xmm10        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  68,15,40,21,175,145,1,0             // movaps        0x191af(%rip),%xmm10        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  65,15,89,234                        // mulps         %xmm10,%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  68,15,40,37,205,122,1,0             // movaps        0x17acd(%rip),%xmm12        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,37,173,145,1,0             // movaps        0x191ad(%rip),%xmm12        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  68,15,40,242                        // movaps        %xmm2,%xmm14
   .byte  69,15,89,244                        // mulps         %xmm12,%xmm14
   .byte  68,15,88,245                        // addps         %xmm5,%xmm14
@@ -36004,7 +36004,7 @@
   .byte  65,15,88,244                        // addps         %xmm12,%xmm6
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  102,68,15,56,20,206                 // blendvps      %xmm0,%xmm6,%xmm9
-  .byte  15,40,5,211,120,1,0                 // movaps        0x178d3(%rip),%xmm0        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,5,179,143,1,0                 // movaps        0x18fb3(%rip),%xmm0        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,92,215                           // subps         %xmm7,%xmm2
   .byte  15,89,226                           // mulps         %xmm2,%xmm4
@@ -36052,20 +36052,20 @@
   .byte  15,133,219,0,0,0                    // jne           1bc65 <_sk_srcover_rgba_8888_sse41+0xfb>
   .byte  243,65,15,111,60,128                // movdqu        (%r8,%rax,4),%xmm7
   .byte  72,133,255                          // test          %rdi,%rdi
-  .byte  102,15,111,37,229,120,1,0           // movdqa        0x178e5(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,37,197,143,1,0           // movdqa        0x18fc5(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,231                      // pand          %xmm7,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
   .byte  102,15,111,239                      // movdqa        %xmm7,%xmm5
-  .byte  102,15,56,0,45,225,120,1,0          // pshufb        0x178e1(%rip),%xmm5        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,45,193,143,1,0          // pshufb        0x18fc1(%rip),%xmm5        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
   .byte  102,15,111,247                      // movdqa        %xmm7,%xmm6
-  .byte  102,15,56,0,53,225,120,1,0          // pshufb        0x178e1(%rip),%xmm6        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,53,193,143,1,0          // pshufb        0x18fc1(%rip),%xmm6        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  102,15,114,215,24                   // psrld         $0x18,%xmm7
   .byte  15,91,255                           // cvtdq2ps      %xmm7,%xmm7
-  .byte  68,15,40,5,254,119,1,0              // movaps        0x177fe(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,222,142,1,0              // movaps        0x18ede(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
-  .byte  68,15,40,37,210,120,1,0             // movaps        0x178d2(%rip),%xmm12        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,37,178,143,1,0             // movaps        0x18fb2(%rip),%xmm12        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,196                        // mulps         %xmm12,%xmm0
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -36145,7 +36145,7 @@
 .globl _sk_clamp_1_sse41
 FUNCTION(_sk_clamp_1_sse41)
 _sk_clamp_1_sse41:
-  .byte  68,15,40,5,205,118,1,0              // movaps        0x176cd(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,173,141,1,0              // movaps        0x18dad(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -36157,7 +36157,7 @@
 .globl _sk_clamp_a_sse41
 FUNCTION(_sk_clamp_a_sse41)
 _sk_clamp_a_sse41:
-  .byte  15,93,29,178,118,1,0                // minps         0x176b2(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,29,146,141,1,0                // minps         0x18d92(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -36168,7 +36168,7 @@
 .globl _sk_clamp_a_dst_sse41
 FUNCTION(_sk_clamp_a_dst_sse41)
 _sk_clamp_a_dst_sse41:
-  .byte  15,93,61,158,118,1,0                // minps         0x1769e(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,61,126,141,1,0                // minps         0x18d7e(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,93,231                           // minps         %xmm7,%xmm4
   .byte  15,93,239                           // minps         %xmm7,%xmm5
   .byte  15,93,247                           // minps         %xmm7,%xmm6
@@ -36203,7 +36203,7 @@
 .globl _sk_invert_sse41
 FUNCTION(_sk_invert_sse41)
 _sk_invert_sse41:
-  .byte  68,15,40,5,90,118,1,0               // movaps        0x1765a(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,58,141,1,0               // movaps        0x18d3a(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,200                        // subps         %xmm0,%xmm9
   .byte  69,15,40,208                        // movaps        %xmm8,%xmm10
@@ -36264,10 +36264,10 @@
 .globl _sk_unpremul_sse41
 FUNCTION(_sk_unpremul_sse41)
 _sk_unpremul_sse41:
-  .byte  68,15,40,5,232,117,1,0              // movaps        0x175e8(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,200,140,1,0              // movaps        0x18cc8(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,94,195                        // divps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
-  .byte  68,15,194,13,199,118,1,0,1          // cmpltps       0x176c7(%rip),%xmm9        # 334c0 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
+  .byte  68,15,194,13,167,141,1,0,1          // cmpltps       0x18da7(%rip),%xmm9        # 34ba0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
   .byte  69,15,84,200                        // andps         %xmm8,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,201                        // mulps         %xmm9,%xmm1
@@ -36279,20 +36279,20 @@
 .globl _sk_from_srgb_sse41
 FUNCTION(_sk_from_srgb_sse41)
 _sk_from_srgb_sse41:
-  .byte  68,15,40,29,187,118,1,0             // movaps        0x176bb(%rip),%xmm11        # 334d0 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
+  .byte  68,15,40,29,155,141,1,0             // movaps        0x18d9b(%rip),%xmm11        # 34bb0 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
   .byte  68,15,40,208                        // movaps        %xmm0,%xmm10
   .byte  69,15,89,210                        // mulps         %xmm10,%xmm10
-  .byte  68,15,40,37,35,118,1,0              // movaps        0x17623(%rip),%xmm12        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,37,3,141,1,0               // movaps        0x18d03(%rip),%xmm12        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,163,118,1,0             // movaps        0x176a3(%rip),%xmm13        # 334e0 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
+  .byte  68,15,40,45,131,141,1,0             // movaps        0x18d83(%rip),%xmm13        # 34bc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
   .byte  69,15,88,197                        // addps         %xmm13,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
-  .byte  68,15,40,53,163,118,1,0             // movaps        0x176a3(%rip),%xmm14        # 334f0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
+  .byte  68,15,40,53,131,141,1,0             // movaps        0x18d83(%rip),%xmm14        # 34bd0 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
   .byte  69,15,88,198                        // addps         %xmm14,%xmm8
-  .byte  68,15,40,61,167,118,1,0             // movaps        0x176a7(%rip),%xmm15        # 33500 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
+  .byte  68,15,40,61,135,141,1,0             // movaps        0x18d87(%rip),%xmm15        # 34be0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
@@ -36329,19 +36329,19 @@
 _sk_from_srgb_dst_sse41:
   .byte  68,15,40,204                        // movaps        %xmm4,%xmm9
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,29,242,117,1,0             // movaps        0x175f2(%rip),%xmm11        # 334d0 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
+  .byte  68,15,40,29,210,140,1,0             // movaps        0x18cd2(%rip),%xmm11        # 34bb0 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  69,15,89,211                        // mulps         %xmm11,%xmm10
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
-  .byte  68,15,40,37,91,117,1,0              // movaps        0x1755b(%rip),%xmm12        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,37,59,140,1,0              // movaps        0x18c3b(%rip),%xmm12        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  65,15,89,228                        // mulps         %xmm12,%xmm4
-  .byte  68,15,40,45,223,117,1,0             // movaps        0x175df(%rip),%xmm13        # 334e0 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
+  .byte  68,15,40,45,191,140,1,0             // movaps        0x18cbf(%rip),%xmm13        # 34bc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
   .byte  65,15,88,229                        // addps         %xmm13,%xmm4
   .byte  15,89,224                           // mulps         %xmm0,%xmm4
-  .byte  68,15,40,53,224,117,1,0             // movaps        0x175e0(%rip),%xmm14        # 334f0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
+  .byte  68,15,40,53,192,140,1,0             // movaps        0x18cc0(%rip),%xmm14        # 34bd0 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
   .byte  65,15,88,230                        // addps         %xmm14,%xmm4
-  .byte  68,15,40,61,228,117,1,0             // movaps        0x175e4(%rip),%xmm15        # 33500 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
+  .byte  68,15,40,61,196,140,1,0             // movaps        0x18cc4(%rip),%xmm15        # 34be0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
   .byte  69,15,194,207,1                     // cmpltps       %xmm15,%xmm9
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  102,65,15,56,20,226                 // blendvps      %xmm0,%xmm10,%xmm4
@@ -36385,22 +36385,22 @@
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  15,40,209                           // movaps        %xmm1,%xmm2
   .byte  68,15,82,192                        // rsqrtps       %xmm0,%xmm8
-  .byte  68,15,40,29,88,117,1,0              // movaps        0x17558(%rip),%xmm11        # 33510 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
+  .byte  68,15,40,29,56,140,1,0              // movaps        0x18c38(%rip),%xmm11        # 34bf0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc0>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
-  .byte  68,15,40,37,88,117,1,0              // movaps        0x17558(%rip),%xmm12        # 33520 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
+  .byte  68,15,40,37,56,140,1,0              // movaps        0x18c38(%rip),%xmm12        # 34c00 <_sk_srcover_rgba_8888_sse2_lowp+0xbd0>
   .byte  69,15,40,248                        // movaps        %xmm8,%xmm15
   .byte  69,15,89,252                        // mulps         %xmm12,%xmm15
-  .byte  68,15,40,21,88,117,1,0              // movaps        0x17558(%rip),%xmm10        # 33530 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
+  .byte  68,15,40,21,56,140,1,0              // movaps        0x18c38(%rip),%xmm10        # 34c10 <_sk_srcover_rgba_8888_sse2_lowp+0xbe0>
   .byte  69,15,88,250                        // addps         %xmm10,%xmm15
   .byte  69,15,89,248                        // mulps         %xmm8,%xmm15
-  .byte  68,15,40,45,88,117,1,0              // movaps        0x17558(%rip),%xmm13        # 33540 <_sk_srcover_rgba_8888_sse2_lowp+0xbc0>
+  .byte  68,15,40,45,56,140,1,0              // movaps        0x18c38(%rip),%xmm13        # 34c20 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
   .byte  69,15,88,253                        // addps         %xmm13,%xmm15
-  .byte  68,15,40,53,92,117,1,0              // movaps        0x1755c(%rip),%xmm14        # 33550 <_sk_srcover_rgba_8888_sse2_lowp+0xbd0>
+  .byte  68,15,40,53,60,140,1,0              // movaps        0x18c3c(%rip),%xmm14        # 34c30 <_sk_srcover_rgba_8888_sse2_lowp+0xc00>
   .byte  69,15,88,198                        // addps         %xmm14,%xmm8
   .byte  69,15,83,192                        // rcpps         %xmm8,%xmm8
   .byte  69,15,89,199                        // mulps         %xmm15,%xmm8
-  .byte  68,15,40,61,88,117,1,0              // movaps        0x17558(%rip),%xmm15        # 33560 <_sk_srcover_rgba_8888_sse2_lowp+0xbe0>
+  .byte  68,15,40,61,56,140,1,0              // movaps        0x18c38(%rip),%xmm15        # 34c40 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  68,15,82,202                        // rsqrtps       %xmm2,%xmm9
@@ -36455,7 +36455,7 @@
   .byte  68,15,93,224                        // minps         %xmm0,%xmm12
   .byte  65,15,40,203                        // movaps        %xmm11,%xmm1
   .byte  65,15,92,204                        // subps         %xmm12,%xmm1
-  .byte  68,15,40,53,255,114,1,0             // movaps        0x172ff(%rip),%xmm14        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,53,223,137,1,0             // movaps        0x189df(%rip),%xmm14        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,94,241                        // divps         %xmm1,%xmm14
   .byte  69,15,40,211                        // movaps        %xmm11,%xmm10
   .byte  69,15,194,208,0                     // cmpeqps       %xmm8,%xmm10
@@ -36464,27 +36464,27 @@
   .byte  65,15,89,198                        // mulps         %xmm14,%xmm0
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  68,15,194,250,1                     // cmpltps       %xmm2,%xmm15
-  .byte  68,15,84,61,118,116,1,0             // andps         0x17476(%rip),%xmm15        # 33570 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
+  .byte  68,15,84,61,86,139,1,0              // andps         0x18b56(%rip),%xmm15        # 34c50 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,194,193,0                     // cmpeqps       %xmm9,%xmm0
   .byte  65,15,92,208                        // subps         %xmm8,%xmm2
   .byte  65,15,89,214                        // mulps         %xmm14,%xmm2
-  .byte  68,15,40,45,105,116,1,0             // movaps        0x17469(%rip),%xmm13        # 33580 <_sk_srcover_rgba_8888_sse2_lowp+0xc00>
+  .byte  68,15,40,45,73,139,1,0              // movaps        0x18b49(%rip),%xmm13        # 34c60 <_sk_srcover_rgba_8888_sse2_lowp+0xc30>
   .byte  65,15,88,213                        // addps         %xmm13,%xmm2
   .byte  69,15,92,193                        // subps         %xmm9,%xmm8
   .byte  69,15,89,198                        // mulps         %xmm14,%xmm8
-  .byte  68,15,88,5,101,116,1,0              // addps         0x17465(%rip),%xmm8        # 33590 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
+  .byte  68,15,88,5,69,139,1,0               // addps         0x18b45(%rip),%xmm8        # 34c70 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
   .byte  102,68,15,56,20,194                 // blendvps      %xmm0,%xmm2,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  102,69,15,56,20,199                 // blendvps      %xmm0,%xmm15,%xmm8
-  .byte  68,15,89,5,93,116,1,0               // mulps         0x1745d(%rip),%xmm8        # 335a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
+  .byte  68,15,89,5,61,139,1,0               // mulps         0x18b3d(%rip),%xmm8        # 34c80 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
   .byte  69,15,40,203                        // movaps        %xmm11,%xmm9
   .byte  69,15,194,204,4                     // cmpneqps      %xmm12,%xmm9
   .byte  69,15,84,193                        // andps         %xmm9,%xmm8
   .byte  69,15,92,235                        // subps         %xmm11,%xmm13
   .byte  69,15,88,220                        // addps         %xmm12,%xmm11
-  .byte  15,40,5,97,114,1,0                  // movaps        0x17261(%rip),%xmm0        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,40,5,65,137,1,0                  // movaps        0x18941(%rip),%xmm0        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,40,211                        // movaps        %xmm11,%xmm2
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
   .byte  15,194,194,1                        // cmpltps       %xmm2,%xmm0
@@ -36506,7 +36506,7 @@
   .byte  15,41,100,36,184                    // movaps        %xmm4,-0x48(%rsp)
   .byte  15,41,92,36,168                     // movaps        %xmm3,-0x58(%rsp)
   .byte  68,15,40,208                        // movaps        %xmm0,%xmm10
-  .byte  68,15,40,13,23,114,1,0              // movaps        0x17217(%rip),%xmm9        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,13,247,136,1,0             // movaps        0x188f7(%rip),%xmm9        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,194,194,2                        // cmpleps       %xmm2,%xmm0
   .byte  15,40,217                           // movaps        %xmm1,%xmm3
@@ -36519,19 +36519,19 @@
   .byte  15,41,84,36,152                     // movaps        %xmm2,-0x68(%rsp)
   .byte  69,15,88,192                        // addps         %xmm8,%xmm8
   .byte  68,15,92,197                        // subps         %xmm5,%xmm8
-  .byte  68,15,40,53,210,115,1,0             // movaps        0x173d2(%rip),%xmm14        # 335b0 <_sk_srcover_rgba_8888_sse2_lowp+0xc30>
+  .byte  68,15,40,53,178,138,1,0             // movaps        0x18ab2(%rip),%xmm14        # 34c90 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  102,65,15,58,8,198,1                // roundps       $0x1,%xmm14,%xmm0
   .byte  68,15,92,240                        // subps         %xmm0,%xmm14
-  .byte  68,15,40,29,203,115,1,0             // movaps        0x173cb(%rip),%xmm11        # 335c0 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
+  .byte  68,15,40,29,171,138,1,0             // movaps        0x18aab(%rip),%xmm11        # 34ca0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  15,40,245                           // movaps        %xmm5,%xmm6
   .byte  65,15,92,240                        // subps         %xmm8,%xmm6
-  .byte  15,40,61,100,115,1,0                // movaps        0x17364(%rip),%xmm7        # 33570 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
+  .byte  15,40,61,68,138,1,0                 // movaps        0x18a44(%rip),%xmm7        # 34c50 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
   .byte  69,15,40,238                        // movaps        %xmm14,%xmm13
   .byte  68,15,89,239                        // mulps         %xmm7,%xmm13
-  .byte  15,40,29,117,115,1,0                // movaps        0x17375(%rip),%xmm3        # 33590 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
+  .byte  15,40,29,85,138,1,0                 // movaps        0x18a55(%rip),%xmm3        # 34c70 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
   .byte  68,15,40,227                        // movaps        %xmm3,%xmm12
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  68,15,89,230                        // mulps         %xmm6,%xmm12
@@ -36541,7 +36541,7 @@
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  68,15,40,253                        // movaps        %xmm5,%xmm15
   .byte  102,69,15,56,20,252                 // blendvps      %xmm0,%xmm12,%xmm15
-  .byte  68,15,40,37,84,115,1,0              // movaps        0x17354(%rip),%xmm12        # 335a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
+  .byte  68,15,40,37,52,138,1,0              // movaps        0x18a34(%rip),%xmm12        # 34c80 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
   .byte  65,15,40,196                        // movaps        %xmm12,%xmm0
   .byte  65,15,194,198,2                     // cmpleps       %xmm14,%xmm0
   .byte  68,15,89,238                        // mulps         %xmm6,%xmm13
@@ -36575,7 +36575,7 @@
   .byte  65,15,40,198                        // movaps        %xmm14,%xmm0
   .byte  15,40,84,36,152                     // movaps        -0x68(%rsp),%xmm2
   .byte  102,15,56,20,202                    // blendvps      %xmm0,%xmm2,%xmm1
-  .byte  68,15,88,21,236,114,1,0             // addps         0x172ec(%rip),%xmm10        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,88,21,204,137,1,0             // addps         0x189cc(%rip),%xmm10        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  102,65,15,58,8,194,1                // roundps       $0x1,%xmm10,%xmm0
   .byte  68,15,92,208                        // subps         %xmm0,%xmm10
   .byte  69,15,194,218,2                     // cmpleps       %xmm10,%xmm11
@@ -36632,9 +36632,9 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,52                              // jne           1c3cb <_sk_scale_u8_sse41+0x4c>
   .byte  102,71,15,56,49,4,16                // pmovzxbd      (%r8,%r10,1),%xmm8
-  .byte  102,68,15,219,5,217,112,1,0         // pand          0x170d9(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,219,5,185,135,1,0         // pand          0x187b9(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,45,114,1,0               // mulps         0x1722d(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,89,5,13,137,1,0               // mulps         0x1890d(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -36678,17 +36678,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,159,0,0,0                    // jne           1c4e5 <_sk_scale_565_sse41+0xc3>
   .byte  102,71,15,56,51,28,80               // pmovzxwd      (%r8,%r10,2),%xmm11
-  .byte  102,15,111,5,155,113,1,0            // movdqa        0x1719b(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,123,136,1,0            // movdqa        0x1887b(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,65,15,219,195                   // pand          %xmm11,%xmm0
   .byte  68,15,91,200                        // cvtdq2ps      %xmm0,%xmm9
-  .byte  68,15,89,13,154,113,1,0             // mulps         0x1719a(%rip),%xmm9        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,5,162,113,1,0            // movdqa        0x171a2(%rip),%xmm0        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  68,15,89,13,122,136,1,0             // mulps         0x1887a(%rip),%xmm9        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,5,130,136,1,0            // movdqa        0x18882(%rip),%xmm0        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,65,15,219,195                   // pand          %xmm11,%xmm0
   .byte  68,15,91,208                        // cvtdq2ps      %xmm0,%xmm10
-  .byte  68,15,89,21,161,113,1,0             // mulps         0x171a1(%rip),%xmm10        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,68,15,219,29,168,113,1,0        // pand          0x171a8(%rip),%xmm11        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  68,15,89,21,129,136,1,0             // mulps         0x18881(%rip),%xmm10        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,219,29,136,136,1,0        // pand          0x18888(%rip),%xmm11        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  69,15,91,219                        // cvtdq2ps      %xmm11,%xmm11
-  .byte  68,15,89,29,172,113,1,0             // mulps         0x171ac(%rip),%xmm11        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  68,15,89,29,140,136,1,0             // mulps         0x1888c(%rip),%xmm11        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  15,40,195                           // movaps        %xmm3,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
   .byte  69,15,40,226                        // movaps        %xmm10,%xmm12
@@ -36765,9 +36765,9 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,72                              // jne           1c5d5 <_sk_lerp_u8_sse41+0x60>
   .byte  102,71,15,56,49,4,16                // pmovzxbd      (%r8,%r10,1),%xmm8
-  .byte  102,68,15,219,5,227,110,1,0         // pand          0x16ee3(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,219,5,195,133,1,0         // pand          0x185c3(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,55,112,1,0               // mulps         0x17037(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,89,5,23,135,1,0               // mulps         0x18717(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -36818,17 +36818,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  15,133,165,0,0,0                    // jne           1c6f8 <_sk_lerp_565_sse41+0xc9>
   .byte  102,71,15,56,51,20,80               // pmovzxwd      (%r8,%r10,2),%xmm10
-  .byte  102,15,111,5,142,111,1,0            // movdqa        0x16f8e(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,110,134,1,0            // movdqa        0x1866e(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,65,15,219,194                   // pand          %xmm10,%xmm0
   .byte  68,15,91,200                        // cvtdq2ps      %xmm0,%xmm9
-  .byte  68,15,89,13,141,111,1,0             // mulps         0x16f8d(%rip),%xmm9        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,5,149,111,1,0            // movdqa        0x16f95(%rip),%xmm0        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  68,15,89,13,109,134,1,0             // mulps         0x1866d(%rip),%xmm9        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,5,117,134,1,0            // movdqa        0x18675(%rip),%xmm0        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,65,15,219,194                   // pand          %xmm10,%xmm0
   .byte  68,15,91,216                        // cvtdq2ps      %xmm0,%xmm11
-  .byte  68,15,89,29,148,111,1,0             // mulps         0x16f94(%rip),%xmm11        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,68,15,219,21,155,111,1,0        // pand          0x16f9b(%rip),%xmm10        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  68,15,89,29,116,134,1,0             // mulps         0x18674(%rip),%xmm11        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,219,21,123,134,1,0        // pand          0x1867b(%rip),%xmm10        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,159,111,1,0             // mulps         0x16f9f(%rip),%xmm10        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  68,15,89,21,127,134,1,0             // mulps         0x1867f(%rip),%xmm10        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  15,40,195                           // movaps        %xmm3,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -36884,7 +36884,7 @@
   .byte  243,69,15,111,4,144                 // movdqu        (%r8,%rdx,4),%xmm8
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
-  .byte  102,15,111,5,17,109,1,0             // movdqa        0x16d11(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,241,131,1,0            // movdqa        0x183f1(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,73,15,58,22,192,1               // pextrq        $0x1,%xmm0,%r8
   .byte  102,73,15,126,193                   // movq          %xmm0,%r9
@@ -36899,7 +36899,7 @@
   .byte  102,66,15,58,33,4,155,32            // insertps      $0x20,(%rbx,%r11,4),%xmm0
   .byte  102,66,15,58,33,4,3,48              // insertps      $0x30,(%rbx,%r8,1),%xmm0
   .byte  102,65,15,111,200                   // movdqa        %xmm8,%xmm1
-  .byte  102,15,56,0,13,204,108,1,0          // pshufb        0x16ccc(%rip),%xmm1        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,13,172,131,1,0          // pshufb        0x183ac(%rip),%xmm1        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  102,73,15,58,22,200,1               // pextrq        $0x1,%xmm1,%r8
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
   .byte  68,15,182,203                       // movzbl        %bl,%r9d
@@ -36914,7 +36914,7 @@
   .byte  102,15,58,33,202,48                 // insertps      $0x30,%xmm2,%xmm1
   .byte  76,139,64,24                        // mov           0x18(%rax),%r8
   .byte  102,65,15,111,208                   // movdqa        %xmm8,%xmm2
-  .byte  102,15,56,0,21,136,108,1,0          // pshufb        0x16c88(%rip),%xmm2        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,21,104,131,1,0          // pshufb        0x18368(%rip),%xmm2        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  102,72,15,58,22,211,1               // pextrq        $0x1,%xmm2,%rbx
   .byte  102,72,15,126,208                   // movq          %xmm2,%rax
   .byte  68,15,182,200                       // movzbl        %al,%r9d
@@ -36929,7 +36929,7 @@
   .byte  102,15,58,33,211,48                 // insertps      $0x30,%xmm3,%xmm2
   .byte  102,65,15,114,208,24                // psrld         $0x18,%xmm8
   .byte  65,15,91,216                        // cvtdq2ps      %xmm8,%xmm3
-  .byte  15,89,29,117,109,1,0                // mulps         0x16d75(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,85,132,1,0                 // mulps         0x18455(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
@@ -36970,7 +36970,7 @@
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
-  .byte  102,68,15,111,5,71,109,1,0          // movdqa        0x16d47(%rip),%xmm8        # 33650 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,111,5,39,132,1,0          // movdqa        0x18427(%rip),%xmm8        # 34d30 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,56,51,192                    // pmovzxwd      %xmm0,%xmm0
@@ -36988,7 +36988,7 @@
   .byte  102,15,58,33,194,32                 // insertps      $0x20,%xmm2,%xmm0
   .byte  243,66,15,16,20,3                   // movss         (%rbx,%r8,1),%xmm2
   .byte  102,15,58,33,194,48                 // insertps      $0x30,%xmm2,%xmm0
-  .byte  102,15,56,0,13,246,108,1,0          // pshufb        0x16cf6(%rip),%xmm1        # 33660 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
+  .byte  102,15,56,0,13,214,131,1,0          // pshufb        0x183d6(%rip),%xmm1        # 34d40 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
   .byte  102,15,56,51,201                    // pmovzxwd      %xmm1,%xmm1
   .byte  102,73,15,58,22,200,1               // pextrq        $0x1,%xmm1,%r8
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
@@ -37024,7 +37024,7 @@
   .byte  102,65,15,235,216                   // por           %xmm8,%xmm3
   .byte  102,15,56,51,219                    // pmovzxwd      %xmm3,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,68,108,1,0                 // mulps         0x16c44(%rip),%xmm3        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  15,89,29,36,131,1,0                 // mulps         0x18324(%rip),%xmm3        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
@@ -37064,7 +37064,7 @@
   .byte  102,15,97,209                       // punpcklwd     %xmm1,%xmm2
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,97,201                    // punpcklwd     %xmm9,%xmm1
-  .byte  102,68,15,111,5,141,107,1,0         // movdqa        0x16b8d(%rip),%xmm8        # 33650 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,111,5,109,130,1,0         // movdqa        0x1826d(%rip),%xmm8        # 34d30 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,56,51,192                    // pmovzxwd      %xmm0,%xmm0
@@ -37082,7 +37082,7 @@
   .byte  102,15,58,33,195,32                 // insertps      $0x20,%xmm3,%xmm0
   .byte  243,66,15,16,28,3                   // movss         (%rbx,%r8,1),%xmm3
   .byte  102,15,58,33,195,48                 // insertps      $0x30,%xmm3,%xmm0
-  .byte  102,15,56,0,13,60,107,1,0           // pshufb        0x16b3c(%rip),%xmm1        # 33660 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
+  .byte  102,15,56,0,13,28,130,1,0           // pshufb        0x1821c(%rip),%xmm1        # 34d40 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
   .byte  102,15,56,51,201                    // pmovzxwd      %xmm1,%xmm1
   .byte  102,73,15,58,22,200,1               // pextrq        $0x1,%xmm1,%r8
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
@@ -37113,7 +37113,7 @@
   .byte  243,65,15,16,28,24                  // movss         (%r8,%rbx,1),%xmm3
   .byte  102,15,58,33,211,48                 // insertps      $0x30,%xmm3,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,7,104,1,0                  // movaps        0x16807(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,231,126,1,0                // movaps        0x17ee7(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
   .byte  255,224                             // jmpq          *%rax
@@ -37145,7 +37145,7 @@
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,116,104,1,0              // movaps        0x16874(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,84,127,1,0               // movaps        0x17f54(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  102,73,15,58,22,192,1               // pextrq        $0x1,%xmm0,%r8
@@ -37168,7 +37168,7 @@
   .byte  102,15,58,32,197,3                  // pinsrb        $0x3,%ebp,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,13,58,105,1,0              // movaps        0x1693a(%rip),%xmm9        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,13,26,128,1,0              // movaps        0x1801a(%rip),%xmm9        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -37270,7 +37270,7 @@
   .byte  102,15,58,32,197,3                  // pinsrb        $0x3,%ebp,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,13,150,103,1,0             // movaps        0x16796(%rip),%xmm9        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,13,118,126,1,0             // movaps        0x17e76(%rip),%xmm9        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -37452,31 +37452,31 @@
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,194                        // cvtdq2ps      %xmm10,%xmm8
-  .byte  68,15,89,5,113,101,1,0              // mulps         0x16571(%rip),%xmm8        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,21,121,101,1,0             // andps         0x16579(%rip),%xmm10        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,21,161,98,1,0              // orps          0x162a1(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,5,121,101,1,0              // addps         0x16579(%rip),%xmm8        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  68,15,40,37,129,101,1,0             // movaps        0x16581(%rip),%xmm12        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,5,81,124,1,0               // mulps         0x17c51(%rip),%xmm8        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,21,89,124,1,0              // andps         0x17c59(%rip),%xmm10        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,21,129,121,1,0             // orps          0x17981(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,5,89,124,1,0               // addps         0x17c59(%rip),%xmm8        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,37,97,124,1,0              // movaps        0x17c61(%rip),%xmm12        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,196                        // subps         %xmm12,%xmm8
-  .byte  68,15,88,21,129,101,1,0             // addps         0x16581(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  68,15,40,37,137,101,1,0             // movaps        0x16589(%rip),%xmm12        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,21,97,124,1,0              // addps         0x17c61(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,40,37,105,124,1,0             // movaps        0x17c69(%rip),%xmm12        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,196                        // subps         %xmm12,%xmm8
   .byte  69,15,89,195                        // mulps         %xmm11,%xmm8
   .byte  102,69,15,58,8,208,1                // roundps       $0x1,%xmm8,%xmm10
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,5,118,101,1,0              // addps         0x16576(%rip),%xmm8        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  68,15,40,21,126,101,1,0             // movaps        0x1657e(%rip),%xmm10        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,5,86,124,1,0               // addps         0x17c56(%rip),%xmm8        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,21,94,124,1,0              // movaps        0x17c5e(%rip),%xmm10        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  69,15,89,211                        // mulps         %xmm11,%xmm10
   .byte  69,15,92,194                        // subps         %xmm10,%xmm8
-  .byte  68,15,40,21,126,101,1,0             // movaps        0x1657e(%rip),%xmm10        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,21,94,124,1,0              // movaps        0x17c5e(%rip),%xmm10        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  68,15,40,29,130,101,1,0             // movaps        0x16582(%rip),%xmm11        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,29,98,124,1,0              // movaps        0x17c62(%rip),%xmm11        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  69,15,94,218                        // divps         %xmm10,%xmm11
   .byte  69,15,88,216                        // addps         %xmm8,%xmm11
-  .byte  68,15,89,29,130,101,1,0             // mulps         0x16582(%rip),%xmm11        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  68,15,89,29,98,124,1,0              // mulps         0x17c62(%rip),%xmm11        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,69,15,91,211                    // cvtps2dq      %xmm11,%xmm10
   .byte  243,68,15,16,64,20                  // movss         0x14(%rax),%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
@@ -37484,7 +37484,7 @@
   .byte  102,69,15,56,20,193                 // blendvps      %xmm0,%xmm9,%xmm8
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
-  .byte  68,15,93,5,9,98,1,0                 // minps         0x16209(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,93,5,233,120,1,0              // minps         0x178e9(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -37514,31 +37514,31 @@
   .byte  68,15,88,217                        // addps         %xmm1,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,74,100,1,0              // mulps         0x1644a(%rip),%xmm12        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,29,82,100,1,0              // andps         0x16452(%rip),%xmm11        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,29,122,97,1,0              // orps          0x1617a(%rip),%xmm11        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,37,82,100,1,0              // addps         0x16452(%rip),%xmm12        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  15,40,13,91,100,1,0                 // movaps        0x1645b(%rip),%xmm1        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,37,42,123,1,0              // mulps         0x17b2a(%rip),%xmm12        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,29,50,123,1,0              // andps         0x17b32(%rip),%xmm11        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,29,90,120,1,0              // orps          0x1785a(%rip),%xmm11        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,37,50,123,1,0              // addps         0x17b32(%rip),%xmm12        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  15,40,13,59,123,1,0                 // movaps        0x17b3b(%rip),%xmm1        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,88,29,91,100,1,0              // addps         0x1645b(%rip),%xmm11        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  15,40,13,100,100,1,0                // movaps        0x16464(%rip),%xmm1        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,29,59,123,1,0              // addps         0x17b3b(%rip),%xmm11        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  15,40,13,68,123,1,0                 // movaps        0x17b44(%rip),%xmm1        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  65,15,94,203                        // divps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,81,100,1,0              // addps         0x16451(%rip),%xmm12        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  15,40,13,90,100,1,0                 // movaps        0x1645a(%rip),%xmm1        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,37,49,123,1,0              // addps         0x17b31(%rip),%xmm12        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,13,58,123,1,0                 // movaps        0x17b3a(%rip),%xmm1        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  68,15,92,225                        // subps         %xmm1,%xmm12
-  .byte  68,15,40,21,90,100,1,0              // movaps        0x1645a(%rip),%xmm10        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,21,58,123,1,0              // movaps        0x17b3a(%rip),%xmm10        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,13,95,100,1,0                 // movaps        0x1645f(%rip),%xmm1        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,13,63,123,1,0                 // movaps        0x17b3f(%rip),%xmm1        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  65,15,94,202                        // divps         %xmm10,%xmm1
   .byte  65,15,88,204                        // addps         %xmm12,%xmm1
-  .byte  15,89,13,96,100,1,0                 // mulps         0x16460(%rip),%xmm1        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  15,89,13,64,123,1,0                 // mulps         0x17b40(%rip),%xmm1        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,68,15,91,209                    // cvtps2dq      %xmm1,%xmm10
   .byte  243,15,16,72,20                     // movss         0x14(%rax),%xmm1
   .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
@@ -37546,7 +37546,7 @@
   .byte  102,65,15,56,20,201                 // blendvps      %xmm0,%xmm9,%xmm1
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,200                           // maxps         %xmm0,%xmm1
-  .byte  15,93,13,235,96,1,0                 // minps         0x160eb(%rip),%xmm1        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,13,203,119,1,0                // minps         0x177cb(%rip),%xmm1        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -37576,31 +37576,31 @@
   .byte  68,15,88,218                        // addps         %xmm2,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,44,99,1,0               // mulps         0x1632c(%rip),%xmm12        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,29,52,99,1,0               // andps         0x16334(%rip),%xmm11        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,29,92,96,1,0               // orps          0x1605c(%rip),%xmm11        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,37,52,99,1,0               // addps         0x16334(%rip),%xmm12        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  15,40,21,61,99,1,0                  // movaps        0x1633d(%rip),%xmm2        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,37,12,122,1,0              // mulps         0x17a0c(%rip),%xmm12        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,29,20,122,1,0              // andps         0x17a14(%rip),%xmm11        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,29,60,119,1,0              // orps          0x1773c(%rip),%xmm11        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,37,20,122,1,0              // addps         0x17a14(%rip),%xmm12        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  15,40,21,29,122,1,0                 // movaps        0x17a1d(%rip),%xmm2        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,88,29,61,99,1,0               // addps         0x1633d(%rip),%xmm11        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  15,40,21,70,99,1,0                  // movaps        0x16346(%rip),%xmm2        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,29,29,122,1,0              // addps         0x17a1d(%rip),%xmm11        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  15,40,21,38,122,1,0                 // movaps        0x17a26(%rip),%xmm2        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  65,15,94,211                        // divps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,51,99,1,0               // addps         0x16333(%rip),%xmm12        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  15,40,21,60,99,1,0                  // movaps        0x1633c(%rip),%xmm2        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,37,19,122,1,0              // addps         0x17a13(%rip),%xmm12        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,21,28,122,1,0                 // movaps        0x17a1c(%rip),%xmm2        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  65,15,89,211                        // mulps         %xmm11,%xmm2
   .byte  68,15,92,226                        // subps         %xmm2,%xmm12
-  .byte  68,15,40,21,60,99,1,0               // movaps        0x1633c(%rip),%xmm10        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,21,28,122,1,0              // movaps        0x17a1c(%rip),%xmm10        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,21,65,99,1,0                  // movaps        0x16341(%rip),%xmm2        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,21,33,122,1,0                 // movaps        0x17a21(%rip),%xmm2        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  65,15,94,210                        // divps         %xmm10,%xmm2
   .byte  65,15,88,212                        // addps         %xmm12,%xmm2
-  .byte  15,89,21,66,99,1,0                  // mulps         0x16342(%rip),%xmm2        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  15,89,21,34,122,1,0                 // mulps         0x17a22(%rip),%xmm2        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,68,15,91,210                    // cvtps2dq      %xmm2,%xmm10
   .byte  243,15,16,80,20                     // movss         0x14(%rax),%xmm2
   .byte  15,198,210,0                        // shufps        $0x0,%xmm2,%xmm2
@@ -37608,7 +37608,7 @@
   .byte  102,65,15,56,20,209                 // blendvps      %xmm0,%xmm9,%xmm2
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,208                           // maxps         %xmm0,%xmm2
-  .byte  15,93,21,205,95,1,0                 // minps         0x15fcd(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,21,173,118,1,0                // minps         0x176ad(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -37638,31 +37638,31 @@
   .byte  68,15,88,219                        // addps         %xmm3,%xmm11
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,91,227                        // cvtdq2ps      %xmm11,%xmm12
-  .byte  68,15,89,37,14,98,1,0               // mulps         0x1620e(%rip),%xmm12        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,29,22,98,1,0               // andps         0x16216(%rip),%xmm11        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,29,62,95,1,0               // orps          0x15f3e(%rip),%xmm11        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,37,22,98,1,0               // addps         0x16216(%rip),%xmm12        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  15,40,29,31,98,1,0                  // movaps        0x1621f(%rip),%xmm3        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,37,238,120,1,0             // mulps         0x178ee(%rip),%xmm12        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,29,246,120,1,0             // andps         0x178f6(%rip),%xmm11        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,29,30,118,1,0              // orps          0x1761e(%rip),%xmm11        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,37,246,120,1,0             // addps         0x178f6(%rip),%xmm12        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  15,40,29,255,120,1,0                // movaps        0x178ff(%rip),%xmm3        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,88,29,31,98,1,0               // addps         0x1621f(%rip),%xmm11        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  15,40,29,40,98,1,0                  // movaps        0x16228(%rip),%xmm3        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,29,255,120,1,0             // addps         0x178ff(%rip),%xmm11        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  15,40,29,8,121,1,0                  // movaps        0x17908(%rip),%xmm3        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  65,15,94,219                        // divps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  102,69,15,58,8,212,1                // roundps       $0x1,%xmm12,%xmm10
   .byte  69,15,40,220                        // movaps        %xmm12,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  68,15,88,37,21,98,1,0               // addps         0x16215(%rip),%xmm12        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  15,40,29,30,98,1,0                  // movaps        0x1621e(%rip),%xmm3        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,37,245,120,1,0             // addps         0x178f5(%rip),%xmm12        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,29,254,120,1,0                // movaps        0x178fe(%rip),%xmm3        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  65,15,89,219                        // mulps         %xmm11,%xmm3
   .byte  68,15,92,227                        // subps         %xmm3,%xmm12
-  .byte  68,15,40,21,30,98,1,0               // movaps        0x1621e(%rip),%xmm10        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,21,254,120,1,0             // movaps        0x178fe(%rip),%xmm10        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
-  .byte  15,40,29,35,98,1,0                  // movaps        0x16223(%rip),%xmm3        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  15,40,29,3,121,1,0                  // movaps        0x17903(%rip),%xmm3        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  65,15,94,218                        // divps         %xmm10,%xmm3
   .byte  65,15,88,220                        // addps         %xmm12,%xmm3
-  .byte  15,89,29,36,98,1,0                  // mulps         0x16224(%rip),%xmm3        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  15,89,29,4,121,1,0                  // mulps         0x17904(%rip),%xmm3        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,68,15,91,211                    // cvtps2dq      %xmm3,%xmm10
   .byte  243,15,16,88,20                     // movss         0x14(%rax),%xmm3
   .byte  15,198,219,0                        // shufps        $0x0,%xmm3,%xmm3
@@ -37670,7 +37670,7 @@
   .byte  102,65,15,56,20,217                 // blendvps      %xmm0,%xmm9,%xmm3
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,95,216                           // maxps         %xmm0,%xmm3
-  .byte  15,93,29,175,94,1,0                 // minps         0x15eaf(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,29,143,117,1,0                // minps         0x1758f(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -37687,20 +37687,20 @@
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,91,194                           // cvtdq2ps      %xmm2,%xmm0
-  .byte  15,89,5,46,97,1,0                   // mulps         0x1612e(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,40,53,54,97,1,0               // movaps        0x16136(%rip),%xmm14        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
+  .byte  15,89,5,14,120,1,0                  // mulps         0x1780e(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,40,53,22,120,1,0              // movaps        0x17816(%rip),%xmm14        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
   .byte  65,15,84,214                        // andps         %xmm14,%xmm2
-  .byte  68,15,40,37,90,94,1,0               // movaps        0x15e5a(%rip),%xmm12        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,37,58,117,1,0              // movaps        0x1753a(%rip),%xmm12        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,86,212                        // orps          %xmm12,%xmm2
-  .byte  68,15,40,21,46,97,1,0               // movaps        0x1612e(%rip),%xmm10        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
+  .byte  68,15,40,21,14,120,1,0              // movaps        0x1780e(%rip),%xmm10        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
   .byte  65,15,88,194                        // addps         %xmm10,%xmm0
-  .byte  68,15,40,29,50,97,1,0               // movaps        0x16132(%rip),%xmm11        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,40,29,18,120,1,0              // movaps        0x17812(%rip),%xmm11        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  15,40,226                           // movaps        %xmm2,%xmm4
   .byte  65,15,89,227                        // mulps         %xmm11,%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  68,15,40,13,48,97,1,0               // movaps        0x16130(%rip),%xmm9        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,40,13,16,120,1,0              // movaps        0x17810(%rip),%xmm9        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
   .byte  65,15,88,209                        // addps         %xmm9,%xmm2
-  .byte  68,15,40,45,52,97,1,0               // movaps        0x16134(%rip),%xmm13        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,45,20,120,1,0              // movaps        0x17814(%rip),%xmm13        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  65,15,40,229                        // movaps        %xmm13,%xmm4
   .byte  15,94,226                           // divps         %xmm2,%xmm4
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -37711,21 +37711,21 @@
   .byte  102,15,58,8,208,1                   // roundps       $0x1,%xmm0,%xmm2
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
   .byte  15,92,226                           // subps         %xmm2,%xmm4
-  .byte  15,40,53,23,97,1,0                  // movaps        0x16117(%rip),%xmm6        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
+  .byte  15,40,53,247,119,1,0                // movaps        0x177f7(%rip),%xmm6        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
   .byte  15,88,198                           // addps         %xmm6,%xmm0
-  .byte  15,40,61,45,97,1,0                  // movaps        0x1612d(%rip),%xmm7        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  15,40,61,13,120,1,0                 // movaps        0x1780d(%rip),%xmm7        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  15,40,239                           // movaps        %xmm7,%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,40,212                           // movaps        %xmm4,%xmm2
-  .byte  15,40,37,13,97,1,0                  // movaps        0x1610d(%rip),%xmm4        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  15,40,37,237,119,1,0                // movaps        0x177ed(%rip),%xmm4        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  15,92,194                           // subps         %xmm2,%xmm0
-  .byte  68,15,40,61,31,97,1,0               // movaps        0x1611f(%rip),%xmm15        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,61,255,119,1,0             // movaps        0x177ff(%rip),%xmm15        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  65,15,40,215                        // movaps        %xmm15,%xmm2
   .byte  15,94,213                           // divps         %xmm5,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  15,89,5,123,96,1,0                  // mulps         0x1607b(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
+  .byte  15,89,5,91,119,1,0                  // mulps         0x1775b(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
   .byte  65,15,84,206                        // andps         %xmm14,%xmm1
   .byte  65,15,86,204                        // orps          %xmm12,%xmm1
   .byte  65,15,88,194                        // addps         %xmm10,%xmm0
@@ -37751,8 +37751,8 @@
   .byte  15,94,206                           // divps         %xmm6,%xmm1
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  15,91,195                           // cvtdq2ps      %xmm3,%xmm0
-  .byte  15,89,5,28,96,1,0                   // mulps         0x1601c(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  15,84,29,37,96,1,0                  // andps         0x16025(%rip),%xmm3        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
+  .byte  15,89,5,252,118,1,0                 // mulps         0x176fc(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  15,84,29,5,119,1,0                  // andps         0x17705(%rip),%xmm3        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
   .byte  65,15,86,220                        // orps          %xmm12,%xmm3
   .byte  65,15,88,194                        // addps         %xmm10,%xmm0
   .byte  68,15,89,219                        // mulps         %xmm3,%xmm11
@@ -37770,7 +37770,7 @@
   .byte  15,92,253                           // subps         %xmm5,%xmm7
   .byte  68,15,94,255                        // divps         %xmm7,%xmm15
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
-  .byte  15,40,5,109,96,1,0                  // movaps        0x1606d(%rip),%xmm0        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  15,40,5,77,119,1,0                  // movaps        0x1774d(%rip),%xmm0        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  68,15,89,248                        // mulps         %xmm0,%xmm15
@@ -37790,29 +37790,29 @@
 FUNCTION(_sk_lab_to_xyz_sse41)
 _sk_lab_to_xyz_sse41:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,89,5,61,96,1,0                // mulps         0x1603d(%rip),%xmm8        # 33730 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
-  .byte  68,15,40,13,181,93,1,0              // movaps        0x15db5(%rip),%xmm9        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,89,5,29,119,1,0               // mulps         0x1771d(%rip),%xmm8        # 34e10 <_sk_srcover_rgba_8888_sse2_lowp+0xde0>
+  .byte  68,15,40,13,149,116,1,0             // movaps        0x17495(%rip),%xmm9        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,201                        // mulps         %xmm9,%xmm1
-  .byte  15,40,5,58,96,1,0                   // movaps        0x1603a(%rip),%xmm0        # 33740 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
+  .byte  15,40,5,26,119,1,0                  // movaps        0x1771a(%rip),%xmm0        # 34e20 <_sk_srcover_rgba_8888_sse2_lowp+0xdf0>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
-  .byte  68,15,88,5,56,96,1,0                // addps         0x16038(%rip),%xmm8        # 33750 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
-  .byte  68,15,89,5,64,96,1,0                // mulps         0x16040(%rip),%xmm8        # 33760 <_sk_srcover_rgba_8888_sse2_lowp+0xde0>
-  .byte  15,89,13,73,96,1,0                  // mulps         0x16049(%rip),%xmm1        # 33770 <_sk_srcover_rgba_8888_sse2_lowp+0xdf0>
+  .byte  68,15,88,5,24,119,1,0               // addps         0x17718(%rip),%xmm8        # 34e30 <_sk_srcover_rgba_8888_sse2_lowp+0xe00>
+  .byte  68,15,89,5,32,119,1,0               // mulps         0x17720(%rip),%xmm8        # 34e40 <_sk_srcover_rgba_8888_sse2_lowp+0xe10>
+  .byte  15,89,13,41,119,1,0                 // mulps         0x17729(%rip),%xmm1        # 34e50 <_sk_srcover_rgba_8888_sse2_lowp+0xe20>
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  15,89,21,78,96,1,0                  // mulps         0x1604e(%rip),%xmm2        # 33780 <_sk_srcover_rgba_8888_sse2_lowp+0xe00>
+  .byte  15,89,21,46,119,1,0                 // mulps         0x1772e(%rip),%xmm2        # 34e60 <_sk_srcover_rgba_8888_sse2_lowp+0xe30>
   .byte  69,15,40,208                        // movaps        %xmm8,%xmm10
   .byte  68,15,92,210                        // subps         %xmm2,%xmm10
   .byte  68,15,40,217                        // movaps        %xmm1,%xmm11
   .byte  69,15,89,219                        // mulps         %xmm11,%xmm11
   .byte  68,15,89,217                        // mulps         %xmm1,%xmm11
-  .byte  68,15,40,13,66,96,1,0               // movaps        0x16042(%rip),%xmm9        # 33790 <_sk_srcover_rgba_8888_sse2_lowp+0xe10>
+  .byte  68,15,40,13,34,119,1,0              // movaps        0x17722(%rip),%xmm9        # 34e70 <_sk_srcover_rgba_8888_sse2_lowp+0xe40>
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  15,40,21,66,96,1,0                  // movaps        0x16042(%rip),%xmm2        # 337a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe20>
+  .byte  15,40,21,34,119,1,0                 // movaps        0x17722(%rip),%xmm2        # 34e80 <_sk_srcover_rgba_8888_sse2_lowp+0xe50>
   .byte  15,88,202                           // addps         %xmm2,%xmm1
-  .byte  68,15,40,37,71,96,1,0               // movaps        0x16047(%rip),%xmm12        # 337b0 <_sk_srcover_rgba_8888_sse2_lowp+0xe30>
+  .byte  68,15,40,37,39,119,1,0              // movaps        0x17727(%rip),%xmm12        # 34e90 <_sk_srcover_rgba_8888_sse2_lowp+0xe60>
   .byte  65,15,89,204                        // mulps         %xmm12,%xmm1
   .byte  102,65,15,56,20,203                 // blendvps      %xmm0,%xmm11,%xmm1
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
@@ -37831,8 +37831,8 @@
   .byte  65,15,89,212                        // mulps         %xmm12,%xmm2
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  102,65,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm2
-  .byte  15,89,13,0,96,1,0                   // mulps         0x16000(%rip),%xmm1        # 337c0 <_sk_srcover_rgba_8888_sse2_lowp+0xe40>
-  .byte  15,89,21,9,96,1,0                   // mulps         0x16009(%rip),%xmm2        # 337d0 <_sk_srcover_rgba_8888_sse2_lowp+0xe50>
+  .byte  15,89,13,224,118,1,0                // mulps         0x176e0(%rip),%xmm1        # 34ea0 <_sk_srcover_rgba_8888_sse2_lowp+0xe70>
+  .byte  15,89,21,233,118,1,0                // mulps         0x176e9(%rip),%xmm2        # 34eb0 <_sk_srcover_rgba_8888_sse2_lowp+0xe80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  65,15,40,200                        // movaps        %xmm8,%xmm1
@@ -37851,9 +37851,9 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,39                              // jne           1d811 <_sk_load_a8_sse41+0x3f>
   .byte  102,67,15,56,49,4,16                // pmovzxbd      (%r8,%r10,1),%xmm0
-  .byte  102,15,219,5,135,92,1,0             // pand          0x15c87(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,5,103,115,1,0            // pand          0x17367(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,221,93,1,0                 // mulps         0x15ddd(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,189,116,1,0                // mulps         0x174bd(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -37893,9 +37893,9 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,39                              // jne           1d89d <_sk_load_a8_dst_sse41+0x3f>
   .byte  102,67,15,56,49,36,16               // pmovzxbd      (%r8,%r10,1),%xmm4
-  .byte  102,15,219,37,251,91,1,0            // pand          0x15bfb(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,37,219,114,1,0           // pand          0x172db(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,252                           // cvtdq2ps      %xmm4,%xmm7
-  .byte  15,89,61,81,93,1,0                  // mulps         0x15d51(%rip),%xmm7        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,61,49,116,1,0                 // mulps         0x17431(%rip),%xmm7        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
@@ -37963,7 +37963,7 @@
   .byte  102,15,58,32,192,3                  // pinsrb        $0x3,%eax,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,73,92,1,0                  // mulps         0x15c49(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,41,115,1,0                 // mulps         0x17329(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -37980,7 +37980,7 @@
   .byte  77,15,175,193                       // imul          %r9,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,210                           // movslq        %edx,%r10
-  .byte  68,15,40,5,240,90,1,0               // movaps        0x15af0(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,208,113,1,0              // movaps        0x171d0(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,89,195                        // mulps         %xmm3,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
   .byte  102,68,15,56,43,192                 // packusdw      %xmm0,%xmm8
@@ -38001,7 +38001,7 @@
   .byte  65,128,249,3                        // cmp           $0x3,%r9b
   .byte  117,221                             // jne           1d9e2 <_sk_store_a8_sse41+0x3d>
   .byte  102,71,15,58,20,68,16,2,8           // pextrb        $0x8,%xmm8,0x2(%r8,%r10,1)
-  .byte  102,68,15,56,0,5,200,93,1,0         // pshufb        0x15dc8(%rip),%xmm8        # 337e0 <_sk_srcover_rgba_8888_sse2_lowp+0xe60>
+  .byte  102,68,15,56,0,5,168,116,1,0        // pshufb        0x174a8(%rip),%xmm8        # 34ec0 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
   .byte  102,71,15,58,21,4,16,0              // pextrw        $0x0,%xmm8,(%r8,%r10,1)
   .byte  235,192                             // jmp           1d9e2 <_sk_store_a8_sse41+0x3d>
   .byte  102,71,15,58,20,4,16,0              // pextrb        $0x0,%xmm8,(%r8,%r10,1)
@@ -38020,11 +38020,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,42                              // jne           1da6e <_sk_load_g8_sse41+0x42>
   .byte  102,67,15,56,49,4,16                // pmovzxbd      (%r8,%r10,1),%xmm0
-  .byte  102,15,219,5,45,90,1,0              // pand          0x15a2d(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,5,13,113,1,0             // pand          0x1710d(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,131,91,1,0                  // mulps         0x15b83(%rip),%xmm0        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,5,99,114,1,0                  // mulps         0x17263(%rip),%xmm0        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,106,89,1,0                 // movaps        0x1596a(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,74,112,1,0                 // movaps        0x1704a(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -38062,11 +38062,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,42                              // jne           1dafd <_sk_load_g8_dst_sse41+0x42>
   .byte  102,67,15,56,49,36,16               // pmovzxbd      (%r8,%r10,1),%xmm4
-  .byte  102,15,219,37,158,89,1,0            // pand          0x1599e(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,37,126,112,1,0           // pand          0x1707e(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,244,90,1,0                 // mulps         0x15af4(%rip),%xmm4        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,37,212,113,1,0                // mulps         0x171d4(%rip),%xmm4        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,219,88,1,0                 // movaps        0x158db(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,61,187,111,1,0                // movaps        0x16fbb(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  255,224                             // jmpq          *%rax
@@ -38132,9 +38132,9 @@
   .byte  102,15,58,32,192,3                  // pinsrb        $0x3,%eax,%xmm0
   .byte  102,15,56,49,192                    // pmovzxbd      %xmm0,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,233,89,1,0                  // mulps         0x159e9(%rip),%xmm0        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,5,201,112,1,0                 // mulps         0x170c9(%rip),%xmm0        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,208,87,1,0                 // movaps        0x157d0(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,176,110,1,0                // movaps        0x16eb0(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -38153,19 +38153,19 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,80                              // jne           1dc73 <_sk_load_565_sse41+0x6b>
   .byte  102,67,15,56,51,20,80               // pmovzxwd      (%r8,%r10,2),%xmm2
-  .byte  102,15,111,5,190,89,1,0             // movdqa        0x159be(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,158,112,1,0            // movdqa        0x1709e(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,192,89,1,0                  // mulps         0x159c0(%rip),%xmm0        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,13,200,89,1,0            // movdqa        0x159c8(%rip),%xmm1        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,5,160,112,1,0                 // mulps         0x170a0(%rip),%xmm0        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,13,168,112,1,0           // movdqa        0x170a8(%rip),%xmm1        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,202,89,1,0                 // mulps         0x159ca(%rip),%xmm1        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,21,210,89,1,0            // pand          0x159d2(%rip),%xmm2        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,13,170,112,1,0                // mulps         0x170aa(%rip),%xmm1        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,21,178,112,1,0           // pand          0x170b2(%rip),%xmm2        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,216,89,1,0                 // mulps         0x159d8(%rip),%xmm2        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,21,184,112,1,0                // mulps         0x170b8(%rip),%xmm2        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,95,87,1,0                  // movaps        0x1575f(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,63,110,1,0                 // movaps        0x16e3f(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,3                        // and           $0x3,%r9b
@@ -38201,19 +38201,19 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,80                              // jne           1dd2e <_sk_load_565_dst_sse41+0x6b>
   .byte  102,67,15,56,51,52,80               // pmovzxwd      (%r8,%r10,2),%xmm6
-  .byte  102,15,111,37,3,89,1,0              // movdqa        0x15903(%rip),%xmm4        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,37,227,111,1,0           // movdqa        0x16fe3(%rip),%xmm4        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,5,89,1,0                   // mulps         0x15905(%rip),%xmm4        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,45,13,89,1,0             // movdqa        0x1590d(%rip),%xmm5        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,37,229,111,1,0                // mulps         0x16fe5(%rip),%xmm4        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,45,237,111,1,0           // movdqa        0x16fed(%rip),%xmm5        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,238                      // pand          %xmm6,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,45,15,89,1,0                  // mulps         0x1590f(%rip),%xmm5        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,53,23,89,1,0             // pand          0x15917(%rip),%xmm6        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,45,239,111,1,0                // mulps         0x16fef(%rip),%xmm5        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,53,247,111,1,0           // pand          0x16ff7(%rip),%xmm6        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,53,29,89,1,0                  // mulps         0x1591d(%rip),%xmm6        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,53,253,111,1,0                // mulps         0x16ffd(%rip),%xmm6        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,164,86,1,0                 // movaps        0x156a4(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,61,132,109,1,0                // movaps        0x16d84(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,3                        // and           $0x3,%r9b
@@ -38275,19 +38275,19 @@
   .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
   .byte  102,15,196,192,3                    // pinsrw        $0x3,%eax,%xmm0
   .byte  102,15,56,51,208                    // pmovzxwd      %xmm0,%xmm2
-  .byte  102,15,111,5,202,87,1,0             // movdqa        0x157ca(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,170,110,1,0            // movdqa        0x16eaa(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,204,87,1,0                  // mulps         0x157cc(%rip),%xmm0        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,13,212,87,1,0            // movdqa        0x157d4(%rip),%xmm1        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,5,172,110,1,0                 // mulps         0x16eac(%rip),%xmm0        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,13,180,110,1,0           // movdqa        0x16eb4(%rip),%xmm1        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,214,87,1,0                 // mulps         0x157d6(%rip),%xmm1        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,21,222,87,1,0            // pand          0x157de(%rip),%xmm2        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,13,182,110,1,0                // mulps         0x16eb6(%rip),%xmm1        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,21,190,110,1,0           // pand          0x16ebe(%rip),%xmm2        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,228,87,1,0                 // mulps         0x157e4(%rip),%xmm2        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,21,196,110,1,0                // mulps         0x16ec4(%rip),%xmm2        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,107,85,1,0                 // movaps        0x1556b(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,75,108,1,0                 // movaps        0x16c4b(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse41
@@ -38301,12 +38301,12 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,107,89,1,0               // movaps        0x1596b(%rip),%xmm8        # 337f0 <_sk_srcover_rgba_8888_sse2_lowp+0xe70>
+  .byte  68,15,40,5,75,112,1,0               // movaps        0x1704b(%rip),%xmm8        # 34ed0 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
   .byte  102,65,15,114,241,11                // pslld         $0xb,%xmm9
-  .byte  68,15,40,21,96,89,1,0               // movaps        0x15960(%rip),%xmm10        # 33800 <_sk_srcover_rgba_8888_sse2_lowp+0xe80>
+  .byte  68,15,40,21,64,112,1,0              // movaps        0x17040(%rip),%xmm10        # 34ee0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
   .byte  68,15,89,209                        // mulps         %xmm1,%xmm10
   .byte  102,69,15,91,210                    // cvtps2dq      %xmm10,%xmm10
   .byte  102,65,15,114,242,5                 // pslld         $0x5,%xmm10
@@ -38350,21 +38350,21 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,95                              // jne           1df91 <_sk_load_4444_sse41+0x7a>
   .byte  102,67,15,56,51,28,80               // pmovzxwd      (%r8,%r10,2),%xmm3
-  .byte  102,15,111,5,207,88,1,0             // movdqa        0x158cf(%rip),%xmm0        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,5,175,111,1,0            // movdqa        0x16faf(%rip),%xmm0        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,209,88,1,0                  // mulps         0x158d1(%rip),%xmm0        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,13,217,88,1,0            // movdqa        0x158d9(%rip),%xmm1        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,5,177,111,1,0                 // mulps         0x16fb1(%rip),%xmm0        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,13,185,111,1,0           // movdqa        0x16fb9(%rip),%xmm1        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,219,88,1,0                 // mulps         0x158db(%rip),%xmm1        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,21,227,88,1,0            // movdqa        0x158e3(%rip),%xmm2        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,13,187,111,1,0                // mulps         0x16fbb(%rip),%xmm1        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,21,195,111,1,0           // movdqa        0x16fc3(%rip),%xmm2        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,229,88,1,0                 // mulps         0x158e5(%rip),%xmm2        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,29,237,88,1,0            // pand          0x158ed(%rip),%xmm3        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,21,197,111,1,0                // mulps         0x16fc5(%rip),%xmm2        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,29,205,111,1,0           // pand          0x16fcd(%rip),%xmm3        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,243,88,1,0                 // mulps         0x158f3(%rip),%xmm3        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,29,211,111,1,0                // mulps         0x16fd3(%rip),%xmm3        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -38401,21 +38401,21 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,95                              // jne           1e05b <_sk_load_4444_dst_sse41+0x7a>
   .byte  102,67,15,56,51,60,80               // pmovzxwd      (%r8,%r10,2),%xmm7
-  .byte  102,15,111,37,5,88,1,0              // movdqa        0x15805(%rip),%xmm4        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,37,229,110,1,0           // movdqa        0x16ee5(%rip),%xmm4        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,231                      // pand          %xmm7,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,7,88,1,0                   // mulps         0x15807(%rip),%xmm4        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,45,15,88,1,0             // movdqa        0x1580f(%rip),%xmm5        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,37,231,110,1,0                // mulps         0x16ee7(%rip),%xmm4        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,45,239,110,1,0           // movdqa        0x16eef(%rip),%xmm5        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,239                      // pand          %xmm7,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,45,17,88,1,0                  // mulps         0x15811(%rip),%xmm5        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,53,25,88,1,0             // movdqa        0x15819(%rip),%xmm6        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,45,241,110,1,0                // mulps         0x16ef1(%rip),%xmm5        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,53,249,110,1,0           // movdqa        0x16ef9(%rip),%xmm6        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,247                      // pand          %xmm7,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,53,27,88,1,0                  // mulps         0x1581b(%rip),%xmm6        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,61,35,88,1,0             // pand          0x15823(%rip),%xmm7        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,53,251,110,1,0                // mulps         0x16efb(%rip),%xmm6        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,61,3,111,1,0             // pand          0x16f03(%rip),%xmm7        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,255                           // cvtdq2ps      %xmm7,%xmm7
-  .byte  15,89,61,41,88,1,0                  // mulps         0x15829(%rip),%xmm7        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,61,9,111,1,0                  // mulps         0x16f09(%rip),%xmm7        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -38478,21 +38478,21 @@
   .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
   .byte  102,15,196,192,3                    // pinsrw        $0x3,%eax,%xmm0
   .byte  102,15,56,51,216                    // pmovzxwd      %xmm0,%xmm3
-  .byte  102,15,111,5,189,86,1,0             // movdqa        0x156bd(%rip),%xmm0        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,5,157,109,1,0            // movdqa        0x16d9d(%rip),%xmm0        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,191,86,1,0                  // mulps         0x156bf(%rip),%xmm0        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,13,199,86,1,0            // movdqa        0x156c7(%rip),%xmm1        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,5,159,109,1,0                 // mulps         0x16d9f(%rip),%xmm0        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,13,167,109,1,0           // movdqa        0x16da7(%rip),%xmm1        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,201,86,1,0                 // mulps         0x156c9(%rip),%xmm1        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,21,209,86,1,0            // movdqa        0x156d1(%rip),%xmm2        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,13,169,109,1,0                // mulps         0x16da9(%rip),%xmm1        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,21,177,109,1,0           // movdqa        0x16db1(%rip),%xmm2        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,211,86,1,0                 // mulps         0x156d3(%rip),%xmm2        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,29,219,86,1,0            // pand          0x156db(%rip),%xmm3        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,21,179,109,1,0                // mulps         0x16db3(%rip),%xmm2        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,29,187,109,1,0           // pand          0x16dbb(%rip),%xmm3        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,225,86,1,0                 // mulps         0x156e1(%rip),%xmm3        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,29,193,109,1,0                // mulps         0x16dc1(%rip),%xmm3        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -38507,7 +38507,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,207,86,1,0               // movaps        0x156cf(%rip),%xmm8        # 33890 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  68,15,40,5,175,109,1,0              // movaps        0x16daf(%rip),%xmm8        # 34f70 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -38561,17 +38561,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,89                              // jne           1e2dc <_sk_load_8888_sse41+0x75>
   .byte  243,65,15,111,28,128                // movdqu        (%r8,%rax,4),%xmm3
-  .byte  102,15,111,5,239,81,1,0             // movdqa        0x151ef(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,207,104,1,0            // movdqa        0x168cf(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,64,83,1,0                // movaps        0x15340(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,32,106,1,0               // movaps        0x16a20(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,223,81,1,0           // pshufb        0x151df(%rip),%xmm1        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,13,191,104,1,0          // pshufb        0x168bf(%rip),%xmm1        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,219,81,1,0           // pshufb        0x151db(%rip),%xmm2        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,21,187,104,1,0          // pshufb        0x168bb(%rip),%xmm2        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -38610,17 +38610,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,89                              // jne           1e396 <_sk_load_8888_dst_sse41+0x75>
   .byte  243,65,15,111,60,128                // movdqu        (%r8,%rax,4),%xmm7
-  .byte  102,15,111,37,53,81,1,0             // movdqa        0x15135(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,37,21,104,1,0            // movdqa        0x16815(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,231                      // pand          %xmm7,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  68,15,40,5,134,82,1,0               // movaps        0x15286(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,102,105,1,0              // movaps        0x16966(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,224                        // mulps         %xmm8,%xmm4
   .byte  102,15,111,239                      // movdqa        %xmm7,%xmm5
-  .byte  102,15,56,0,45,37,81,1,0            // pshufb        0x15125(%rip),%xmm5        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,45,5,104,1,0            // pshufb        0x16805(%rip),%xmm5        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
   .byte  65,15,89,232                        // mulps         %xmm8,%xmm5
   .byte  102,15,111,247                      // movdqa        %xmm7,%xmm6
-  .byte  102,15,56,0,53,33,81,1,0            // pshufb        0x15121(%rip),%xmm6        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,53,1,104,1,0            // pshufb        0x16801(%rip),%xmm6        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  65,15,89,240                        // mulps         %xmm8,%xmm6
   .byte  102,15,114,215,24                   // psrld         $0x18,%xmm7
@@ -38680,17 +38680,17 @@
   .byte  102,65,15,58,34,28,128,1            // pinsrd        $0x1,(%r8,%rax,4),%xmm3
   .byte  102,67,15,58,34,28,152,2            // pinsrd        $0x2,(%r8,%r11,4),%xmm3
   .byte  102,67,15,58,34,28,144,3            // pinsrd        $0x3,(%r8,%r10,4),%xmm3
-  .byte  102,15,111,5,14,80,1,0              // movdqa        0x1500e(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,238,102,1,0            // movdqa        0x166ee(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,95,81,1,0                // movaps        0x1515f(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,63,104,1,0               // movaps        0x1683f(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
-  .byte  102,15,56,0,13,254,79,1,0           // pshufb        0x14ffe(%rip),%xmm1        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,13,222,102,1,0          // pshufb        0x166de(%rip),%xmm1        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,211                      // movdqa        %xmm3,%xmm2
-  .byte  102,15,56,0,21,250,79,1,0           // pshufb        0x14ffa(%rip),%xmm2        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,21,218,102,1,0          // pshufb        0x166da(%rip),%xmm2        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -38710,7 +38710,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,212,79,1,0               // movaps        0x14fd4(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,180,102,1,0              // movaps        0x166b4(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -38761,17 +38761,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,89                              // jne           1e5e3 <_sk_load_bgra_sse41+0x75>
   .byte  243,65,15,111,28,128                // movdqu        (%r8,%rax,4),%xmm3
-  .byte  102,15,111,5,232,78,1,0             // movdqa        0x14ee8(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,200,101,1,0            // movdqa        0x165c8(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,208                           // cvtdq2ps      %xmm0,%xmm2
-  .byte  68,15,40,5,57,80,1,0                // movaps        0x15039(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,25,103,1,0               // movaps        0x16719(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,15,56,0,5,216,78,1,0            // pshufb        0x14ed8(%rip),%xmm0        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,5,184,101,1,0           // pshufb        0x165b8(%rip),%xmm0        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,200                           // cvtdq2ps      %xmm0,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,15,56,0,5,212,78,1,0            // pshufb        0x14ed4(%rip),%xmm0        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,5,180,101,1,0           // pshufb        0x165b4(%rip),%xmm0        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -38810,17 +38810,17 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,89                              // jne           1e69d <_sk_load_bgra_dst_sse41+0x75>
   .byte  243,65,15,111,60,128                // movdqu        (%r8,%rax,4),%xmm7
-  .byte  102,15,111,37,46,78,1,0             // movdqa        0x14e2e(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,37,14,101,1,0            // movdqa        0x1650e(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,231                      // pand          %xmm7,%xmm4
   .byte  15,91,244                           // cvtdq2ps      %xmm4,%xmm6
-  .byte  68,15,40,5,127,79,1,0               // movaps        0x14f7f(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,95,102,1,0               // movaps        0x1665f(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,240                        // mulps         %xmm8,%xmm6
   .byte  102,15,111,231                      // movdqa        %xmm7,%xmm4
-  .byte  102,15,56,0,37,30,78,1,0            // pshufb        0x14e1e(%rip),%xmm4        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,37,254,100,1,0          // pshufb        0x164fe(%rip),%xmm4        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,236                           // cvtdq2ps      %xmm4,%xmm5
   .byte  65,15,89,232                        // mulps         %xmm8,%xmm5
   .byte  102,15,111,231                      // movdqa        %xmm7,%xmm4
-  .byte  102,15,56,0,37,26,78,1,0            // pshufb        0x14e1a(%rip),%xmm4        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,37,250,100,1,0          // pshufb        0x164fa(%rip),%xmm4        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
   .byte  65,15,89,224                        // mulps         %xmm8,%xmm4
   .byte  102,15,114,215,24                   // psrld         $0x18,%xmm7
@@ -38880,17 +38880,17 @@
   .byte  102,65,15,58,34,28,128,1            // pinsrd        $0x1,(%r8,%rax,4),%xmm3
   .byte  102,67,15,58,34,28,152,2            // pinsrd        $0x2,(%r8,%r11,4),%xmm3
   .byte  102,67,15,58,34,28,144,3            // pinsrd        $0x3,(%r8,%r10,4),%xmm3
-  .byte  102,15,111,5,7,77,1,0               // movdqa        0x14d07(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,231,99,1,0             // movdqa        0x163e7(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,208                           // cvtdq2ps      %xmm0,%xmm2
-  .byte  68,15,40,5,88,78,1,0                // movaps        0x14e58(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,56,101,1,0               // movaps        0x16538(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,15,56,0,5,247,76,1,0            // pshufb        0x14cf7(%rip),%xmm0        # 33490 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
+  .byte  102,15,56,0,5,215,99,1,0            // pshufb        0x163d7(%rip),%xmm0        # 34b70 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
   .byte  15,91,200                           // cvtdq2ps      %xmm0,%xmm1
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,15,56,0,5,243,76,1,0            // pshufb        0x14cf3(%rip),%xmm0        # 334a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  102,15,56,0,5,211,99,1,0            // pshufb        0x163d3(%rip),%xmm0        # 34b80 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
@@ -38910,7 +38910,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,205,76,1,0               // movaps        0x14ccd(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,173,99,1,0               // movaps        0x163ad(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,40,202                        // movaps        %xmm2,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -38969,19 +38969,19 @@
   .byte  102,68,15,97,224                    // punpcklwd     %xmm0,%xmm12
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
   .byte  102,65,15,56,51,212                 // pmovzxwd      %xmm12,%xmm2
-  .byte  102,68,15,111,5,210,79,1,0          // movdqa        0x14fd2(%rip),%xmm8        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,5,178,102,1,0         // movdqa        0x166b2(%rip),%xmm8        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,219,200                   // pand          %xmm8,%xmm1
-  .byte  102,68,15,111,21,208,79,1,0         // movdqa        0x14fd0(%rip),%xmm10        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,21,176,102,1,0        // movdqa        0x166b0(%rip),%xmm10        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,210                   // pand          %xmm10,%xmm2
-  .byte  102,15,111,29,211,79,1,0            // movdqa        0x14fd3(%rip),%xmm3        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,29,179,102,1,0           // movdqa        0x166b3(%rip),%xmm3        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,15,114,241,16                   // pslld         $0x10,%xmm1
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
   .byte  102,15,56,63,195                    // pmaxud        %xmm3,%xmm0
   .byte  102,15,118,194                      // pcmpeqd       %xmm2,%xmm0
   .byte  102,15,114,242,13                   // pslld         $0xd,%xmm2
   .byte  102,15,235,209                      // por           %xmm1,%xmm2
-  .byte  102,68,15,111,29,191,79,1,0         // movdqa        0x14fbf(%rip),%xmm11        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,29,159,102,1,0        // movdqa        0x1669f(%rip),%xmm11        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,211                   // paddd         %xmm11,%xmm2
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  102,65,15,112,204,238               // pshufd        $0xee,%xmm12,%xmm1
@@ -39057,19 +39057,19 @@
   .byte  102,68,15,97,228                    // punpcklwd     %xmm4,%xmm12
   .byte  102,68,15,105,204                   // punpckhwd     %xmm4,%xmm9
   .byte  102,65,15,56,51,244                 // pmovzxwd      %xmm12,%xmm6
-  .byte  102,68,15,111,5,52,78,1,0           // movdqa        0x14e34(%rip),%xmm8        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,5,20,101,1,0          // movdqa        0x16514(%rip),%xmm8        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,15,111,238                      // movdqa        %xmm6,%xmm5
   .byte  102,65,15,219,232                   // pand          %xmm8,%xmm5
-  .byte  102,68,15,111,21,50,78,1,0          // movdqa        0x14e32(%rip),%xmm10        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,21,18,101,1,0         // movdqa        0x16512(%rip),%xmm10        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,242                   // pand          %xmm10,%xmm6
-  .byte  102,15,111,61,53,78,1,0             // movdqa        0x14e35(%rip),%xmm7        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,61,21,101,1,0            // movdqa        0x16515(%rip),%xmm7        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,15,114,245,16                   // pslld         $0x10,%xmm5
   .byte  102,15,111,230                      // movdqa        %xmm6,%xmm4
   .byte  102,15,56,63,231                    // pmaxud        %xmm7,%xmm4
   .byte  102,15,118,230                      // pcmpeqd       %xmm6,%xmm4
   .byte  102,15,114,246,13                   // pslld         $0xd,%xmm6
   .byte  102,15,235,245                      // por           %xmm5,%xmm6
-  .byte  102,68,15,111,29,33,78,1,0          // movdqa        0x14e21(%rip),%xmm11        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,29,1,101,1,0          // movdqa        0x16501(%rip),%xmm11        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,243                   // paddd         %xmm11,%xmm6
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  102,65,15,112,236,238               // pshufd        $0xee,%xmm12,%xmm5
@@ -39167,19 +39167,19 @@
   .byte  102,68,15,97,226                    // punpcklwd     %xmm2,%xmm12
   .byte  102,68,15,105,202                   // punpckhwd     %xmm2,%xmm9
   .byte  102,65,15,56,51,212                 // pmovzxwd      %xmm12,%xmm2
-  .byte  102,68,15,111,5,50,76,1,0           // movdqa        0x14c32(%rip),%xmm8        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,5,18,99,1,0           // movdqa        0x16312(%rip),%xmm8        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,219,200                   // pand          %xmm8,%xmm1
-  .byte  102,68,15,111,21,48,76,1,0          // movdqa        0x14c30(%rip),%xmm10        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,21,16,99,1,0          // movdqa        0x16310(%rip),%xmm10        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,210                   // pand          %xmm10,%xmm2
-  .byte  102,15,111,29,51,76,1,0             // movdqa        0x14c33(%rip),%xmm3        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,29,19,99,1,0             // movdqa        0x16313(%rip),%xmm3        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,15,114,241,16                   // pslld         $0x10,%xmm1
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
   .byte  102,15,56,63,195                    // pmaxud        %xmm3,%xmm0
   .byte  102,15,118,194                      // pcmpeqd       %xmm2,%xmm0
   .byte  102,15,114,242,13                   // pslld         $0xd,%xmm2
   .byte  102,15,235,209                      // por           %xmm1,%xmm2
-  .byte  102,68,15,111,29,31,76,1,0          // movdqa        0x14c1f(%rip),%xmm11        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,29,255,98,1,0         // movdqa        0x162ff(%rip),%xmm11        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,211                   // paddd         %xmm11,%xmm2
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  102,65,15,112,204,238               // pshufd        $0xee,%xmm12,%xmm1
@@ -39232,20 +39232,20 @@
   .byte  73,193,224,3                        // shl           $0x3,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  102,68,15,111,13,71,75,1,0          // movdqa        0x14b47(%rip),%xmm9        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
+  .byte  102,68,15,111,13,39,98,1,0          // movdqa        0x16227(%rip),%xmm9        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
   .byte  102,68,15,111,224                   // movdqa        %xmm0,%xmm12
   .byte  102,69,15,219,225                   // pand          %xmm9,%xmm12
-  .byte  102,68,15,111,29,68,75,1,0          // movdqa        0x14b44(%rip),%xmm11        # 338f0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
+  .byte  102,68,15,111,29,36,98,1,0          // movdqa        0x16224(%rip),%xmm11        # 34fd0 <_sk_srcover_rgba_8888_sse2_lowp+0xfa0>
   .byte  102,68,15,111,232                   // movdqa        %xmm0,%xmm13
   .byte  102,69,15,219,235                   // pand          %xmm11,%xmm13
-  .byte  102,68,15,111,21,65,75,1,0          // movdqa        0x14b41(%rip),%xmm10        # 33900 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
+  .byte  102,68,15,111,21,33,98,1,0          // movdqa        0x16221(%rip),%xmm10        # 34fe0 <_sk_srcover_rgba_8888_sse2_lowp+0xfb0>
   .byte  102,65,15,114,212,16                // psrld         $0x10,%xmm12
   .byte  102,69,15,111,197                   // movdqa        %xmm13,%xmm8
   .byte  102,69,15,56,63,194                 // pmaxud        %xmm10,%xmm8
   .byte  102,69,15,118,197                   // pcmpeqd       %xmm13,%xmm8
   .byte  102,65,15,114,213,13                // psrld         $0xd,%xmm13
   .byte  102,69,15,254,236                   // paddd         %xmm12,%xmm13
-  .byte  102,68,15,111,37,39,75,1,0          // movdqa        0x14b27(%rip),%xmm12        # 33910 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
+  .byte  102,68,15,111,37,7,98,1,0           // movdqa        0x16207(%rip),%xmm12        # 34ff0 <_sk_srcover_rgba_8888_sse2_lowp+0xfc0>
   .byte  102,69,15,254,236                   // paddd         %xmm12,%xmm13
   .byte  102,69,15,219,197                   // pand          %xmm13,%xmm8
   .byte  102,68,15,56,43,192                 // packusdw      %xmm0,%xmm8
@@ -39335,7 +39335,7 @@
   .byte  102,15,235,200                      // por           %xmm0,%xmm1
   .byte  102,15,56,51,193                    // pmovzxwd      %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,239,70,1,0               // movaps        0x146ef(%rip),%xmm8        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  68,15,40,5,207,93,1,0               // movaps        0x15dcf(%rip),%xmm8        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -39409,7 +39409,7 @@
   .byte  102,15,235,200                      // por           %xmm0,%xmm1
   .byte  102,15,56,51,193                    // pmovzxwd      %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,190,69,1,0               // movaps        0x145be(%rip),%xmm8        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  68,15,40,5,158,92,1,0               // movaps        0x15c9e(%rip),%xmm8        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -39426,7 +39426,7 @@
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,213,66,1,0                 // movaps        0x142d5(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,181,89,1,0                 // movaps        0x159b5(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,20,64                 // movd          (%r8,%rax,2),%xmm2
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -39460,7 +39460,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  73,99,193                           // movslq        %r9d,%rax
-  .byte  68,15,40,21,158,71,1,0              // movaps        0x1479e(%rip),%xmm10        # 33920 <_sk_srcover_rgba_8888_sse2_lowp+0xfa0>
+  .byte  68,15,40,21,126,94,1,0              // movaps        0x15e7e(%rip),%xmm10        # 35000 <_sk_srcover_rgba_8888_sse2_lowp+0xfd0>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
@@ -39691,7 +39691,7 @@
   .byte  65,15,92,194                        // subps         %xmm10,%xmm0
   .byte  243,69,15,88,192                    // addss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
-  .byte  243,68,15,89,13,104,55,1,0          // mulss         0x13768(%rip),%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,13,24,78,1,0           // mulss         0x14e18(%rip),%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
   .byte  102,69,15,58,8,201,1                // roundps       $0x1,%xmm9,%xmm9
@@ -39716,7 +39716,7 @@
   .byte  65,15,92,202                        // subps         %xmm10,%xmm1
   .byte  243,69,15,88,192                    // addss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
-  .byte  243,68,15,89,13,15,55,1,0           // mulss         0x1370f(%rip),%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,13,191,77,1,0          // mulss         0x14dbf(%rip),%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,89,201                        // mulps         %xmm1,%xmm9
   .byte  102,69,15,58,8,201,1                // roundps       $0x1,%xmm9,%xmm9
@@ -39735,7 +39735,7 @@
 _sk_clamp_x_1_sse41:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
   .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
-  .byte  68,15,93,5,119,62,1,0               // minps         0x13e77(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,93,5,87,85,1,0                // minps         0x15557(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -39748,7 +39748,7 @@
   .byte  65,15,92,192                        // subps         %xmm8,%xmm0
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
   .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
-  .byte  68,15,93,5,84,62,1,0                // minps         0x13e54(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,93,5,52,85,1,0                // minps         0x15534(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -39757,9 +39757,9 @@
 .globl _sk_mirror_x_1_sse41
 FUNCTION(_sk_mirror_x_1_sse41)
 _sk_mirror_x_1_sse41:
-  .byte  68,15,40,5,164,62,1,0               // movaps        0x13ea4(%rip),%xmm8        # 33430 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
+  .byte  68,15,40,5,132,85,1,0               // movaps        0x15584(%rip),%xmm8        # 34b10 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,13,40,62,1,0               // movaps        0x13e28(%rip),%xmm9        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,13,8,85,1,0                // movaps        0x15508(%rip),%xmm9        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
   .byte  102,69,15,58,8,201,1                // roundps       $0x1,%xmm9,%xmm9
   .byte  69,15,88,201                        // addps         %xmm9,%xmm9
@@ -39770,7 +39770,7 @@
   .byte  68,15,92,200                        // subps         %xmm0,%xmm9
   .byte  68,15,84,200                        // andps         %xmm0,%xmm9
   .byte  69,15,95,193                        // maxps         %xmm9,%xmm8
-  .byte  68,15,93,5,5,62,1,0                 // minps         0x13e05(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,93,5,229,84,1,0               // minps         0x154e5(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -39780,10 +39780,10 @@
 FUNCTION(_sk_luminance_to_alpha_sse41)
 _sk_luminance_to_alpha_sse41:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,83,67,1,0                   // mulps         0x14353(%rip),%xmm0        # 33930 <_sk_srcover_rgba_8888_sse2_lowp+0xfb0>
-  .byte  15,89,13,92,67,1,0                  // mulps         0x1435c(%rip),%xmm1        # 33940 <_sk_srcover_rgba_8888_sse2_lowp+0xfc0>
+  .byte  15,89,5,51,90,1,0                   // mulps         0x15a33(%rip),%xmm0        # 35010 <_sk_srcover_rgba_8888_sse2_lowp+0xfe0>
+  .byte  15,89,13,60,90,1,0                  // mulps         0x15a3c(%rip),%xmm1        # 35020 <_sk_srcover_rgba_8888_sse2_lowp+0xff0>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,98,67,1,0                  // mulps         0x14362(%rip),%xmm3        # 33950 <_sk_srcover_rgba_8888_sse2_lowp+0xfd0>
+  .byte  15,89,29,66,90,1,0                  // mulps         0x15a42(%rip),%xmm3        # 35030 <_sk_srcover_rgba_8888_sse2_lowp+0x1000>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -40186,15 +40186,15 @@
 .globl _sk_gauss_a_to_rgba_sse41
 FUNCTION(_sk_gauss_a_to_rgba_sse41)
 _sk_gauss_a_to_rgba_sse41:
-  .byte  15,40,5,17,61,1,0                   // movaps        0x13d11(%rip),%xmm0        # 33960 <_sk_srcover_rgba_8888_sse2_lowp+0xfe0>
+  .byte  15,40,5,241,83,1,0                  // movaps        0x153f1(%rip),%xmm0        # 35040 <_sk_srcover_rgba_8888_sse2_lowp+0x1010>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,23,61,1,0                   // addps         0x13d17(%rip),%xmm0        # 33970 <_sk_srcover_rgba_8888_sse2_lowp+0xff0>
+  .byte  15,88,5,247,83,1,0                  // addps         0x153f7(%rip),%xmm0        # 35050 <_sk_srcover_rgba_8888_sse2_lowp+0x1020>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,29,61,1,0                   // addps         0x13d1d(%rip),%xmm0        # 33980 <_sk_srcover_rgba_8888_sse2_lowp+0x1000>
+  .byte  15,88,5,253,83,1,0                  // addps         0x153fd(%rip),%xmm0        # 35060 <_sk_srcover_rgba_8888_sse2_lowp+0x1030>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,35,61,1,0                   // addps         0x13d23(%rip),%xmm0        # 33990 <_sk_srcover_rgba_8888_sse2_lowp+0x1010>
+  .byte  15,88,5,3,84,1,0                    // addps         0x15403(%rip),%xmm0        # 35070 <_sk_srcover_rgba_8888_sse2_lowp+0x1040>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,41,61,1,0                   // addps         0x13d29(%rip),%xmm0        # 339a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1020>
+  .byte  15,88,5,9,84,1,0                    // addps         0x15409(%rip),%xmm0        # 35080 <_sk_srcover_rgba_8888_sse2_lowp+0x1050>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
@@ -40351,26 +40351,26 @@
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,40,236                        // movaps        %xmm12,%xmm13
   .byte  69,15,89,237                        // mulps         %xmm13,%xmm13
-  .byte  68,15,40,21,160,58,1,0              // movaps        0x13aa0(%rip),%xmm10        # 339b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1030>
+  .byte  68,15,40,21,128,81,1,0              // movaps        0x15180(%rip),%xmm10        # 35090 <_sk_srcover_rgba_8888_sse2_lowp+0x1060>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,164,58,1,0              // addps         0x13aa4(%rip),%xmm10        # 339c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1040>
+  .byte  68,15,88,21,132,81,1,0              // addps         0x15184(%rip),%xmm10        # 350a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1070>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,168,58,1,0              // addps         0x13aa8(%rip),%xmm10        # 339d0 <_sk_srcover_rgba_8888_sse2_lowp+0x1050>
+  .byte  68,15,88,21,136,81,1,0              // addps         0x15188(%rip),%xmm10        # 350b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
   .byte  69,15,89,213                        // mulps         %xmm13,%xmm10
-  .byte  68,15,88,21,172,58,1,0              // addps         0x13aac(%rip),%xmm10        # 339e0 <_sk_srcover_rgba_8888_sse2_lowp+0x1060>
+  .byte  68,15,88,21,140,81,1,0              // addps         0x1518c(%rip),%xmm10        # 350c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,194,195,1                     // cmpltps       %xmm11,%xmm0
-  .byte  68,15,40,29,171,58,1,0              // movaps        0x13aab(%rip),%xmm11        # 339f0 <_sk_srcover_rgba_8888_sse2_lowp+0x1070>
+  .byte  68,15,40,29,139,81,1,0              // movaps        0x1518b(%rip),%xmm11        # 350d0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  102,69,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm10
   .byte  69,15,194,200,1                     // cmpltps       %xmm8,%xmm9
-  .byte  68,15,40,29,100,52,1,0              // movaps        0x13464(%rip),%xmm11        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,29,68,75,1,0               // movaps        0x14b44(%rip),%xmm11        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  102,69,15,56,20,211                 // blendvps      %xmm0,%xmm11,%xmm10
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  65,15,194,192,1                     // cmpltps       %xmm8,%xmm0
-  .byte  68,15,40,13,86,52,1,0               // movaps        0x13456(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,54,75,1,0               // movaps        0x14b36(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  102,69,15,56,20,209                 // blendvps      %xmm0,%xmm9,%xmm10
   .byte  69,15,194,194,7                     // cmpordps      %xmm10,%xmm8
@@ -40404,7 +40404,7 @@
   .byte  243,69,15,89,203                    // mulss         %xmm11,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,31,58,1,0               // mulps         0x13a1f(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,255,80,1,0              // mulps         0x150ff(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
@@ -40412,7 +40412,7 @@
   .byte  243,69,15,89,219                    // mulss         %xmm11,%xmm11
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,92,227                        // subps         %xmm11,%xmm12
-  .byte  68,15,89,21,10,58,1,0               // mulps         0x13a0a(%rip),%xmm10        # 33a10 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
+  .byte  68,15,89,21,234,80,1,0              // mulps         0x150ea(%rip),%xmm10        # 350f0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
@@ -40421,8 +40421,8 @@
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,92,193                        // subps         %xmm9,%xmm0
-  .byte  68,15,87,13,178,56,1,0              // xorps         0x138b2(%rip),%xmm9        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
-  .byte  68,15,89,5,138,51,1,0               // mulps         0x1338a(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,87,13,146,79,1,0              // xorps         0x14f92(%rip),%xmm9        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
+  .byte  68,15,89,5,106,74,1,0               // mulps         0x14a6a(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -40443,7 +40443,7 @@
   .byte  243,69,15,89,203                    // mulss         %xmm11,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,129,57,1,0              // mulps         0x13981(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,97,80,1,0               // mulps         0x15061(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
@@ -40451,7 +40451,7 @@
   .byte  243,69,15,89,219                    // mulss         %xmm11,%xmm11
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,92,227                        // subps         %xmm11,%xmm12
-  .byte  68,15,89,21,108,57,1,0              // mulps         0x1396c(%rip),%xmm10        # 33a10 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
+  .byte  68,15,89,21,76,80,1,0               // mulps         0x1504c(%rip),%xmm10        # 350f0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
@@ -40460,8 +40460,8 @@
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,92,193                        // subps         %xmm9,%xmm0
-  .byte  68,15,87,13,20,56,1,0               // xorps         0x13814(%rip),%xmm9        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
-  .byte  68,15,89,5,236,50,1,0               // mulps         0x132ec(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,87,13,244,78,1,0              // xorps         0x14ef4(%rip),%xmm9        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
+  .byte  68,15,89,5,204,73,1,0               // mulps         0x149cc(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -40479,7 +40479,7 @@
   .byte  243,69,15,89,200                    // mulss         %xmm8,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,244,56,1,0              // mulps         0x138f4(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,212,79,1,0              // mulps         0x14fd4(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  69,15,89,210                        // mulps         %xmm10,%xmm10
@@ -40487,7 +40487,7 @@
   .byte  243,69,15,89,192                    // mulss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,92,192                        // subps         %xmm8,%xmm0
-  .byte  15,87,5,176,55,1,0                  // xorps         0x137b0(%rip),%xmm0        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
+  .byte  15,87,5,144,78,1,0                  // xorps         0x14e90(%rip),%xmm0        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
   .byte  65,15,94,193                        // divps         %xmm9,%xmm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40535,7 +40535,7 @@
 FUNCTION(_sk_save_xy_sse41)
 _sk_save_xy_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,15,50,1,0                // movaps        0x1320f(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,239,72,1,0               // movaps        0x148ef(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -40579,8 +40579,8 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,216,55,1,0                  // addps         0x137d8(%rip),%xmm0        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,128,49,1,0              // movaps        0x13180(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,184,78,1,0                  // addps         0x14eb8(%rip),%xmm0        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,96,72,1,0               // movaps        0x14860(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,0,1,0,0                // movups        %xmm9,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40593,7 +40593,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,76,49,1,0                   // addps         0x1314c(%rip),%xmm0        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,5,44,72,1,0                   // addps         0x1482c(%rip),%xmm0        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40605,8 +40605,8 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,139,55,1,0                 // addps         0x1378b(%rip),%xmm1        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,51,49,1,0               // movaps        0x13133(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,107,78,1,0                 // addps         0x14e6b(%rip),%xmm1        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,19,72,1,0               // movaps        0x14813(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,64,1,0,0               // movups        %xmm9,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40619,7 +40619,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,254,48,1,0                 // addps         0x130fe(%rip),%xmm1        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,13,222,71,1,0                 // addps         0x147de(%rip),%xmm1        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40631,13 +40631,13 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,78,55,1,0                   // addps         0x1374e(%rip),%xmm0        # 33a30 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
-  .byte  68,15,40,13,230,48,1,0              // movaps        0x130e6(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,46,78,1,0                   // addps         0x14e2e(%rip),%xmm0        # 35110 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,40,13,198,71,1,0              // movaps        0x147c6(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,66,55,1,0               // mulps         0x13742(%rip),%xmm9        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,13,202,50,1,0              // addps         0x132ca(%rip),%xmm9        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,13,34,78,1,0               // mulps         0x14e22(%rip),%xmm9        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,13,170,73,1,0              // addps         0x149aa(%rip),%xmm9        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,0,1,0,0                // movups        %xmm9,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40650,16 +40650,16 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,246,54,1,0                  // addps         0x136f6(%rip),%xmm0        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,158,48,1,0              // movaps        0x1309e(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,214,77,1,0                  // addps         0x14dd6(%rip),%xmm0        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,126,71,1,0              // movaps        0x1477e(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,18,55,1,0                // movaps        0x13712(%rip),%xmm8        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,5,242,77,1,0               // movaps        0x14df2(%rip),%xmm8        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,22,55,1,0                // addps         0x13716(%rip),%xmm8        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,5,246,77,1,0               // addps         0x14df6(%rip),%xmm8        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,106,48,1,0               // addps         0x1306a(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,88,5,74,71,1,0                // addps         0x1474a(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,14,55,1,0                // addps         0x1370e(%rip),%xmm8        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,238,77,1,0               // addps         0x14dee(%rip),%xmm8        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40669,17 +40669,17 @@
 FUNCTION(_sk_bicubic_p1x_sse41)
 _sk_bicubic_p1x_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,72,48,1,0                // movaps        0x13048(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,40,71,1,0                // movaps        0x14728(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,136,128,0,0,0              // movups        0x80(%rax),%xmm9
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,21,193,54,1,0              // movaps        0x136c1(%rip),%xmm10        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,21,161,77,1,0              // movaps        0x14da1(%rip),%xmm10        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,197,54,1,0              // addps         0x136c5(%rip),%xmm10        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,21,165,77,1,0              // addps         0x14da5(%rip),%xmm10        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,193,54,1,0              // addps         0x136c1(%rip),%xmm10        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,21,161,77,1,0              // addps         0x14da1(%rip),%xmm10        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,144,0,1,0,0                // movups        %xmm10,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40691,11 +40691,11 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,145,54,1,0                  // addps         0x13691(%rip),%xmm0        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  15,88,5,113,77,1,0                  // addps         0x14d71(%rip),%xmm0        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,97,54,1,0                // mulps         0x13661(%rip),%xmm8        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,5,233,49,1,0               // addps         0x131e9(%rip),%xmm8        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,5,65,77,1,0                // mulps         0x14d41(%rip),%xmm8        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,201,72,1,0               // addps         0x148c9(%rip),%xmm8        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40708,13 +40708,13 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,36,54,1,0                  // addps         0x13624(%rip),%xmm1        # 33a30 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
-  .byte  68,15,40,13,188,47,1,0              // movaps        0x12fbc(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,4,77,1,0                   // addps         0x14d04(%rip),%xmm1        # 35110 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,40,13,156,70,1,0              // movaps        0x1469c(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,24,54,1,0               // mulps         0x13618(%rip),%xmm9        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,13,160,49,1,0              // addps         0x131a0(%rip),%xmm9        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,13,248,76,1,0              // mulps         0x14cf8(%rip),%xmm9        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,13,128,72,1,0              // addps         0x14880(%rip),%xmm9        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,64,1,0,0               // movups        %xmm9,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40727,16 +40727,16 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,203,53,1,0                 // addps         0x135cb(%rip),%xmm1        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,115,47,1,0              // movaps        0x12f73(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,171,76,1,0                 // addps         0x14cab(%rip),%xmm1        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,83,70,1,0               // movaps        0x14653(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,231,53,1,0               // movaps        0x135e7(%rip),%xmm8        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,5,199,76,1,0               // movaps        0x14cc7(%rip),%xmm8        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,235,53,1,0               // addps         0x135eb(%rip),%xmm8        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,5,203,76,1,0               // addps         0x14ccb(%rip),%xmm8        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,63,47,1,0                // addps         0x12f3f(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,88,5,31,70,1,0                // addps         0x1461f(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,227,53,1,0               // addps         0x135e3(%rip),%xmm8        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,195,76,1,0               // addps         0x14cc3(%rip),%xmm8        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40746,17 +40746,17 @@
 FUNCTION(_sk_bicubic_p1y_sse41)
 _sk_bicubic_p1y_sse41:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,29,47,1,0                // movaps        0x12f1d(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,253,69,1,0               // movaps        0x145fd(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,136,192,0,0,0              // movups        0xc0(%rax),%xmm9
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  68,15,40,21,149,53,1,0              // movaps        0x13595(%rip),%xmm10        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,21,117,76,1,0              // movaps        0x14c75(%rip),%xmm10        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,153,53,1,0              // addps         0x13599(%rip),%xmm10        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,21,121,76,1,0              // addps         0x14c79(%rip),%xmm10        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,149,53,1,0              // addps         0x13595(%rip),%xmm10        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,21,117,76,1,0              // addps         0x14c75(%rip),%xmm10        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,144,64,1,0,0               // movups        %xmm10,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -40768,11 +40768,11 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,100,53,1,0                 // addps         0x13564(%rip),%xmm1        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  15,88,13,68,76,1,0                  // addps         0x14c44(%rip),%xmm1        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,52,53,1,0                // mulps         0x13534(%rip),%xmm8        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,5,188,48,1,0               // addps         0x130bc(%rip),%xmm8        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,5,20,76,1,0                // mulps         0x14c14(%rip),%xmm8        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,156,71,1,0               // addps         0x1479c(%rip),%xmm8        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -40883,7 +40883,7 @@
   .byte  15,89,194                           // mulps         %xmm2,%xmm0
   .byte  15,41,68,36,208                     // movaps        %xmm0,-0x30(%rsp)
   .byte  243,15,91,240                       // cvttps2dq     %xmm0,%xmm6
-  .byte  15,40,37,31,52,1,0                  // movaps        0x1341f(%rip),%xmm4        # 33a80 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
+  .byte  15,40,37,255,74,1,0                 // movaps        0x14aff(%rip),%xmm4        # 35160 <_sk_srcover_rgba_8888_sse2_lowp+0x1130>
   .byte  15,88,196                           // addps         %xmm4,%xmm0
   .byte  15,41,68,36,176                     // movaps        %xmm0,-0x50(%rsp)
   .byte  102,65,15,110,208                   // movd          %r8d,%xmm2
@@ -40922,7 +40922,7 @@
   .byte  102,68,15,56,64,192                 // pmulld        %xmm0,%xmm8
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  102,15,254,218                      // paddd         %xmm2,%xmm3
-  .byte  102,68,15,111,37,118,51,1,0         // movdqa        0x13376(%rip),%xmm12        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,68,15,111,37,86,74,1,0          // movdqa        0x14a56(%rip),%xmm12        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,65,15,56,64,220                 // pmulld        %xmm12,%xmm3
   .byte  102,15,118,228                      // pcmpeqd       %xmm4,%xmm4
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
@@ -40945,7 +40945,7 @@
   .byte  102,65,15,58,22,217,1               // pextrd        $0x1,%xmm3,%r9d
   .byte  102,65,15,58,22,218,2               // pextrd        $0x2,%xmm3,%r10d
   .byte  102,65,15,58,22,219,3               // pextrd        $0x3,%xmm3,%r11d
-  .byte  102,15,111,37,106,44,1,0            // movdqa        0x12c6a(%rip),%xmm4        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,37,74,67,1,0             // movdqa        0x1434a(%rip),%xmm4        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,220                      // paddd         %xmm4,%xmm3
   .byte  102,15,111,252                      // movdqa        %xmm4,%xmm7
   .byte  102,73,15,58,22,222,1               // pextrq        $0x1,%xmm3,%r14
@@ -41118,7 +41118,7 @@
   .byte  102,68,15,254,202                   // paddd         %xmm2,%xmm9
   .byte  102,65,15,111,192                   // movdqa        %xmm8,%xmm0
   .byte  102,65,15,254,193                   // paddd         %xmm9,%xmm0
-  .byte  102,15,111,45,133,47,1,0            // movdqa        0x12f85(%rip),%xmm5        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,45,101,70,1,0            // movdqa        0x14665(%rip),%xmm5        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,56,64,197                    // pmulld        %xmm5,%xmm0
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,118,246                      // pcmpeqd       %xmm6,%xmm6
@@ -41139,7 +41139,7 @@
   .byte  102,65,15,58,22,193,1               // pextrd        $0x1,%xmm0,%r9d
   .byte  102,65,15,58,22,194,2               // pextrd        $0x2,%xmm0,%r10d
   .byte  102,65,15,58,22,195,3               // pextrd        $0x3,%xmm0,%r11d
-  .byte  102,15,111,61,134,40,1,0            // movdqa        0x12886(%rip),%xmm7        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,61,102,63,1,0            // movdqa        0x13f66(%rip),%xmm7        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,199                      // paddd         %xmm7,%xmm0
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
   .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
@@ -41205,7 +41205,7 @@
   .byte  15,88,248                           // addps         %xmm0,%xmm7
   .byte  102,68,15,254,210                   // paddd         %xmm2,%xmm10
   .byte  102,69,15,254,194                   // paddd         %xmm10,%xmm8
-  .byte  102,15,111,13,193,45,1,0            // movdqa        0x12dc1(%rip),%xmm1        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,13,161,68,1,0            // movdqa        0x144a1(%rip),%xmm1        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,68,15,56,64,193                 // pmulld        %xmm1,%xmm8
   .byte  102,65,15,111,192                   // movdqa        %xmm8,%xmm0
   .byte  102,15,118,237                      // pcmpeqd       %xmm5,%xmm5
@@ -41226,7 +41226,7 @@
   .byte  102,69,15,58,22,193,1               // pextrd        $0x1,%xmm8,%r9d
   .byte  102,69,15,58,22,194,2               // pextrd        $0x2,%xmm8,%r10d
   .byte  102,69,15,58,22,195,3               // pextrd        $0x3,%xmm8,%r11d
-  .byte  102,15,111,21,188,38,1,0            // movdqa        0x126bc(%rip),%xmm2        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,21,156,61,1,0            // movdqa        0x13d9c(%rip),%xmm2        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,68,15,254,194                   // paddd         %xmm2,%xmm8
   .byte  102,76,15,58,22,193,1               // pextrq        $0x1,%xmm8,%rcx
   .byte  102,76,15,126,195                   // movq          %xmm8,%rbx
@@ -41353,7 +41353,7 @@
   .byte  15,89,203                           // mulps         %xmm3,%xmm1
   .byte  15,41,76,36,48                      // movaps        %xmm1,0x30(%rsp)
   .byte  243,15,91,249                       // cvttps2dq     %xmm1,%xmm7
-  .byte  15,40,29,31,43,1,0                  // movaps        0x12b1f(%rip),%xmm3        # 33a80 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
+  .byte  15,40,29,255,65,1,0                 // movaps        0x141ff(%rip),%xmm3        # 35160 <_sk_srcover_rgba_8888_sse2_lowp+0x1130>
   .byte  15,88,203                           // addps         %xmm3,%xmm1
   .byte  15,41,76,36,160                     // movaps        %xmm1,-0x60(%rsp)
   .byte  102,65,15,110,232                   // movd          %r8d,%xmm5
@@ -41410,7 +41410,7 @@
   .byte  102,68,15,56,64,253                 // pmulld        %xmm5,%xmm15
   .byte  102,65,15,111,223                   // movdqa        %xmm15,%xmm3
   .byte  102,15,254,217                      // paddd         %xmm1,%xmm3
-  .byte  102,68,15,111,29,30,42,1,0          // movdqa        0x12a1e(%rip),%xmm11        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,68,15,111,29,254,64,1,0         // movdqa        0x140fe(%rip),%xmm11        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,65,15,56,64,219                 // pmulld        %xmm11,%xmm3
   .byte  102,15,118,192                      // pcmpeqd       %xmm0,%xmm0
   .byte  102,15,111,243                      // movdqa        %xmm3,%xmm6
@@ -41432,7 +41432,7 @@
   .byte  102,65,15,58,22,217,1               // pextrd        $0x1,%xmm3,%r9d
   .byte  102,65,15,58,22,218,2               // pextrd        $0x2,%xmm3,%r10d
   .byte  102,65,15,58,22,219,3               // pextrd        $0x3,%xmm3,%r11d
-  .byte  102,68,15,111,21,22,35,1,0          // movdqa        0x12316(%rip),%xmm10        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,21,246,57,1,0         // movdqa        0x139f6(%rip),%xmm10        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,218                   // paddd         %xmm10,%xmm3
   .byte  102,73,15,58,22,222,1               // pextrq        $0x1,%xmm3,%r14
   .byte  102,72,15,126,219                   // movq          %xmm3,%rbx
@@ -41450,7 +41450,7 @@
   .byte  102,65,15,254,201                   // paddd         %xmm9,%xmm1
   .byte  102,65,15,56,64,203                 // pmulld        %xmm11,%xmm1
   .byte  102,15,111,217                      // movdqa        %xmm1,%xmm3
-  .byte  102,15,250,29,84,41,1,0             // psubd         0x12954(%rip),%xmm3        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,29,52,64,1,0             // psubd         0x14034(%rip),%xmm3        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,73,15,58,22,222,1               // pextrq        $0x1,%xmm3,%r14
   .byte  102,72,15,126,219                   // movq          %xmm3,%rbx
   .byte  65,137,223                          // mov           %ebx,%r15d
@@ -41505,7 +41505,7 @@
   .byte  102,65,15,254,199                   // paddd         %xmm15,%xmm0
   .byte  102,65,15,56,64,195                 // pmulld        %xmm11,%xmm0
   .byte  102,15,111,232                      // movdqa        %xmm0,%xmm5
-  .byte  102,15,250,45,52,40,1,0             // psubd         0x12834(%rip),%xmm5        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,45,20,63,1,0             // psubd         0x13f14(%rip),%xmm5        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,69,15,118,192                   // pcmpeqd       %xmm8,%xmm8
   .byte  102,72,15,58,22,233,1               // pextrq        $0x1,%xmm5,%rcx
   .byte  102,72,15,126,237                   // movq          %xmm5,%rbp
@@ -41630,7 +41630,7 @@
   .byte  72,193,233,32                       // shr           $0x20,%rcx
   .byte  243,15,16,36,136                    // movss         (%rax,%rcx,4),%xmm4
   .byte  102,68,15,58,33,220,48              // insertps      $0x30,%xmm4,%xmm11
-  .byte  102,15,111,61,8,31,1,0              // movdqa        0x11f08(%rip),%xmm7        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,61,232,53,1,0            // movdqa        0x135e8(%rip),%xmm7        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,223                      // paddd         %xmm7,%xmm3
   .byte  102,72,15,58,22,217,1               // pextrq        $0x1,%xmm3,%rcx
   .byte  102,72,15,126,221                   // movq          %xmm3,%rbp
@@ -41716,7 +41716,7 @@
   .byte  102,65,15,58,22,201,1               // pextrd        $0x1,%xmm1,%r9d
   .byte  102,65,15,58,22,202,2               // pextrd        $0x2,%xmm1,%r10d
   .byte  102,65,15,58,22,203,3               // pextrd        $0x3,%xmm1,%r11d
-  .byte  102,68,15,111,21,76,29,1,0          // movdqa        0x11d4c(%rip),%xmm10        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,21,44,52,1,0          // movdqa        0x1342c(%rip),%xmm10        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,202                   // paddd         %xmm10,%xmm1
   .byte  102,72,15,58,22,201,1               // pextrq        $0x1,%xmm1,%rcx
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
@@ -41812,7 +41812,7 @@
   .byte  102,15,111,124,36,192               // movdqa        -0x40(%rsp),%xmm7
   .byte  102,15,111,199                      // movdqa        %xmm7,%xmm0
   .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
-  .byte  102,15,111,21,246,33,1,0            // movdqa        0x121f6(%rip),%xmm2        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,21,214,56,1,0            // movdqa        0x138d6(%rip),%xmm2        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,56,64,194                    // pmulld        %xmm2,%xmm0
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,69,15,118,192                   // pcmpeqd       %xmm8,%xmm8
@@ -41833,7 +41833,7 @@
   .byte  102,65,15,58,22,193,1               // pextrd        $0x1,%xmm0,%r9d
   .byte  102,65,15,58,22,194,2               // pextrd        $0x2,%xmm0,%r10d
   .byte  102,65,15,58,22,195,3               // pextrd        $0x3,%xmm0,%r11d
-  .byte  102,15,111,37,245,26,1,0            // movdqa        0x11af5(%rip),%xmm4        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,37,213,49,1,0            // movdqa        0x131d5(%rip),%xmm4        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,196                      // paddd         %xmm4,%xmm0
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
   .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
@@ -41905,7 +41905,7 @@
   .byte  102,68,15,111,226                   // movdqa        %xmm2,%xmm12
   .byte  102,65,15,56,64,204                 // pmulld        %xmm12,%xmm1
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
-  .byte  102,15,250,21,33,32,1,0             // psubd         0x12021(%rip),%xmm2        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,21,1,55,1,0              // psubd         0x13701(%rip),%xmm2        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,72,15,58,22,209,1               // pextrq        $0x1,%xmm2,%rcx
   .byte  102,72,15,126,213                   // movq          %xmm2,%rbp
   .byte  137,235                             // mov           %ebp,%ebx
@@ -41939,7 +41939,7 @@
   .byte  102,65,15,111,249                   // movdqa        %xmm9,%xmm7
   .byte  102,69,15,56,64,236                 // pmulld        %xmm12,%xmm13
   .byte  102,65,15,111,205                   // movdqa        %xmm13,%xmm1
-  .byte  102,15,250,13,109,31,1,0            // psubd         0x11f6d(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,77,54,1,0             // psubd         0x1364d(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,72,15,58,22,201,1               // pextrq        $0x1,%xmm1,%rcx
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
   .byte  137,221                             // mov           %ebx,%ebp
@@ -42004,7 +42004,7 @@
   .byte  102,65,15,111,206                   // movdqa        %xmm14,%xmm1
   .byte  102,15,111,108,36,192               // movdqa        -0x40(%rsp),%xmm5
   .byte  102,15,254,205                      // paddd         %xmm5,%xmm1
-  .byte  102,15,111,37,4,30,1,0              // movdqa        0x11e04(%rip),%xmm4        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,37,228,52,1,0            // movdqa        0x134e4(%rip),%xmm4        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,56,64,204                    // pmulld        %xmm4,%xmm1
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,118,246                      // pcmpeqd       %xmm6,%xmm6
@@ -42025,7 +42025,7 @@
   .byte  102,65,15,58,22,201,1               // pextrd        $0x1,%xmm1,%r9d
   .byte  102,65,15,58,22,202,2               // pextrd        $0x2,%xmm1,%r10d
   .byte  102,65,15,58,22,203,3               // pextrd        $0x3,%xmm1,%r11d
-  .byte  102,15,111,29,5,23,1,0              // movdqa        0x11705(%rip),%xmm3        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,29,229,45,1,0            // movdqa        0x12de5(%rip),%xmm3        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
   .byte  102,72,15,58,22,201,1               // pextrq        $0x1,%xmm1,%rcx
   .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
@@ -42097,7 +42097,7 @@
   .byte  102,15,56,64,239                    // pmulld        %xmm7,%xmm5
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
-  .byte  102,15,250,5,53,28,1,0              // psubd         0x11c35(%rip),%xmm0        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,5,21,51,1,0              // psubd         0x13315(%rip),%xmm0        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
   .byte  102,72,15,126,197                   // movq          %xmm0,%rbp
   .byte  137,235                             // mov           %ebp,%ebx
@@ -42115,7 +42115,7 @@
   .byte  102,65,15,58,22,193,1               // pextrd        $0x1,%xmm0,%r9d
   .byte  102,65,15,58,22,194,2               // pextrd        $0x2,%xmm0,%r10d
   .byte  102,65,15,58,22,195,3               // pextrd        $0x3,%xmm0,%r11d
-  .byte  102,15,254,5,51,21,1,0              // paddd         0x11533(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,5,19,44,1,0              // paddd         0x12c13(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
   .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
   .byte  137,221                             // mov           %ebx,%ebp
@@ -42131,7 +42131,7 @@
   .byte  102,15,254,84,36,176                // paddd         -0x50(%rsp),%xmm2
   .byte  102,15,56,64,215                    // pmulld        %xmm7,%xmm2
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
-  .byte  102,15,250,5,124,27,1,0             // psubd         0x11b7c(%rip),%xmm0        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,5,92,50,1,0              // psubd         0x1325c(%rip),%xmm0        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,72,15,58,22,193,1               // pextrq        $0x1,%xmm0,%rcx
   .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
   .byte  137,221                             // mov           %ebx,%ebp
@@ -42149,7 +42149,7 @@
   .byte  102,65,15,58,22,199,1               // pextrd        $0x1,%xmm0,%r15d
   .byte  102,65,15,58,22,196,2               // pextrd        $0x2,%xmm0,%r12d
   .byte  102,65,15,58,22,198,3               // pextrd        $0x3,%xmm0,%r14d
-  .byte  102,15,254,5,122,20,1,0             // paddd         0x1147a(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,5,90,43,1,0              // paddd         0x12b5a(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,72,15,58,22,195,1               // pextrq        $0x1,%xmm0,%rbx
   .byte  102,72,15,126,197                   // movq          %xmm0,%rbp
   .byte  137,233                             // mov           %ebp,%ecx
@@ -42212,7 +42212,7 @@
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  65,15,88,211                        // addps         %xmm11,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,55,19,1,0                  // movaps        0x11337(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,23,42,1,0                  // movaps        0x12a17(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,139,76,36,248                    // mov           -0x8(%rsp),%rcx
   .byte  15,40,100,36,64                     // movaps        0x40(%rsp),%xmm4
   .byte  15,40,108,36,80                     // movaps        0x50(%rsp),%xmm5
@@ -42319,9 +42319,9 @@
   .byte  102,15,110,201                      // movd          %ecx,%xmm1
   .byte  102,15,112,201,0                    // pshufd        $0x0,%xmm1,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,88,13,14,18,1,0                  // addps         0x1120e(%rip),%xmm1        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,13,238,40,1,0                 // addps         0x128ee(%rip),%xmm1        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,21,21,18,1,0                  // movaps        0x11215(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,245,40,1,0                 // movaps        0x128f5(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
@@ -42336,18 +42336,18 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  102,68,15,110,194                   // movd          %edx,%xmm8
   .byte  102,69,15,112,192,0                 // pshufd        $0x0,%xmm8,%xmm8
-  .byte  102,68,15,254,5,190,17,1,0          // paddd         0x111be(%rip),%xmm8        # 333a0 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,68,15,254,5,158,40,1,0          // paddd         0x1289e(%rip),%xmm8        # 34a80 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
   .byte  102,68,15,110,201                   // movd          %ecx,%xmm9
   .byte  102,69,15,112,201,0                 // pshufd        $0x0,%xmm9,%xmm9
   .byte  102,69,15,239,200                   // pxor          %xmm8,%xmm9
-  .byte  102,68,15,111,21,229,17,1,0         // movdqa        0x111e5(%rip),%xmm10        # 333e0 <_sk_srcover_rgba_8888_sse2_lowp+0xa60>
+  .byte  102,68,15,111,21,197,40,1,0         // movdqa        0x128c5(%rip),%xmm10        # 34ac0 <_sk_srcover_rgba_8888_sse2_lowp+0xa90>
   .byte  102,69,15,111,217                   // movdqa        %xmm9,%xmm11
   .byte  102,69,15,219,218                   // pand          %xmm10,%xmm11
   .byte  102,65,15,114,243,5                 // pslld         $0x5,%xmm11
   .byte  102,69,15,219,208                   // pand          %xmm8,%xmm10
   .byte  102,65,15,114,242,4                 // pslld         $0x4,%xmm10
-  .byte  102,68,15,111,37,209,17,1,0         // movdqa        0x111d1(%rip),%xmm12        # 333f0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
-  .byte  102,68,15,111,45,216,17,1,0         // movdqa        0x111d8(%rip),%xmm13        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,37,177,40,1,0         // movdqa        0x128b1(%rip),%xmm12        # 34ad0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa0>
+  .byte  102,68,15,111,45,184,40,1,0         // movdqa        0x128b8(%rip),%xmm13        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,69,15,111,241                   // movdqa        %xmm9,%xmm14
   .byte  102,69,15,219,245                   // pand          %xmm13,%xmm14
   .byte  102,65,15,114,246,2                 // pslld         $0x2,%xmm14
@@ -42363,8 +42363,8 @@
   .byte  102,69,15,235,198                   // por           %xmm14,%xmm8
   .byte  102,69,15,235,193                   // por           %xmm9,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,147,17,1,0               // mulps         0x11193(%rip),%xmm8        # 33410 <_sk_srcover_rgba_8888_sse2_lowp+0xa90>
-  .byte  68,15,88,5,155,17,1,0               // addps         0x1119b(%rip),%xmm8        # 33420 <_sk_srcover_rgba_8888_sse2_lowp+0xaa0>
+  .byte  68,15,89,5,115,40,1,0               // mulps         0x12873(%rip),%xmm8        # 34af0 <_sk_srcover_rgba_8888_sse2_lowp+0xac0>
+  .byte  68,15,88,5,123,40,1,0               // addps         0x1287b(%rip),%xmm8        # 34b00 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
   .byte  243,68,15,16,16                     // movss         (%rax),%xmm10
   .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
@@ -42406,7 +42406,7 @@
 FUNCTION(_sk_black_color_sse2)
 _sk_black_color_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,209,16,1,0                 // movaps        0x110d1(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,177,39,1,0                 // movaps        0x127b1(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -42417,7 +42417,7 @@
 FUNCTION(_sk_white_color_sse2)
 _sk_white_color_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,5,189,16,1,0                  // movaps        0x110bd(%rip),%xmm0        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,5,157,39,1,0                  // movaps        0x1279d(%rip),%xmm0        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -42463,7 +42463,7 @@
 FUNCTION(_sk_srcatop_sse2)
 _sk_srcatop_sse2:
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  68,15,40,5,109,16,1,0               // movaps        0x1106d(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,77,39,1,0                // movaps        0x1274d(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -42488,7 +42488,7 @@
 _sk_dstatop_sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
   .byte  68,15,89,196                        // mulps         %xmm4,%xmm8
-  .byte  68,15,40,13,32,16,1,0               // movaps        0x11020(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,0,39,1,0                // movaps        0x12700(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
@@ -42535,7 +42535,7 @@
 .globl _sk_srcout_sse2
 FUNCTION(_sk_srcout_sse2)
 _sk_srcout_sse2:
-  .byte  68,15,40,5,180,15,1,0               // movaps        0x10fb4(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,148,38,1,0               // movaps        0x12694(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
@@ -42548,7 +42548,7 @@
 .globl _sk_dstout_sse2
 FUNCTION(_sk_dstout_sse2)
 _sk_dstout_sse2:
-  .byte  68,15,40,5,148,15,1,0               // movaps        0x10f94(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,116,38,1,0               // movaps        0x12674(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -42565,7 +42565,7 @@
 .globl _sk_srcover_sse2
 FUNCTION(_sk_srcover_sse2)
 _sk_srcover_sse2:
-  .byte  68,15,40,5,103,15,1,0               // movaps        0x10f67(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,71,38,1,0                // movaps        0x12647(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -42585,7 +42585,7 @@
 .globl _sk_dstover_sse2
 FUNCTION(_sk_dstover_sse2)
 _sk_dstover_sse2:
-  .byte  68,15,40,5,43,15,1,0                // movaps        0x10f2b(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,11,38,1,0                // movaps        0x1260b(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -42613,7 +42613,7 @@
 .globl _sk_multiply_sse2
 FUNCTION(_sk_multiply_sse2)
 _sk_multiply_sse2:
-  .byte  68,15,40,5,239,14,1,0               // movaps        0x10eef(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,207,37,1,0               // movaps        0x125cf(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
@@ -42651,7 +42651,7 @@
 FUNCTION(_sk_plus__sse2)
 _sk_plus__sse2:
   .byte  15,88,196                           // addps         %xmm4,%xmm0
-  .byte  68,15,40,5,112,14,1,0               // movaps        0x10e70(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,80,37,1,0                // movaps        0x12550(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  15,88,205                           // addps         %xmm5,%xmm1
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
@@ -42693,7 +42693,7 @@
 FUNCTION(_sk_xor__sse2)
 _sk_xor__sse2:
   .byte  68,15,40,195                        // movaps        %xmm3,%xmm8
-  .byte  15,40,29,252,13,1,0                 // movaps        0x10dfc(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,220,36,1,0                 // movaps        0x124dc(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,40,203                        // movaps        %xmm3,%xmm9
   .byte  68,15,92,207                        // subps         %xmm7,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
@@ -42741,7 +42741,7 @@
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,95,209                        // maxps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,87,13,1,0                  // movaps        0x10d57(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,55,36,1,0                  // movaps        0x12437(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -42775,7 +42775,7 @@
   .byte  68,15,89,206                        // mulps         %xmm6,%xmm9
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,236,12,1,0                 // movaps        0x10cec(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,204,35,1,0                 // movaps        0x123cc(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -42812,7 +42812,7 @@
   .byte  65,15,93,209                        // minps         %xmm9,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,118,12,1,0                 // movaps        0x10c76(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,86,35,1,0                  // movaps        0x12356(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -42840,7 +42840,7 @@
   .byte  15,89,214                           // mulps         %xmm6,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  68,15,92,194                        // subps         %xmm2,%xmm8
-  .byte  15,40,21,38,12,1,0                  // movaps        0x10c26(%rip),%xmm2        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,21,6,35,1,0                   // movaps        0x12306(%rip),%xmm2        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,215                           // mulps         %xmm7,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -42853,7 +42853,7 @@
 FUNCTION(_sk_colorburn_sse2)
 _sk_colorburn_sse2:
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
-  .byte  68,15,40,13,9,12,1,0                // movaps        0x10c09(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,233,34,1,0              // movaps        0x122e9(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -42948,7 +42948,7 @@
 FUNCTION(_sk_colordodge_sse2)
 _sk_colordodge_sse2:
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
-  .byte  68,15,40,21,177,10,1,0              // movaps        0x10ab1(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,145,33,1,0              // movaps        0x12191(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
   .byte  68,15,92,223                        // subps         %xmm7,%xmm11
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
@@ -43040,7 +43040,7 @@
   .byte  15,41,116,36,232                    // movaps        %xmm6,-0x18(%rsp)
   .byte  15,40,245                           // movaps        %xmm5,%xmm6
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
-  .byte  68,15,40,29,90,9,1,0                // movaps        0x1095a(%rip),%xmm11        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,29,58,32,1,0               // movaps        0x1203a(%rip),%xmm11        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,211                        // movaps        %xmm11,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,194                        // movaps        %xmm10,%xmm8
@@ -43128,7 +43128,7 @@
 _sk_overlay_sse2:
   .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
   .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
-  .byte  68,15,40,13,24,8,1,0                // movaps        0x10818(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,248,30,1,0              // movaps        0x11ef8(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  68,15,92,215                        // subps         %xmm7,%xmm10
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -43219,7 +43219,7 @@
   .byte  68,15,40,213                        // movaps        %xmm5,%xmm10
   .byte  68,15,94,215                        // divps         %xmm7,%xmm10
   .byte  69,15,84,212                        // andps         %xmm12,%xmm10
-  .byte  68,15,40,13,197,6,1,0               // movaps        0x106c5(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,165,29,1,0              // movaps        0x11da5(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,249                        // movaps        %xmm9,%xmm15
   .byte  69,15,92,250                        // subps         %xmm10,%xmm15
   .byte  69,15,40,218                        // movaps        %xmm10,%xmm11
@@ -43232,10 +43232,10 @@
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  65,15,88,194                        // addps         %xmm10,%xmm0
-  .byte  68,15,40,53,239,6,1,0               // movaps        0x106ef(%rip),%xmm14        # 33430 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
+  .byte  68,15,40,53,207,29,1,0              // movaps        0x11dcf(%rip),%xmm14        # 34b10 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
   .byte  69,15,88,222                        // addps         %xmm14,%xmm11
   .byte  68,15,89,216                        // mulps         %xmm0,%xmm11
-  .byte  68,15,40,21,239,6,1,0               // movaps        0x106ef(%rip),%xmm10        # 33440 <_sk_srcover_rgba_8888_sse2_lowp+0xac0>
+  .byte  68,15,40,21,207,29,1,0              // movaps        0x11dcf(%rip),%xmm10        # 34b20 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
   .byte  69,15,89,234                        // mulps         %xmm10,%xmm13
   .byte  69,15,88,235                        // addps         %xmm11,%xmm13
   .byte  15,88,228                           // addps         %xmm4,%xmm4
@@ -43380,7 +43380,7 @@
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  68,15,89,211                        // mulps         %xmm3,%xmm10
-  .byte  68,15,40,5,130,4,1,0                // movaps        0x10482(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,98,27,1,0                // movaps        0x11b62(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,216                        // movaps        %xmm8,%xmm11
   .byte  15,40,207                           // movaps        %xmm7,%xmm1
   .byte  68,15,92,217                        // subps         %xmm1,%xmm11
@@ -43428,12 +43428,12 @@
   .byte  69,15,84,206                        // andps         %xmm14,%xmm9
   .byte  69,15,84,214                        // andps         %xmm14,%xmm10
   .byte  65,15,84,214                        // andps         %xmm14,%xmm2
-  .byte  68,15,40,61,63,4,1,0                // movaps        0x1043f(%rip),%xmm15        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,61,31,27,1,0               // movaps        0x11b1f(%rip),%xmm15        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  65,15,89,231                        // mulps         %xmm15,%xmm4
-  .byte  15,40,5,68,4,1,0                    // movaps        0x10444(%rip),%xmm0        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  15,40,5,36,27,1,0                   // movaps        0x11b24(%rip),%xmm0        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  15,89,240                           // mulps         %xmm0,%xmm6
   .byte  15,88,244                           // addps         %xmm4,%xmm6
-  .byte  68,15,40,53,70,4,1,0                // movaps        0x10446(%rip),%xmm14        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,53,38,27,1,0               // movaps        0x11b26(%rip),%xmm14        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  68,15,40,239                        // movaps        %xmm7,%xmm13
   .byte  69,15,89,238                        // mulps         %xmm14,%xmm13
   .byte  68,15,88,238                        // addps         %xmm6,%xmm13
@@ -43611,14 +43611,14 @@
   .byte  68,15,84,211                        // andps         %xmm3,%xmm10
   .byte  68,15,84,203                        // andps         %xmm3,%xmm9
   .byte  15,84,195                           // andps         %xmm3,%xmm0
-  .byte  68,15,40,5,152,1,1,0                // movaps        0x10198(%rip),%xmm8        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,5,120,24,1,0               // movaps        0x11878(%rip),%xmm8        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  15,40,214                           // movaps        %xmm6,%xmm2
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
-  .byte  15,40,13,154,1,1,0                  // movaps        0x1019a(%rip),%xmm1        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  15,40,13,122,24,1,0                 // movaps        0x1187a(%rip),%xmm1        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  15,40,221                           // movaps        %xmm5,%xmm3
   .byte  15,89,217                           // mulps         %xmm1,%xmm3
   .byte  15,88,218                           // addps         %xmm2,%xmm3
-  .byte  68,15,40,37,153,1,1,0               // movaps        0x10199(%rip),%xmm12        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,37,121,24,1,0              // movaps        0x11879(%rip),%xmm12        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
   .byte  68,15,88,235                        // addps         %xmm3,%xmm13
   .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
@@ -43663,7 +43663,7 @@
   .byte  15,40,223                           // movaps        %xmm7,%xmm3
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,221                           // mulps         %xmm5,%xmm3
-  .byte  68,15,40,5,78,0,1,0                 // movaps        0x1004e(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,46,23,1,0                // movaps        0x1172e(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  68,15,92,199                        // subps         %xmm7,%xmm8
   .byte  15,88,253                           // addps         %xmm5,%xmm7
@@ -43764,14 +43764,14 @@
   .byte  68,15,40,213                        // movaps        %xmm5,%xmm10
   .byte  69,15,89,208                        // mulps         %xmm8,%xmm10
   .byte  65,15,40,208                        // movaps        %xmm8,%xmm2
-  .byte  68,15,40,45,92,255,0,0              // movaps        0xff5c(%rip),%xmm13        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,45,60,22,1,0               // movaps        0x1163c(%rip),%xmm13        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,40,198                        // movaps        %xmm6,%xmm8
   .byte  69,15,89,197                        // mulps         %xmm13,%xmm8
-  .byte  68,15,40,53,92,255,0,0              // movaps        0xff5c(%rip),%xmm14        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  68,15,40,53,60,22,1,0               // movaps        0x1163c(%rip),%xmm14        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  65,15,40,195                        // movaps        %xmm11,%xmm0
   .byte  65,15,89,198                        // mulps         %xmm14,%xmm0
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,29,88,255,0,0              // movaps        0xff58(%rip),%xmm11        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
+  .byte  68,15,40,29,56,22,1,0               // movaps        0x11638(%rip),%xmm11        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
   .byte  69,15,89,227                        // mulps         %xmm11,%xmm12
   .byte  68,15,88,224                        // addps         %xmm0,%xmm12
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
@@ -43779,7 +43779,7 @@
   .byte  69,15,40,250                        // movaps        %xmm10,%xmm15
   .byte  69,15,89,254                        // mulps         %xmm14,%xmm15
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
-  .byte  68,15,40,5,148,254,0,0              // movaps        0xfe94(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,116,21,1,0               // movaps        0x11574(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,40,224                        // movaps        %xmm8,%xmm4
   .byte  15,92,226                           // subps         %xmm2,%xmm4
   .byte  15,89,252                           // mulps         %xmm4,%xmm7
@@ -43915,15 +43915,15 @@
   .byte  68,15,40,205                        // movaps        %xmm5,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
   .byte  15,89,222                           // mulps         %xmm6,%xmm3
-  .byte  68,15,40,37,27,253,0,0              // movaps        0xfd1b(%rip),%xmm12        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,37,251,19,1,0              // movaps        0x113fb(%rip),%xmm12        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,40,199                        // movaps        %xmm7,%xmm8
   .byte  69,15,89,196                        // mulps         %xmm12,%xmm8
-  .byte  68,15,40,45,27,253,0,0              // movaps        0xfd1b(%rip),%xmm13        # 33460 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
+  .byte  68,15,40,45,251,19,1,0              // movaps        0x113fb(%rip),%xmm13        # 34b40 <_sk_srcover_rgba_8888_sse2_lowp+0xb10>
   .byte  68,15,40,241                        // movaps        %xmm1,%xmm14
   .byte  69,15,89,245                        // mulps         %xmm13,%xmm14
   .byte  69,15,88,240                        // addps         %xmm8,%xmm14
-  .byte  68,15,40,29,23,253,0,0              // movaps        0xfd17(%rip),%xmm11        # 33470 <_sk_srcover_rgba_8888_sse2_lowp+0xaf0>
-  .byte  68,15,40,5,111,252,0,0              // movaps        0xfc6f(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,29,247,19,1,0              // movaps        0x113f7(%rip),%xmm11        # 34b50 <_sk_srcover_rgba_8888_sse2_lowp+0xb20>
+  .byte  68,15,40,5,79,19,1,0                // movaps        0x1134f(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,248                        // movaps        %xmm8,%xmm15
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  68,15,92,248                        // subps         %xmm0,%xmm15
@@ -44067,7 +44067,7 @@
   .byte  15,133,228,0,0,0                    // jne           23a4c <_sk_srcover_rgba_8888_sse2+0x104>
   .byte  243,69,15,111,4,128                 // movdqu        (%r8,%rax,4),%xmm8
   .byte  72,133,255                          // test          %rdi,%rdi
-  .byte  102,15,111,53,7,251,0,0             // movdqa        0xfb07(%rip),%xmm6        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,53,231,17,1,0            // movdqa        0x111e7(%rip),%xmm6        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,224                   // movdqa        %xmm8,%xmm4
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
@@ -44081,9 +44081,9 @@
   .byte  15,91,247                           // cvtdq2ps      %xmm7,%xmm6
   .byte  102,65,15,114,208,24                // psrld         $0x18,%xmm8
   .byte  65,15,91,248                        // cvtdq2ps      %xmm8,%xmm7
-  .byte  68,15,40,5,23,250,0,0               // movaps        0xfa17(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,247,16,1,0               // movaps        0x110f7(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,92,195                        // subps         %xmm3,%xmm8
-  .byte  68,15,40,37,235,250,0,0             // movaps        0xfaeb(%rip),%xmm12        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,37,203,17,1,0              // movaps        0x111cb(%rip),%xmm12        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,196                        // mulps         %xmm12,%xmm0
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,89,204                        // mulps         %xmm4,%xmm9
@@ -44163,7 +44163,7 @@
 .globl _sk_clamp_1_sse2
 FUNCTION(_sk_clamp_1_sse2)
 _sk_clamp_1_sse2:
-  .byte  68,15,40,5,231,248,0,0              // movaps        0xf8e7(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,199,15,1,0               // movaps        0x10fc7(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,93,192                        // minps         %xmm8,%xmm0
   .byte  65,15,93,200                        // minps         %xmm8,%xmm1
   .byte  65,15,93,208                        // minps         %xmm8,%xmm2
@@ -44175,7 +44175,7 @@
 .globl _sk_clamp_a_sse2
 FUNCTION(_sk_clamp_a_sse2)
 _sk_clamp_a_sse2:
-  .byte  15,93,29,204,248,0,0                // minps         0xf8cc(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,29,172,15,1,0                 // minps         0x10fac(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -44186,7 +44186,7 @@
 .globl _sk_clamp_a_dst_sse2
 FUNCTION(_sk_clamp_a_dst_sse2)
 _sk_clamp_a_dst_sse2:
-  .byte  15,93,61,184,248,0,0                // minps         0xf8b8(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,93,61,152,15,1,0                 // minps         0x10f98(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,93,231                           // minps         %xmm7,%xmm4
   .byte  15,93,239                           // minps         %xmm7,%xmm5
   .byte  15,93,247                           // minps         %xmm7,%xmm6
@@ -44221,7 +44221,7 @@
 .globl _sk_invert_sse2
 FUNCTION(_sk_invert_sse2)
 _sk_invert_sse2:
-  .byte  68,15,40,5,116,248,0,0              // movaps        0xf874(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,84,15,1,0                // movaps        0x10f54(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  68,15,92,200                        // subps         %xmm0,%xmm9
   .byte  69,15,40,208                        // movaps        %xmm8,%xmm10
@@ -44282,10 +44282,10 @@
 .globl _sk_unpremul_sse2
 FUNCTION(_sk_unpremul_sse2)
 _sk_unpremul_sse2:
-  .byte  68,15,40,5,2,248,0,0                // movaps        0xf802(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,5,226,14,1,0               // movaps        0x10ee2(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,94,195                        // divps         %xmm3,%xmm8
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
-  .byte  68,15,194,13,225,248,0,0,1          // cmpltps       0xf8e1(%rip),%xmm9        # 334c0 <_sk_srcover_rgba_8888_sse2_lowp+0xb40>
+  .byte  68,15,194,13,193,15,1,0,1           // cmpltps       0x10fc1(%rip),%xmm9        # 34ba0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
   .byte  69,15,84,200                        // andps         %xmm8,%xmm9
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  65,15,89,201                        // mulps         %xmm9,%xmm1
@@ -44297,20 +44297,20 @@
 .globl _sk_from_srgb_sse2
 FUNCTION(_sk_from_srgb_sse2)
 _sk_from_srgb_sse2:
-  .byte  68,15,40,5,213,248,0,0              // movaps        0xf8d5(%rip),%xmm8        # 334d0 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
+  .byte  68,15,40,5,181,15,1,0               // movaps        0x10fb5(%rip),%xmm8        # 34bb0 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
   .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
   .byte  69,15,89,232                        // mulps         %xmm8,%xmm13
   .byte  68,15,40,216                        // movaps        %xmm0,%xmm11
   .byte  69,15,89,219                        // mulps         %xmm11,%xmm11
-  .byte  68,15,40,13,61,248,0,0              // movaps        0xf83d(%rip),%xmm9        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,13,29,15,1,0               // movaps        0x10f1d(%rip),%xmm9        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,241                        // mulps         %xmm9,%xmm14
-  .byte  68,15,40,21,189,248,0,0             // movaps        0xf8bd(%rip),%xmm10        # 334e0 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
+  .byte  68,15,40,21,157,15,1,0              // movaps        0x10f9d(%rip),%xmm10        # 34bc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  69,15,89,243                        // mulps         %xmm11,%xmm14
-  .byte  68,15,40,29,189,248,0,0             // movaps        0xf8bd(%rip),%xmm11        # 334f0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
+  .byte  68,15,40,29,157,15,1,0              // movaps        0x10f9d(%rip),%xmm11        # 34bd0 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
   .byte  69,15,88,243                        // addps         %xmm11,%xmm14
-  .byte  68,15,40,37,193,248,0,0             // movaps        0xf8c1(%rip),%xmm12        # 33500 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
+  .byte  68,15,40,37,161,15,1,0              // movaps        0x10fa1(%rip),%xmm12        # 34be0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
   .byte  65,15,194,196,1                     // cmpltps       %xmm12,%xmm0
   .byte  68,15,84,232                        // andps         %xmm0,%xmm13
   .byte  65,15,85,198                        // andnps        %xmm14,%xmm0
@@ -44346,20 +44346,20 @@
 .globl _sk_from_srgb_dst_sse2
 FUNCTION(_sk_from_srgb_dst_sse2)
 _sk_from_srgb_dst_sse2:
-  .byte  68,15,40,5,18,248,0,0               // movaps        0xf812(%rip),%xmm8        # 334d0 <_sk_srcover_rgba_8888_sse2_lowp+0xb50>
+  .byte  68,15,40,5,242,14,1,0               // movaps        0x10ef2(%rip),%xmm8        # 34bb0 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
   .byte  68,15,40,236                        // movaps        %xmm4,%xmm13
   .byte  69,15,89,232                        // mulps         %xmm8,%xmm13
   .byte  68,15,40,220                        // movaps        %xmm4,%xmm11
   .byte  69,15,89,219                        // mulps         %xmm11,%xmm11
-  .byte  68,15,40,13,122,247,0,0             // movaps        0xf77a(%rip),%xmm9        # 33450 <_sk_srcover_rgba_8888_sse2_lowp+0xad0>
+  .byte  68,15,40,13,90,14,1,0               // movaps        0x10e5a(%rip),%xmm9        # 34b30 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
   .byte  68,15,40,244                        // movaps        %xmm4,%xmm14
   .byte  69,15,89,241                        // mulps         %xmm9,%xmm14
-  .byte  68,15,40,21,250,247,0,0             // movaps        0xf7fa(%rip),%xmm10        # 334e0 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
+  .byte  68,15,40,21,218,14,1,0              // movaps        0x10eda(%rip),%xmm10        # 34bc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
   .byte  69,15,88,242                        // addps         %xmm10,%xmm14
   .byte  69,15,89,243                        // mulps         %xmm11,%xmm14
-  .byte  68,15,40,29,250,247,0,0             // movaps        0xf7fa(%rip),%xmm11        # 334f0 <_sk_srcover_rgba_8888_sse2_lowp+0xb70>
+  .byte  68,15,40,29,218,14,1,0              // movaps        0x10eda(%rip),%xmm11        # 34bd0 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
   .byte  69,15,88,243                        // addps         %xmm11,%xmm14
-  .byte  68,15,40,37,254,247,0,0             // movaps        0xf7fe(%rip),%xmm12        # 33500 <_sk_srcover_rgba_8888_sse2_lowp+0xb80>
+  .byte  68,15,40,37,222,14,1,0              // movaps        0x10ede(%rip),%xmm12        # 34be0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
   .byte  65,15,194,228,1                     // cmpltps       %xmm12,%xmm4
   .byte  68,15,84,236                        // andps         %xmm4,%xmm13
   .byte  65,15,85,230                        // andnps        %xmm14,%xmm4
@@ -44396,22 +44396,22 @@
 FUNCTION(_sk_to_srgb_sse2)
 _sk_to_srgb_sse2:
   .byte  68,15,82,232                        // rsqrtps       %xmm0,%xmm13
-  .byte  68,15,40,5,139,247,0,0              // movaps        0xf78b(%rip),%xmm8        # 33510 <_sk_srcover_rgba_8888_sse2_lowp+0xb90>
+  .byte  68,15,40,5,107,14,1,0               // movaps        0x10e6b(%rip),%xmm8        # 34bf0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc0>
   .byte  68,15,40,240                        // movaps        %xmm0,%xmm14
   .byte  69,15,89,240                        // mulps         %xmm8,%xmm14
-  .byte  68,15,40,13,139,247,0,0             // movaps        0xf78b(%rip),%xmm9        # 33520 <_sk_srcover_rgba_8888_sse2_lowp+0xba0>
+  .byte  68,15,40,13,107,14,1,0              // movaps        0x10e6b(%rip),%xmm9        # 34c00 <_sk_srcover_rgba_8888_sse2_lowp+0xbd0>
   .byte  69,15,40,253                        // movaps        %xmm13,%xmm15
   .byte  69,15,89,249                        // mulps         %xmm9,%xmm15
-  .byte  68,15,40,21,139,247,0,0             // movaps        0xf78b(%rip),%xmm10        # 33530 <_sk_srcover_rgba_8888_sse2_lowp+0xbb0>
+  .byte  68,15,40,21,107,14,1,0              // movaps        0x10e6b(%rip),%xmm10        # 34c10 <_sk_srcover_rgba_8888_sse2_lowp+0xbe0>
   .byte  69,15,88,250                        // addps         %xmm10,%xmm15
   .byte  69,15,89,253                        // mulps         %xmm13,%xmm15
-  .byte  68,15,40,29,139,247,0,0             // movaps        0xf78b(%rip),%xmm11        # 33540 <_sk_srcover_rgba_8888_sse2_lowp+0xbc0>
+  .byte  68,15,40,29,107,14,1,0              // movaps        0x10e6b(%rip),%xmm11        # 34c20 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
   .byte  69,15,88,251                        // addps         %xmm11,%xmm15
-  .byte  68,15,40,37,143,247,0,0             // movaps        0xf78f(%rip),%xmm12        # 33550 <_sk_srcover_rgba_8888_sse2_lowp+0xbd0>
+  .byte  68,15,40,37,111,14,1,0              // movaps        0x10e6f(%rip),%xmm12        # 34c30 <_sk_srcover_rgba_8888_sse2_lowp+0xc00>
   .byte  69,15,88,236                        // addps         %xmm12,%xmm13
   .byte  69,15,83,237                        // rcpps         %xmm13,%xmm13
   .byte  69,15,89,239                        // mulps         %xmm15,%xmm13
-  .byte  68,15,40,61,139,247,0,0             // movaps        0xf78b(%rip),%xmm15        # 33560 <_sk_srcover_rgba_8888_sse2_lowp+0xbe0>
+  .byte  68,15,40,61,107,14,1,0              // movaps        0x10e6b(%rip),%xmm15        # 34c40 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
   .byte  65,15,194,199,1                     // cmpltps       %xmm15,%xmm0
   .byte  68,15,84,240                        // andps         %xmm0,%xmm14
   .byte  65,15,85,197                        // andnps        %xmm13,%xmm0
@@ -44463,7 +44463,7 @@
   .byte  68,15,93,216                        // minps         %xmm0,%xmm11
   .byte  65,15,40,202                        // movaps        %xmm10,%xmm1
   .byte  65,15,92,203                        // subps         %xmm11,%xmm1
-  .byte  68,15,40,45,62,245,0,0              // movaps        0xf53e(%rip),%xmm13        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,45,30,12,1,0               // movaps        0x10c1e(%rip),%xmm13        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  68,15,94,233                        // divps         %xmm1,%xmm13
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,194,192,0                     // cmpeqps       %xmm8,%xmm0
@@ -44472,30 +44472,30 @@
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,40,241                        // movaps        %xmm9,%xmm14
   .byte  68,15,194,242,1                     // cmpltps       %xmm2,%xmm14
-  .byte  68,15,84,53,180,246,0,0             // andps         0xf6b4(%rip),%xmm14        # 33570 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
+  .byte  68,15,84,53,148,13,1,0              // andps         0x10d94(%rip),%xmm14        # 34c50 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
   .byte  69,15,88,244                        // addps         %xmm12,%xmm14
   .byte  69,15,40,250                        // movaps        %xmm10,%xmm15
   .byte  69,15,194,249,0                     // cmpeqps       %xmm9,%xmm15
   .byte  65,15,92,208                        // subps         %xmm8,%xmm2
   .byte  65,15,89,213                        // mulps         %xmm13,%xmm2
-  .byte  68,15,40,37,167,246,0,0             // movaps        0xf6a7(%rip),%xmm12        # 33580 <_sk_srcover_rgba_8888_sse2_lowp+0xc00>
+  .byte  68,15,40,37,135,13,1,0              // movaps        0x10d87(%rip),%xmm12        # 34c60 <_sk_srcover_rgba_8888_sse2_lowp+0xc30>
   .byte  65,15,88,212                        // addps         %xmm12,%xmm2
   .byte  69,15,92,193                        // subps         %xmm9,%xmm8
   .byte  69,15,89,197                        // mulps         %xmm13,%xmm8
-  .byte  68,15,88,5,163,246,0,0              // addps         0xf6a3(%rip),%xmm8        # 33590 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
+  .byte  68,15,88,5,131,13,1,0               // addps         0x10d83(%rip),%xmm8        # 34c70 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
   .byte  65,15,84,215                        // andps         %xmm15,%xmm2
   .byte  69,15,85,248                        // andnps        %xmm8,%xmm15
   .byte  68,15,86,250                        // orps          %xmm2,%xmm15
   .byte  68,15,84,240                        // andps         %xmm0,%xmm14
   .byte  65,15,85,199                        // andnps        %xmm15,%xmm0
   .byte  65,15,86,198                        // orps          %xmm14,%xmm0
-  .byte  15,89,5,148,246,0,0                 // mulps         0xf694(%rip),%xmm0        # 335a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
+  .byte  15,89,5,116,13,1,0                  // mulps         0x10d74(%rip),%xmm0        # 34c80 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
   .byte  69,15,40,194                        // movaps        %xmm10,%xmm8
   .byte  69,15,194,195,4                     // cmpneqps      %xmm11,%xmm8
   .byte  65,15,84,192                        // andps         %xmm8,%xmm0
   .byte  69,15,92,226                        // subps         %xmm10,%xmm12
   .byte  69,15,88,211                        // addps         %xmm11,%xmm10
-  .byte  68,15,40,13,151,244,0,0             // movaps        0xf497(%rip),%xmm9        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,13,119,11,1,0              // movaps        0x10b77(%rip),%xmm9        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  68,15,194,202,1                     // cmpltps       %xmm2,%xmm9
@@ -44519,7 +44519,7 @@
   .byte  15,41,92,36,168                     // movaps        %xmm3,-0x58(%rsp)
   .byte  68,15,40,218                        // movaps        %xmm2,%xmm11
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
-  .byte  68,15,40,13,70,244,0,0              // movaps        0xf446(%rip),%xmm9        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,13,38,11,1,0               // movaps        0x10b26(%rip),%xmm9        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,40,209                        // movaps        %xmm9,%xmm10
   .byte  69,15,194,211,2                     // cmpleps       %xmm11,%xmm10
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
@@ -44536,28 +44536,28 @@
   .byte  69,15,88,211                        // addps         %xmm11,%xmm10
   .byte  69,15,88,219                        // addps         %xmm11,%xmm11
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
-  .byte  15,40,5,239,245,0,0                 // movaps        0xf5ef(%rip),%xmm0        # 335b0 <_sk_srcover_rgba_8888_sse2_lowp+0xc30>
+  .byte  15,40,5,207,12,1,0                  // movaps        0x10ccf(%rip),%xmm0        # 34c90 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
   .byte  15,88,198                           // addps         %xmm6,%xmm0
   .byte  243,15,91,200                       // cvttps2dq     %xmm0,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
   .byte  15,194,217,1                        // cmpltps       %xmm1,%xmm3
-  .byte  15,84,29,247,243,0,0                // andps         0xf3f7(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,84,29,215,10,1,0                 // andps         0x10ad7(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,92,193                           // subps         %xmm1,%xmm0
-  .byte  68,15,40,45,217,245,0,0             // movaps        0xf5d9(%rip),%xmm13        # 335c0 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
+  .byte  68,15,40,45,185,12,1,0              // movaps        0x10cb9(%rip),%xmm13        # 34ca0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
   .byte  69,15,40,197                        // movaps        %xmm13,%xmm8
   .byte  68,15,194,192,2                     // cmpleps       %xmm0,%xmm8
   .byte  69,15,40,242                        // movaps        %xmm10,%xmm14
   .byte  69,15,92,243                        // subps         %xmm11,%xmm14
   .byte  65,15,40,217                        // movaps        %xmm9,%xmm3
   .byte  15,194,216,2                        // cmpleps       %xmm0,%xmm3
-  .byte  15,40,21,153,245,0,0                // movaps        0xf599(%rip),%xmm2        # 335a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
+  .byte  15,40,21,121,12,1,0                 // movaps        0x10c79(%rip),%xmm2        # 34c80 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
   .byte  68,15,40,250                        // movaps        %xmm2,%xmm15
   .byte  68,15,194,248,2                     // cmpleps       %xmm0,%xmm15
-  .byte  15,40,13,89,245,0,0                 // movaps        0xf559(%rip),%xmm1        # 33570 <_sk_srcover_rgba_8888_sse2_lowp+0xbf0>
+  .byte  15,40,13,57,12,1,0                  // movaps        0x10c39(%rip),%xmm1        # 34c50 <_sk_srcover_rgba_8888_sse2_lowp+0xc20>
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,45,111,245,0,0                // movaps        0xf56f(%rip),%xmm5        # 33590 <_sk_srcover_rgba_8888_sse2_lowp+0xc10>
+  .byte  15,40,45,79,12,1,0                  // movaps        0x10c4f(%rip),%xmm5        # 34c70 <_sk_srcover_rgba_8888_sse2_lowp+0xc40>
   .byte  15,40,229                           // movaps        %xmm5,%xmm4
   .byte  15,92,224                           // subps         %xmm0,%xmm4
   .byte  65,15,89,230                        // mulps         %xmm14,%xmm4
@@ -44580,7 +44580,7 @@
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,222                           // movaps        %xmm6,%xmm3
   .byte  15,194,216,1                        // cmpltps       %xmm0,%xmm3
-  .byte  15,84,29,84,243,0,0                 // andps         0xf354(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,84,29,52,10,1,0                  // andps         0x10a34(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  68,15,40,230                        // movaps        %xmm6,%xmm12
   .byte  68,15,92,224                        // subps         %xmm0,%xmm12
@@ -44610,12 +44610,12 @@
   .byte  15,40,124,36,136                    // movaps        -0x78(%rsp),%xmm7
   .byte  15,40,231                           // movaps        %xmm7,%xmm4
   .byte  15,85,227                           // andnps        %xmm3,%xmm4
-  .byte  15,88,53,220,244,0,0                // addps         0xf4dc(%rip),%xmm6        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  15,88,53,188,11,1,0                 // addps         0x10bbc(%rip),%xmm6        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  243,15,91,198                       // cvttps2dq     %xmm6,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,222                           // movaps        %xmm6,%xmm3
   .byte  15,194,216,1                        // cmpltps       %xmm0,%xmm3
-  .byte  15,84,29,199,242,0,0                // andps         0xf2c7(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,84,29,167,9,1,0                  // andps         0x109a7(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,92,195                           // subps         %xmm3,%xmm0
   .byte  15,92,240                           // subps         %xmm0,%xmm6
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
@@ -44683,9 +44683,9 @@
   .byte  102,71,15,110,4,16                  // movd          (%r8,%r10,1),%xmm8
   .byte  102,68,15,96,192                    // punpcklbw     %xmm0,%xmm8
   .byte  102,68,15,97,192                    // punpcklwd     %xmm0,%xmm8
-  .byte  102,68,15,219,5,154,242,0,0         // pand          0xf29a(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,219,5,122,9,1,0           // pand          0x1097a(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,238,243,0,0              // mulps         0xf3ee(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,89,5,206,10,1,0               // mulps         0x10ace(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
@@ -44730,17 +44730,17 @@
   .byte  15,133,173,0,0,0                    // jne           2432f <_sk_scale_565_sse2+0xcc>
   .byte  243,71,15,126,20,80                 // movq          (%r8,%r10,2),%xmm10
   .byte  102,68,15,97,208                    // punpcklwd     %xmm0,%xmm10
-  .byte  102,68,15,111,5,90,243,0,0          // movdqa        0xf35a(%rip),%xmm8        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,68,15,111,5,58,10,1,0           // movdqa        0x10a3a(%rip),%xmm8        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,69,15,219,194                   // pand          %xmm10,%xmm8
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,89,243,0,0               // mulps         0xf359(%rip),%xmm8        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,68,15,111,13,96,243,0,0         // movdqa        0xf360(%rip),%xmm9        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  68,15,89,5,57,10,1,0                // mulps         0x10a39(%rip),%xmm8        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,68,15,111,13,64,10,1,0          // movdqa        0x10a40(%rip),%xmm9        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,69,15,219,202                   // pand          %xmm10,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,95,243,0,0              // mulps         0xf35f(%rip),%xmm9        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,68,15,219,21,102,243,0,0        // pand          0xf366(%rip),%xmm10        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  68,15,89,13,63,10,1,0               // mulps         0x10a3f(%rip),%xmm9        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,219,21,70,10,1,0          // pand          0x10a46(%rip),%xmm10        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
-  .byte  68,15,89,21,106,243,0,0             // mulps         0xf36a(%rip),%xmm10        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  68,15,89,21,74,10,1,0               // mulps         0x10a4a(%rip),%xmm10        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  68,15,40,219                        // movaps        %xmm3,%xmm11
   .byte  68,15,194,223,1                     // cmpltps       %xmm7,%xmm11
   .byte  69,15,40,225                        // movaps        %xmm9,%xmm12
@@ -44821,9 +44821,9 @@
   .byte  102,71,15,110,4,16                  // movd          (%r8,%r10,1),%xmm8
   .byte  102,68,15,96,192                    // punpcklbw     %xmm0,%xmm8
   .byte  102,68,15,97,192                    // punpcklwd     %xmm0,%xmm8
-  .byte  102,68,15,219,5,144,240,0,0         // pand          0xf090(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,219,5,112,7,1,0           // pand          0x10770(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,228,241,0,0              // mulps         0xf1e4(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,89,5,196,8,1,0                // mulps         0x108c4(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  15,88,196                           // addps         %xmm4,%xmm0
@@ -44875,17 +44875,17 @@
   .byte  15,133,173,0,0,0                    // jne           24550 <_sk_lerp_565_sse2+0xcc>
   .byte  243,71,15,126,4,80                  // movq          (%r8,%r10,2),%xmm8
   .byte  102,68,15,97,192                    // punpcklwd     %xmm0,%xmm8
-  .byte  102,68,15,111,13,57,241,0,0         // movdqa        0xf139(%rip),%xmm9        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,68,15,111,13,25,8,1,0           // movdqa        0x10819(%rip),%xmm9        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,69,15,219,200                   // pand          %xmm8,%xmm9
   .byte  69,15,91,201                        // cvtdq2ps      %xmm9,%xmm9
-  .byte  68,15,89,13,56,241,0,0              // mulps         0xf138(%rip),%xmm9        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,68,15,111,21,63,241,0,0         // movdqa        0xf13f(%rip),%xmm10        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  68,15,89,13,24,8,1,0                // mulps         0x10818(%rip),%xmm9        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,68,15,111,21,31,8,1,0           // movdqa        0x1081f(%rip),%xmm10        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,69,15,219,208                   // pand          %xmm8,%xmm10
   .byte  69,15,91,218                        // cvtdq2ps      %xmm10,%xmm11
-  .byte  68,15,89,29,62,241,0,0              // mulps         0xf13e(%rip),%xmm11        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,68,15,219,5,69,241,0,0          // pand          0xf145(%rip),%xmm8        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  68,15,89,29,30,8,1,0                // mulps         0x1081e(%rip),%xmm11        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,219,5,37,8,1,0            // pand          0x10825(%rip),%xmm8        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  69,15,91,192                        // cvtdq2ps      %xmm8,%xmm8
-  .byte  68,15,89,5,73,241,0,0               // mulps         0xf149(%rip),%xmm8        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  68,15,89,5,41,8,1,0                 // mulps         0x10829(%rip),%xmm8        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  68,15,40,211                        // movaps        %xmm3,%xmm10
   .byte  68,15,194,215,1                     // cmpltps       %xmm7,%xmm10
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
@@ -44942,7 +44942,7 @@
   .byte  243,69,15,111,12,144                // movdqu        (%r8,%rdx,4),%xmm9
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
-  .byte  102,68,15,111,5,184,238,0,0         // movdqa        0xeeb8(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,111,5,152,5,1,0           // movdqa        0x10598(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
@@ -44998,7 +44998,7 @@
   .byte  102,15,20,211                       // unpcklpd      %xmm3,%xmm2
   .byte  102,65,15,114,209,24                // psrld         $0x18,%xmm9
   .byte  65,15,91,217                        // cvtdq2ps      %xmm9,%xmm3
-  .byte  15,89,29,15,239,0,0                 // mulps         0xef0f(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,239,5,1,0                  // mulps         0x105ef(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
@@ -45038,7 +45038,7 @@
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
   .byte  102,68,15,105,200                   // punpckhwd     %xmm0,%xmm9
-  .byte  102,15,111,21,233,238,0,0           // movdqa        0xeee9(%rip),%xmm2        # 33650 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,111,21,201,5,1,0             // movdqa        0x105c9(%rip),%xmm2        # 34d30 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
   .byte  102,15,112,217,238                  // pshufd        $0xee,%xmm1,%xmm3
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
@@ -45099,7 +45099,7 @@
   .byte  102,65,15,235,217                   // por           %xmm9,%xmm3
   .byte  102,65,15,97,216                    // punpcklwd     %xmm8,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,231,237,0,0                // mulps         0xede7(%rip),%xmm3        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  15,89,29,199,4,1,0                  // mulps         0x104c7(%rip),%xmm3        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
@@ -45140,7 +45140,7 @@
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
   .byte  102,65,15,97,194                    // punpcklwd     %xmm10,%xmm0
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
-  .byte  102,68,15,111,5,43,237,0,0          // movdqa        0xed2b(%rip),%xmm8        # 33650 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,68,15,111,5,11,4,1,0            // movdqa        0x1040b(%rip),%xmm8        # 34d30 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
   .byte  102,65,15,219,192                   // pand          %xmm8,%xmm0
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,65,15,97,193                    // punpcklwd     %xmm9,%xmm0
@@ -45195,7 +45195,7 @@
   .byte  65,15,20,216                        // unpcklps      %xmm8,%xmm3
   .byte  102,15,20,211                       // unpcklpd      %xmm3,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,168,233,0,0                // movaps        0xe9a8(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,136,0,1,0                  // movaps        0x10088(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  91                                  // pop           %rbx
   .byte  65,94                               // pop           %r14
   .byte  255,224                             // jmpq          *%rax
@@ -45227,7 +45227,7 @@
   .byte  65,86                               // push          %r14
   .byte  83                                  // push          %rbx
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,21,234,0,0               // movaps        0xea15(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,245,0,1,0                // movaps        0x100f5(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  102,73,15,126,192                   // movq          %xmm0,%r8
@@ -45255,7 +45255,7 @@
   .byte  102,65,15,96,193                    // punpcklbw     %xmm9,%xmm0
   .byte  102,65,15,97,193                    // punpcklwd     %xmm9,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,21,208,234,0,0             // movaps        0xead0(%rip),%xmm10        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,21,176,1,1,0               // movaps        0x101b0(%rip),%xmm10        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -45374,7 +45374,7 @@
   .byte  102,65,15,96,193                    // punpcklbw     %xmm9,%xmm0
   .byte  102,65,15,97,193                    // punpcklwd     %xmm9,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,21,16,233,0,0              // movaps        0xe910(%rip),%xmm10        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,21,240,255,0,0             // movaps        0xfff0(%rip),%xmm10        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,194                        // mulps         %xmm10,%xmm0
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
@@ -45572,15 +45572,15 @@
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,220,230,0,0             // mulps         0xe6dc(%rip),%xmm9        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,21,228,230,0,0             // andps         0xe6e4(%rip),%xmm10        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,21,12,228,0,0              // orps          0xe40c(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,13,228,230,0,0             // addps         0xe6e4(%rip),%xmm9        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  68,15,40,37,236,230,0,0             // movaps        0xe6ec(%rip),%xmm12        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,13,188,253,0,0             // mulps         0xfdbc(%rip),%xmm9        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,21,196,253,0,0             // andps         0xfdc4(%rip),%xmm10        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,21,236,250,0,0             // orps          0xfaec(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,13,196,253,0,0             // addps         0xfdc4(%rip),%xmm9        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,37,204,253,0,0             // movaps        0xfdcc(%rip),%xmm12        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,236,230,0,0             // addps         0xe6ec(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  68,15,40,37,244,230,0,0             // movaps        0xe6f4(%rip),%xmm12        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,21,204,253,0,0             // addps         0xfdcc(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,40,37,212,253,0,0             // movaps        0xfdd4(%rip),%xmm12        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -45588,22 +45588,22 @@
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,206,227,0,0             // movaps        0xe3ce(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,174,250,0,0             // movaps        0xfaae(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,194,230,0,0             // addps         0xe6c2(%rip),%xmm9        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  68,15,40,37,202,230,0,0             // movaps        0xe6ca(%rip),%xmm12        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,13,162,253,0,0             // addps         0xfda2(%rip),%xmm9        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,37,170,253,0,0             // movaps        0xfdaa(%rip),%xmm12        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,202,230,0,0             // movaps        0xe6ca(%rip),%xmm12        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,37,170,253,0,0             // movaps        0xfdaa(%rip),%xmm12        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,206,230,0,0             // movaps        0xe6ce(%rip),%xmm13        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,45,174,253,0,0             // movaps        0xfdae(%rip),%xmm13        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,206,230,0,0             // mulps         0xe6ce(%rip),%xmm13        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  68,15,89,45,174,253,0,0             // mulps         0xfdae(%rip),%xmm13        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -45639,15 +45639,15 @@
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,158,229,0,0             // mulps         0xe59e(%rip),%xmm9        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,21,166,229,0,0             // andps         0xe5a6(%rip),%xmm10        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,21,206,226,0,0             // orps          0xe2ce(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,13,166,229,0,0             // addps         0xe5a6(%rip),%xmm9        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  68,15,40,37,174,229,0,0             // movaps        0xe5ae(%rip),%xmm12        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,13,126,252,0,0             // mulps         0xfc7e(%rip),%xmm9        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,21,134,252,0,0             // andps         0xfc86(%rip),%xmm10        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,21,174,249,0,0             // orps          0xf9ae(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,13,134,252,0,0             // addps         0xfc86(%rip),%xmm9        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,37,142,252,0,0             // movaps        0xfc8e(%rip),%xmm12        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,174,229,0,0             // addps         0xe5ae(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  68,15,40,37,182,229,0,0             // movaps        0xe5b6(%rip),%xmm12        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,21,142,252,0,0             // addps         0xfc8e(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,40,37,150,252,0,0             // movaps        0xfc96(%rip),%xmm12        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -45655,22 +45655,22 @@
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,144,226,0,0             // movaps        0xe290(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,112,249,0,0             // movaps        0xf970(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,132,229,0,0             // addps         0xe584(%rip),%xmm9        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  68,15,40,37,140,229,0,0             // movaps        0xe58c(%rip),%xmm12        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,13,100,252,0,0             // addps         0xfc64(%rip),%xmm9        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,37,108,252,0,0             // movaps        0xfc6c(%rip),%xmm12        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,140,229,0,0             // movaps        0xe58c(%rip),%xmm12        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,37,108,252,0,0             // movaps        0xfc6c(%rip),%xmm12        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,144,229,0,0             // movaps        0xe590(%rip),%xmm13        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,45,112,252,0,0             // movaps        0xfc70(%rip),%xmm13        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,144,229,0,0             // mulps         0xe590(%rip),%xmm13        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  68,15,89,45,112,252,0,0             // mulps         0xfc70(%rip),%xmm13        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -45706,15 +45706,15 @@
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,96,228,0,0              // mulps         0xe460(%rip),%xmm9        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,21,104,228,0,0             // andps         0xe468(%rip),%xmm10        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,21,144,225,0,0             // orps          0xe190(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,13,104,228,0,0             // addps         0xe468(%rip),%xmm9        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  68,15,40,37,112,228,0,0             // movaps        0xe470(%rip),%xmm12        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,13,64,251,0,0              // mulps         0xfb40(%rip),%xmm9        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,21,72,251,0,0              // andps         0xfb48(%rip),%xmm10        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,21,112,248,0,0             // orps          0xf870(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,13,72,251,0,0              // addps         0xfb48(%rip),%xmm9        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,37,80,251,0,0              // movaps        0xfb50(%rip),%xmm12        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,112,228,0,0             // addps         0xe470(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  68,15,40,37,120,228,0,0             // movaps        0xe478(%rip),%xmm12        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,21,80,251,0,0              // addps         0xfb50(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,40,37,88,251,0,0              // movaps        0xfb58(%rip),%xmm12        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -45722,22 +45722,22 @@
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,82,225,0,0              // movaps        0xe152(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,50,248,0,0              // movaps        0xf832(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,70,228,0,0              // addps         0xe446(%rip),%xmm9        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  68,15,40,37,78,228,0,0              // movaps        0xe44e(%rip),%xmm12        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,13,38,251,0,0              // addps         0xfb26(%rip),%xmm9        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,37,46,251,0,0              // movaps        0xfb2e(%rip),%xmm12        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,78,228,0,0              // movaps        0xe44e(%rip),%xmm12        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,37,46,251,0,0              // movaps        0xfb2e(%rip),%xmm12        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,82,228,0,0              // movaps        0xe452(%rip),%xmm13        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,45,50,251,0,0              // movaps        0xfb32(%rip),%xmm13        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,82,228,0,0              // mulps         0xe452(%rip),%xmm13        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  68,15,89,45,50,251,0,0              // mulps         0xfb32(%rip),%xmm13        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -45773,15 +45773,15 @@
   .byte  69,15,88,209                        // addps         %xmm9,%xmm10
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,91,202                        // cvtdq2ps      %xmm10,%xmm9
-  .byte  68,15,89,13,34,227,0,0              // mulps         0xe322(%rip),%xmm9        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  68,15,84,21,42,227,0,0              // andps         0xe32a(%rip),%xmm10        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,86,21,82,224,0,0              // orps          0xe052(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
-  .byte  68,15,88,13,42,227,0,0              // addps         0xe32a(%rip),%xmm9        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
-  .byte  68,15,40,37,50,227,0,0              // movaps        0xe332(%rip),%xmm12        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,89,13,2,250,0,0               // mulps         0xfa02(%rip),%xmm9        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,84,21,10,250,0,0              // andps         0xfa0a(%rip),%xmm10        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,86,21,50,247,0,0              // orps          0xf732(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
+  .byte  68,15,88,13,10,250,0,0              // addps         0xfa0a(%rip),%xmm9        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,37,18,250,0,0              // movaps        0xfa12(%rip),%xmm12        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  69,15,89,226                        // mulps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,88,21,50,227,0,0              // addps         0xe332(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
-  .byte  68,15,40,37,58,227,0,0              // movaps        0xe33a(%rip),%xmm12        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,88,21,18,250,0,0              // addps         0xfa12(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,40,37,26,250,0,0              // movaps        0xfa1a(%rip),%xmm12        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  69,15,94,226                        // divps         %xmm10,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
   .byte  69,15,89,203                        // mulps         %xmm11,%xmm9
@@ -45789,22 +45789,22 @@
   .byte  69,15,91,226                        // cvtdq2ps      %xmm10,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,194,236,1                     // cmpltps       %xmm12,%xmm13
-  .byte  68,15,40,21,20,224,0,0              // movaps        0xe014(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,244,246,0,0             // movaps        0xf6f4(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,234                        // andps         %xmm10,%xmm13
   .byte  69,15,87,219                        // xorps         %xmm11,%xmm11
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
   .byte  69,15,40,233                        // movaps        %xmm9,%xmm13
   .byte  69,15,92,236                        // subps         %xmm12,%xmm13
-  .byte  68,15,88,13,8,227,0,0               // addps         0xe308(%rip),%xmm9        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
-  .byte  68,15,40,37,16,227,0,0              // movaps        0xe310(%rip),%xmm12        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  68,15,88,13,232,249,0,0             // addps         0xf9e8(%rip),%xmm9        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,37,240,249,0,0             // movaps        0xf9f0(%rip),%xmm12        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  69,15,89,229                        // mulps         %xmm13,%xmm12
   .byte  69,15,92,204                        // subps         %xmm12,%xmm9
-  .byte  68,15,40,37,16,227,0,0              // movaps        0xe310(%rip),%xmm12        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  68,15,40,37,240,249,0,0             // movaps        0xf9f0(%rip),%xmm12        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  69,15,92,229                        // subps         %xmm13,%xmm12
-  .byte  68,15,40,45,20,227,0,0              // movaps        0xe314(%rip),%xmm13        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,45,244,249,0,0             // movaps        0xf9f4(%rip),%xmm13        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  69,15,94,236                        // divps         %xmm12,%xmm13
   .byte  69,15,88,233                        // addps         %xmm9,%xmm13
-  .byte  68,15,89,45,20,227,0,0              // mulps         0xe314(%rip),%xmm13        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  68,15,89,45,244,249,0,0             // mulps         0xf9f4(%rip),%xmm13        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  102,69,15,91,205                    // cvtps2dq      %xmm13,%xmm9
   .byte  243,68,15,16,96,20                  // movss         0x14(%rax),%xmm12
   .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
@@ -45829,19 +45829,19 @@
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,91,194                           // cvtdq2ps      %xmm2,%xmm0
-  .byte  15,89,5,31,226,0,0                  // mulps         0xe21f(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  15,84,21,40,226,0,0                 // andps         0xe228(%rip),%xmm2        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  68,15,40,53,80,223,0,0              // movaps        0xdf50(%rip),%xmm14        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,89,5,255,248,0,0                 // mulps         0xf8ff(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  15,84,21,8,249,0,0                  // andps         0xf908(%rip),%xmm2        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,40,53,48,246,0,0              // movaps        0xf630(%rip),%xmm14        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,86,214                        // orps          %xmm14,%xmm2
-  .byte  68,15,40,37,36,226,0,0              // movaps        0xe224(%rip),%xmm12        # 336a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
+  .byte  68,15,40,37,4,249,0,0               // movaps        0xf904(%rip),%xmm12        # 34d80 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
   .byte  65,15,88,196                        // addps         %xmm12,%xmm0
-  .byte  68,15,40,29,40,226,0,0              // movaps        0xe228(%rip),%xmm11        # 336b0 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  68,15,40,29,8,249,0,0               // movaps        0xf908(%rip),%xmm11        # 34d90 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
   .byte  15,40,226                           // movaps        %xmm2,%xmm4
   .byte  65,15,89,227                        // mulps         %xmm11,%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  68,15,40,21,38,226,0,0              // movaps        0xe226(%rip),%xmm10        # 336c0 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  68,15,40,21,6,249,0,0               // movaps        0xf906(%rip),%xmm10        # 34da0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
   .byte  65,15,88,210                        // addps         %xmm10,%xmm2
-  .byte  68,15,40,45,42,226,0,0              // movaps        0xe22a(%rip),%xmm13        # 336d0 <_sk_srcover_rgba_8888_sse2_lowp+0xd50>
+  .byte  68,15,40,45,10,249,0,0              // movaps        0xf90a(%rip),%xmm13        # 34db0 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
   .byte  65,15,40,229                        // movaps        %xmm13,%xmm4
   .byte  15,94,226                           // divps         %xmm2,%xmm4
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -45853,27 +45853,27 @@
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
   .byte  15,194,226,1                        // cmpltps       %xmm2,%xmm4
-  .byte  68,15,40,13,250,222,0,0             // movaps        0xdefa(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,13,218,245,0,0             // movaps        0xf5da(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  65,15,84,225                        // andps         %xmm9,%xmm4
   .byte  15,92,212                           // subps         %xmm4,%xmm2
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
   .byte  15,92,226                           // subps         %xmm2,%xmm4
-  .byte  15,40,53,246,225,0,0                // movaps        0xe1f6(%rip),%xmm6        # 336e0 <_sk_srcover_rgba_8888_sse2_lowp+0xd60>
+  .byte  15,40,53,214,248,0,0                // movaps        0xf8d6(%rip),%xmm6        # 34dc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
   .byte  15,88,198                           // addps         %xmm6,%xmm0
-  .byte  15,40,61,12,226,0,0                 // movaps        0xe20c(%rip),%xmm7        # 33700 <_sk_srcover_rgba_8888_sse2_lowp+0xd80>
+  .byte  15,40,61,236,248,0,0                // movaps        0xf8ec(%rip),%xmm7        # 34de0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
   .byte  15,40,239                           // movaps        %xmm7,%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,40,212                           // movaps        %xmm4,%xmm2
-  .byte  15,40,37,236,225,0,0                // movaps        0xe1ec(%rip),%xmm4        # 336f0 <_sk_srcover_rgba_8888_sse2_lowp+0xd70>
+  .byte  15,40,37,204,248,0,0                // movaps        0xf8cc(%rip),%xmm4        # 34dd0 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  15,92,194                           // subps         %xmm2,%xmm0
-  .byte  68,15,40,61,254,225,0,0             // movaps        0xe1fe(%rip),%xmm15        # 33710 <_sk_srcover_rgba_8888_sse2_lowp+0xd90>
+  .byte  68,15,40,61,222,248,0,0             // movaps        0xf8de(%rip),%xmm15        # 34df0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
   .byte  65,15,40,215                        // movaps        %xmm15,%xmm2
   .byte  15,94,213                           // divps         %xmm5,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  15,89,5,90,225,0,0                  // mulps         0xe15a(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  15,84,13,99,225,0,0                 // andps         0xe163(%rip),%xmm1        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
+  .byte  15,89,5,58,248,0,0                  // mulps         0xf83a(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  15,84,13,67,248,0,0                 // andps         0xf843(%rip),%xmm1        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
   .byte  65,15,86,206                        // orps          %xmm14,%xmm1
   .byte  65,15,88,196                        // addps         %xmm12,%xmm0
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
@@ -45903,9 +45903,9 @@
   .byte  15,94,206                           // divps         %xmm6,%xmm1
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  15,91,195                           // cvtdq2ps      %xmm3,%xmm0
-  .byte  15,89,5,233,224,0,0                 // mulps         0xe0e9(%rip),%xmm0        # 33680 <_sk_srcover_rgba_8888_sse2_lowp+0xd00>
-  .byte  15,84,29,242,224,0,0                // andps         0xe0f2(%rip),%xmm3        # 33690 <_sk_srcover_rgba_8888_sse2_lowp+0xd10>
-  .byte  15,86,29,27,222,0,0                 // orps          0xde1b(%rip),%xmm3        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,89,5,201,247,0,0                 // mulps         0xf7c9(%rip),%xmm0        # 34d60 <_sk_srcover_rgba_8888_sse2_lowp+0xd30>
+  .byte  15,84,29,210,247,0,0                // andps         0xf7d2(%rip),%xmm3        # 34d70 <_sk_srcover_rgba_8888_sse2_lowp+0xd40>
+  .byte  15,86,29,251,244,0,0                // orps          0xf4fb(%rip),%xmm3        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,88,196                        // addps         %xmm12,%xmm0
   .byte  68,15,89,219                        // mulps         %xmm3,%xmm11
   .byte  65,15,92,195                        // subps         %xmm11,%xmm0
@@ -45927,7 +45927,7 @@
   .byte  15,92,253                           // subps         %xmm5,%xmm7
   .byte  68,15,94,255                        // divps         %xmm7,%xmm15
   .byte  68,15,88,248                        // addps         %xmm0,%xmm15
-  .byte  15,40,5,40,225,0,0                  // movaps        0xe128(%rip),%xmm0        # 33720 <_sk_srcover_rgba_8888_sse2_lowp+0xda0>
+  .byte  15,40,5,8,248,0,0                   // movaps        0xf808(%rip),%xmm0        # 34e00 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  68,15,89,248                        // mulps         %xmm0,%xmm15
@@ -45946,29 +45946,29 @@
 .globl _sk_lab_to_xyz_sse2
 FUNCTION(_sk_lab_to_xyz_sse2)
 _sk_lab_to_xyz_sse2:
-  .byte  15,89,5,253,224,0,0                 // mulps         0xe0fd(%rip),%xmm0        # 33730 <_sk_srcover_rgba_8888_sse2_lowp+0xdb0>
-  .byte  68,15,40,5,117,222,0,0              // movaps        0xde75(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  15,89,5,221,247,0,0                 // mulps         0xf7dd(%rip),%xmm0        # 34e10 <_sk_srcover_rgba_8888_sse2_lowp+0xde0>
+  .byte  68,15,40,5,85,245,0,0               // movaps        0xf555(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
-  .byte  68,15,40,13,249,224,0,0             // movaps        0xe0f9(%rip),%xmm9        # 33740 <_sk_srcover_rgba_8888_sse2_lowp+0xdc0>
+  .byte  68,15,40,13,217,247,0,0             // movaps        0xf7d9(%rip),%xmm9        # 34e20 <_sk_srcover_rgba_8888_sse2_lowp+0xdf0>
   .byte  65,15,88,201                        // addps         %xmm9,%xmm1
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  65,15,88,209                        // addps         %xmm9,%xmm2
-  .byte  15,88,5,246,224,0,0                 // addps         0xe0f6(%rip),%xmm0        # 33750 <_sk_srcover_rgba_8888_sse2_lowp+0xdd0>
-  .byte  15,89,5,255,224,0,0                 // mulps         0xe0ff(%rip),%xmm0        # 33760 <_sk_srcover_rgba_8888_sse2_lowp+0xde0>
-  .byte  15,89,13,8,225,0,0                  // mulps         0xe108(%rip),%xmm1        # 33770 <_sk_srcover_rgba_8888_sse2_lowp+0xdf0>
+  .byte  15,88,5,214,247,0,0                 // addps         0xf7d6(%rip),%xmm0        # 34e30 <_sk_srcover_rgba_8888_sse2_lowp+0xe00>
+  .byte  15,89,5,223,247,0,0                 // mulps         0xf7df(%rip),%xmm0        # 34e40 <_sk_srcover_rgba_8888_sse2_lowp+0xe10>
+  .byte  15,89,13,232,247,0,0                // mulps         0xf7e8(%rip),%xmm1        # 34e50 <_sk_srcover_rgba_8888_sse2_lowp+0xe20>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,21,14,225,0,0                 // mulps         0xe10e(%rip),%xmm2        # 33780 <_sk_srcover_rgba_8888_sse2_lowp+0xe00>
+  .byte  15,89,21,238,247,0,0                // mulps         0xf7ee(%rip),%xmm2        # 34e60 <_sk_srcover_rgba_8888_sse2_lowp+0xe30>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  68,15,92,202                        // subps         %xmm2,%xmm9
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
   .byte  68,15,89,225                        // mulps         %xmm1,%xmm12
-  .byte  15,40,21,3,225,0,0                  // movaps        0xe103(%rip),%xmm2        # 33790 <_sk_srcover_rgba_8888_sse2_lowp+0xe10>
+  .byte  15,40,21,227,247,0,0                // movaps        0xf7e3(%rip),%xmm2        # 34e70 <_sk_srcover_rgba_8888_sse2_lowp+0xe40>
   .byte  68,15,40,194                        // movaps        %xmm2,%xmm8
   .byte  69,15,194,196,1                     // cmpltps       %xmm12,%xmm8
-  .byte  68,15,40,21,2,225,0,0               // movaps        0xe102(%rip),%xmm10        # 337a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe20>
+  .byte  68,15,40,21,226,247,0,0             // movaps        0xf7e2(%rip),%xmm10        # 34e80 <_sk_srcover_rgba_8888_sse2_lowp+0xe50>
   .byte  65,15,88,202                        // addps         %xmm10,%xmm1
-  .byte  68,15,40,29,6,225,0,0               // movaps        0xe106(%rip),%xmm11        # 337b0 <_sk_srcover_rgba_8888_sse2_lowp+0xe30>
+  .byte  68,15,40,29,230,247,0,0             // movaps        0xf7e6(%rip),%xmm11        # 34e90 <_sk_srcover_rgba_8888_sse2_lowp+0xe60>
   .byte  65,15,89,203                        // mulps         %xmm11,%xmm1
   .byte  69,15,84,224                        // andps         %xmm8,%xmm12
   .byte  68,15,85,193                        // andnps        %xmm1,%xmm8
@@ -45992,8 +45992,8 @@
   .byte  15,84,194                           // andps         %xmm2,%xmm0
   .byte  65,15,85,209                        // andnps        %xmm9,%xmm2
   .byte  15,86,208                           // orps          %xmm0,%xmm2
-  .byte  68,15,89,5,182,224,0,0              // mulps         0xe0b6(%rip),%xmm8        # 337c0 <_sk_srcover_rgba_8888_sse2_lowp+0xe40>
-  .byte  15,89,21,191,224,0,0                // mulps         0xe0bf(%rip),%xmm2        # 337d0 <_sk_srcover_rgba_8888_sse2_lowp+0xe50>
+  .byte  68,15,89,5,150,247,0,0              // mulps         0xf796(%rip),%xmm8        # 34ea0 <_sk_srcover_rgba_8888_sse2_lowp+0xe70>
+  .byte  15,89,21,159,247,0,0                // mulps         0xf79f(%rip),%xmm2        # 34eb0 <_sk_srcover_rgba_8888_sse2_lowp+0xe80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -46013,9 +46013,9 @@
   .byte  102,67,15,110,4,16                  // movd          (%r8,%r10,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  102,15,97,192                       // punpcklwd     %xmm0,%xmm0
-  .byte  102,15,219,5,57,221,0,0             // pand          0xdd39(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,5,25,244,0,0             // pand          0xf419(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,143,222,0,0                // mulps         0xde8f(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,111,245,0,0                // mulps         0xf56f(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,87,201                       // xorpd         %xmm1,%xmm1
@@ -46058,9 +46058,9 @@
   .byte  102,67,15,110,36,16                 // movd          (%r8,%r10,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,37,165,220,0,0           // pand          0xdca5(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,37,133,243,0,0           // pand          0xf385(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,252                           // cvtdq2ps      %xmm4,%xmm7
-  .byte  15,89,61,251,221,0,0                // mulps         0xddfb(%rip),%xmm7        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,61,219,244,0,0                // mulps         0xf4db(%rip),%xmm7        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  102,15,87,237                       // xorpd         %xmm5,%xmm5
@@ -46139,7 +46139,7 @@
   .byte  102,65,15,96,192                    // punpcklbw     %xmm8,%xmm0
   .byte  102,65,15,97,192                    // punpcklwd     %xmm8,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,29,213,220,0,0                // mulps         0xdcd5(%rip),%xmm3        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,29,181,243,0,0                // mulps         0xf3b5(%rip),%xmm3        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -46156,7 +46156,7 @@
   .byte  77,15,175,193                       // imul          %r9,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,210                           // movslq        %edx,%r10
-  .byte  68,15,40,5,124,219,0,0              // movaps        0xdb7c(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,92,242,0,0               // movaps        0xf25c(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,89,195                        // mulps         %xmm3,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
   .byte  102,65,15,114,240,16                // pslld         $0x10,%xmm8
@@ -46181,7 +46181,7 @@
   .byte  117,217                             // jne           25961 <_sk_store_a8_sse2+0x48>
   .byte  102,65,15,197,192,4                 // pextrw        $0x4,%xmm8,%eax
   .byte  67,136,68,16,2                      // mov           %al,0x2(%r8,%r10,1)
-  .byte  102,68,15,219,5,228,218,0,0         // pand          0xdae4(%rip),%xmm8        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,68,15,219,5,196,241,0,0         // pand          0xf1c4(%rip),%xmm8        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,69,15,103,192                   // packuswb      %xmm8,%xmm8
   .byte  102,69,15,103,192                   // packuswb      %xmm8,%xmm8
   .byte  102,68,15,126,192                   // movd          %xmm8,%eax
@@ -46206,11 +46206,11 @@
   .byte  102,67,15,110,4,16                  // movd          (%r8,%r10,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  102,15,97,192                       // punpcklwd     %xmm0,%xmm0
-  .byte  102,15,219,5,149,218,0,0            // pand          0xda95(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,5,117,241,0,0            // pand          0xf175(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,235,219,0,0                 // mulps         0xdbeb(%rip),%xmm0        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,5,203,242,0,0                 // mulps         0xf2cb(%rip),%xmm0        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,210,217,0,0                // movaps        0xd9d2(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,178,240,0,0                // movaps        0xf0b2(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -46251,11 +46251,11 @@
   .byte  102,67,15,110,36,16                 // movd          (%r8,%r10,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,37,254,217,0,0           // pand          0xd9fe(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,219,37,222,240,0,0           // pand          0xf0de(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,84,219,0,0                 // mulps         0xdb54(%rip),%xmm4        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,37,52,242,0,0                 // mulps         0xf234(%rip),%xmm4        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,59,217,0,0                 // movaps        0xd93b(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,61,27,240,0,0                 // movaps        0xf01b(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  255,224                             // jmpq          *%rax
@@ -46332,9 +46332,9 @@
   .byte  102,65,15,96,192                    // punpcklbw     %xmm8,%xmm0
   .byte  102,65,15,97,192                    // punpcklwd     %xmm8,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,43,218,0,0                  // mulps         0xda2b(%rip),%xmm0        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  15,89,5,11,241,0,0                  // mulps         0xf10b(%rip),%xmm0        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,18,216,0,0                 // movaps        0xd812(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,242,238,0,0                // movaps        0xeef2(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
@@ -46354,19 +46354,19 @@
   .byte  117,83                              // jne           25c34 <_sk_load_565_sse2+0x6e>
   .byte  243,67,15,126,20,80                 // movq          (%r8,%r10,2),%xmm2
   .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
-  .byte  102,15,111,5,253,217,0,0            // movdqa        0xd9fd(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,221,240,0,0            // movdqa        0xf0dd(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,255,217,0,0                 // mulps         0xd9ff(%rip),%xmm0        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,13,7,218,0,0             // movdqa        0xda07(%rip),%xmm1        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,5,223,240,0,0                 // mulps         0xf0df(%rip),%xmm0        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,13,231,240,0,0           // movdqa        0xf0e7(%rip),%xmm1        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,9,218,0,0                  // mulps         0xda09(%rip),%xmm1        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,21,17,218,0,0            // pand          0xda11(%rip),%xmm2        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,13,233,240,0,0                // mulps         0xf0e9(%rip),%xmm1        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,21,241,240,0,0           // pand          0xf0f1(%rip),%xmm2        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,23,218,0,0                 // mulps         0xda17(%rip),%xmm2        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,21,247,240,0,0                // mulps         0xf0f7(%rip),%xmm2        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,158,215,0,0                // movaps        0xd79e(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,126,238,0,0                // movaps        0xee7e(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,3                        // and           $0x3,%r9b
@@ -46403,19 +46403,19 @@
   .byte  117,83                              // jne           25cf0 <_sk_load_565_dst_sse2+0x6e>
   .byte  243,67,15,126,52,80                 // movq          (%r8,%r10,2),%xmm6
   .byte  102,15,97,240                       // punpcklwd     %xmm0,%xmm6
-  .byte  102,15,111,37,65,217,0,0            // movdqa        0xd941(%rip),%xmm4        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,37,33,240,0,0            // movdqa        0xf021(%rip),%xmm4        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,67,217,0,0                 // mulps         0xd943(%rip),%xmm4        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,45,75,217,0,0            // movdqa        0xd94b(%rip),%xmm5        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,37,35,240,0,0                 // mulps         0xf023(%rip),%xmm4        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,45,43,240,0,0            // movdqa        0xf02b(%rip),%xmm5        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,238                      // pand          %xmm6,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,45,77,217,0,0                 // mulps         0xd94d(%rip),%xmm5        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,53,85,217,0,0            // pand          0xd955(%rip),%xmm6        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,45,45,240,0,0                 // mulps         0xf02d(%rip),%xmm5        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,53,53,240,0,0            // pand          0xf035(%rip),%xmm6        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,53,91,217,0,0                 // mulps         0xd95b(%rip),%xmm6        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,53,59,240,0,0                 // mulps         0xf03b(%rip),%xmm6        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,226,214,0,0                // movaps        0xd6e2(%rip),%xmm7        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,61,194,237,0,0                // movaps        0xedc2(%rip),%xmm7        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
   .byte  65,128,225,3                        // and           $0x3,%r9b
@@ -46484,19 +46484,19 @@
   .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
   .byte  102,15,196,208,3                    // pinsrw        $0x3,%eax,%xmm2
   .byte  102,65,15,97,208                    // punpcklwd     %xmm8,%xmm2
-  .byte  102,15,111,5,240,215,0,0            // movdqa        0xd7f0(%rip),%xmm0        # 335f0 <_sk_srcover_rgba_8888_sse2_lowp+0xc70>
+  .byte  102,15,111,5,208,238,0,0            // movdqa        0xeed0(%rip),%xmm0        # 34cd0 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,242,215,0,0                 // mulps         0xd7f2(%rip),%xmm0        # 33600 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
-  .byte  102,15,111,13,250,215,0,0           // movdqa        0xd7fa(%rip),%xmm1        # 33610 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
+  .byte  15,89,5,210,238,0,0                 // mulps         0xeed2(%rip),%xmm0        # 34ce0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  102,15,111,13,218,238,0,0           // movdqa        0xeeda(%rip),%xmm1        # 34cf0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,252,215,0,0                // mulps         0xd7fc(%rip),%xmm1        # 33620 <_sk_srcover_rgba_8888_sse2_lowp+0xca0>
-  .byte  102,15,219,21,4,216,0,0             // pand          0xd804(%rip),%xmm2        # 33630 <_sk_srcover_rgba_8888_sse2_lowp+0xcb0>
+  .byte  15,89,13,220,238,0,0                // mulps         0xeedc(%rip),%xmm1        # 34d00 <_sk_srcover_rgba_8888_sse2_lowp+0xcd0>
+  .byte  102,15,219,21,228,238,0,0           // pand          0xeee4(%rip),%xmm2        # 34d10 <_sk_srcover_rgba_8888_sse2_lowp+0xce0>
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,10,216,0,0                 // mulps         0xd80a(%rip),%xmm2        # 33640 <_sk_srcover_rgba_8888_sse2_lowp+0xcc0>
+  .byte  15,89,21,234,238,0,0                // mulps         0xeeea(%rip),%xmm2        # 34d20 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,145,213,0,0                // movaps        0xd591(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,113,236,0,0                // movaps        0xec71(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
 
 HIDDEN _sk_store_565_sse2
@@ -46510,12 +46510,12 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,210                           // movslq        %edx,%r10
-  .byte  68,15,40,5,145,217,0,0              // movaps        0xd991(%rip),%xmm8        # 337f0 <_sk_srcover_rgba_8888_sse2_lowp+0xe70>
+  .byte  68,15,40,5,113,240,0,0              // movaps        0xf071(%rip),%xmm8        # 34ed0 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
   .byte  102,65,15,114,241,11                // pslld         $0xb,%xmm9
-  .byte  68,15,40,21,134,217,0,0             // movaps        0xd986(%rip),%xmm10        # 33800 <_sk_srcover_rgba_8888_sse2_lowp+0xe80>
+  .byte  68,15,40,21,102,240,0,0             // movaps        0xf066(%rip),%xmm10        # 34ee0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
   .byte  68,15,89,209                        // mulps         %xmm1,%xmm10
   .byte  102,69,15,91,210                    // cvtps2dq      %xmm10,%xmm10
   .byte  102,65,15,114,242,5                 // pslld         $0x5,%xmm10
@@ -46564,21 +46564,21 @@
   .byte  117,98                              // jne           25f7d <_sk_load_4444_sse2+0x7d>
   .byte  243,67,15,126,28,80                 // movq          (%r8,%r10,2),%xmm3
   .byte  102,15,97,216                       // punpcklwd     %xmm0,%xmm3
-  .byte  102,15,111,5,227,216,0,0            // movdqa        0xd8e3(%rip),%xmm0        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,5,195,239,0,0            // movdqa        0xefc3(%rip),%xmm0        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,229,216,0,0                 // mulps         0xd8e5(%rip),%xmm0        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,13,237,216,0,0           // movdqa        0xd8ed(%rip),%xmm1        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,5,197,239,0,0                 // mulps         0xefc5(%rip),%xmm0        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,13,205,239,0,0           // movdqa        0xefcd(%rip),%xmm1        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,239,216,0,0                // mulps         0xd8ef(%rip),%xmm1        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,21,247,216,0,0           // movdqa        0xd8f7(%rip),%xmm2        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,13,207,239,0,0                // mulps         0xefcf(%rip),%xmm1        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,21,215,239,0,0           // movdqa        0xefd7(%rip),%xmm2        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,249,216,0,0                // mulps         0xd8f9(%rip),%xmm2        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,29,1,217,0,0             // pand          0xd901(%rip),%xmm3        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,21,217,239,0,0                // mulps         0xefd9(%rip),%xmm2        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,29,225,239,0,0           // pand          0xefe1(%rip),%xmm3        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,7,217,0,0                  // mulps         0xd907(%rip),%xmm3        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,29,231,239,0,0                // mulps         0xefe7(%rip),%xmm3        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -46616,21 +46616,21 @@
   .byte  117,98                              // jne           26048 <_sk_load_4444_dst_sse2+0x7d>
   .byte  243,67,15,126,60,80                 // movq          (%r8,%r10,2),%xmm7
   .byte  102,15,97,248                       // punpcklwd     %xmm0,%xmm7
-  .byte  102,15,111,37,24,216,0,0            // movdqa        0xd818(%rip),%xmm4        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,37,248,238,0,0           // movdqa        0xeef8(%rip),%xmm4        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,231                      // pand          %xmm7,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,37,26,216,0,0                 // mulps         0xd81a(%rip),%xmm4        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,45,34,216,0,0            // movdqa        0xd822(%rip),%xmm5        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,37,250,238,0,0                // mulps         0xeefa(%rip),%xmm4        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,45,2,239,0,0             // movdqa        0xef02(%rip),%xmm5        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,239                      // pand          %xmm7,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,45,36,216,0,0                 // mulps         0xd824(%rip),%xmm5        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,53,44,216,0,0            // movdqa        0xd82c(%rip),%xmm6        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,45,4,239,0,0                  // mulps         0xef04(%rip),%xmm5        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,53,12,239,0,0            // movdqa        0xef0c(%rip),%xmm6        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,247                      // pand          %xmm7,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,53,46,216,0,0                 // mulps         0xd82e(%rip),%xmm6        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,61,54,216,0,0            // pand          0xd836(%rip),%xmm7        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,53,14,239,0,0                 // mulps         0xef0e(%rip),%xmm6        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,61,22,239,0,0            // pand          0xef16(%rip),%xmm7        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,255                           // cvtdq2ps      %xmm7,%xmm7
-  .byte  15,89,61,60,216,0,0                 // mulps         0xd83c(%rip),%xmm7        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,61,28,239,0,0                 // mulps         0xef1c(%rip),%xmm7        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,137,249                          // mov           %edi,%r9d
@@ -46700,21 +46700,21 @@
   .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
   .byte  102,15,196,216,3                    // pinsrw        $0x3,%eax,%xmm3
   .byte  102,65,15,97,216                    // punpcklwd     %xmm8,%xmm3
-  .byte  102,15,111,5,184,214,0,0            // movdqa        0xd6b8(%rip),%xmm0        # 33810 <_sk_srcover_rgba_8888_sse2_lowp+0xe90>
+  .byte  102,15,111,5,152,237,0,0            // movdqa        0xed98(%rip),%xmm0        # 34ef0 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,5,186,214,0,0                 // mulps         0xd6ba(%rip),%xmm0        # 33820 <_sk_srcover_rgba_8888_sse2_lowp+0xea0>
-  .byte  102,15,111,13,194,214,0,0           // movdqa        0xd6c2(%rip),%xmm1        # 33830 <_sk_srcover_rgba_8888_sse2_lowp+0xeb0>
+  .byte  15,89,5,154,237,0,0                 // mulps         0xed9a(%rip),%xmm0        # 34f00 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  102,15,111,13,162,237,0,0           // movdqa        0xeda2(%rip),%xmm1        # 34f10 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,13,196,214,0,0                // mulps         0xd6c4(%rip),%xmm1        # 33840 <_sk_srcover_rgba_8888_sse2_lowp+0xec0>
-  .byte  102,15,111,21,204,214,0,0           // movdqa        0xd6cc(%rip),%xmm2        # 33850 <_sk_srcover_rgba_8888_sse2_lowp+0xed0>
+  .byte  15,89,13,164,237,0,0                // mulps         0xeda4(%rip),%xmm1        # 34f20 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  102,15,111,21,172,237,0,0           // movdqa        0xedac(%rip),%xmm2        # 34f30 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,21,206,214,0,0                // mulps         0xd6ce(%rip),%xmm2        # 33860 <_sk_srcover_rgba_8888_sse2_lowp+0xee0>
-  .byte  102,15,219,29,214,214,0,0           // pand          0xd6d6(%rip),%xmm3        # 33870 <_sk_srcover_rgba_8888_sse2_lowp+0xef0>
+  .byte  15,89,21,174,237,0,0                // mulps         0xedae(%rip),%xmm2        # 34f40 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  102,15,219,29,182,237,0,0           // pand          0xedb6(%rip),%xmm3        # 34f50 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,29,220,214,0,0                // mulps         0xd6dc(%rip),%xmm3        # 33880 <_sk_srcover_rgba_8888_sse2_lowp+0xf00>
+  .byte  15,89,29,188,237,0,0                // mulps         0xedbc(%rip),%xmm3        # 34f60 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
@@ -46729,7 +46729,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  76,99,210                           // movslq        %edx,%r10
-  .byte  68,15,40,5,202,214,0,0              // movaps        0xd6ca(%rip),%xmm8        # 33890 <_sk_srcover_rgba_8888_sse2_lowp+0xf10>
+  .byte  68,15,40,5,170,237,0,0              // movaps        0xedaa(%rip),%xmm8        # 34f70 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -46787,11 +46787,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,98                              // jne           262f9 <_sk_load_8888_sse2+0x7e>
   .byte  243,69,15,111,12,128                // movdqu        (%r8,%rax,4),%xmm9
-  .byte  102,15,111,21,219,209,0,0           // movdqa        0xd1db(%rip),%xmm2        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,21,187,232,0,0           // movdqa        0xe8bb(%rip),%xmm2        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,39,211,0,0               // movaps        0xd327(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,7,234,0,0                // movaps        0xea07(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -46838,11 +46838,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,98                              // jne           263b8 <_sk_load_8888_dst_sse2+0x7e>
   .byte  243,69,15,111,12,128                // movdqu        (%r8,%rax,4),%xmm9
-  .byte  102,15,111,53,28,209,0,0            // movdqa        0xd11c(%rip),%xmm6        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,53,252,231,0,0           // movdqa        0xe7fc(%rip),%xmm6        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,225                   // movdqa        %xmm9,%xmm4
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  68,15,40,5,104,210,0,0              // movaps        0xd268(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,72,233,0,0               // movaps        0xe948(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,224                        // mulps         %xmm8,%xmm4
   .byte  102,65,15,111,233                   // movdqa        %xmm9,%xmm5
   .byte  102,15,114,213,8                    // psrld         $0x8,%xmm5
@@ -46919,11 +46919,11 @@
   .byte  102,67,15,110,12,136                // movd          (%r8,%r9,4),%xmm1
   .byte  102,15,98,200                       // punpckldq     %xmm0,%xmm1
   .byte  102,68,15,108,201                   // punpcklqdq    %xmm1,%xmm9
-  .byte  102,15,111,21,207,207,0,0           // movdqa        0xcfcf(%rip),%xmm2        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,21,175,230,0,0           // movdqa        0xe6af(%rip),%xmm2        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  68,15,40,5,27,209,0,0               // movaps        0xd11b(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,251,231,0,0              // movaps        0xe7fb(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -46952,7 +46952,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,140,207,0,0              // movaps        0xcf8c(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,108,230,0,0              // movaps        0xe66c(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -47004,11 +47004,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,98                              // jne           26638 <_sk_load_bgra_sse2+0x7e>
   .byte  243,69,15,111,12,128                // movdqu        (%r8,%rax,4),%xmm9
-  .byte  102,15,111,5,156,206,0,0            // movdqa        0xce9c(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,124,229,0,0            // movdqa        0xe57c(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
   .byte  15,91,209                           // cvtdq2ps      %xmm1,%xmm2
-  .byte  68,15,40,5,232,207,0,0              // movaps        0xcfe8(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,200,230,0,0              // movaps        0xe6c8(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -47055,11 +47055,11 @@
   .byte  72,133,255                          // test          %rdi,%rdi
   .byte  117,98                              // jne           266f7 <_sk_load_bgra_dst_sse2+0x7e>
   .byte  243,69,15,111,12,128                // movdqu        (%r8,%rax,4),%xmm9
-  .byte  102,15,111,37,221,205,0,0           // movdqa        0xcddd(%rip),%xmm4        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,37,189,228,0,0           // movdqa        0xe4bd(%rip),%xmm4        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,233                   // movdqa        %xmm9,%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,245                           // cvtdq2ps      %xmm5,%xmm6
-  .byte  68,15,40,5,41,207,0,0               // movaps        0xcf29(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,9,230,0,0                // movaps        0xe609(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,240                        // mulps         %xmm8,%xmm6
   .byte  102,65,15,111,233                   // movdqa        %xmm9,%xmm5
   .byte  102,15,114,213,8                    // psrld         $0x8,%xmm5
@@ -47136,11 +47136,11 @@
   .byte  102,67,15,110,12,136                // movd          (%r8,%r9,4),%xmm1
   .byte  102,15,98,200                       // punpckldq     %xmm0,%xmm1
   .byte  102,68,15,108,201                   // punpcklqdq    %xmm1,%xmm9
-  .byte  102,15,111,5,144,204,0,0            // movdqa        0xcc90(%rip),%xmm0        # 33480 <_sk_srcover_rgba_8888_sse2_lowp+0xb00>
+  .byte  102,15,111,5,112,227,0,0            // movdqa        0xe370(%rip),%xmm0        # 34b60 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
   .byte  15,91,209                           // cvtdq2ps      %xmm1,%xmm2
-  .byte  68,15,40,5,220,205,0,0              // movaps        0xcddc(%rip),%xmm8        # 335e0 <_sk_srcover_rgba_8888_sse2_lowp+0xc60>
+  .byte  68,15,40,5,188,228,0,0              // movaps        0xe4bc(%rip),%xmm8        # 34cc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc90>
   .byte  65,15,89,208                        // mulps         %xmm8,%xmm2
   .byte  102,65,15,111,201                   // movdqa        %xmm9,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -47169,7 +47169,7 @@
   .byte  73,193,224,2                        // shl           $0x2,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  72,99,194                           // movslq        %edx,%rax
-  .byte  68,15,40,5,77,204,0,0               // movaps        0xcc4d(%rip),%xmm8        # 334b0 <_sk_srcover_rgba_8888_sse2_lowp+0xb30>
+  .byte  68,15,40,5,45,227,0,0               // movaps        0xe32d(%rip),%xmm8        # 34b90 <_sk_srcover_rgba_8888_sse2_lowp+0xb60>
   .byte  68,15,40,202                        // movaps        %xmm2,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  102,69,15,91,201                    // cvtps2dq      %xmm9,%xmm9
@@ -47231,18 +47231,18 @@
   .byte  102,69,15,239,210                   // pxor          %xmm10,%xmm10
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,97,202                    // punpcklwd     %xmm10,%xmm1
-  .byte  102,68,15,111,13,71,207,0,0         // movdqa        0xcf47(%rip),%xmm9        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,13,39,230,0,0         // movdqa        0xe627(%rip),%xmm9        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,68,15,111,225                   // movdqa        %xmm1,%xmm12
   .byte  102,69,15,219,225                   // pand          %xmm9,%xmm12
-  .byte  102,68,15,111,29,68,207,0,0         // movdqa        0xcf44(%rip),%xmm11        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,29,36,230,0,0         // movdqa        0xe624(%rip),%xmm11        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,203                   // pand          %xmm11,%xmm1
-  .byte  102,15,111,29,71,207,0,0            // movdqa        0xcf47(%rip),%xmm3        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,29,39,230,0,0            // movdqa        0xe627(%rip),%xmm3        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,65,15,114,244,16                // pslld         $0x10,%xmm12
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,102,193                      // pcmpgtd       %xmm1,%xmm0
   .byte  102,15,114,241,13                   // pslld         $0xd,%xmm1
   .byte  102,65,15,235,204                   // por           %xmm12,%xmm1
-  .byte  102,68,15,111,37,54,207,0,0         // movdqa        0xcf36(%rip),%xmm12        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,37,22,230,0,0         // movdqa        0xe616(%rip),%xmm12        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,204                   // paddd         %xmm12,%xmm1
   .byte  102,15,223,193                      // pandn         %xmm1,%xmm0
   .byte  102,65,15,105,210                   // punpckhwd     %xmm10,%xmm2
@@ -47316,18 +47316,18 @@
   .byte  102,69,15,239,210                   // pxor          %xmm10,%xmm10
   .byte  102,15,111,238                      // movdqa        %xmm6,%xmm5
   .byte  102,65,15,97,234                    // punpcklwd     %xmm10,%xmm5
-  .byte  102,68,15,111,13,190,205,0,0        // movdqa        0xcdbe(%rip),%xmm9        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,13,158,228,0,0        // movdqa        0xe49e(%rip),%xmm9        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,68,15,111,229                   // movdqa        %xmm5,%xmm12
   .byte  102,69,15,219,225                   // pand          %xmm9,%xmm12
-  .byte  102,68,15,111,29,187,205,0,0        // movdqa        0xcdbb(%rip),%xmm11        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,29,155,228,0,0        // movdqa        0xe49b(%rip),%xmm11        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,235                   // pand          %xmm11,%xmm5
-  .byte  102,15,111,61,190,205,0,0           // movdqa        0xcdbe(%rip),%xmm7        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,61,158,228,0,0           // movdqa        0xe49e(%rip),%xmm7        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,65,15,114,244,16                // pslld         $0x10,%xmm12
   .byte  102,15,111,231                      // movdqa        %xmm7,%xmm4
   .byte  102,15,102,229                      // pcmpgtd       %xmm5,%xmm4
   .byte  102,15,114,245,13                   // pslld         $0xd,%xmm5
   .byte  102,65,15,235,236                   // por           %xmm12,%xmm5
-  .byte  102,68,15,111,37,173,205,0,0        // movdqa        0xcdad(%rip),%xmm12        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,37,141,228,0,0        // movdqa        0xe48d(%rip),%xmm12        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,236                   // paddd         %xmm12,%xmm5
   .byte  102,15,223,229                      // pandn         %xmm5,%xmm4
   .byte  102,65,15,105,242                   // punpckhwd     %xmm10,%xmm6
@@ -47429,18 +47429,18 @@
   .byte  102,68,15,105,201                   // punpckhwd     %xmm1,%xmm9
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,65,15,97,200                    // punpcklwd     %xmm8,%xmm1
-  .byte  102,68,15,111,21,188,203,0,0        // movdqa        0xcbbc(%rip),%xmm10        # 338a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf20>
+  .byte  102,68,15,111,21,156,226,0,0        // movdqa        0xe29c(%rip),%xmm10        # 34f80 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
   .byte  102,68,15,111,225                   // movdqa        %xmm1,%xmm12
   .byte  102,69,15,219,226                   // pand          %xmm10,%xmm12
-  .byte  102,68,15,111,29,185,203,0,0        // movdqa        0xcbb9(%rip),%xmm11        # 338b0 <_sk_srcover_rgba_8888_sse2_lowp+0xf30>
+  .byte  102,68,15,111,29,153,226,0,0        // movdqa        0xe299(%rip),%xmm11        # 34f90 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
   .byte  102,65,15,219,203                   // pand          %xmm11,%xmm1
-  .byte  102,15,111,29,188,203,0,0           // movdqa        0xcbbc(%rip),%xmm3        # 338c0 <_sk_srcover_rgba_8888_sse2_lowp+0xf40>
+  .byte  102,15,111,29,156,226,0,0           // movdqa        0xe29c(%rip),%xmm3        # 34fa0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
   .byte  102,65,15,114,244,16                // pslld         $0x10,%xmm12
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,102,193                      // pcmpgtd       %xmm1,%xmm0
   .byte  102,15,114,241,13                   // pslld         $0xd,%xmm1
   .byte  102,65,15,235,204                   // por           %xmm12,%xmm1
-  .byte  102,68,15,111,37,171,203,0,0        // movdqa        0xcbab(%rip),%xmm12        # 338d0 <_sk_srcover_rgba_8888_sse2_lowp+0xf50>
+  .byte  102,68,15,111,37,139,226,0,0        // movdqa        0xe28b(%rip),%xmm12        # 34fb0 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
   .byte  102,65,15,254,204                   // paddd         %xmm12,%xmm1
   .byte  102,15,223,193                      // pandn         %xmm1,%xmm0
   .byte  102,65,15,105,208                   // punpckhwd     %xmm8,%xmm2
@@ -47483,19 +47483,19 @@
 FUNCTION(_sk_store_f16_sse2)
 _sk_store_f16_sse2:
   .byte  15,41,124,36,232                    // movaps        %xmm7,-0x18(%rsp)
-  .byte  102,68,15,111,29,255,202,0,0        // movdqa        0xcaff(%rip),%xmm11        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
+  .byte  102,68,15,111,29,223,225,0,0        // movdqa        0xe1df(%rip),%xmm11        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
   .byte  102,69,15,219,195                   // pand          %xmm11,%xmm8
-  .byte  102,68,15,111,21,252,202,0,0        // movdqa        0xcafc(%rip),%xmm10        # 338f0 <_sk_srcover_rgba_8888_sse2_lowp+0xf70>
+  .byte  102,68,15,111,21,220,225,0,0        // movdqa        0xe1dc(%rip),%xmm10        # 34fd0 <_sk_srcover_rgba_8888_sse2_lowp+0xfa0>
   .byte  102,68,15,111,240                   // movdqa        %xmm0,%xmm14
   .byte  102,69,15,219,242                   // pand          %xmm10,%xmm14
-  .byte  102,15,111,61,250,202,0,0           // movdqa        0xcafa(%rip),%xmm7        # 33900 <_sk_srcover_rgba_8888_sse2_lowp+0xf80>
-  .byte  102,68,15,111,37,161,204,0,0        // movdqa        0xcca1(%rip),%xmm12        # 33ab0 <_sk_srcover_rgba_8888_sse2_lowp+0x1130>
+  .byte  102,15,111,61,218,225,0,0           // movdqa        0xe1da(%rip),%xmm7        # 34fe0 <_sk_srcover_rgba_8888_sse2_lowp+0xfb0>
+  .byte  102,68,15,111,37,129,227,0,0        // movdqa        0xe381(%rip),%xmm12        # 35190 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
   .byte  102,68,15,111,248                   // movdqa        %xmm0,%xmm15
   .byte  102,65,15,114,247,3                 // pslld         $0x3,%xmm15
   .byte  102,69,15,219,252                   // pand          %xmm12,%xmm15
   .byte  102,69,15,254,248                   // paddd         %xmm8,%xmm15
-  .byte  102,68,15,111,45,83,199,0,0         // movdqa        0xc753(%rip),%xmm13        # 33580 <_sk_srcover_rgba_8888_sse2_lowp+0xc00>
+  .byte  102,68,15,111,45,51,222,0,0         // movdqa        0xde33(%rip),%xmm13        # 34c60 <_sk_srcover_rgba_8888_sse2_lowp+0xc30>
   .byte  102,69,15,254,253                   // paddd         %xmm13,%xmm15
   .byte  102,65,15,114,231,16                // psrad         $0x10,%xmm15
   .byte  102,68,15,111,199                   // movdqa        %xmm7,%xmm8
@@ -47602,7 +47602,7 @@
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,65,15,97,201                    // punpcklwd     %xmm9,%xmm1
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  68,15,40,5,104,198,0,0              // movaps        0xc668(%rip),%xmm8        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  68,15,40,5,72,221,0,0               // movaps        0xdd48(%rip),%xmm8        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -47677,7 +47677,7 @@
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,65,15,97,200                    // punpcklwd     %xmm8,%xmm1
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  68,15,40,13,50,197,0,0              // movaps        0xc532(%rip),%xmm9        # 33670 <_sk_srcover_rgba_8888_sse2_lowp+0xcf0>
+  .byte  68,15,40,13,18,220,0,0              // movaps        0xdc12(%rip),%xmm9        # 34d50 <_sk_srcover_rgba_8888_sse2_lowp+0xd20>
   .byte  65,15,89,193                        // mulps         %xmm9,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -47694,7 +47694,7 @@
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,73,194,0,0                 // movaps        0xc249(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,41,217,0,0                 // movaps        0xd929(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,20,64                 // movd          (%r8,%rax,2),%xmm2
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
@@ -47728,7 +47728,7 @@
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  73,99,193                           // movslq        %r9d,%rax
-  .byte  68,15,40,21,18,199,0,0              // movaps        0xc712(%rip),%xmm10        # 33920 <_sk_srcover_rgba_8888_sse2_lowp+0xfa0>
+  .byte  68,15,40,21,242,221,0,0             // movaps        0xddf2(%rip),%xmm10        # 35000 <_sk_srcover_rgba_8888_sse2_lowp+0xfd0>
   .byte  68,15,40,192                        // movaps        %xmm0,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
   .byte  102,69,15,91,192                    // cvtps2dq      %xmm8,%xmm8
@@ -47935,7 +47935,7 @@
   .byte  243,69,15,91,209                    // cvttps2dq     %xmm9,%xmm10
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,84,13,177,190,0,0             // andps         0xbeb1(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,84,13,145,213,0,0             // andps         0xd591(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
@@ -47955,7 +47955,7 @@
   .byte  243,69,15,91,209                    // cvttps2dq     %xmm9,%xmm10
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,84,13,112,190,0,0             // andps         0xbe70(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,84,13,80,213,0,0              // andps         0xd550(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  69,15,89,194                        // mulps         %xmm10,%xmm8
@@ -47975,13 +47975,13 @@
   .byte  65,15,92,194                        // subps         %xmm10,%xmm0
   .byte  243,69,15,88,192                    // addss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
-  .byte  243,68,15,89,13,138,182,0,0         // mulss         0xb68a(%rip),%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,13,58,205,0,0          // mulss         0xcd3a(%rip),%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
   .byte  243,69,15,91,217                    // cvttps2dq     %xmm9,%xmm11
   .byte  69,15,91,219                        // cvtdq2ps      %xmm11,%xmm11
   .byte  69,15,194,203,1                     // cmpltps       %xmm11,%xmm9
-  .byte  68,15,84,13,15,190,0,0              // andps         0xbe0f(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,84,13,239,212,0,0             // andps         0xd4ef(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,217                        // subps         %xmm9,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -48004,13 +48004,13 @@
   .byte  65,15,92,202                        // subps         %xmm10,%xmm1
   .byte  243,69,15,88,192                    // addss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
-  .byte  243,68,15,89,13,30,182,0,0          // mulss         0xb61e(%rip),%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,13,206,204,0,0         // mulss         0xccce(%rip),%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,89,201                        // mulps         %xmm1,%xmm9
   .byte  243,69,15,91,217                    // cvttps2dq     %xmm9,%xmm11
   .byte  69,15,91,219                        // cvtdq2ps      %xmm11,%xmm11
   .byte  69,15,194,203,1                     // cmpltps       %xmm11,%xmm9
-  .byte  68,15,84,13,163,189,0,0             // andps         0xbda3(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,84,13,131,212,0,0             // andps         0xd483(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,87,228                        // xorps         %xmm12,%xmm12
   .byte  69,15,92,217                        // subps         %xmm9,%xmm11
   .byte  69,15,89,216                        // mulps         %xmm8,%xmm11
@@ -48027,7 +48027,7 @@
 _sk_clamp_x_1_sse2:
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
   .byte  68,15,95,192                        // maxps         %xmm0,%xmm8
-  .byte  68,15,93,5,115,189,0,0              // minps         0xbd73(%rip),%xmm8        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,93,5,83,212,0,0               // minps         0xd453(%rip),%xmm8        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  65,15,40,192                        // movaps        %xmm8,%xmm0
   .byte  255,224                             // jmpq          *%rax
@@ -48040,7 +48040,7 @@
   .byte  69,15,91,200                        // cvtdq2ps      %xmm8,%xmm9
   .byte  68,15,40,208                        // movaps        %xmm0,%xmm10
   .byte  69,15,194,209,1                     // cmpltps       %xmm9,%xmm10
-  .byte  68,15,40,29,81,189,0,0              // movaps        0xbd51(%rip),%xmm11        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,29,49,212,0,0              // movaps        0xd431(%rip),%xmm11        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,211                        // andps         %xmm11,%xmm10
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
@@ -48055,14 +48055,14 @@
 .globl _sk_mirror_x_1_sse2
 FUNCTION(_sk_mirror_x_1_sse2)
 _sk_mirror_x_1_sse2:
-  .byte  68,15,40,13,137,189,0,0             // movaps        0xbd89(%rip),%xmm9        # 33430 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
+  .byte  68,15,40,13,105,212,0,0             // movaps        0xd469(%rip),%xmm9        # 34b10 <_sk_srcover_rgba_8888_sse2_lowp+0xae0>
   .byte  65,15,88,193                        // addps         %xmm9,%xmm0
-  .byte  68,15,40,21,13,189,0,0              // movaps        0xbd0d(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,21,237,211,0,0             // movaps        0xd3ed(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,89,208                        // mulps         %xmm0,%xmm10
   .byte  243,69,15,91,194                    // cvttps2dq     %xmm10,%xmm8
   .byte  69,15,91,216                        // cvtdq2ps      %xmm8,%xmm11
   .byte  69,15,194,211,1                     // cmpltps       %xmm11,%xmm10
-  .byte  68,15,40,37,3,189,0,0               // movaps        0xbd03(%rip),%xmm12        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,37,227,211,0,0             // movaps        0xd3e3(%rip),%xmm12        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,212                        // andps         %xmm12,%xmm10
   .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
   .byte  69,15,92,218                        // subps         %xmm10,%xmm11
@@ -48083,10 +48083,10 @@
 FUNCTION(_sk_luminance_to_alpha_sse2)
 _sk_luminance_to_alpha_sse2:
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
-  .byte  15,89,5,37,194,0,0                  // mulps         0xc225(%rip),%xmm0        # 33930 <_sk_srcover_rgba_8888_sse2_lowp+0xfb0>
-  .byte  15,89,13,46,194,0,0                 // mulps         0xc22e(%rip),%xmm1        # 33940 <_sk_srcover_rgba_8888_sse2_lowp+0xfc0>
+  .byte  15,89,5,5,217,0,0                   // mulps         0xd905(%rip),%xmm0        # 35010 <_sk_srcover_rgba_8888_sse2_lowp+0xfe0>
+  .byte  15,89,13,14,217,0,0                 // mulps         0xd90e(%rip),%xmm1        # 35020 <_sk_srcover_rgba_8888_sse2_lowp+0xff0>
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,29,52,194,0,0                 // mulps         0xc234(%rip),%xmm3        # 33950 <_sk_srcover_rgba_8888_sse2_lowp+0xfd0>
+  .byte  15,89,29,20,217,0,0                 // mulps         0xd914(%rip),%xmm3        # 35030 <_sk_srcover_rgba_8888_sse2_lowp+0x1000>
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -48498,15 +48498,15 @@
 .globl _sk_gauss_a_to_rgba_sse2
 FUNCTION(_sk_gauss_a_to_rgba_sse2)
 _sk_gauss_a_to_rgba_sse2:
-  .byte  15,40,5,252,187,0,0                 // movaps        0xbbfc(%rip),%xmm0        # 33960 <_sk_srcover_rgba_8888_sse2_lowp+0xfe0>
+  .byte  15,40,5,220,210,0,0                 // movaps        0xd2dc(%rip),%xmm0        # 35040 <_sk_srcover_rgba_8888_sse2_lowp+0x1010>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,2,188,0,0                   // addps         0xbc02(%rip),%xmm0        # 33970 <_sk_srcover_rgba_8888_sse2_lowp+0xff0>
+  .byte  15,88,5,226,210,0,0                 // addps         0xd2e2(%rip),%xmm0        # 35050 <_sk_srcover_rgba_8888_sse2_lowp+0x1020>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,8,188,0,0                   // addps         0xbc08(%rip),%xmm0        # 33980 <_sk_srcover_rgba_8888_sse2_lowp+0x1000>
+  .byte  15,88,5,232,210,0,0                 // addps         0xd2e8(%rip),%xmm0        # 35060 <_sk_srcover_rgba_8888_sse2_lowp+0x1030>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,14,188,0,0                  // addps         0xbc0e(%rip),%xmm0        # 33990 <_sk_srcover_rgba_8888_sse2_lowp+0x1010>
+  .byte  15,88,5,238,210,0,0                 // addps         0xd2ee(%rip),%xmm0        # 35070 <_sk_srcover_rgba_8888_sse2_lowp+0x1040>
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,5,20,188,0,0                  // addps         0xbc14(%rip),%xmm0        # 339a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1020>
+  .byte  15,88,5,244,210,0,0                 // addps         0xd2f4(%rip),%xmm0        # 35080 <_sk_srcover_rgba_8888_sse2_lowp+0x1050>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
@@ -48672,29 +48672,29 @@
   .byte  69,15,94,220                        // divps         %xmm12,%xmm11
   .byte  69,15,40,227                        // movaps        %xmm11,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
-  .byte  68,15,40,45,164,185,0,0             // movaps        0xb9a4(%rip),%xmm13        # 339b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1030>
+  .byte  68,15,40,45,132,208,0,0             // movaps        0xd084(%rip),%xmm13        # 35090 <_sk_srcover_rgba_8888_sse2_lowp+0x1060>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,168,185,0,0             // addps         0xb9a8(%rip),%xmm13        # 339c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1040>
+  .byte  68,15,88,45,136,208,0,0             // addps         0xd088(%rip),%xmm13        # 350a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1070>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,172,185,0,0             // addps         0xb9ac(%rip),%xmm13        # 339d0 <_sk_srcover_rgba_8888_sse2_lowp+0x1050>
+  .byte  68,15,88,45,140,208,0,0             // addps         0xd08c(%rip),%xmm13        # 350b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
   .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
-  .byte  68,15,88,45,176,185,0,0             // addps         0xb9b0(%rip),%xmm13        # 339e0 <_sk_srcover_rgba_8888_sse2_lowp+0x1060>
+  .byte  68,15,88,45,144,208,0,0             // addps         0xd090(%rip),%xmm13        # 350c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
   .byte  69,15,89,235                        // mulps         %xmm11,%xmm13
   .byte  69,15,194,202,1                     // cmpltps       %xmm10,%xmm9
-  .byte  68,15,40,21,175,185,0,0             // movaps        0xb9af(%rip),%xmm10        # 339f0 <_sk_srcover_rgba_8888_sse2_lowp+0x1070>
+  .byte  68,15,40,21,143,208,0,0             // movaps        0xd08f(%rip),%xmm10        # 350d0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
   .byte  69,15,92,213                        // subps         %xmm13,%xmm10
   .byte  69,15,84,209                        // andps         %xmm9,%xmm10
   .byte  69,15,85,205                        // andnps        %xmm13,%xmm9
   .byte  69,15,86,202                        // orps          %xmm10,%xmm9
   .byte  68,15,194,192,1                     // cmpltps       %xmm0,%xmm8
-  .byte  68,15,40,21,98,179,0,0              // movaps        0xb362(%rip),%xmm10        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,21,66,202,0,0              // movaps        0xca42(%rip),%xmm10        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,92,209                        // subps         %xmm9,%xmm10
   .byte  69,15,84,208                        // andps         %xmm8,%xmm10
   .byte  69,15,85,193                        // andnps        %xmm9,%xmm8
   .byte  69,15,86,194                        // orps          %xmm10,%xmm8
   .byte  68,15,40,201                        // movaps        %xmm1,%xmm9
   .byte  68,15,194,200,1                     // cmpltps       %xmm0,%xmm9
-  .byte  68,15,40,21,81,179,0,0              // movaps        0xb351(%rip),%xmm10        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,21,49,202,0,0              // movaps        0xca31(%rip),%xmm10        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,208                        // subps         %xmm8,%xmm10
   .byte  69,15,84,209                        // andps         %xmm9,%xmm10
   .byte  69,15,85,200                        // andnps        %xmm8,%xmm9
@@ -48729,7 +48729,7 @@
   .byte  243,69,15,89,203                    // mulss         %xmm11,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,24,185,0,0              // mulps         0xb918(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,248,207,0,0             // mulps         0xcff8(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
@@ -48737,7 +48737,7 @@
   .byte  243,69,15,89,219                    // mulss         %xmm11,%xmm11
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,92,227                        // subps         %xmm11,%xmm12
-  .byte  68,15,89,21,3,185,0,0               // mulps         0xb903(%rip),%xmm10        # 33a10 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
+  .byte  68,15,89,21,227,207,0,0             // mulps         0xcfe3(%rip),%xmm10        # 350f0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
@@ -48746,8 +48746,8 @@
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,92,193                        // subps         %xmm9,%xmm0
-  .byte  68,15,87,13,171,183,0,0             // xorps         0xb7ab(%rip),%xmm9        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
-  .byte  68,15,89,5,131,178,0,0              // mulps         0xb283(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,87,13,139,206,0,0             // xorps         0xce8b(%rip),%xmm9        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
+  .byte  68,15,89,5,99,201,0,0               // mulps         0xc963(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -48768,7 +48768,7 @@
   .byte  243,69,15,89,203                    // mulss         %xmm11,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,122,184,0,0             // mulps         0xb87a(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,90,207,0,0              // mulps         0xcf5a(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,225                        // movaps        %xmm1,%xmm12
   .byte  69,15,89,228                        // mulps         %xmm12,%xmm12
@@ -48776,7 +48776,7 @@
   .byte  243,69,15,89,219                    // mulss         %xmm11,%xmm11
   .byte  69,15,198,219,0                     // shufps        $0x0,%xmm11,%xmm11
   .byte  69,15,92,227                        // subps         %xmm11,%xmm12
-  .byte  68,15,89,21,101,184,0,0             // mulps         0xb865(%rip),%xmm10        # 33a10 <_sk_srcover_rgba_8888_sse2_lowp+0x1090>
+  .byte  68,15,89,21,69,207,0,0              // mulps         0xcf45(%rip),%xmm10        # 350f0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
   .byte  69,15,89,212                        // mulps         %xmm12,%xmm10
   .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
@@ -48785,8 +48785,8 @@
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,40,194                        // movaps        %xmm10,%xmm0
   .byte  65,15,92,193                        // subps         %xmm9,%xmm0
-  .byte  68,15,87,13,13,183,0,0              // xorps         0xb70d(%rip),%xmm9        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
-  .byte  68,15,89,5,229,177,0,0              // mulps         0xb1e5(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,87,13,237,205,0,0             // xorps         0xcded(%rip),%xmm9        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
+  .byte  68,15,89,5,197,200,0,0              // mulps         0xc8c5(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
@@ -48804,7 +48804,7 @@
   .byte  243,69,15,89,200                    // mulss         %xmm8,%xmm9
   .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
   .byte  68,15,88,200                        // addps         %xmm0,%xmm9
-  .byte  68,15,89,13,237,183,0,0             // mulps         0xb7ed(%rip),%xmm9        # 33a00 <_sk_srcover_rgba_8888_sse2_lowp+0x1080>
+  .byte  68,15,89,13,205,206,0,0             // mulps         0xcecd(%rip),%xmm9        # 350e0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  68,15,40,209                        // movaps        %xmm1,%xmm10
   .byte  69,15,89,210                        // mulps         %xmm10,%xmm10
@@ -48812,7 +48812,7 @@
   .byte  243,69,15,89,192                    // mulss         %xmm8,%xmm8
   .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
   .byte  65,15,92,192                        // subps         %xmm8,%xmm0
-  .byte  15,87,5,169,182,0,0                 // xorps         0xb6a9(%rip),%xmm0        # 338e0 <_sk_srcover_rgba_8888_sse2_lowp+0xf60>
+  .byte  15,87,5,137,205,0,0                 // xorps         0xcd89(%rip),%xmm0        # 34fc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf90>
   .byte  65,15,94,193                        // divps         %xmm9,%xmm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -48858,7 +48858,7 @@
 FUNCTION(_sk_save_xy_sse2)
 _sk_save_xy_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,16,177,0,0               // movaps        0xb110(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,240,199,0,0              // movaps        0xc7f0(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,17,0                             // movups        %xmm0,(%rax)
   .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
   .byte  69,15,88,200                        // addps         %xmm8,%xmm9
@@ -48866,7 +48866,7 @@
   .byte  69,15,91,210                        // cvtdq2ps      %xmm10,%xmm10
   .byte  69,15,40,217                        // movaps        %xmm9,%xmm11
   .byte  69,15,194,218,1                     // cmpltps       %xmm10,%xmm11
-  .byte  68,15,40,37,251,176,0,0             // movaps        0xb0fb(%rip),%xmm12        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  68,15,40,37,219,199,0,0             // movaps        0xc7db(%rip),%xmm12        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,84,220                        // andps         %xmm12,%xmm11
   .byte  69,15,92,211                        // subps         %xmm11,%xmm10
   .byte  69,15,92,202                        // subps         %xmm10,%xmm9
@@ -48913,8 +48913,8 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,171,182,0,0                 // addps         0xb6ab(%rip),%xmm0        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,83,176,0,0              // movaps        0xb053(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,139,205,0,0                 // addps         0xcd8b(%rip),%xmm0        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,51,199,0,0              // movaps        0xc733(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,0,1,0,0                // movups        %xmm9,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -48927,7 +48927,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,31,176,0,0                  // addps         0xb01f(%rip),%xmm0        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,5,255,198,0,0                 // addps         0xc6ff(%rip),%xmm0        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -48939,8 +48939,8 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,94,182,0,0                 // addps         0xb65e(%rip),%xmm1        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,6,176,0,0               // movaps        0xb006(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,62,205,0,0                 // addps         0xcd3e(%rip),%xmm1        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,230,198,0,0             // movaps        0xc6e6(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  68,15,17,136,64,1,0,0               // movups        %xmm9,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -48953,7 +48953,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,209,175,0,0                // addps         0xafd1(%rip),%xmm1        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,13,177,198,0,0                // addps         0xc6b1(%rip),%xmm1        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -48965,13 +48965,13 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,33,182,0,0                  // addps         0xb621(%rip),%xmm0        # 33a30 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
-  .byte  68,15,40,13,185,175,0,0             // movaps        0xafb9(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,1,205,0,0                   // addps         0xcd01(%rip),%xmm0        # 35110 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,40,13,153,198,0,0             // movaps        0xc699(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,21,182,0,0              // mulps         0xb615(%rip),%xmm9        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,13,157,177,0,0             // addps         0xb19d(%rip),%xmm9        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,13,245,204,0,0             // mulps         0xccf5(%rip),%xmm9        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,13,125,200,0,0             // addps         0xc87d(%rip),%xmm9        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,0,1,0,0                // movups        %xmm9,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -48984,16 +48984,16 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,201,181,0,0                 // addps         0xb5c9(%rip),%xmm0        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,113,175,0,0             // movaps        0xaf71(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,5,169,204,0,0                 // addps         0xcca9(%rip),%xmm0        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,81,198,0,0              // movaps        0xc651(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,229,181,0,0              // movaps        0xb5e5(%rip),%xmm8        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,5,197,204,0,0              // movaps        0xccc5(%rip),%xmm8        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,233,181,0,0              // addps         0xb5e9(%rip),%xmm8        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,5,201,204,0,0              // addps         0xccc9(%rip),%xmm8        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,61,175,0,0               // addps         0xaf3d(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,88,5,29,198,0,0               // addps         0xc61d(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,225,181,0,0              // addps         0xb5e1(%rip),%xmm8        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,193,204,0,0              // addps         0xccc1(%rip),%xmm8        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -49003,17 +49003,17 @@
 FUNCTION(_sk_bicubic_p1x_sse2)
 _sk_bicubic_p1x_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,27,175,0,0               // movaps        0xaf1b(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,251,197,0,0              // movaps        0xc5fb(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,136,128,0,0,0              // movups        0x80(%rax),%xmm9
   .byte  65,15,88,192                        // addps         %xmm8,%xmm0
-  .byte  68,15,40,21,148,181,0,0             // movaps        0xb594(%rip),%xmm10        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,21,116,204,0,0             // movaps        0xcc74(%rip),%xmm10        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,152,181,0,0             // addps         0xb598(%rip),%xmm10        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,21,120,204,0,0             // addps         0xcc78(%rip),%xmm10        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,148,181,0,0             // addps         0xb594(%rip),%xmm10        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,21,116,204,0,0             // addps         0xcc74(%rip),%xmm10        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,144,0,1,0,0                // movups        %xmm10,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -49025,11 +49025,11 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,0                             // movups        (%rax),%xmm0
   .byte  68,15,16,128,128,0,0,0              // movups        0x80(%rax),%xmm8
-  .byte  15,88,5,100,181,0,0                 // addps         0xb564(%rip),%xmm0        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  15,88,5,68,204,0,0                  // addps         0xcc44(%rip),%xmm0        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,52,181,0,0               // mulps         0xb534(%rip),%xmm8        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,5,188,176,0,0              // addps         0xb0bc(%rip),%xmm8        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,5,20,204,0,0               // mulps         0xcc14(%rip),%xmm8        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,156,199,0,0              // addps         0xc79c(%rip),%xmm8        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,0,1,0,0                // movups        %xmm8,0x100(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -49042,13 +49042,13 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,247,180,0,0                // addps         0xb4f7(%rip),%xmm1        # 33a30 <_sk_srcover_rgba_8888_sse2_lowp+0x10b0>
-  .byte  68,15,40,13,143,174,0,0             // movaps        0xae8f(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,215,203,0,0                // addps         0xcbd7(%rip),%xmm1        # 35110 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,40,13,111,197,0,0             // movaps        0xc56f(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
   .byte  69,15,40,193                        // movaps        %xmm9,%xmm8
   .byte  69,15,89,192                        // mulps         %xmm8,%xmm8
-  .byte  68,15,89,13,235,180,0,0             // mulps         0xb4eb(%rip),%xmm9        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,13,115,176,0,0             // addps         0xb073(%rip),%xmm9        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,13,203,203,0,0             // mulps         0xcbcb(%rip),%xmm9        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,13,83,199,0,0              // addps         0xc753(%rip),%xmm9        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,200                        // mulps         %xmm8,%xmm9
   .byte  68,15,17,136,64,1,0,0               // movups        %xmm9,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -49061,16 +49061,16 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,158,180,0,0                // addps         0xb49e(%rip),%xmm1        # 33a20 <_sk_srcover_rgba_8888_sse2_lowp+0x10a0>
-  .byte  68,15,40,13,70,174,0,0              // movaps        0xae46(%rip),%xmm9        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,88,13,126,203,0,0                // addps         0xcb7e(%rip),%xmm1        # 35100 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,13,38,197,0,0              // movaps        0xc526(%rip),%xmm9        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  69,15,92,200                        // subps         %xmm8,%xmm9
-  .byte  68,15,40,5,186,180,0,0              // movaps        0xb4ba(%rip),%xmm8        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,5,154,203,0,0              // movaps        0xcb9a(%rip),%xmm8        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,190,180,0,0              // addps         0xb4be(%rip),%xmm8        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,5,158,203,0,0              // addps         0xcb9e(%rip),%xmm8        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,18,174,0,0               // addps         0xae12(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,88,5,242,196,0,0              // addps         0xc4f2(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
-  .byte  68,15,88,5,182,180,0,0              // addps         0xb4b6(%rip),%xmm8        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,150,203,0,0              // addps         0xcb96(%rip),%xmm8        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -49080,17 +49080,17 @@
 FUNCTION(_sk_bicubic_p1y_sse2)
 _sk_bicubic_p1y_sse2:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  68,15,40,5,240,173,0,0              // movaps        0xadf0(%rip),%xmm8        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  68,15,40,5,208,196,0,0              // movaps        0xc4d0(%rip),%xmm8        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,136,192,0,0,0              // movups        0xc0(%rax),%xmm9
   .byte  65,15,88,200                        // addps         %xmm8,%xmm1
-  .byte  68,15,40,21,104,180,0,0             // movaps        0xb468(%rip),%xmm10        # 33a50 <_sk_srcover_rgba_8888_sse2_lowp+0x10d0>
+  .byte  68,15,40,21,72,203,0,0              // movaps        0xcb48(%rip),%xmm10        # 35130 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,108,180,0,0             // addps         0xb46c(%rip),%xmm10        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  68,15,88,21,76,203,0,0              // addps         0xcb4c(%rip),%xmm10        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
   .byte  69,15,88,208                        // addps         %xmm8,%xmm10
   .byte  69,15,89,209                        // mulps         %xmm9,%xmm10
-  .byte  68,15,88,21,104,180,0,0             // addps         0xb468(%rip),%xmm10        # 33a70 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,21,72,203,0,0              // addps         0xcb48(%rip),%xmm10        # 35150 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
   .byte  68,15,17,144,64,1,0,0               // movups        %xmm10,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -49102,11 +49102,11 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,16,72,64                         // movups        0x40(%rax),%xmm1
   .byte  68,15,16,128,192,0,0,0              // movups        0xc0(%rax),%xmm8
-  .byte  15,88,13,55,180,0,0                 // addps         0xb437(%rip),%xmm1        # 33a60 <_sk_srcover_rgba_8888_sse2_lowp+0x10e0>
+  .byte  15,88,13,23,203,0,0                 // addps         0xcb17(%rip),%xmm1        # 35140 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
   .byte  69,15,40,200                        // movaps        %xmm8,%xmm9
   .byte  69,15,89,201                        // mulps         %xmm9,%xmm9
-  .byte  68,15,89,5,7,180,0,0                // mulps         0xb407(%rip),%xmm8        # 33a40 <_sk_srcover_rgba_8888_sse2_lowp+0x10c0>
-  .byte  68,15,88,5,143,175,0,0              // addps         0xaf8f(%rip),%xmm8        # 335d0 <_sk_srcover_rgba_8888_sse2_lowp+0xc50>
+  .byte  68,15,89,5,231,202,0,0              // mulps         0xcae7(%rip),%xmm8        # 35120 <_sk_srcover_rgba_8888_sse2_lowp+0x10f0>
+  .byte  68,15,88,5,111,198,0,0              // addps         0xc66f(%rip),%xmm8        # 34cb0 <_sk_srcover_rgba_8888_sse2_lowp+0xc80>
   .byte  69,15,89,193                        // mulps         %xmm9,%xmm8
   .byte  68,15,17,128,64,1,0,0               // movups        %xmm8,0x140(%rax)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -49251,7 +49251,7 @@
   .byte  102,65,15,254,205                   // paddd         %xmm13,%xmm1
   .byte  102,68,15,127,108,36,208            // movdqa        %xmm13,-0x30(%rsp)
   .byte  102,15,254,217                      // paddd         %xmm1,%xmm3
-  .byte  102,68,15,111,5,78,178,0,0          // movdqa        0xb24e(%rip),%xmm8        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,68,15,111,5,46,201,0,0          // movdqa        0xc92e(%rip),%xmm8        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,112,195,245                  // pshufd        $0xf5,%xmm3,%xmm0
   .byte  102,65,15,244,216                   // pmuludq       %xmm8,%xmm3
   .byte  102,65,15,244,192                   // pmuludq       %xmm8,%xmm0
@@ -49290,7 +49290,7 @@
   .byte  243,66,15,16,44,139                 // movss         (%rbx,%r9,4),%xmm5
   .byte  15,20,236                           // unpcklps      %xmm4,%xmm5
   .byte  102,15,20,221                       // unpcklpd      %xmm5,%xmm3
-  .byte  102,68,15,111,37,5,171,0,0          // movdqa        0xab05(%rip),%xmm12        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,37,229,193,0,0        // movdqa        0xc1e5(%rip),%xmm12        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,196                   // paddd         %xmm12,%xmm0
   .byte  102,15,112,224,78                   // pshufd        $0x4e,%xmm0,%xmm4
   .byte  102,73,15,126,224                   // movq          %xmm4,%r8
@@ -49306,7 +49306,7 @@
   .byte  243,66,15,16,36,139                 // movss         (%rbx,%r9,4),%xmm4
   .byte  15,20,224                           // unpcklps      %xmm0,%xmm4
   .byte  102,15,20,236                       // unpcklpd      %xmm4,%xmm5
-  .byte  15,40,37,58,177,0,0                 // movaps        0xb13a(%rip),%xmm4        # 33a80 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
+  .byte  15,40,37,26,200,0,0                 // movaps        0xc81a(%rip),%xmm4        # 35160 <_sk_srcover_rgba_8888_sse2_lowp+0x1130>
   .byte  68,15,88,220                        // addps         %xmm4,%xmm11
   .byte  68,15,41,92,36,192                  // movaps        %xmm11,-0x40(%rsp)
   .byte  68,15,40,223                        // movaps        %xmm7,%xmm11
@@ -49342,7 +49342,7 @@
   .byte  15,20,249                           // unpcklps      %xmm1,%xmm7
   .byte  102,15,20,252                       // unpcklpd      %xmm4,%xmm7
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
-  .byte  102,15,250,13,175,176,0,0           // psubd         0xb0af(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,143,199,0,0           // psubd         0xc78f(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,225,78                   // pshufd        $0x4e,%xmm1,%xmm4
   .byte  102,73,15,126,224                   // movq          %xmm4,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -49432,7 +49432,7 @@
   .byte  243,15,16,44,131                    // movss         (%rbx,%rax,4),%xmm5
   .byte  15,20,233                           // unpcklps      %xmm1,%xmm5
   .byte  102,15,20,221                       // unpcklpd      %xmm5,%xmm3
-  .byte  102,68,15,111,45,119,168,0,0        // movdqa        0xa877(%rip),%xmm13        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,45,87,191,0,0         // movdqa        0xbf57(%rip),%xmm13        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,197                   // paddd         %xmm13,%xmm0
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
   .byte  102,73,15,126,200                   // movq          %xmm1,%r8
@@ -49529,7 +49529,7 @@
   .byte  102,15,111,206                      // movdqa        %xmm6,%xmm1
   .byte  102,65,15,254,202                   // paddd         %xmm10,%xmm1
   .byte  102,15,112,209,245                  // pshufd        $0xf5,%xmm1,%xmm2
-  .byte  102,15,111,29,74,173,0,0            // movdqa        0xad4a(%rip),%xmm3        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,29,42,196,0,0            // movdqa        0xc42a(%rip),%xmm3        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,203                      // pmuludq       %xmm3,%xmm1
   .byte  102,15,244,211                      // pmuludq       %xmm3,%xmm2
   .byte  102,15,111,251                      // movdqa        %xmm3,%xmm7
@@ -49567,7 +49567,7 @@
   .byte  243,15,16,44,131                    // movss         (%rbx,%rax,4),%xmm5
   .byte  15,20,236                           // unpcklps      %xmm4,%xmm5
   .byte  102,15,20,213                       // unpcklpd      %xmm5,%xmm2
-  .byte  102,15,111,5,10,166,0,0             // movdqa        0xa60a(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,111,5,234,188,0,0            // movdqa        0xbcea(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,254,216                      // paddd         %xmm0,%xmm3
   .byte  102,15,112,227,78                   // pshufd        $0x4e,%xmm3,%xmm4
   .byte  102,73,15,126,224                   // movq          %xmm4,%r8
@@ -49648,7 +49648,7 @@
   .byte  102,15,111,214                      // movdqa        %xmm6,%xmm2
   .byte  102,65,15,254,212                   // paddd         %xmm12,%xmm2
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,13,39,171,0,0            // movdqa        0xab27(%rip),%xmm1        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,13,7,194,0,0             // movdqa        0xc207(%rip),%xmm1        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,209                      // pmuludq       %xmm1,%xmm2
   .byte  102,15,244,193                      // pmuludq       %xmm1,%xmm0
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
@@ -49670,7 +49670,7 @@
   .byte  68,15,20,209                        // unpcklps      %xmm1,%xmm10
   .byte  102,68,15,20,210                    // unpcklpd      %xmm2,%xmm10
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  102,15,250,13,204,170,0,0           // psubd         0xaacc(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,172,193,0,0           // psubd         0xc1ac(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
   .byte  102,73,15,126,208                   // movq          %xmm2,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -49685,7 +49685,7 @@
   .byte  243,15,16,20,131                    // movss         (%rbx,%rax,4),%xmm2
   .byte  15,20,209                           // unpcklps      %xmm1,%xmm2
   .byte  102,15,20,226                       // unpcklpd      %xmm2,%xmm4
-  .byte  102,15,254,5,231,163,0,0            // paddd         0xa3e7(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,5,199,186,0,0            // paddd         0xbac7(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
   .byte  102,73,15,126,200                   // movq          %xmm1,%r8
   .byte  102,72,15,126,192                   // movq          %xmm0,%rax
@@ -49722,7 +49722,7 @@
   .byte  15,20,199                           // unpcklps      %xmm7,%xmm0
   .byte  102,15,20,193                       // unpcklpd      %xmm1,%xmm0
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
-  .byte  102,15,250,13,214,169,0,0           // psubd         0xa9d6(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,182,192,0,0           // psubd         0xc0b6(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,249,78                   // pshufd        $0x4e,%xmm1,%xmm7
   .byte  102,73,15,126,248                   // movq          %xmm7,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -49737,7 +49737,7 @@
   .byte  243,15,16,52,131                    // movss         (%rbx,%rax,4),%xmm6
   .byte  15,20,247                           // unpcklps      %xmm7,%xmm6
   .byte  102,15,20,206                       // unpcklpd      %xmm6,%xmm1
-  .byte  102,15,254,21,241,162,0,0           // paddd         0xa2f1(%rip),%xmm2        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,21,209,185,0,0           // paddd         0xb9d1(%rip),%xmm2        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,242,78                   // pshufd        $0x4e,%xmm2,%xmm6
   .byte  102,73,15,126,240                   // movq          %xmm6,%r8
   .byte  102,72,15,126,208                   // movq          %xmm2,%rax
@@ -49883,7 +49883,7 @@
   .byte  102,65,15,254,210                   // paddd         %xmm10,%xmm2
   .byte  102,15,111,220                      // movdqa        %xmm4,%xmm3
   .byte  102,15,254,218                      // paddd         %xmm2,%xmm3
-  .byte  102,15,111,37,25,167,0,0            // movdqa        0xa719(%rip),%xmm4        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,37,249,189,0,0           // movdqa        0xbdf9(%rip),%xmm4        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,112,195,245                  // pshufd        $0xf5,%xmm3,%xmm0
   .byte  102,15,244,220                      // pmuludq       %xmm4,%xmm3
   .byte  102,15,244,196                      // pmuludq       %xmm4,%xmm0
@@ -49922,7 +49922,7 @@
   .byte  243,66,15,16,52,139                 // movss         (%rbx,%r9,4),%xmm6
   .byte  15,20,244                           // unpcklps      %xmm4,%xmm6
   .byte  102,15,20,238                       // unpcklpd      %xmm6,%xmm5
-  .byte  102,15,254,5,211,159,0,0            // paddd         0x9fd3(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,5,179,182,0,0            // paddd         0xb6b3(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,224,78                   // pshufd        $0x4e,%xmm0,%xmm4
   .byte  102,73,15,126,224                   // movq          %xmm4,%r8
   .byte  102,73,15,126,193                   // movq          %xmm0,%r9
@@ -49937,7 +49937,7 @@
   .byte  243,66,15,16,52,139                 // movss         (%rbx,%r9,4),%xmm6
   .byte  15,20,244                           // unpcklps      %xmm4,%xmm6
   .byte  102,15,20,222                       // unpcklpd      %xmm6,%xmm3
-  .byte  15,40,53,13,166,0,0                 // movaps        0xa60d(%rip),%xmm6        # 33a80 <_sk_srcover_rgba_8888_sse2_lowp+0x1100>
+  .byte  15,40,53,237,188,0,0                // movaps        0xbced(%rip),%xmm6        # 35160 <_sk_srcover_rgba_8888_sse2_lowp+0x1130>
   .byte  68,15,88,198                        // addps         %xmm6,%xmm8
   .byte  68,15,41,68,36,176                  // movaps        %xmm8,-0x50(%rsp)
   .byte  68,15,88,254                        // addps         %xmm6,%xmm15
@@ -49955,7 +49955,7 @@
   .byte  102,65,15,254,215                   // paddd         %xmm15,%xmm2
   .byte  102,68,15,127,124,36,224            // movdqa        %xmm15,-0x20(%rsp)
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,13,195,165,0,0           // movdqa        0xa5c3(%rip),%xmm1        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,13,163,188,0,0           // movdqa        0xbca3(%rip),%xmm1        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,209                      // pmuludq       %xmm1,%xmm2
   .byte  102,15,244,193                      // pmuludq       %xmm1,%xmm0
   .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
@@ -49976,7 +49976,7 @@
   .byte  15,20,240                           // unpcklps      %xmm0,%xmm6
   .byte  102,15,20,242                       // unpcklpd      %xmm2,%xmm6
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
-  .byte  102,15,250,5,110,165,0,0            // psubd         0xa56e(%rip),%xmm0        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,5,78,188,0,0             // psubd         0xbc4e(%rip),%xmm0        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,208,78                   // pshufd        $0x4e,%xmm0,%xmm2
   .byte  102,73,15,126,208                   // movq          %xmm2,%r8
   .byte  102,72,15,126,192                   // movq          %xmm0,%rax
@@ -49991,7 +49991,7 @@
   .byte  243,68,15,16,28,131                 // movss         (%rbx,%rax,4),%xmm11
   .byte  68,15,20,216                        // unpcklps      %xmm0,%xmm11
   .byte  102,65,15,20,211                    // unpcklpd      %xmm11,%xmm2
-  .byte  102,15,254,13,134,158,0,0           // paddd         0x9e86(%rip),%xmm1        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,13,102,181,0,0           // paddd         0xb566(%rip),%xmm1        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,193,78                   // pshufd        $0x4e,%xmm1,%xmm0
   .byte  102,73,15,126,192                   // movq          %xmm0,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -50029,7 +50029,7 @@
   .byte  102,65,15,111,194                   // movdqa        %xmm10,%xmm0
   .byte  102,15,254,68,36,160                // paddd         -0x60(%rsp),%xmm0
   .byte  102,15,112,216,245                  // pshufd        $0xf5,%xmm0,%xmm3
-  .byte  102,15,111,37,109,164,0,0           // movdqa        0xa46d(%rip),%xmm4        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,37,77,187,0,0            // movdqa        0xbb4d(%rip),%xmm4        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,196                      // pmuludq       %xmm4,%xmm0
   .byte  102,15,244,220                      // pmuludq       %xmm4,%xmm3
   .byte  102,68,15,111,220                   // movdqa        %xmm4,%xmm11
@@ -50067,7 +50067,7 @@
   .byte  243,15,16,60,131                    // movss         (%rbx,%rax,4),%xmm7
   .byte  15,20,253                           // unpcklps      %xmm5,%xmm7
   .byte  102,15,20,199                       // unpcklpd      %xmm7,%xmm0
-  .byte  102,68,15,111,5,41,157,0,0          // movdqa        0x9d29(%rip),%xmm8        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,5,9,180,0,0           // movdqa        0xb409(%rip),%xmm8        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,224                   // paddd         %xmm8,%xmm4
   .byte  102,15,112,236,78                   // pshufd        $0x4e,%xmm4,%xmm5
   .byte  102,73,15,126,232                   // movq          %xmm5,%r8
@@ -50213,7 +50213,7 @@
   .byte  243,15,16,36,131                    // movss         (%rbx,%rax,4),%xmm4
   .byte  15,20,226                           // unpcklps      %xmm2,%xmm4
   .byte  102,15,20,252                       // unpcklpd      %xmm4,%xmm7
-  .byte  102,68,15,111,61,131,154,0,0        // movdqa        0x9a83(%rip),%xmm15        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,61,99,177,0,0         // movdqa        0xb163(%rip),%xmm15        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,207                   // paddd         %xmm15,%xmm1
   .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
   .byte  102,73,15,126,208                   // movq          %xmm2,%r8
@@ -50296,7 +50296,7 @@
   .byte  102,65,15,111,217                   // movdqa        %xmm9,%xmm3
   .byte  102,65,15,254,216                   // paddd         %xmm8,%xmm3
   .byte  102,15,112,211,245                  // pshufd        $0xf5,%xmm3,%xmm2
-  .byte  102,15,111,45,162,159,0,0           // movdqa        0x9fa2(%rip),%xmm5        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,45,130,182,0,0           // movdqa        0xb682(%rip),%xmm5        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,221                      // pmuludq       %xmm5,%xmm3
   .byte  102,15,244,213                      // pmuludq       %xmm5,%xmm2
   .byte  102,15,112,234,232                  // pshufd        $0xe8,%xmm2,%xmm5
@@ -50317,7 +50317,7 @@
   .byte  68,15,20,198                        // unpcklps      %xmm6,%xmm8
   .byte  102,68,15,20,197                    // unpcklpd      %xmm5,%xmm8
   .byte  102,15,111,234                      // movdqa        %xmm2,%xmm5
-  .byte  102,15,250,45,75,159,0,0            // psubd         0x9f4b(%rip),%xmm5        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,45,43,182,0,0            // psubd         0xb62b(%rip),%xmm5        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,245,78                   // pshufd        $0x4e,%xmm5,%xmm6
   .byte  102,73,15,126,240                   // movq          %xmm6,%r8
   .byte  102,72,15,126,232                   // movq          %xmm5,%rax
@@ -50349,7 +50349,7 @@
   .byte  102,15,20,214                       // unpcklpd      %xmm6,%xmm2
   .byte  102,69,15,254,205                   // paddd         %xmm13,%xmm9
   .byte  102,65,15,112,233,245               // pshufd        $0xf5,%xmm9,%xmm5
-  .byte  102,15,111,29,169,158,0,0           // movdqa        0x9ea9(%rip),%xmm3        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,29,137,181,0,0           // movdqa        0xb589(%rip),%xmm3        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,68,15,244,203                   // pmuludq       %xmm3,%xmm9
   .byte  102,15,244,235                      // pmuludq       %xmm3,%xmm5
   .byte  102,15,112,237,232                  // pshufd        $0xe8,%xmm5,%xmm5
@@ -50370,7 +50370,7 @@
   .byte  69,15,20,233                        // unpcklps      %xmm9,%xmm13
   .byte  102,68,15,20,237                    // unpcklpd      %xmm5,%xmm13
   .byte  102,65,15,111,239                   // movdqa        %xmm15,%xmm5
-  .byte  102,15,250,45,73,158,0,0            // psubd         0x9e49(%rip),%xmm5        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,45,41,181,0,0            // psubd         0xb529(%rip),%xmm5        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,221,78                   // pshufd        $0x4e,%xmm5,%xmm3
   .byte  102,73,15,126,216                   // movq          %xmm3,%r8
   .byte  102,72,15,126,232                   // movq          %xmm5,%rax
@@ -50385,7 +50385,7 @@
   .byte  243,15,16,44,131                    // movss         (%rbx,%rax,4),%xmm5
   .byte  15,20,235                           // unpcklps      %xmm3,%xmm5
   .byte  102,15,20,245                       // unpcklpd      %xmm5,%xmm6
-  .byte  102,68,15,111,13,99,151,0,0         // movdqa        0x9763(%rip),%xmm9        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,13,67,174,0,0         // movdqa        0xae43(%rip),%xmm9        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,69,15,254,249                   // paddd         %xmm9,%xmm15
   .byte  102,65,15,112,223,78                // pshufd        $0x4e,%xmm15,%xmm3
   .byte  102,73,15,126,216                   // movq          %xmm3,%r8
@@ -50447,7 +50447,7 @@
   .byte  102,65,15,111,195                   // movdqa        %xmm11,%xmm0
   .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
   .byte  102,15,112,200,245                  // pshufd        $0xf5,%xmm0,%xmm1
-  .byte  102,15,111,21,228,156,0,0           // movdqa        0x9ce4(%rip),%xmm2        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,21,196,179,0,0           // movdqa        0xb3c4(%rip),%xmm2        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
   .byte  102,15,244,202                      // pmuludq       %xmm2,%xmm1
   .byte  102,68,15,111,234                   // movdqa        %xmm2,%xmm13
@@ -50640,7 +50640,7 @@
   .byte  68,15,20,238                        // unpcklps      %xmm6,%xmm13
   .byte  102,68,15,20,234                    // unpcklpd      %xmm2,%xmm13
   .byte  102,15,111,212                      // movdqa        %xmm4,%xmm2
-  .byte  102,15,250,21,128,153,0,0           // psubd         0x9980(%rip),%xmm2        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,21,96,176,0,0            // psubd         0xb060(%rip),%xmm2        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,242,78                   // pshufd        $0x4e,%xmm2,%xmm6
   .byte  102,73,15,126,240                   // movq          %xmm6,%r8
   .byte  102,72,15,126,208                   // movq          %xmm2,%rax
@@ -50700,7 +50700,7 @@
   .byte  102,68,15,111,116,36,160            // movdqa        -0x60(%rsp),%xmm14
   .byte  102,65,15,254,198                   // paddd         %xmm14,%xmm0
   .byte  102,15,112,200,245                  // pshufd        $0xf5,%xmm0,%xmm1
-  .byte  102,68,15,111,5,95,152,0,0          // movdqa        0x985f(%rip),%xmm8        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,68,15,111,5,63,175,0,0          // movdqa        0xaf3f(%rip),%xmm8        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,65,15,244,192                   // pmuludq       %xmm8,%xmm0
   .byte  102,65,15,244,200                   // pmuludq       %xmm8,%xmm1
   .byte  102,65,15,111,240                   // movdqa        %xmm8,%xmm6
@@ -50738,7 +50738,7 @@
   .byte  243,15,16,44,131                    // movss         (%rbx,%rax,4),%xmm5
   .byte  15,20,235                           // unpcklps      %xmm3,%xmm5
   .byte  102,15,20,205                       // unpcklpd      %xmm5,%xmm1
-  .byte  102,68,15,111,5,29,145,0,0          // movdqa        0x911d(%rip),%xmm8        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,68,15,111,5,253,167,0,0         // movdqa        0xa7fd(%rip),%xmm8        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,65,15,254,208                   // paddd         %xmm8,%xmm2
   .byte  102,15,112,218,78                   // pshufd        $0x4e,%xmm2,%xmm3
   .byte  102,73,15,126,216                   // movq          %xmm3,%r8
@@ -50821,7 +50821,7 @@
   .byte  102,15,254,208                      // paddd         %xmm0,%xmm2
   .byte  102,15,111,240                      // movdqa        %xmm0,%xmm6
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,37,57,150,0,0            // movdqa        0x9639(%rip),%xmm4        # 33a90 <_sk_srcover_rgba_8888_sse2_lowp+0x1110>
+  .byte  102,15,111,37,25,173,0,0            // movdqa        0xad19(%rip),%xmm4        # 35170 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
   .byte  102,15,244,212                      // pmuludq       %xmm4,%xmm2
   .byte  102,15,244,196                      // pmuludq       %xmm4,%xmm0
   .byte  102,15,112,200,232                  // pshufd        $0xe8,%xmm0,%xmm1
@@ -50842,7 +50842,7 @@
   .byte  68,15,20,193                        // unpcklps      %xmm1,%xmm8
   .byte  102,68,15,20,194                    // unpcklpd      %xmm2,%xmm8
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  102,15,250,13,226,149,0,0           // psubd         0x95e2(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,194,172,0,0           // psubd         0xacc2(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
   .byte  102,73,15,126,208                   // movq          %xmm2,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -50857,7 +50857,7 @@
   .byte  243,15,16,20,131                    // movss         (%rbx,%rax,4),%xmm2
   .byte  15,20,209                           // unpcklps      %xmm1,%xmm2
   .byte  102,68,15,20,242                    // unpcklpd      %xmm2,%xmm14
-  .byte  102,15,254,5,251,142,0,0            // paddd         0x8efb(%rip),%xmm0        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,5,219,165,0,0            // paddd         0xa5db(%rip),%xmm0        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
   .byte  102,73,15,126,200                   // movq          %xmm1,%r8
   .byte  102,72,15,126,192                   // movq          %xmm0,%rax
@@ -50895,7 +50895,7 @@
   .byte  15,20,198                           // unpcklps      %xmm6,%xmm0
   .byte  102,15,20,193                       // unpcklpd      %xmm1,%xmm0
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
-  .byte  102,15,250,13,234,148,0,0           // psubd         0x94ea(%rip),%xmm1        # 33aa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1120>
+  .byte  102,15,250,13,202,171,0,0           // psubd         0xabca(%rip),%xmm1        # 35180 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
   .byte  102,15,112,241,78                   // pshufd        $0x4e,%xmm1,%xmm6
   .byte  102,73,15,126,240                   // movq          %xmm6,%r8
   .byte  102,72,15,126,200                   // movq          %xmm1,%rax
@@ -50910,7 +50910,7 @@
   .byte  243,15,16,36,131                    // movss         (%rbx,%rax,4),%xmm4
   .byte  15,20,230                           // unpcklps      %xmm6,%xmm4
   .byte  102,15,20,204                       // unpcklpd      %xmm4,%xmm1
-  .byte  102,15,254,21,5,142,0,0             // paddd         0x8e05(%rip),%xmm2        # 33400 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
+  .byte  102,15,254,21,229,164,0,0           // paddd         0xa4e5(%rip),%xmm2        # 34ae0 <_sk_srcover_rgba_8888_sse2_lowp+0xab0>
   .byte  102,15,112,226,78                   // pshufd        $0x4e,%xmm2,%xmm4
   .byte  102,73,15,126,224                   // movq          %xmm4,%r8
   .byte  102,72,15,126,208                   // movq          %xmm2,%rax
@@ -50969,7 +50969,7 @@
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  65,15,88,215                        // addps         %xmm15,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,241,140,0,0                // movaps        0x8cf1(%rip),%xmm3        # 333d0 <_sk_srcover_rgba_8888_sse2_lowp+0xa50>
+  .byte  15,40,29,209,163,0,0                // movaps        0xa3d1(%rip),%xmm3        # 34ab0 <_sk_srcover_rgba_8888_sse2_lowp+0xa80>
   .byte  15,40,100,36,96                     // movaps        0x60(%rsp),%xmm4
   .byte  15,40,108,36,112                    // movaps        0x70(%rsp),%xmm5
   .byte  15,40,180,36,128,0,0,0              // movaps        0x80(%rsp),%xmm6
@@ -51072,12 +51072,42 @@
   .byte  197,249,110,209                     // vmovd         %ecx,%xmm2
   .byte  196,226,125,88,210                  // vpbroadcastd  %xmm2,%ymm2
   .byte  197,252,91,210                      // vcvtdq2ps     %ymm2,%ymm2
-  .byte  196,226,125,24,29,29,132,0,0        // vbroadcastss  0x841d(%rip),%ymm3        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  196,226,125,24,29,205,154,0,0       // vbroadcastss  0x9acd(%rip),%ymm3        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  197,236,88,211                      // vaddps        %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,252,40,218                      // vmovaps       %ymm2,%ymm3
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_translate_hsw_lowp
+.globl _sk_matrix_translate_hsw_lowp
+FUNCTION(_sk_matrix_translate_hsw_lowp)
+_sk_matrix_translate_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,24,0                     // vbroadcastss  (%rax),%ymm8
+  .byte  196,193,124,88,192                  // vaddps        %ymm8,%ymm0,%ymm0
+  .byte  196,193,116,88,200                  // vaddps        %ymm8,%ymm1,%ymm1
+  .byte  196,98,125,24,64,4                  // vbroadcastss  0x4(%rax),%ymm8
+  .byte  196,193,108,88,208                  // vaddps        %ymm8,%ymm2,%ymm2
+  .byte  196,193,100,88,216                  // vaddps        %ymm8,%ymm3,%ymm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_matrix_scale_translate_hsw_lowp
+.globl _sk_matrix_scale_translate_hsw_lowp
+FUNCTION(_sk_matrix_scale_translate_hsw_lowp)
+_sk_matrix_scale_translate_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,24,0                     // vbroadcastss  (%rax),%ymm8
+  .byte  196,98,125,24,72,8                  // vbroadcastss  0x8(%rax),%ymm9
+  .byte  196,194,61,168,193                  // vfmadd213ps   %ymm9,%ymm8,%ymm0
+  .byte  196,194,61,168,201                  // vfmadd213ps   %ymm9,%ymm8,%ymm1
+  .byte  196,98,125,24,64,4                  // vbroadcastss  0x4(%rax),%ymm8
+  .byte  196,98,125,24,72,12                 // vbroadcastss  0xc(%rax),%ymm9
+  .byte  196,194,61,168,209                  // vfmadd213ps   %ymm9,%ymm8,%ymm2
+  .byte  196,194,61,168,217                  // vfmadd213ps   %ymm9,%ymm8,%ymm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_matrix_2x3_hsw_lowp
 .globl _sk_matrix_2x3_hsw_lowp
 FUNCTION(_sk_matrix_2x3_hsw_lowp)
@@ -51104,6 +51134,43 @@
   .byte  197,124,41,194                      // vmovaps       %ymm8,%ymm2
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_perspective_hsw_lowp
+.globl _sk_matrix_perspective_hsw_lowp
+FUNCTION(_sk_matrix_perspective_hsw_lowp)
+_sk_matrix_perspective_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,24,8                     // vbroadcastss  (%rax),%ymm9
+  .byte  196,98,125,24,64,4                  // vbroadcastss  0x4(%rax),%ymm8
+  .byte  196,98,125,24,80,8                  // vbroadcastss  0x8(%rax),%ymm10
+  .byte  196,65,124,40,216                   // vmovaps       %ymm8,%ymm11
+  .byte  196,66,109,168,218                  // vfmadd213ps   %ymm10,%ymm2,%ymm11
+  .byte  196,66,101,168,194                  // vfmadd213ps   %ymm10,%ymm3,%ymm8
+  .byte  196,66,117,184,193                  // vfmadd231ps   %ymm9,%ymm1,%ymm8
+  .byte  196,66,125,184,217                  // vfmadd231ps   %ymm9,%ymm0,%ymm11
+  .byte  196,98,125,24,72,12                 // vbroadcastss  0xc(%rax),%ymm9
+  .byte  196,98,125,24,80,16                 // vbroadcastss  0x10(%rax),%ymm10
+  .byte  196,98,125,24,96,20                 // vbroadcastss  0x14(%rax),%ymm12
+  .byte  196,65,124,40,234                   // vmovaps       %ymm10,%ymm13
+  .byte  196,66,109,168,236                  // vfmadd213ps   %ymm12,%ymm2,%ymm13
+  .byte  196,66,101,168,212                  // vfmadd213ps   %ymm12,%ymm3,%ymm10
+  .byte  196,66,117,184,209                  // vfmadd231ps   %ymm9,%ymm1,%ymm10
+  .byte  196,66,125,184,233                  // vfmadd231ps   %ymm9,%ymm0,%ymm13
+  .byte  196,98,125,24,72,24                 // vbroadcastss  0x18(%rax),%ymm9
+  .byte  196,98,125,24,96,28                 // vbroadcastss  0x1c(%rax),%ymm12
+  .byte  196,98,125,24,112,32                // vbroadcastss  0x20(%rax),%ymm14
+  .byte  196,194,29,168,214                  // vfmadd213ps   %ymm14,%ymm12,%ymm2
+  .byte  196,66,101,168,230                  // vfmadd213ps   %ymm14,%ymm3,%ymm12
+  .byte  196,194,53,168,204                  // vfmadd213ps   %ymm12,%ymm9,%ymm1
+  .byte  196,194,125,184,209                 // vfmadd231ps   %ymm9,%ymm0,%ymm2
+  .byte  197,252,83,210                      // vrcpps        %ymm2,%ymm2
+  .byte  197,252,83,217                      // vrcpps        %ymm1,%ymm3
+  .byte  197,164,89,194                      // vmulps        %ymm2,%ymm11,%ymm0
+  .byte  197,188,89,203                      // vmulps        %ymm3,%ymm8,%ymm1
+  .byte  197,148,89,210                      // vmulps        %ymm2,%ymm13,%ymm2
+  .byte  197,172,89,219                      // vmulps        %ymm3,%ymm10,%ymm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_uniform_color_hsw_lowp
 .globl _sk_uniform_color_hsw_lowp
 FUNCTION(_sk_uniform_color_hsw_lowp)
@@ -51121,7 +51188,7 @@
 FUNCTION(_sk_black_color_hsw_lowp)
 _sk_black_color_hsw_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,29,123,137,0,0      // vpbroadcastw  0x897b(%rip),%ymm3        # 33220 <_sk_srcover_rgba_8888_sse2_lowp+0x8a0>
+  .byte  196,226,125,121,29,44,159,0,0       // vpbroadcastw  0x9f2c(%rip),%ymm3        # 348c0 <_sk_srcover_rgba_8888_sse2_lowp+0x890>
   .byte  197,252,87,192                      // vxorps        %ymm0,%ymm0,%ymm0
   .byte  197,244,87,201                      // vxorps        %ymm1,%ymm1,%ymm1
   .byte  197,236,87,210                      // vxorps        %ymm2,%ymm2,%ymm2
@@ -51132,7 +51199,7 @@
 FUNCTION(_sk_white_color_hsw_lowp)
 _sk_white_color_hsw_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,5,100,137,0,0       // vpbroadcastw  0x8964(%rip),%ymm0        # 33222 <_sk_srcover_rgba_8888_sse2_lowp+0x8a2>
+  .byte  196,226,125,121,5,21,159,0,0        // vpbroadcastw  0x9f15(%rip),%ymm0        # 348c2 <_sk_srcover_rgba_8888_sse2_lowp+0x892>
   .byte  197,253,111,200                     // vmovdqa       %ymm0,%ymm1
   .byte  197,253,111,208                     // vmovdqa       %ymm0,%ymm2
   .byte  197,253,111,216                     // vmovdqa       %ymm0,%ymm3
@@ -51143,8 +51210,8 @@
 FUNCTION(_sk_set_rgb_hsw_lowp)
 _sk_set_rgb_hsw_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  197,122,16,5,86,131,0,0             // vmovss        0x8356(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  .byte  197,250,16,21,126,131,0,0           // vmovss        0x837e(%rip),%xmm2        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  197,122,16,5,23,153,0,0             // vmovss        0x9917(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,250,16,21,63,153,0,0            // vmovss        0x993f(%rip),%xmm2        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  197,250,16,0                        // vmovss        (%rax),%xmm0
   .byte  196,226,57,153,194                  // vfmadd132ss   %xmm2,%xmm8,%xmm0
   .byte  197,122,44,192                      // vcvttss2si    %xmm0,%r8d
@@ -51167,7 +51234,7 @@
 FUNCTION(_sk_premul_hsw_lowp)
 _sk_premul_hsw_lowp:
   .byte  197,229,213,192                     // vpmullw       %ymm0,%ymm3,%ymm0
-  .byte  196,98,125,121,5,243,136,0,0        // vpbroadcastw  0x88f3(%rip),%ymm8        # 33224 <_sk_srcover_rgba_8888_sse2_lowp+0x8a4>
+  .byte  196,98,125,121,5,164,158,0,0        // vpbroadcastw  0x9ea4(%rip),%ymm8        # 348c4 <_sk_srcover_rgba_8888_sse2_lowp+0x894>
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,229,213,201                     // vpmullw       %ymm1,%ymm3,%ymm1
@@ -51215,7 +51282,7 @@
 .globl _sk_invert_hsw_lowp
 FUNCTION(_sk_invert_hsw_lowp)
 _sk_invert_hsw_lowp:
-  .byte  196,98,125,121,5,138,136,0,0        // vpbroadcastw  0x888a(%rip),%ymm8        # 33226 <_sk_srcover_rgba_8888_sse2_lowp+0x8a6>
+  .byte  196,98,125,121,5,59,158,0,0         // vpbroadcastw  0x9e3b(%rip),%ymm8        # 348c6 <_sk_srcover_rgba_8888_sse2_lowp+0x896>
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
   .byte  197,189,249,201                     // vpsubw        %ymm1,%ymm8,%ymm1
   .byte  197,189,249,210                     // vpsubw        %ymm2,%ymm8,%ymm2
@@ -51239,7 +51306,7 @@
 FUNCTION(_sk_srcatop_hsw_lowp)
 _sk_srcatop_hsw_lowp:
   .byte  197,197,213,192                     // vpmullw       %ymm0,%ymm7,%ymm0
-  .byte  196,98,125,121,5,87,136,0,0         // vpbroadcastw  0x8857(%rip),%ymm8        # 33228 <_sk_srcover_rgba_8888_sse2_lowp+0x8a8>
+  .byte  196,98,125,121,5,8,158,0,0          // vpbroadcastw  0x9e08(%rip),%ymm8        # 348c8 <_sk_srcover_rgba_8888_sse2_lowp+0x898>
   .byte  197,189,249,219                     // vpsubw        %ymm3,%ymm8,%ymm3
   .byte  197,101,213,204                     // vpmullw       %ymm4,%ymm3,%ymm9
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
@@ -51266,7 +51333,7 @@
 FUNCTION(_sk_dstatop_hsw_lowp)
 _sk_dstatop_hsw_lowp:
   .byte  197,93,213,195                      // vpmullw       %ymm3,%ymm4,%ymm8
-  .byte  196,98,125,121,13,247,135,0,0       // vpbroadcastw  0x87f7(%rip),%ymm9        # 3322a <_sk_srcover_rgba_8888_sse2_lowp+0x8aa>
+  .byte  196,98,125,121,13,168,157,0,0       // vpbroadcastw  0x9da8(%rip),%ymm9        # 348ca <_sk_srcover_rgba_8888_sse2_lowp+0x89a>
   .byte  197,53,249,215                      // vpsubw        %ymm7,%ymm9,%ymm10
   .byte  197,173,213,192                     // vpmullw       %ymm0,%ymm10,%ymm0
   .byte  196,65,61,253,193                   // vpaddw        %ymm9,%ymm8,%ymm8
@@ -51293,7 +51360,7 @@
 FUNCTION(_sk_srcin_hsw_lowp)
 _sk_srcin_hsw_lowp:
   .byte  197,197,213,192                     // vpmullw       %ymm0,%ymm7,%ymm0
-  .byte  196,98,125,121,5,148,135,0,0        // vpbroadcastw  0x8794(%rip),%ymm8        # 3322c <_sk_srcover_rgba_8888_sse2_lowp+0x8ac>
+  .byte  196,98,125,121,5,69,157,0,0         // vpbroadcastw  0x9d45(%rip),%ymm8        # 348cc <_sk_srcover_rgba_8888_sse2_lowp+0x89c>
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,197,213,201                     // vpmullw       %ymm1,%ymm7,%ymm1
@@ -51313,7 +51380,7 @@
 FUNCTION(_sk_dstin_hsw_lowp)
 _sk_dstin_hsw_lowp:
   .byte  197,221,213,195                     // vpmullw       %ymm3,%ymm4,%ymm0
-  .byte  196,98,125,121,5,81,135,0,0         // vpbroadcastw  0x8751(%rip),%ymm8        # 3322e <_sk_srcover_rgba_8888_sse2_lowp+0x8ae>
+  .byte  196,98,125,121,5,2,157,0,0          // vpbroadcastw  0x9d02(%rip),%ymm8        # 348ce <_sk_srcover_rgba_8888_sse2_lowp+0x89e>
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,213,213,203                     // vpmullw       %ymm3,%ymm5,%ymm1
@@ -51332,7 +51399,7 @@
 .globl _sk_srcout_hsw_lowp
 FUNCTION(_sk_srcout_hsw_lowp)
 _sk_srcout_hsw_lowp:
-  .byte  196,98,125,121,5,18,135,0,0         // vpbroadcastw  0x8712(%rip),%ymm8        # 33230 <_sk_srcover_rgba_8888_sse2_lowp+0x8b0>
+  .byte  196,98,125,121,5,195,156,0,0        // vpbroadcastw  0x9cc3(%rip),%ymm8        # 348d0 <_sk_srcover_rgba_8888_sse2_lowp+0x8a0>
   .byte  197,61,249,207                      // vpsubw        %ymm7,%ymm8,%ymm9
   .byte  197,181,213,192                     // vpmullw       %ymm0,%ymm9,%ymm0
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
@@ -51353,7 +51420,7 @@
 .globl _sk_dstout_hsw_lowp
 FUNCTION(_sk_dstout_hsw_lowp)
 _sk_dstout_hsw_lowp:
-  .byte  196,98,125,121,5,203,134,0,0        // vpbroadcastw  0x86cb(%rip),%ymm8        # 33232 <_sk_srcover_rgba_8888_sse2_lowp+0x8b2>
+  .byte  196,98,125,121,5,124,156,0,0        // vpbroadcastw  0x9c7c(%rip),%ymm8        # 348d2 <_sk_srcover_rgba_8888_sse2_lowp+0x8a2>
   .byte  197,189,249,219                     // vpsubw        %ymm3,%ymm8,%ymm3
   .byte  197,229,213,196                     // vpmullw       %ymm4,%ymm3,%ymm0
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
@@ -51374,7 +51441,7 @@
 .globl _sk_srcover_hsw_lowp
 FUNCTION(_sk_srcover_hsw_lowp)
 _sk_srcover_hsw_lowp:
-  .byte  196,98,125,121,5,132,134,0,0        // vpbroadcastw  0x8684(%rip),%ymm8        # 33234 <_sk_srcover_rgba_8888_sse2_lowp+0x8b4>
+  .byte  196,98,125,121,5,53,156,0,0         // vpbroadcastw  0x9c35(%rip),%ymm8        # 348d4 <_sk_srcover_rgba_8888_sse2_lowp+0x8a4>
   .byte  197,61,249,203                      // vpsubw        %ymm3,%ymm8,%ymm9
   .byte  197,53,213,212                      // vpmullw       %ymm4,%ymm9,%ymm10
   .byte  196,65,45,253,208                   // vpaddw        %ymm8,%ymm10,%ymm10
@@ -51399,7 +51466,7 @@
 .globl _sk_dstover_hsw_lowp
 FUNCTION(_sk_dstover_hsw_lowp)
 _sk_dstover_hsw_lowp:
-  .byte  196,98,125,121,5,41,134,0,0         // vpbroadcastw  0x8629(%rip),%ymm8        # 33236 <_sk_srcover_rgba_8888_sse2_lowp+0x8b6>
+  .byte  196,98,125,121,5,218,155,0,0        // vpbroadcastw  0x9bda(%rip),%ymm8        # 348d6 <_sk_srcover_rgba_8888_sse2_lowp+0x8a6>
   .byte  197,61,249,207                      // vpsubw        %ymm7,%ymm8,%ymm9
   .byte  197,181,213,192                     // vpmullw       %ymm0,%ymm9,%ymm0
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
@@ -51425,7 +51492,7 @@
 FUNCTION(_sk_modulate_hsw_lowp)
 _sk_modulate_hsw_lowp:
   .byte  197,221,213,192                     // vpmullw       %ymm0,%ymm4,%ymm0
-  .byte  196,98,125,121,5,206,133,0,0        // vpbroadcastw  0x85ce(%rip),%ymm8        # 33238 <_sk_srcover_rgba_8888_sse2_lowp+0x8b8>
+  .byte  196,98,125,121,5,127,155,0,0        // vpbroadcastw  0x9b7f(%rip),%ymm8        # 348d8 <_sk_srcover_rgba_8888_sse2_lowp+0x8a8>
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,213,213,201                     // vpmullw       %ymm1,%ymm5,%ymm1
@@ -51444,7 +51511,7 @@
 .globl _sk_multiply_hsw_lowp
 FUNCTION(_sk_multiply_hsw_lowp)
 _sk_multiply_hsw_lowp:
-  .byte  196,98,125,121,5,143,133,0,0        // vpbroadcastw  0x858f(%rip),%ymm8        # 3323a <_sk_srcover_rgba_8888_sse2_lowp+0x8ba>
+  .byte  196,98,125,121,5,64,155,0,0         // vpbroadcastw  0x9b40(%rip),%ymm8        # 348da <_sk_srcover_rgba_8888_sse2_lowp+0x8aa>
   .byte  197,61,249,203                      // vpsubw        %ymm3,%ymm8,%ymm9
   .byte  197,53,213,212                      // vpmullw       %ymm4,%ymm9,%ymm10
   .byte  196,65,93,253,216                   // vpaddw        %ymm8,%ymm4,%ymm11
@@ -51480,7 +51547,7 @@
 FUNCTION(_sk_plus__hsw_lowp)
 _sk_plus__hsw_lowp:
   .byte  197,221,253,192                     // vpaddw        %ymm0,%ymm4,%ymm0
-  .byte  196,98,125,121,5,5,133,0,0          // vpbroadcastw  0x8505(%rip),%ymm8        # 3323c <_sk_srcover_rgba_8888_sse2_lowp+0x8bc>
+  .byte  196,98,125,121,5,182,154,0,0        // vpbroadcastw  0x9ab6(%rip),%ymm8        # 348dc <_sk_srcover_rgba_8888_sse2_lowp+0x8ac>
   .byte  196,194,125,58,192                  // vpminuw       %ymm8,%ymm0,%ymm0
   .byte  197,213,253,201                     // vpaddw        %ymm1,%ymm5,%ymm1
   .byte  196,194,117,58,200                  // vpminuw       %ymm8,%ymm1,%ymm1
@@ -51497,7 +51564,7 @@
 _sk_screen_hsw_lowp:
   .byte  197,93,253,192                      // vpaddw        %ymm0,%ymm4,%ymm8
   .byte  197,221,213,192                     // vpmullw       %ymm0,%ymm4,%ymm0
-  .byte  196,98,125,121,13,210,132,0,0       // vpbroadcastw  0x84d2(%rip),%ymm9        # 3323e <_sk_srcover_rgba_8888_sse2_lowp+0x8be>
+  .byte  196,98,125,121,13,131,154,0,0       // vpbroadcastw  0x9a83(%rip),%ymm9        # 348de <_sk_srcover_rgba_8888_sse2_lowp+0x8ae>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
@@ -51523,7 +51590,7 @@
 .globl _sk_xor__hsw_lowp
 FUNCTION(_sk_xor__hsw_lowp)
 _sk_xor__hsw_lowp:
-  .byte  196,98,125,121,5,119,132,0,0        // vpbroadcastw  0x8477(%rip),%ymm8        # 33240 <_sk_srcover_rgba_8888_sse2_lowp+0x8c0>
+  .byte  196,98,125,121,5,40,154,0,0         // vpbroadcastw  0x9a28(%rip),%ymm8        # 348e0 <_sk_srcover_rgba_8888_sse2_lowp+0x8b0>
   .byte  197,61,249,207                      // vpsubw        %ymm7,%ymm8,%ymm9
   .byte  197,181,213,192                     // vpmullw       %ymm0,%ymm9,%ymm0
   .byte  197,61,249,211                      // vpsubw        %ymm3,%ymm8,%ymm10
@@ -51557,7 +51624,7 @@
   .byte  197,197,213,192                     // vpmullw       %ymm0,%ymm7,%ymm0
   .byte  197,93,213,203                      // vpmullw       %ymm3,%ymm4,%ymm9
   .byte  196,194,125,62,193                  // vpmaxuw       %ymm9,%ymm0,%ymm0
-  .byte  196,98,125,121,13,251,131,0,0       // vpbroadcastw  0x83fb(%rip),%ymm9        # 33242 <_sk_srcover_rgba_8888_sse2_lowp+0x8c2>
+  .byte  196,98,125,121,13,172,153,0,0       // vpbroadcastw  0x99ac(%rip),%ymm9        # 348e2 <_sk_srcover_rgba_8888_sse2_lowp+0x8b2>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
@@ -51591,7 +51658,7 @@
   .byte  197,197,213,192                     // vpmullw       %ymm0,%ymm7,%ymm0
   .byte  197,93,213,203                      // vpmullw       %ymm3,%ymm4,%ymm9
   .byte  196,194,125,58,193                  // vpminuw       %ymm9,%ymm0,%ymm0
-  .byte  196,98,125,121,13,124,131,0,0       // vpbroadcastw  0x837c(%rip),%ymm9        # 33244 <_sk_srcover_rgba_8888_sse2_lowp+0x8c4>
+  .byte  196,98,125,121,13,45,153,0,0        // vpbroadcastw  0x992d(%rip),%ymm9        # 348e4 <_sk_srcover_rgba_8888_sse2_lowp+0x8b4>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
@@ -51625,10 +51692,10 @@
   .byte  197,197,213,192                     // vpmullw       %ymm0,%ymm7,%ymm0
   .byte  197,93,213,203                      // vpmullw       %ymm3,%ymm4,%ymm9
   .byte  196,194,125,58,193                  // vpminuw       %ymm9,%ymm0,%ymm0
-  .byte  196,98,125,121,13,253,130,0,0       // vpbroadcastw  0x82fd(%rip),%ymm9        # 33246 <_sk_srcover_rgba_8888_sse2_lowp+0x8c6>
+  .byte  196,98,125,121,13,174,152,0,0       // vpbroadcastw  0x98ae(%rip),%ymm9        # 348e6 <_sk_srcover_rgba_8888_sse2_lowp+0x8b6>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,7                   // vpsrlw        $0x7,%ymm0,%ymm0
-  .byte  196,98,125,121,21,236,130,0,0       // vpbroadcastw  0x82ec(%rip),%ymm10        # 33248 <_sk_srcover_rgba_8888_sse2_lowp+0x8c8>
+  .byte  196,98,125,121,21,157,152,0,0       // vpbroadcastw  0x989d(%rip),%ymm10        # 348e8 <_sk_srcover_rgba_8888_sse2_lowp+0x8b8>
   .byte  196,193,125,219,194                 // vpand         %ymm10,%ymm0,%ymm0
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
   .byte  197,85,253,193                      // vpaddw        %ymm1,%ymm5,%ymm8
@@ -51661,10 +51728,10 @@
 _sk_exclusion_hsw_lowp:
   .byte  197,93,253,192                      // vpaddw        %ymm0,%ymm4,%ymm8
   .byte  197,221,213,192                     // vpmullw       %ymm0,%ymm4,%ymm0
-  .byte  196,98,125,121,13,113,130,0,0       // vpbroadcastw  0x8271(%rip),%ymm9        # 3324a <_sk_srcover_rgba_8888_sse2_lowp+0x8ca>
+  .byte  196,98,125,121,13,34,152,0,0        // vpbroadcastw  0x9822(%rip),%ymm9        # 348ea <_sk_srcover_rgba_8888_sse2_lowp+0x8ba>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,7                   // vpsrlw        $0x7,%ymm0,%ymm0
-  .byte  196,98,125,121,21,96,130,0,0        // vpbroadcastw  0x8260(%rip),%ymm10        # 3324c <_sk_srcover_rgba_8888_sse2_lowp+0x8cc>
+  .byte  196,98,125,121,21,17,152,0,0        // vpbroadcastw  0x9811(%rip),%ymm10        # 348ec <_sk_srcover_rgba_8888_sse2_lowp+0x8bc>
   .byte  196,193,125,219,194                 // vpand         %ymm10,%ymm0,%ymm0
   .byte  197,189,249,192                     // vpsubw        %ymm0,%ymm8,%ymm0
   .byte  197,85,253,193                      // vpaddw        %ymm1,%ymm5,%ymm8
@@ -51692,7 +51759,7 @@
 FUNCTION(_sk_hardlight_hsw_lowp)
 _sk_hardlight_hsw_lowp:
   .byte  197,125,253,192                     // vpaddw        %ymm0,%ymm0,%ymm8
-  .byte  196,98,125,121,13,253,129,0,0       // vpbroadcastw  0x81fd(%rip),%ymm9        # 33250 <_sk_srcover_rgba_8888_sse2_lowp+0x8d0>
+  .byte  196,98,125,121,13,174,151,0,0       // vpbroadcastw  0x97ae(%rip),%ymm9        # 348f0 <_sk_srcover_rgba_8888_sse2_lowp+0x8c0>
   .byte  196,65,101,239,209                  // vpxor         %ymm9,%ymm3,%ymm10
   .byte  196,65,61,239,217                   // vpxor         %ymm9,%ymm8,%ymm11
   .byte  196,65,37,101,218                   // vpcmpgtw      %ymm10,%ymm11,%ymm11
@@ -51720,7 +51787,7 @@
   .byte  197,101,249,210                     // vpsubw        %ymm2,%ymm3,%ymm10
   .byte  197,69,249,246                      // vpsubw        %ymm6,%ymm7,%ymm14
   .byte  196,65,45,213,214                   // vpmullw       %ymm14,%ymm10,%ymm10
-  .byte  196,98,125,121,53,116,129,0,0       // vpbroadcastw  0x8174(%rip),%ymm14        # 3324e <_sk_srcover_rgba_8888_sse2_lowp+0x8ce>
+  .byte  196,98,125,121,53,37,151,0,0        // vpbroadcastw  0x9725(%rip),%ymm14        # 348ee <_sk_srcover_rgba_8888_sse2_lowp+0x8be>
   .byte  196,65,45,253,210                   // vpaddw        %ymm10,%ymm10,%ymm10
   .byte  196,65,61,249,194                   // vpsubw        %ymm10,%ymm8,%ymm8
   .byte  197,13,249,215                      // vpsubw        %ymm7,%ymm14,%ymm10
@@ -51757,7 +51824,7 @@
 FUNCTION(_sk_overlay_hsw_lowp)
 _sk_overlay_hsw_lowp:
   .byte  197,93,253,196                      // vpaddw        %ymm4,%ymm4,%ymm8
-  .byte  196,98,125,121,13,233,128,0,0       // vpbroadcastw  0x80e9(%rip),%ymm9        # 33254 <_sk_srcover_rgba_8888_sse2_lowp+0x8d4>
+  .byte  196,98,125,121,13,154,150,0,0       // vpbroadcastw  0x969a(%rip),%ymm9        # 348f4 <_sk_srcover_rgba_8888_sse2_lowp+0x8c4>
   .byte  196,65,69,239,209                   // vpxor         %ymm9,%ymm7,%ymm10
   .byte  196,65,61,239,193                   // vpxor         %ymm9,%ymm8,%ymm8
   .byte  196,65,61,101,218                   // vpcmpgtw      %ymm10,%ymm8,%ymm11
@@ -51781,7 +51848,7 @@
   .byte  196,65,21,253,237                   // vpaddw        %ymm13,%ymm13,%ymm13
   .byte  196,65,61,249,237                   // vpsubw        %ymm13,%ymm8,%ymm13
   .byte  196,67,13,76,229,192                // vpblendvb     %ymm12,%ymm13,%ymm14,%ymm12
-  .byte  196,98,125,121,45,113,128,0,0       // vpbroadcastw  0x8071(%rip),%ymm13        # 33252 <_sk_srcover_rgba_8888_sse2_lowp+0x8d2>
+  .byte  196,98,125,121,45,34,150,0,0        // vpbroadcastw  0x9622(%rip),%ymm13        # 348f2 <_sk_srcover_rgba_8888_sse2_lowp+0x8c2>
   .byte  197,77,253,246                      // vpaddw        %ymm6,%ymm6,%ymm14
   .byte  196,65,13,239,201                   // vpxor         %ymm9,%ymm14,%ymm9
   .byte  196,65,53,101,202                   // vpcmpgtw      %ymm10,%ymm9,%ymm9
@@ -51833,10 +51900,10 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,55                              // ja            2b2d7 <_sk_load_8888_hsw_lowp+0x52>
+  .byte  119,55                              // ja            2b3c6 <_sk_load_8888_hsw_lowp+0x52>
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,162,1,0,0                 // lea           0x1a2(%rip),%r9        # 2b450 <_sk_load_8888_hsw_lowp+0x1cb>
+  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2b540 <_sk_load_8888_hsw_lowp+0x1cc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -51846,10 +51913,10 @@
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,12,144              // vmovd         (%r8,%rdx,4),%xmm1
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
-  .byte  233,39,1,0,0                        // jmpq          2b3fe <_sk_load_8888_hsw_lowp+0x179>
+  .byte  233,39,1,0,0                        // jmpq          2b4ed <_sk_load_8888_hsw_lowp+0x179>
   .byte  196,193,126,111,12,144              // vmovdqu       (%r8,%rdx,4),%ymm1
   .byte  196,193,126,111,68,144,32           // vmovdqu       0x20(%r8,%rdx,4),%ymm0
-  .byte  233,21,1,0,0                        // jmpq          2b3fe <_sk_load_8888_hsw_lowp+0x179>
+  .byte  233,21,1,0,0                        // jmpq          2b4ed <_sk_load_8888_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,8            // vmovd         0x8(%r8,%rdx,4),%xmm0
   .byte  196,226,121,89,192                  // vpbroadcastq  %xmm0,%xmm0
   .byte  197,229,239,219                     // vpxor         %ymm3,%ymm3,%ymm3
@@ -51858,7 +51925,7 @@
   .byte  197,249,112,201,232                 // vpshufd       $0xe8,%xmm1,%xmm1
   .byte  196,227,125,2,201,3                 // vpblendd      $0x3,%ymm1,%ymm0,%ymm1
   .byte  197,253,111,195                     // vmovdqa       %ymm3,%ymm0
-  .byte  233,229,0,0,0                       // jmpq          2b3fe <_sk_load_8888_hsw_lowp+0x179>
+  .byte  233,229,0,0,0                       // jmpq          2b4ed <_sk_load_8888_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,24           // vmovd         0x18(%r8,%rdx,4),%xmm0
   .byte  196,226,125,89,192                  // vpbroadcastq  %xmm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -51874,7 +51941,7 @@
   .byte  196,193,122,111,12,144              // vmovdqu       (%r8,%rdx,4),%xmm1
   .byte  196,227,117,2,200,240               // vpblendd      $0xf0,%ymm0,%ymm1,%ymm1
   .byte  197,125,127,192                     // vmovdqa       %ymm8,%ymm0
-  .byte  233,138,0,0,0                       // jmpq          2b3fe <_sk_load_8888_hsw_lowp+0x179>
+  .byte  233,138,0,0,0                       // jmpq          2b4ed <_sk_load_8888_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,40           // vmovd         0x28(%r8,%rdx,4),%xmm0
   .byte  196,226,121,89,192                  // vpbroadcastq  %xmm0,%xmm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -51884,7 +51951,7 @@
   .byte  196,193,121,110,76,144,32           // vmovd         0x20(%r8,%rdx,4),%xmm1
   .byte  196,227,125,2,193,1                 // vpblendd      $0x1,%ymm1,%ymm0,%ymm0
   .byte  196,193,126,111,12,144              // vmovdqu       (%r8,%rdx,4),%ymm1
-  .byte  235,81                              // jmp           2b3fe <_sk_load_8888_hsw_lowp+0x179>
+  .byte  235,81                              // jmp           2b4ed <_sk_load_8888_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,56           // vmovd         0x38(%r8,%rdx,4),%xmm0
   .byte  196,226,125,89,192                  // vpbroadcastq  %xmm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -51904,7 +51971,7 @@
   .byte  196,227,101,14,200,170              // vpblendw      $0xaa,%ymm0,%ymm3,%ymm1
   .byte  196,227,109,14,192,170              // vpblendw      $0xaa,%ymm0,%ymm2,%ymm0
   .byte  196,226,125,43,201                  // vpackusdw     %ymm1,%ymm0,%ymm1
-  .byte  197,125,111,5,57,126,0,0            // vmovdqa       0x7e39(%rip),%ymm8        # 33260 <_sk_srcover_rgba_8888_sse2_lowp+0x8e0>
+  .byte  197,125,111,5,234,147,0,0           // vmovdqa       0x93ea(%rip),%ymm8        # 34900 <_sk_srcover_rgba_8888_sse2_lowp+0x8d0>
   .byte  196,193,117,219,192                 // vpand         %ymm8,%ymm1,%ymm0
   .byte  197,245,113,209,8                   // vpsrlw        $0x8,%ymm1,%ymm1
   .byte  197,229,114,211,16                  // vpsrld        $0x10,%ymm3,%ymm3
@@ -51914,40 +51981,39 @@
   .byte  197,229,113,211,8                   // vpsrlw        $0x8,%ymm3,%ymm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
-  .byte  102,144                             // xchg          %ax,%ax
-  .byte  120,254                             // js            2b450 <_sk_load_8888_hsw_lowp+0x1cb>
+  .byte  15,31,0                             // nopl          (%rax)
+  .byte  119,254                             // ja            2b540 <_sk_load_8888_hsw_lowp+0x1cc>
   .byte  255                                 // (bad)
-  .byte  255,175,254,255,255,153             // ljmp          *-0x66000002(%rdi)
+  .byte  255,174,254,255,255,152             // ljmp          *-0x67000002(%rsi)
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,15                              // decl          (%rdi)
+  .byte  255,14                              // decl          (%rsi)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,247                             // push          %rdi
+  .byte  255,246                             // push          %rsi
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  223,254                             // (bad)
+  .byte  222,254                             // fdivrp        %st,%st(6)
   .byte  255                                 // (bad)
-  .byte  255,201                             // dec           %ecx
+  .byte  255,200                             // dec           %eax
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,85,255                          // callq         *-0x1(%rbp)
-  .byte  255                                 // (bad)
-  .byte  255,72,255                          // decl          -0x1(%rax)
+  .byte  255,84,255,255                      // callq         *-0x1(%rdi,%rdi,8)
+  .byte  255,71,255                          // incl          -0x1(%rdi)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  58,255                              // cmp           %bh,%bh
+  .byte  57,255                              // cmp           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,36,255                          // jmpq          *(%rdi,%rdi,8)
-  .byte  255                                 // (bad)
-  .byte  255,155,255,255,255,135             // lcall         *-0x78000001(%rbx)
+  .byte  255,35                              // jmpq          *(%rbx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,115,255                         // pushq         -0x1(%rbx)
+  .byte  255,154,255,255,255,134             // lcall         *-0x79000001(%rdx)
   .byte  255                                 // (bad)
-  .byte  255,93,255                          // lcall         *-0x1(%rbp)
   .byte  255                                 // (bad)
+  .byte  255,114,255                         // pushq         -0x1(%rdx)
+  .byte  255                                 // (bad)
+  .byte  255,92,255,255                      // lcall         *-0x1(%rdi,%rdi,8)
   .byte  255                                 // .byte         0xff
 
 HIDDEN _sk_load_8888_dst_hsw_lowp
@@ -51963,10 +52029,10 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,55                              // ja            2b4de <_sk_load_8888_dst_hsw_lowp+0x52>
+  .byte  119,55                              // ja            2b5ce <_sk_load_8888_dst_hsw_lowp+0x52>
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2b658 <_sk_load_8888_dst_hsw_lowp+0x1cc>
+  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2b748 <_sk_load_8888_dst_hsw_lowp+0x1cc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -51976,10 +52042,10 @@
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,44,144              // vmovd         (%r8,%rdx,4),%xmm5
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
-  .byte  233,39,1,0,0                        // jmpq          2b605 <_sk_load_8888_dst_hsw_lowp+0x179>
+  .byte  233,39,1,0,0                        // jmpq          2b6f5 <_sk_load_8888_dst_hsw_lowp+0x179>
   .byte  196,193,126,111,44,144              // vmovdqu       (%r8,%rdx,4),%ymm5
   .byte  196,193,126,111,100,144,32          // vmovdqu       0x20(%r8,%rdx,4),%ymm4
-  .byte  233,21,1,0,0                        // jmpq          2b605 <_sk_load_8888_dst_hsw_lowp+0x179>
+  .byte  233,21,1,0,0                        // jmpq          2b6f5 <_sk_load_8888_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,8           // vmovd         0x8(%r8,%rdx,4),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,197,239,255                     // vpxor         %ymm7,%ymm7,%ymm7
@@ -51988,7 +52054,7 @@
   .byte  197,249,112,237,232                 // vpshufd       $0xe8,%xmm5,%xmm5
   .byte  196,227,93,2,237,3                  // vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   .byte  197,253,111,231                     // vmovdqa       %ymm7,%ymm4
-  .byte  233,229,0,0,0                       // jmpq          2b605 <_sk_load_8888_dst_hsw_lowp+0x179>
+  .byte  233,229,0,0,0                       // jmpq          2b6f5 <_sk_load_8888_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,24          // vmovd         0x18(%r8,%rdx,4),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52004,7 +52070,7 @@
   .byte  196,193,122,111,44,144              // vmovdqu       (%r8,%rdx,4),%xmm5
   .byte  196,227,85,2,236,240                // vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   .byte  197,125,127,196                     // vmovdqa       %ymm8,%ymm4
-  .byte  233,138,0,0,0                       // jmpq          2b605 <_sk_load_8888_dst_hsw_lowp+0x179>
+  .byte  233,138,0,0,0                       // jmpq          2b6f5 <_sk_load_8888_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,40          // vmovd         0x28(%r8,%rdx,4),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52014,7 +52080,7 @@
   .byte  196,193,121,110,108,144,32          // vmovd         0x20(%r8,%rdx,4),%xmm5
   .byte  196,227,93,2,229,1                  // vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   .byte  196,193,126,111,44,144              // vmovdqu       (%r8,%rdx,4),%ymm5
-  .byte  235,81                              // jmp           2b605 <_sk_load_8888_dst_hsw_lowp+0x179>
+  .byte  235,81                              // jmp           2b6f5 <_sk_load_8888_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,56          // vmovd         0x38(%r8,%rdx,4),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52034,7 +52100,7 @@
   .byte  196,227,69,14,236,170               // vpblendw      $0xaa,%ymm4,%ymm7,%ymm5
   .byte  196,227,77,14,228,170               // vpblendw      $0xaa,%ymm4,%ymm6,%ymm4
   .byte  196,226,93,43,237                   // vpackusdw     %ymm5,%ymm4,%ymm5
-  .byte  197,125,111,5,82,124,0,0            // vmovdqa       0x7c52(%rip),%ymm8        # 33280 <_sk_srcover_rgba_8888_sse2_lowp+0x900>
+  .byte  197,125,111,5,2,146,0,0             // vmovdqa       0x9202(%rip),%ymm8        # 34920 <_sk_srcover_rgba_8888_sse2_lowp+0x8f0>
   .byte  196,193,85,219,224                  // vpand         %ymm8,%ymm5,%ymm4
   .byte  197,213,113,213,8                   // vpsrlw        $0x8,%ymm5,%ymm5
   .byte  197,197,114,215,16                  // vpsrld        $0x10,%ymm7,%ymm7
@@ -52045,7 +52111,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  119,254                             // ja            2b658 <_sk_load_8888_dst_hsw_lowp+0x1cc>
+  .byte  119,254                             // ja            2b748 <_sk_load_8888_dst_hsw_lowp+0x1cc>
   .byte  255                                 // (bad)
   .byte  255,174,254,255,255,152             // ljmp          *-0x67000002(%rsi)
   .byte  254                                 // (bad)
@@ -52106,20 +52172,20 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,30                              // ja            2b715 <_sk_store_8888_hsw_lowp+0x81>
+  .byte  119,30                              // ja            2b805 <_sk_store_8888_hsw_lowp+0x81>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2b7c4 <_sk_store_8888_hsw_lowp+0x130>
+  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2b8b4 <_sk_store_8888_hsw_lowp+0x130>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,65,121,126,4,144                // vmovd         %xmm8,(%r8,%rdx,4)
-  .byte  233,171,0,0,0                       // jmpq          2b7c0 <_sk_store_8888_hsw_lowp+0x12c>
+  .byte  233,171,0,0,0                       // jmpq          2b8b0 <_sk_store_8888_hsw_lowp+0x12c>
   .byte  196,65,126,127,4,144                // vmovdqu       %ymm8,(%r8,%rdx,4)
   .byte  196,65,126,127,76,144,32            // vmovdqu       %ymm9,0x20(%r8,%rdx,4)
-  .byte  233,153,0,0,0                       // jmpq          2b7c0 <_sk_store_8888_hsw_lowp+0x12c>
+  .byte  233,153,0,0,0                       // jmpq          2b8b0 <_sk_store_8888_hsw_lowp+0x12c>
   .byte  196,67,121,22,68,144,8,2            // vpextrd       $0x2,%xmm8,0x8(%r8,%rdx,4)
   .byte  196,65,121,214,4,144                // vmovq         %xmm8,(%r8,%rdx,4)
-  .byte  233,134,0,0,0                       // jmpq          2b7c0 <_sk_store_8888_hsw_lowp+0x12c>
+  .byte  233,134,0,0,0                       // jmpq          2b8b0 <_sk_store_8888_hsw_lowp+0x12c>
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,22,76,144,24,2           // vpextrd       $0x2,%xmm9,0x18(%r8,%rdx,4)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
@@ -52127,12 +52193,12 @@
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,65,121,126,76,144,16            // vmovd         %xmm9,0x10(%r8,%rdx,4)
   .byte  196,65,122,127,4,144                // vmovdqu       %xmm8,(%r8,%rdx,4)
-  .byte  235,85                              // jmp           2b7c0 <_sk_store_8888_hsw_lowp+0x12c>
+  .byte  235,85                              // jmp           2b8b0 <_sk_store_8888_hsw_lowp+0x12c>
   .byte  196,67,121,22,76,144,40,2           // vpextrd       $0x2,%xmm9,0x28(%r8,%rdx,4)
   .byte  196,67,121,22,76,144,36,1           // vpextrd       $0x1,%xmm9,0x24(%r8,%rdx,4)
   .byte  196,65,121,126,76,144,32            // vmovd         %xmm9,0x20(%r8,%rdx,4)
   .byte  196,65,126,127,4,144                // vmovdqu       %ymm8,(%r8,%rdx,4)
-  .byte  235,54                              // jmp           2b7c0 <_sk_store_8888_hsw_lowp+0x12c>
+  .byte  235,54                              // jmp           2b8b0 <_sk_store_8888_hsw_lowp+0x12c>
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
   .byte  196,67,121,22,84,144,56,2           // vpextrd       $0x2,%xmm10,0x38(%r8,%rdx,4)
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
@@ -52189,10 +52255,10 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,55                              // ja            2b852 <_sk_load_bgra_hsw_lowp+0x52>
+  .byte  119,55                              // ja            2b942 <_sk_load_bgra_hsw_lowp+0x52>
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2b9cc <_sk_load_bgra_hsw_lowp+0x1cc>
+  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2babc <_sk_load_bgra_hsw_lowp+0x1cc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -52202,10 +52268,10 @@
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,12,144              // vmovd         (%r8,%rdx,4),%xmm1
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
-  .byte  233,39,1,0,0                        // jmpq          2b979 <_sk_load_bgra_hsw_lowp+0x179>
+  .byte  233,39,1,0,0                        // jmpq          2ba69 <_sk_load_bgra_hsw_lowp+0x179>
   .byte  196,193,126,111,12,144              // vmovdqu       (%r8,%rdx,4),%ymm1
   .byte  196,193,126,111,68,144,32           // vmovdqu       0x20(%r8,%rdx,4),%ymm0
-  .byte  233,21,1,0,0                        // jmpq          2b979 <_sk_load_bgra_hsw_lowp+0x179>
+  .byte  233,21,1,0,0                        // jmpq          2ba69 <_sk_load_bgra_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,8            // vmovd         0x8(%r8,%rdx,4),%xmm0
   .byte  196,226,121,89,192                  // vpbroadcastq  %xmm0,%xmm0
   .byte  197,229,239,219                     // vpxor         %ymm3,%ymm3,%ymm3
@@ -52214,7 +52280,7 @@
   .byte  197,249,112,201,232                 // vpshufd       $0xe8,%xmm1,%xmm1
   .byte  196,227,125,2,201,3                 // vpblendd      $0x3,%ymm1,%ymm0,%ymm1
   .byte  197,253,111,195                     // vmovdqa       %ymm3,%ymm0
-  .byte  233,229,0,0,0                       // jmpq          2b979 <_sk_load_bgra_hsw_lowp+0x179>
+  .byte  233,229,0,0,0                       // jmpq          2ba69 <_sk_load_bgra_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,24           // vmovd         0x18(%r8,%rdx,4),%xmm0
   .byte  196,226,125,89,192                  // vpbroadcastq  %xmm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -52230,7 +52296,7 @@
   .byte  196,193,122,111,12,144              // vmovdqu       (%r8,%rdx,4),%xmm1
   .byte  196,227,117,2,200,240               // vpblendd      $0xf0,%ymm0,%ymm1,%ymm1
   .byte  197,125,127,192                     // vmovdqa       %ymm8,%ymm0
-  .byte  233,138,0,0,0                       // jmpq          2b979 <_sk_load_bgra_hsw_lowp+0x179>
+  .byte  233,138,0,0,0                       // jmpq          2ba69 <_sk_load_bgra_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,40           // vmovd         0x28(%r8,%rdx,4),%xmm0
   .byte  196,226,121,89,192                  // vpbroadcastq  %xmm0,%xmm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -52240,7 +52306,7 @@
   .byte  196,193,121,110,76,144,32           // vmovd         0x20(%r8,%rdx,4),%xmm1
   .byte  196,227,125,2,193,1                 // vpblendd      $0x1,%ymm1,%ymm0,%ymm0
   .byte  196,193,126,111,12,144              // vmovdqu       (%r8,%rdx,4),%ymm1
-  .byte  235,81                              // jmp           2b979 <_sk_load_bgra_hsw_lowp+0x179>
+  .byte  235,81                              // jmp           2ba69 <_sk_load_bgra_hsw_lowp+0x179>
   .byte  196,193,121,110,68,144,56           // vmovd         0x38(%r8,%rdx,4),%xmm0
   .byte  196,226,125,89,192                  // vpbroadcastq  %xmm0,%ymm0
   .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
@@ -52260,7 +52326,7 @@
   .byte  196,227,125,14,209,170              // vpblendw      $0xaa,%ymm1,%ymm0,%ymm2
   .byte  196,227,101,14,201,170              // vpblendw      $0xaa,%ymm1,%ymm3,%ymm1
   .byte  196,226,117,43,202                  // vpackusdw     %ymm2,%ymm1,%ymm1
-  .byte  197,125,111,5,254,120,0,0           // vmovdqa       0x78fe(%rip),%ymm8        # 332a0 <_sk_srcover_rgba_8888_sse2_lowp+0x920>
+  .byte  197,125,111,5,174,142,0,0           // vmovdqa       0x8eae(%rip),%ymm8        # 34940 <_sk_srcover_rgba_8888_sse2_lowp+0x910>
   .byte  196,193,117,219,208                 // vpand         %ymm8,%ymm1,%ymm2
   .byte  197,245,113,209,8                   // vpsrlw        $0x8,%ymm1,%ymm1
   .byte  197,253,114,208,16                  // vpsrld        $0x10,%ymm0,%ymm0
@@ -52271,7 +52337,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  119,254                             // ja            2b9cc <_sk_load_bgra_hsw_lowp+0x1cc>
+  .byte  119,254                             // ja            2babc <_sk_load_bgra_hsw_lowp+0x1cc>
   .byte  255                                 // (bad)
   .byte  255,174,254,255,255,152             // ljmp          *-0x67000002(%rsi)
   .byte  254                                 // (bad)
@@ -52318,10 +52384,10 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,55                              // ja            2ba5a <_sk_load_bgra_dst_hsw_lowp+0x52>
+  .byte  119,55                              // ja            2bb4a <_sk_load_bgra_dst_hsw_lowp+0x52>
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2bbd4 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
+  .byte  76,141,13,163,1,0,0                 // lea           0x1a3(%rip),%r9        # 2bcc4 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52331,10 +52397,10 @@
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,44,144              // vmovd         (%r8,%rdx,4),%xmm5
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
-  .byte  233,39,1,0,0                        // jmpq          2bb81 <_sk_load_bgra_dst_hsw_lowp+0x179>
+  .byte  233,39,1,0,0                        // jmpq          2bc71 <_sk_load_bgra_dst_hsw_lowp+0x179>
   .byte  196,193,126,111,44,144              // vmovdqu       (%r8,%rdx,4),%ymm5
   .byte  196,193,126,111,100,144,32          // vmovdqu       0x20(%r8,%rdx,4),%ymm4
-  .byte  233,21,1,0,0                        // jmpq          2bb81 <_sk_load_bgra_dst_hsw_lowp+0x179>
+  .byte  233,21,1,0,0                        // jmpq          2bc71 <_sk_load_bgra_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,8           // vmovd         0x8(%r8,%rdx,4),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,197,239,255                     // vpxor         %ymm7,%ymm7,%ymm7
@@ -52343,7 +52409,7 @@
   .byte  197,249,112,237,232                 // vpshufd       $0xe8,%xmm5,%xmm5
   .byte  196,227,93,2,237,3                  // vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   .byte  197,253,111,231                     // vmovdqa       %ymm7,%ymm4
-  .byte  233,229,0,0,0                       // jmpq          2bb81 <_sk_load_bgra_dst_hsw_lowp+0x179>
+  .byte  233,229,0,0,0                       // jmpq          2bc71 <_sk_load_bgra_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,24          // vmovd         0x18(%r8,%rdx,4),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52359,7 +52425,7 @@
   .byte  196,193,122,111,44,144              // vmovdqu       (%r8,%rdx,4),%xmm5
   .byte  196,227,85,2,236,240                // vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   .byte  197,125,127,196                     // vmovdqa       %ymm8,%ymm4
-  .byte  233,138,0,0,0                       // jmpq          2bb81 <_sk_load_bgra_dst_hsw_lowp+0x179>
+  .byte  233,138,0,0,0                       // jmpq          2bc71 <_sk_load_bgra_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,40          // vmovd         0x28(%r8,%rdx,4),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52369,7 +52435,7 @@
   .byte  196,193,121,110,108,144,32          // vmovd         0x20(%r8,%rdx,4),%xmm5
   .byte  196,227,93,2,229,1                  // vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   .byte  196,193,126,111,44,144              // vmovdqu       (%r8,%rdx,4),%ymm5
-  .byte  235,81                              // jmp           2bb81 <_sk_load_bgra_dst_hsw_lowp+0x179>
+  .byte  235,81                              // jmp           2bc71 <_sk_load_bgra_dst_hsw_lowp+0x179>
   .byte  196,193,121,110,100,144,56          // vmovd         0x38(%r8,%rdx,4),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -52389,7 +52455,7 @@
   .byte  196,227,93,14,245,170               // vpblendw      $0xaa,%ymm5,%ymm4,%ymm6
   .byte  196,227,69,14,237,170               // vpblendw      $0xaa,%ymm5,%ymm7,%ymm5
   .byte  196,226,85,43,238                   // vpackusdw     %ymm6,%ymm5,%ymm5
-  .byte  197,125,111,5,22,119,0,0            // vmovdqa       0x7716(%rip),%ymm8        # 332c0 <_sk_srcover_rgba_8888_sse2_lowp+0x940>
+  .byte  197,125,111,5,198,140,0,0           // vmovdqa       0x8cc6(%rip),%ymm8        # 34960 <_sk_srcover_rgba_8888_sse2_lowp+0x930>
   .byte  196,193,85,219,240                  // vpand         %ymm8,%ymm5,%ymm6
   .byte  197,213,113,213,8                   // vpsrlw        $0x8,%ymm5,%ymm5
   .byte  197,221,114,212,16                  // vpsrld        $0x10,%ymm4,%ymm4
@@ -52400,7 +52466,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  119,254                             // ja            2bbd4 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
+  .byte  119,254                             // ja            2bcc4 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
   .byte  255                                 // (bad)
   .byte  255,174,254,255,255,152             // ljmp          *-0x67000002(%rsi)
   .byte  254                                 // (bad)
@@ -52461,20 +52527,20 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,30                              // ja            2bc91 <_sk_store_bgra_hsw_lowp+0x81>
+  .byte  119,30                              // ja            2bd81 <_sk_store_bgra_hsw_lowp+0x81>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2bd40 <_sk_store_bgra_hsw_lowp+0x130>
+  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2be30 <_sk_store_bgra_hsw_lowp+0x130>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,65,121,126,4,144                // vmovd         %xmm8,(%r8,%rdx,4)
-  .byte  233,171,0,0,0                       // jmpq          2bd3c <_sk_store_bgra_hsw_lowp+0x12c>
+  .byte  233,171,0,0,0                       // jmpq          2be2c <_sk_store_bgra_hsw_lowp+0x12c>
   .byte  196,65,126,127,4,144                // vmovdqu       %ymm8,(%r8,%rdx,4)
   .byte  196,65,126,127,76,144,32            // vmovdqu       %ymm9,0x20(%r8,%rdx,4)
-  .byte  233,153,0,0,0                       // jmpq          2bd3c <_sk_store_bgra_hsw_lowp+0x12c>
+  .byte  233,153,0,0,0                       // jmpq          2be2c <_sk_store_bgra_hsw_lowp+0x12c>
   .byte  196,67,121,22,68,144,8,2            // vpextrd       $0x2,%xmm8,0x8(%r8,%rdx,4)
   .byte  196,65,121,214,4,144                // vmovq         %xmm8,(%r8,%rdx,4)
-  .byte  233,134,0,0,0                       // jmpq          2bd3c <_sk_store_bgra_hsw_lowp+0x12c>
+  .byte  233,134,0,0,0                       // jmpq          2be2c <_sk_store_bgra_hsw_lowp+0x12c>
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,22,76,144,24,2           // vpextrd       $0x2,%xmm9,0x18(%r8,%rdx,4)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
@@ -52482,12 +52548,12 @@
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,65,121,126,76,144,16            // vmovd         %xmm9,0x10(%r8,%rdx,4)
   .byte  196,65,122,127,4,144                // vmovdqu       %xmm8,(%r8,%rdx,4)
-  .byte  235,85                              // jmp           2bd3c <_sk_store_bgra_hsw_lowp+0x12c>
+  .byte  235,85                              // jmp           2be2c <_sk_store_bgra_hsw_lowp+0x12c>
   .byte  196,67,121,22,76,144,40,2           // vpextrd       $0x2,%xmm9,0x28(%r8,%rdx,4)
   .byte  196,67,121,22,76,144,36,1           // vpextrd       $0x1,%xmm9,0x24(%r8,%rdx,4)
   .byte  196,65,121,126,76,144,32            // vmovd         %xmm9,0x20(%r8,%rdx,4)
   .byte  196,65,126,127,4,144                // vmovdqu       %ymm8,(%r8,%rdx,4)
-  .byte  235,54                              // jmp           2bd3c <_sk_store_bgra_hsw_lowp+0x12c>
+  .byte  235,54                              // jmp           2be2c <_sk_store_bgra_hsw_lowp+0x12c>
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
   .byte  196,67,121,22,84,144,56,2           // vpextrd       $0x2,%xmm10,0x38(%r8,%rdx,4)
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
@@ -52609,7 +52675,7 @@
   .byte  196,195,109,14,192,170              // vpblendw      $0xaa,%ymm8,%ymm2,%ymm0
   .byte  196,195,101,14,200,170              // vpblendw      $0xaa,%ymm8,%ymm3,%ymm1
   .byte  196,226,117,43,200                  // vpackusdw     %ymm0,%ymm1,%ymm1
-  .byte  197,125,111,5,226,115,0,0           // vmovdqa       0x73e2(%rip),%ymm8        # 332e0 <_sk_srcover_rgba_8888_sse2_lowp+0x960>
+  .byte  197,125,111,5,146,137,0,0           // vmovdqa       0x8992(%rip),%ymm8        # 34980 <_sk_srcover_rgba_8888_sse2_lowp+0x950>
   .byte  196,193,117,219,192                 // vpand         %ymm8,%ymm1,%ymm0
   .byte  197,245,113,209,8                   // vpsrlw        $0x8,%ymm1,%ymm1
   .byte  197,237,114,210,16                  // vpsrld        $0x10,%ymm2,%ymm2
@@ -52620,6 +52686,95 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_bgra_hsw_lowp
+.globl _sk_gather_bgra_hsw_lowp
+FUNCTION(_sk_gather_bgra_hsw_lowp)
+_sk_gather_bgra_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,88,64,12                 // vpbroadcastd  0xc(%rax),%ymm8
+  .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
+  .byte  196,65,61,254,209                   // vpaddd        %ymm9,%ymm8,%ymm10
+  .byte  196,65,61,239,192                   // vpxor         %ymm8,%ymm8,%ymm8
+  .byte  196,193,124,95,192                  // vmaxps        %ymm8,%ymm0,%ymm0
+  .byte  196,193,116,95,200                  // vmaxps        %ymm8,%ymm1,%ymm1
+  .byte  196,193,116,93,202                  // vminps        %ymm10,%ymm1,%ymm1
+  .byte  196,193,124,93,194                  // vminps        %ymm10,%ymm0,%ymm0
+  .byte  196,98,125,88,80,16                 // vpbroadcastd  0x10(%rax),%ymm10
+  .byte  196,65,45,254,201                   // vpaddd        %ymm9,%ymm10,%ymm9
+  .byte  196,193,100,95,216                  // vmaxps        %ymm8,%ymm3,%ymm3
+  .byte  196,193,108,95,208                  // vmaxps        %ymm8,%ymm2,%ymm2
+  .byte  196,193,108,93,209                  // vminps        %ymm9,%ymm2,%ymm2
+  .byte  196,193,100,93,217                  // vminps        %ymm9,%ymm3,%ymm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  197,254,91,219                      // vcvttps2dq    %ymm3,%ymm3
+  .byte  197,254,91,210                      // vcvttps2dq    %ymm2,%ymm2
+  .byte  196,98,125,88,72,8                  // vpbroadcastd  0x8(%rax),%ymm9
+  .byte  196,226,53,64,210                   // vpmulld       %ymm2,%ymm9,%ymm2
+  .byte  196,226,53,64,219                   // vpmulld       %ymm3,%ymm9,%ymm3
+  .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
+  .byte  197,109,254,208                     // vpaddd        %ymm0,%ymm2,%ymm10
+  .byte  197,254,91,201                      // vcvttps2dq    %ymm1,%ymm1
+  .byte  197,229,254,201                     // vpaddd        %ymm1,%ymm3,%ymm1
+  .byte  196,67,125,57,209,1                 // vextracti128  $0x1,%ymm10,%xmm9
+  .byte  196,227,125,57,203,1                // vextracti128  $0x1,%ymm1,%xmm3
+  .byte  196,225,249,126,216                 // vmovq         %xmm3,%rax
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  196,129,121,110,20,136              // vmovd         (%r8,%r9,4),%xmm2
+  .byte  196,195,249,22,217,1                // vpextrq       $0x1,%xmm3,%r9
+  .byte  196,195,105,34,20,128,1             // vpinsrd       $0x1,(%r8,%rax,4),%xmm2,%xmm2
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  196,195,105,34,20,128,2             // vpinsrd       $0x2,(%r8,%rax,4),%xmm2,%xmm2
+  .byte  196,225,249,126,200                 // vmovq         %xmm1,%rax
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  196,131,105,34,20,136,3             // vpinsrd       $0x3,(%r8,%r9,4),%xmm2,%xmm2
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  196,129,121,110,28,136              // vmovd         (%r8,%r9,4),%xmm3
+  .byte  196,195,249,22,201,1                // vpextrq       $0x1,%xmm1,%r9
+  .byte  196,195,97,34,12,128,1              // vpinsrd       $0x1,(%r8,%rax,4),%xmm3,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  196,195,113,34,12,128,2             // vpinsrd       $0x2,(%r8,%rax,4),%xmm1,%xmm1
+  .byte  196,97,249,126,200                  // vmovq         %xmm9,%rax
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  196,131,113,34,12,136,3             // vpinsrd       $0x3,(%r8,%r9,4),%xmm1,%xmm1
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  196,129,121,110,28,136              // vmovd         (%r8,%r9,4),%xmm3
+  .byte  196,67,249,22,201,1                 // vpextrq       $0x1,%xmm9,%r9
+  .byte  196,195,97,34,28,128,1              // vpinsrd       $0x1,(%r8,%rax,4),%xmm3,%xmm3
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  196,195,97,34,28,128,2              // vpinsrd       $0x2,(%r8,%rax,4),%xmm3,%xmm3
+  .byte  196,97,249,126,208                  // vmovq         %xmm10,%rax
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  196,131,97,34,28,136,3              // vpinsrd       $0x3,(%r8,%r9,4),%xmm3,%xmm3
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  196,129,121,110,4,136               // vmovd         (%r8,%r9,4),%xmm0
+  .byte  196,67,249,22,209,1                 // vpextrq       $0x1,%xmm10,%r9
+  .byte  196,195,121,34,4,128,1              // vpinsrd       $0x1,(%r8,%rax,4),%xmm0,%xmm0
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  196,195,121,34,4,128,2              // vpinsrd       $0x2,(%r8,%rax,4),%xmm0,%xmm0
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  196,131,121,34,4,136,3              // vpinsrd       $0x3,(%r8,%r9,4),%xmm0,%xmm0
+  .byte  196,227,117,56,210,1                // vinserti128   $0x1,%xmm2,%ymm1,%ymm2
+  .byte  196,227,125,56,195,1                // vinserti128   $0x1,%xmm3,%ymm0,%ymm0
+  .byte  196,227,125,56,217,1                // vinserti128   $0x1,%xmm1,%ymm0,%ymm3
+  .byte  196,227,125,70,194,49               // vperm2i128    $0x31,%ymm2,%ymm0,%ymm0
+  .byte  196,195,125,14,200,170              // vpblendw      $0xaa,%ymm8,%ymm0,%ymm1
+  .byte  196,195,101,14,208,170              // vpblendw      $0xaa,%ymm8,%ymm3,%ymm2
+  .byte  196,226,109,43,201                  // vpackusdw     %ymm1,%ymm2,%ymm1
+  .byte  197,125,111,5,9,136,0,0             // vmovdqa       0x8809(%rip),%ymm8        # 349a0 <_sk_srcover_rgba_8888_sse2_lowp+0x970>
+  .byte  196,193,117,219,208                 // vpand         %ymm8,%ymm1,%ymm2
+  .byte  197,245,113,209,8                   // vpsrlw        $0x8,%ymm1,%ymm1
+  .byte  197,253,114,208,16                  // vpsrld        $0x10,%ymm0,%ymm0
+  .byte  197,229,114,211,16                  // vpsrld        $0x10,%ymm3,%ymm3
+  .byte  196,226,101,43,216                  // vpackusdw     %ymm0,%ymm3,%ymm3
+  .byte  196,193,101,219,192                 // vpand         %ymm8,%ymm3,%ymm0
+  .byte  197,229,113,211,8                   // vpsrlw        $0x8,%ymm3,%ymm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_565_hsw_lowp
 .globl _sk_load_565_hsw_lowp
 FUNCTION(_sk_load_565_hsw_lowp)
@@ -52633,24 +52788,24 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2bf64 <_sk_load_565_hsw_lowp+0x3f>
+  .byte  119,37                              // ja            2c1fd <_sk_load_565_hsw_lowp+0x3f>
   .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,131,1,0,0                 // lea           0x183(%rip),%r9        # 2c0d0 <_sk_load_565_hsw_lowp+0x1ab>
+  .byte  76,141,13,130,1,0,0                 // lea           0x182(%rip),%r9        # 2c368 <_sk_load_565_hsw_lowp+0x1aa>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  197,249,110,192                     // vmovd         %eax,%xmm0
-  .byte  233,8,1,0,0                         // jmpq          2c06c <_sk_load_565_hsw_lowp+0x147>
+  .byte  233,8,1,0,0                         // jmpq          2c305 <_sk_load_565_hsw_lowp+0x147>
   .byte  196,193,126,111,4,80                // vmovdqu       (%r8,%rdx,2),%ymm0
-  .byte  233,253,0,0,0                       // jmpq          2c06c <_sk_load_565_hsw_lowp+0x147>
+  .byte  233,253,0,0,0                       // jmpq          2c305 <_sk_load_565_hsw_lowp+0x147>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,193,121,196,76,80,4,2           // vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm0,%xmm1
   .byte  196,227,117,56,192,1                // vinserti128   $0x1,%xmm0,%ymm1,%ymm0
   .byte  196,193,121,110,12,80               // vmovd         (%r8,%rdx,2),%xmm1
   .byte  196,227,125,2,193,1                 // vpblendd      $0x1,%ymm1,%ymm0,%ymm0
-  .byte  233,218,0,0,0                       // jmpq          2c06c <_sk_load_565_hsw_lowp+0x147>
+  .byte  233,218,0,0,0                       // jmpq          2c305 <_sk_load_565_hsw_lowp+0x147>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,193,121,196,76,80,12,6          // vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm0,%xmm1
   .byte  196,227,117,56,192,1                // vinserti128   $0x1,%xmm0,%ymm1,%ymm0
@@ -52660,7 +52815,7 @@
   .byte  196,227,125,2,193,15                // vpblendd      $0xf,%ymm1,%ymm0,%ymm0
   .byte  196,193,122,126,12,80               // vmovq         (%r8,%rdx,2),%xmm1
   .byte  196,227,125,2,193,3                 // vpblendd      $0x3,%ymm1,%ymm0,%ymm0
-  .byte  233,155,0,0,0                       // jmpq          2c06c <_sk_load_565_hsw_lowp+0x147>
+  .byte  233,155,0,0,0                       // jmpq          2c305 <_sk_load_565_hsw_lowp+0x147>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,193,121,196,76,80,20,2          // vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm0,%xmm1
   .byte  196,227,125,56,193,1                // vinserti128   $0x1,%xmm1,%ymm0,%ymm0
@@ -52672,7 +52827,7 @@
   .byte  196,227,125,56,193,1                // vinserti128   $0x1,%xmm1,%ymm0,%ymm0
   .byte  196,193,122,111,12,80               // vmovdqu       (%r8,%rdx,2),%xmm1
   .byte  196,227,117,2,192,240               // vpblendd      $0xf0,%ymm0,%ymm1,%ymm0
-  .byte  235,83                              // jmp           2c06c <_sk_load_565_hsw_lowp+0x147>
+  .byte  235,83                              // jmp           2c305 <_sk_load_565_hsw_lowp+0x147>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,193,121,196,76,80,28,6          // vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm0,%xmm1
   .byte  196,227,125,56,193,1                // vinserti128   $0x1,%xmm1,%ymm0,%ymm0
@@ -52686,13 +52841,13 @@
   .byte  196,193,122,111,20,80               // vmovdqu       (%r8,%rdx,2),%xmm2
   .byte  196,227,109,56,201,1                // vinserti128   $0x1,%xmm1,%ymm2,%ymm1
   .byte  196,227,117,2,192,192               // vpblendd      $0xc0,%ymm0,%ymm1,%ymm0
-  .byte  196,226,125,121,13,139,114,0,0      // vpbroadcastw  0x728b(%rip),%ymm1        # 33300 <_sk_srcover_rgba_8888_sse2_lowp+0x980>
+  .byte  196,226,125,121,13,178,134,0,0      // vpbroadcastw  0x86b2(%rip),%ymm1        # 349c0 <_sk_srcover_rgba_8888_sse2_lowp+0x990>
   .byte  197,237,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm2
   .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
   .byte  197,237,113,208,5                   // vpsrlw        $0x5,%ymm0,%ymm2
-  .byte  196,226,125,121,29,118,114,0,0      // vpbroadcastw  0x7276(%rip),%ymm3        # 33302 <_sk_srcover_rgba_8888_sse2_lowp+0x982>
+  .byte  196,226,125,121,29,157,134,0,0      // vpbroadcastw  0x869d(%rip),%ymm3        # 349c2 <_sk_srcover_rgba_8888_sse2_lowp+0x992>
   .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
-  .byte  196,226,125,121,29,107,114,0,0      // vpbroadcastw  0x726b(%rip),%ymm3        # 33304 <_sk_srcover_rgba_8888_sse2_lowp+0x984>
+  .byte  196,226,125,121,29,146,134,0,0      // vpbroadcastw  0x8692(%rip),%ymm3        # 349c4 <_sk_srcover_rgba_8888_sse2_lowp+0x994>
   .byte  197,253,219,219                     // vpand         %ymm3,%ymm0,%ymm3
   .byte  197,253,113,208,13                  // vpsrlw        $0xd,%ymm0,%ymm0
   .byte  197,245,235,192                     // vpor          %ymm0,%ymm1,%ymm0
@@ -52703,44 +52858,40 @@
   .byte  197,229,113,211,2                   // vpsrlw        $0x2,%ymm3,%ymm3
   .byte  197,237,235,211                     // vpor          %ymm3,%ymm2,%ymm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,29,57,114,0,0       // vpbroadcastw  0x7239(%rip),%ymm3        # 33306 <_sk_srcover_rgba_8888_sse2_lowp+0x986>
+  .byte  196,226,125,121,29,96,134,0,0       // vpbroadcastw  0x8660(%rip),%ymm3        # 349c6 <_sk_srcover_rgba_8888_sse2_lowp+0x996>
   .byte  255,224                             // jmpq          *%rax
-  .byte  144                                 // nop
-  .byte  134,254                             // xchg          %bh,%dh
+  .byte  135,254                             // xchg          %edi,%esi
   .byte  255                                 // (bad)
-  .byte  255,177,254,255,255,159             // pushq         -0x60000002(%rcx)
+  .byte  255,178,254,255,255,160             // pushq         -0x5f000002(%rdx)
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,240                             // push          %rax
+  .byte  255,241                             // push          %rcx
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,226                             // jmpq          *%rdx
+  .byte  255,227                             // jmpq          *%rbx
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,212                             // callq         *%rsp
+  .byte  255,213                             // callq         *%rbp
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,194                             // inc           %edx
+  .byte  255,195                             // inc           %ebx
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  59,255                              // cmp           %edi,%edi
+  .byte  60,255                              // cmp           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  255,39                              // jmpq          *(%rdi)
+  .byte  255,40                              // ljmp          *(%rax)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,19                              // callq         *(%rbx)
+  .byte  255,20,255                          // callq         *(%rdi,%rdi,8)
+  .byte  255                                 // (bad)
+  .byte  255,2                               // incl          (%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,1                               // incl          (%rcx)
+  .byte  255,132,255,255,255,112,255         // incl          -0x8f0001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
-  .byte  255                                 // (bad)
-  .byte  255,131,255,255,255,111             // incl          0x6fffffff(%rbx)
-  .byte  255                                 // (bad)
-  .byte  255                                 // (bad)
-  .byte  255,91,255                          // lcall         *-0x1(%rbx)
-  .byte  255                                 // (bad)
-  .byte  255,73,255                          // decl          -0x1(%rcx)
+  .byte  255,92,255,255                      // lcall         *-0x1(%rdi,%rdi,8)
+  .byte  255,74,255                          // decl          -0x1(%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
@@ -52757,24 +52908,24 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2c14b <_sk_load_565_dst_hsw_lowp+0x3f>
+  .byte  119,37                              // ja            2c3e3 <_sk_load_565_dst_hsw_lowp+0x3f>
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,132,1,0,0                 // lea           0x184(%rip),%r9        # 2c2b8 <_sk_load_565_dst_hsw_lowp+0x1ac>
+  .byte  76,141,13,132,1,0,0                 // lea           0x184(%rip),%r9        # 2c550 <_sk_load_565_dst_hsw_lowp+0x1ac>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  197,249,110,224                     // vmovd         %eax,%xmm4
-  .byte  233,8,1,0,0                         // jmpq          2c253 <_sk_load_565_dst_hsw_lowp+0x147>
+  .byte  233,8,1,0,0                         // jmpq          2c4eb <_sk_load_565_dst_hsw_lowp+0x147>
   .byte  196,193,126,111,36,80               // vmovdqu       (%r8,%rdx,2),%ymm4
-  .byte  233,253,0,0,0                       // jmpq          2c253 <_sk_load_565_dst_hsw_lowp+0x147>
+  .byte  233,253,0,0,0                       // jmpq          2c4eb <_sk_load_565_dst_hsw_lowp+0x147>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,193,89,196,108,80,4,2           // vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm4,%xmm5
   .byte  196,227,85,56,228,1                 // vinserti128   $0x1,%xmm4,%ymm5,%ymm4
   .byte  196,193,121,110,44,80               // vmovd         (%r8,%rdx,2),%xmm5
   .byte  196,227,93,2,229,1                  // vpblendd      $0x1,%ymm5,%ymm4,%ymm4
-  .byte  233,218,0,0,0                       // jmpq          2c253 <_sk_load_565_dst_hsw_lowp+0x147>
+  .byte  233,218,0,0,0                       // jmpq          2c4eb <_sk_load_565_dst_hsw_lowp+0x147>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,193,89,196,108,80,12,6          // vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm4,%xmm5
   .byte  196,227,85,56,228,1                 // vinserti128   $0x1,%xmm4,%ymm5,%ymm4
@@ -52784,7 +52935,7 @@
   .byte  196,227,93,2,229,15                 // vpblendd      $0xf,%ymm5,%ymm4,%ymm4
   .byte  196,193,122,126,44,80               // vmovq         (%r8,%rdx,2),%xmm5
   .byte  196,227,93,2,229,3                  // vpblendd      $0x3,%ymm5,%ymm4,%ymm4
-  .byte  233,155,0,0,0                       // jmpq          2c253 <_sk_load_565_dst_hsw_lowp+0x147>
+  .byte  233,155,0,0,0                       // jmpq          2c4eb <_sk_load_565_dst_hsw_lowp+0x147>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,193,89,196,108,80,20,2          // vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm4,%xmm5
   .byte  196,227,93,56,229,1                 // vinserti128   $0x1,%xmm5,%ymm4,%ymm4
@@ -52796,7 +52947,7 @@
   .byte  196,227,93,56,229,1                 // vinserti128   $0x1,%xmm5,%ymm4,%ymm4
   .byte  196,193,122,111,44,80               // vmovdqu       (%r8,%rdx,2),%xmm5
   .byte  196,227,85,2,228,240                // vpblendd      $0xf0,%ymm4,%ymm5,%ymm4
-  .byte  235,83                              // jmp           2c253 <_sk_load_565_dst_hsw_lowp+0x147>
+  .byte  235,83                              // jmp           2c4eb <_sk_load_565_dst_hsw_lowp+0x147>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,193,89,196,108,80,28,6          // vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm4,%xmm5
   .byte  196,227,93,56,229,1                 // vinserti128   $0x1,%xmm5,%ymm4,%ymm4
@@ -52810,13 +52961,13 @@
   .byte  196,193,122,111,52,80               // vmovdqu       (%r8,%rdx,2),%xmm6
   .byte  196,227,77,56,237,1                 // vinserti128   $0x1,%xmm5,%ymm6,%ymm5
   .byte  196,227,85,2,228,192                // vpblendd      $0xc0,%ymm4,%ymm5,%ymm4
-  .byte  196,226,125,121,45,172,112,0,0      // vpbroadcastw  0x70ac(%rip),%ymm5        # 33308 <_sk_srcover_rgba_8888_sse2_lowp+0x988>
+  .byte  196,226,125,121,45,212,132,0,0      // vpbroadcastw  0x84d4(%rip),%ymm5        # 349c8 <_sk_srcover_rgba_8888_sse2_lowp+0x998>
   .byte  197,205,113,212,8                   // vpsrlw        $0x8,%ymm4,%ymm6
   .byte  197,205,219,237                     // vpand         %ymm5,%ymm6,%ymm5
   .byte  197,205,113,212,5                   // vpsrlw        $0x5,%ymm4,%ymm6
-  .byte  196,226,125,121,61,151,112,0,0      // vpbroadcastw  0x7097(%rip),%ymm7        # 3330a <_sk_srcover_rgba_8888_sse2_lowp+0x98a>
+  .byte  196,226,125,121,61,191,132,0,0      // vpbroadcastw  0x84bf(%rip),%ymm7        # 349ca <_sk_srcover_rgba_8888_sse2_lowp+0x99a>
   .byte  197,205,219,247                     // vpand         %ymm7,%ymm6,%ymm6
-  .byte  196,226,125,121,61,140,112,0,0      // vpbroadcastw  0x708c(%rip),%ymm7        # 3330c <_sk_srcover_rgba_8888_sse2_lowp+0x98c>
+  .byte  196,226,125,121,61,180,132,0,0      // vpbroadcastw  0x84b4(%rip),%ymm7        # 349cc <_sk_srcover_rgba_8888_sse2_lowp+0x99c>
   .byte  197,221,219,255                     // vpand         %ymm7,%ymm4,%ymm7
   .byte  197,221,113,212,13                  // vpsrlw        $0xd,%ymm4,%ymm4
   .byte  197,213,235,228                     // vpor          %ymm4,%ymm5,%ymm4
@@ -52827,7 +52978,7 @@
   .byte  197,197,113,215,2                   // vpsrlw        $0x2,%ymm7,%ymm7
   .byte  197,205,235,247                     // vpor          %ymm7,%ymm6,%ymm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,61,90,112,0,0       // vpbroadcastw  0x705a(%rip),%ymm7        # 3330e <_sk_srcover_rgba_8888_sse2_lowp+0x98e>
+  .byte  196,226,125,121,61,130,132,0,0      // vpbroadcastw  0x8482(%rip),%ymm7        # 349ce <_sk_srcover_rgba_8888_sse2_lowp+0x99e>
   .byte  255,224                             // jmpq          *%rax
   .byte  102,144                             // xchg          %ax,%ax
   .byte  133,254                             // test          %edi,%esi
@@ -52878,10 +53029,10 @@
   .byte  76,15,175,193                       // imul          %rcx,%r8
   .byte  77,1,192                            // add           %r8,%r8
   .byte  76,3,0                              // add           (%rax),%r8
-  .byte  196,98,125,121,5,3,112,0,0          // vpbroadcastw  0x7003(%rip),%ymm8        # 33310 <_sk_srcover_rgba_8888_sse2_lowp+0x990>
+  .byte  196,98,125,121,5,43,132,0,0         // vpbroadcastw  0x842b(%rip),%ymm8        # 349d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9a0>
   .byte  197,181,113,240,8                   // vpsllw        $0x8,%ymm0,%ymm9
   .byte  196,65,53,219,192                   // vpand         %ymm8,%ymm9,%ymm8
-  .byte  196,98,125,121,13,242,111,0,0       // vpbroadcastw  0x6ff2(%rip),%ymm9        # 33312 <_sk_srcover_rgba_8888_sse2_lowp+0x992>
+  .byte  196,98,125,121,13,26,132,0,0        // vpbroadcastw  0x841a(%rip),%ymm9        # 349d2 <_sk_srcover_rgba_8888_sse2_lowp+0x9a2>
   .byte  197,173,113,241,3                   // vpsllw        $0x3,%ymm1,%ymm10
   .byte  196,65,45,219,201                   // vpand         %ymm9,%ymm10,%ymm9
   .byte  196,65,53,235,192                   // vpor          %ymm8,%ymm9,%ymm8
@@ -52891,31 +53042,31 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,31                              // ja            2c362 <_sk_store_565_hsw_lowp+0x6e>
+  .byte  119,31                              // ja            2c5fa <_sk_store_565_hsw_lowp+0x6e>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2c410 <_sk_store_565_hsw_lowp+0x11c>
+  .byte  76,141,13,195,0,0,0                 // lea           0xc3(%rip),%r9        # 2c6a8 <_sk_store_565_hsw_lowp+0x11c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,67,121,21,4,80,0                // vpextrw       $0x0,%xmm8,(%r8,%rdx,2)
-  .byte  233,167,0,0,0                       // jmpq          2c409 <_sk_store_565_hsw_lowp+0x115>
+  .byte  233,167,0,0,0                       // jmpq          2c6a1 <_sk_store_565_hsw_lowp+0x115>
   .byte  196,65,126,127,4,80                 // vmovdqu       %ymm8,(%r8,%rdx,2)
-  .byte  233,156,0,0,0                       // jmpq          2c409 <_sk_store_565_hsw_lowp+0x115>
+  .byte  233,156,0,0,0                       // jmpq          2c6a1 <_sk_store_565_hsw_lowp+0x115>
   .byte  196,67,121,21,68,80,4,2             // vpextrw       $0x2,%xmm8,0x4(%r8,%rdx,2)
   .byte  196,65,121,126,4,80                 // vmovd         %xmm8,(%r8,%rdx,2)
-  .byte  233,137,0,0,0                       // jmpq          2c409 <_sk_store_565_hsw_lowp+0x115>
+  .byte  233,137,0,0,0                       // jmpq          2c6a1 <_sk_store_565_hsw_lowp+0x115>
   .byte  196,67,121,21,68,80,12,6            // vpextrw       $0x6,%xmm8,0xc(%r8,%rdx,2)
   .byte  196,67,121,21,68,80,10,5            // vpextrw       $0x5,%xmm8,0xa(%r8,%rdx,2)
   .byte  196,67,121,21,68,80,8,4             // vpextrw       $0x4,%xmm8,0x8(%r8,%rdx,2)
   .byte  196,65,121,214,4,80                 // vmovq         %xmm8,(%r8,%rdx,2)
-  .byte  235,105                             // jmp           2c409 <_sk_store_565_hsw_lowp+0x115>
+  .byte  235,105                             // jmp           2c6a1 <_sk_store_565_hsw_lowp+0x115>
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,21,76,80,20,2            // vpextrw       $0x2,%xmm9,0x14(%r8,%rdx,2)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,21,76,80,18,1            // vpextrw       $0x1,%xmm9,0x12(%r8,%rdx,2)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,21,76,80,16,0            // vpextrw       $0x0,%xmm9,0x10(%r8,%rdx,2)
-  .byte  235,55                              // jmp           2c403 <_sk_store_565_hsw_lowp+0x10f>
+  .byte  235,55                              // jmp           2c69b <_sk_store_565_hsw_lowp+0x10f>
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,21,76,80,28,6            // vpextrw       $0x6,%xmm9,0x1c(%r8,%rdx,2)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
@@ -52938,7 +53089,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  120,255                             // js            2c425 <_sk_store_565_hsw_lowp+0x131>
+  .byte  120,255                             // js            2c6bd <_sk_store_565_hsw_lowp+0x131>
   .byte  255                                 // (bad)
   .byte  255,112,255                         // pushq         -0x1(%rax)
   .byte  255                                 // (bad)
@@ -52962,6 +53113,114 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_565_hsw_lowp
+.globl _sk_gather_565_hsw_lowp
+FUNCTION(_sk_gather_565_hsw_lowp)
+_sk_gather_565_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,88,64,12                 // vpbroadcastd  0xc(%rax),%ymm8
+  .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
+  .byte  196,65,61,254,193                   // vpaddd        %ymm9,%ymm8,%ymm8
+  .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
+  .byte  196,193,124,95,194                  // vmaxps        %ymm10,%ymm0,%ymm0
+  .byte  196,193,116,95,202                  // vmaxps        %ymm10,%ymm1,%ymm1
+  .byte  196,193,116,93,200                  // vminps        %ymm8,%ymm1,%ymm1
+  .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
+  .byte  196,98,125,88,64,16                 // vpbroadcastd  0x10(%rax),%ymm8
+  .byte  196,65,61,254,193                   // vpaddd        %ymm9,%ymm8,%ymm8
+  .byte  196,193,100,95,218                  // vmaxps        %ymm10,%ymm3,%ymm3
+  .byte  196,193,108,95,210                  // vmaxps        %ymm10,%ymm2,%ymm2
+  .byte  196,193,108,93,208                  // vminps        %ymm8,%ymm2,%ymm2
+  .byte  196,193,100,93,216                  // vminps        %ymm8,%ymm3,%ymm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  197,254,91,219                      // vcvttps2dq    %ymm3,%ymm3
+  .byte  197,254,91,210                      // vcvttps2dq    %ymm2,%ymm2
+  .byte  196,98,125,88,64,8                  // vpbroadcastd  0x8(%rax),%ymm8
+  .byte  196,226,61,64,210                   // vpmulld       %ymm2,%ymm8,%ymm2
+  .byte  196,226,61,64,219                   // vpmulld       %ymm3,%ymm8,%ymm3
+  .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
+  .byte  197,237,254,192                     // vpaddd        %ymm0,%ymm2,%ymm0
+  .byte  197,254,91,201                      // vcvttps2dq    %ymm1,%ymm1
+  .byte  197,229,254,201                     // vpaddd        %ymm1,%ymm3,%ymm1
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,249,110,208                     // vmovd         %eax,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,183,4,72                      // movzwl        (%r8,%r9,2),%eax
+  .byte  197,233,196,208,1                   // vpinsrw       $0x1,%eax,%xmm2,%xmm2
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,201,1                // vextracti128  $0x1,%ymm1,%xmm1
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,233,196,208,2                   // vpinsrw       $0x2,%eax,%xmm2,%xmm2
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  67,15,183,4,80                      // movzwl        (%r8,%r10,2),%eax
+  .byte  197,233,196,208,3                   // vpinsrw       $0x3,%eax,%xmm2,%xmm2
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,233,196,208,4                   // vpinsrw       $0x4,%eax,%xmm2,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,183,4,72                      // movzwl        (%r8,%r9,2),%eax
+  .byte  197,233,196,200,5                   // vpinsrw       $0x5,%eax,%xmm2,%xmm1
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,241,196,200,6                   // vpinsrw       $0x6,%eax,%xmm1,%xmm1
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  67,15,183,4,80                      // movzwl        (%r8,%r10,2),%eax
+  .byte  197,241,196,200,7                   // vpinsrw       $0x7,%eax,%xmm1,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,249,110,208                     // vmovd         %eax,%xmm2
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,183,4,72                      // movzwl        (%r8,%r9,2),%eax
+  .byte  197,233,196,208,1                   // vpinsrw       $0x1,%eax,%xmm2,%xmm2
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,192,1                // vextracti128  $0x1,%ymm0,%xmm0
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,233,196,208,2                   // vpinsrw       $0x2,%eax,%xmm2,%xmm2
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  67,15,183,4,80                      // movzwl        (%r8,%r10,2),%eax
+  .byte  197,233,196,208,3                   // vpinsrw       $0x3,%eax,%xmm2,%xmm2
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,233,196,208,4                   // vpinsrw       $0x4,%eax,%xmm2,%xmm2
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,183,4,72                      // movzwl        (%r8,%r9,2),%eax
+  .byte  197,233,196,192,5                   // vpinsrw       $0x5,%eax,%xmm2,%xmm0
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  197,249,196,192,6                   // vpinsrw       $0x6,%eax,%xmm0,%xmm0
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  67,15,183,4,80                      // movzwl        (%r8,%r10,2),%eax
+  .byte  197,249,196,192,7                   // vpinsrw       $0x7,%eax,%xmm0,%xmm0
+  .byte  196,227,125,56,193,1                // vinserti128   $0x1,%xmm1,%ymm0,%ymm0
+  .byte  196,226,125,121,13,94,129,0,0       // vpbroadcastw  0x815e(%rip),%ymm1        # 349d4 <_sk_srcover_rgba_8888_sse2_lowp+0x9a4>
+  .byte  197,237,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm2
+  .byte  197,237,219,201                     // vpand         %ymm1,%ymm2,%ymm1
+  .byte  197,237,113,208,5                   // vpsrlw        $0x5,%ymm0,%ymm2
+  .byte  196,226,125,121,29,73,129,0,0       // vpbroadcastw  0x8149(%rip),%ymm3        # 349d6 <_sk_srcover_rgba_8888_sse2_lowp+0x9a6>
+  .byte  197,237,219,211                     // vpand         %ymm3,%ymm2,%ymm2
+  .byte  196,226,125,121,29,62,129,0,0       // vpbroadcastw  0x813e(%rip),%ymm3        # 349d8 <_sk_srcover_rgba_8888_sse2_lowp+0x9a8>
+  .byte  197,253,219,219                     // vpand         %ymm3,%ymm0,%ymm3
+  .byte  197,253,113,208,13                  // vpsrlw        $0xd,%ymm0,%ymm0
+  .byte  197,245,235,192                     // vpor          %ymm0,%ymm1,%ymm0
+  .byte  197,245,113,242,2                   // vpsllw        $0x2,%ymm2,%ymm1
+  .byte  197,237,113,210,4                   // vpsrlw        $0x4,%ymm2,%ymm2
+  .byte  197,245,235,202                     // vpor          %ymm2,%ymm1,%ymm1
+  .byte  197,237,113,243,3                   // vpsllw        $0x3,%ymm3,%ymm2
+  .byte  197,229,113,211,2                   // vpsrlw        $0x2,%ymm3,%ymm3
+  .byte  197,237,235,211                     // vpor          %ymm3,%ymm2,%ymm2
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,226,125,121,29,12,129,0,0       // vpbroadcastw  0x810c(%rip),%ymm3        # 349da <_sk_srcover_rgba_8888_sse2_lowp+0x9aa>
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_a8_hsw_lowp
 .globl _sk_load_a8_hsw_lowp
 FUNCTION(_sk_load_a8_hsw_lowp)
@@ -52974,38 +53233,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2c488 <_sk_load_a8_hsw_lowp+0x3c>
+  .byte  119,37                              // ja            2c90c <_sk_load_a8_hsw_lowp+0x3c>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,219,0,0,0                 // lea           0xdb(%rip),%r9        # 2c54c <_sk_load_a8_hsw_lowp+0x100>
+  .byte  76,141,13,219,0,0,0                 // lea           0xdb(%rip),%r9        # 2c9d0 <_sk_load_a8_hsw_lowp+0x100>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,249,110,192                     // vmovd         %eax,%xmm0
-  .byte  233,175,0,0,0                       // jmpq          2c537 <_sk_load_a8_hsw_lowp+0xeb>
+  .byte  233,175,0,0,0                       // jmpq          2c9bb <_sk_load_a8_hsw_lowp+0xeb>
   .byte  196,193,122,111,4,16                // vmovdqu       (%r8,%rdx,1),%xmm0
-  .byte  233,164,0,0,0                       // jmpq          2c537 <_sk_load_a8_hsw_lowp+0xeb>
+  .byte  233,164,0,0,0                       // jmpq          2c9bb <_sk_load_a8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,2,2            // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm0,%xmm0
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,249,110,200                     // vmovd         %eax,%xmm1
   .byte  196,227,121,14,193,1                // vpblendw      $0x1,%xmm1,%xmm0,%xmm0
-  .byte  233,132,0,0,0                       // jmpq          2c537 <_sk_load_a8_hsw_lowp+0xeb>
+  .byte  233,132,0,0,0                       // jmpq          2c9bb <_sk_load_a8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,6,6            // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,5,5            // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,4,4            // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,193,121,110,12,16               // vmovd         (%r8,%rdx,1),%xmm1
   .byte  196,227,121,2,193,1                 // vpblendd      $0x1,%xmm1,%xmm0,%xmm0
-  .byte  235,90                              // jmp           2c537 <_sk_load_a8_hsw_lowp+0xeb>
+  .byte  235,90                              // jmp           2c9bb <_sk_load_a8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,10,10          // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,9,9            // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,8,8            // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,193,122,126,12,16               // vmovq         (%r8,%rdx,1),%xmm1
   .byte  196,227,113,2,192,12                // vpblendd      $0xc,%xmm0,%xmm1,%xmm0
-  .byte  235,48                              // jmp           2c537 <_sk_load_a8_hsw_lowp+0xeb>
+  .byte  235,48                              // jmp           2c9bb <_sk_load_a8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,14,14          // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,13,13          // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm0,%xmm0
@@ -53065,38 +53324,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2c5c4 <_sk_load_a8_dst_hsw_lowp+0x3c>
+  .byte  119,37                              // ja            2ca48 <_sk_load_a8_dst_hsw_lowp+0x3c>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,219,0,0,0                 // lea           0xdb(%rip),%r9        # 2c688 <_sk_load_a8_dst_hsw_lowp+0x100>
+  .byte  76,141,13,219,0,0,0                 // lea           0xdb(%rip),%r9        # 2cb0c <_sk_load_a8_dst_hsw_lowp+0x100>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,249,110,224                     // vmovd         %eax,%xmm4
-  .byte  233,175,0,0,0                       // jmpq          2c673 <_sk_load_a8_dst_hsw_lowp+0xeb>
+  .byte  233,175,0,0,0                       // jmpq          2caf7 <_sk_load_a8_dst_hsw_lowp+0xeb>
   .byte  196,193,122,111,36,16               // vmovdqu       (%r8,%rdx,1),%xmm4
-  .byte  233,164,0,0,0                       // jmpq          2c673 <_sk_load_a8_dst_hsw_lowp+0xeb>
+  .byte  233,164,0,0,0                       // jmpq          2caf7 <_sk_load_a8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,2,2            // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm4,%xmm4
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,249,110,232                     // vmovd         %eax,%xmm5
   .byte  196,227,89,14,229,1                 // vpblendw      $0x1,%xmm5,%xmm4,%xmm4
-  .byte  233,132,0,0,0                       // jmpq          2c673 <_sk_load_a8_dst_hsw_lowp+0xeb>
+  .byte  233,132,0,0,0                       // jmpq          2caf7 <_sk_load_a8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,6,6            // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,5,5            // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,4,4            // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,193,121,110,44,16               // vmovd         (%r8,%rdx,1),%xmm5
   .byte  196,227,89,2,229,1                  // vpblendd      $0x1,%xmm5,%xmm4,%xmm4
-  .byte  235,90                              // jmp           2c673 <_sk_load_a8_dst_hsw_lowp+0xeb>
+  .byte  235,90                              // jmp           2caf7 <_sk_load_a8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,10,10          // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,9,9            // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,8,8            // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,193,122,126,44,16               // vmovq         (%r8,%rdx,1),%xmm5
   .byte  196,227,81,2,228,12                 // vpblendd      $0xc,%xmm4,%xmm5,%xmm4
-  .byte  235,48                              // jmp           2c673 <_sk_load_a8_dst_hsw_lowp+0xeb>
+  .byte  235,48                              // jmp           2caf7 <_sk_load_a8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,14,14          // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,13,13          // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm4,%xmm4
@@ -53153,7 +53412,7 @@
   .byte  76,15,175,193                       // imul          %rcx,%r8
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  196,195,125,57,216,1                // vextracti128  $0x1,%ymm3,%xmm8
-  .byte  197,121,111,13,225,115,0,0          // vmovdqa       0x73e1(%rip),%xmm9        # 33ac0 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
+  .byte  197,121,111,13,61,134,0,0           // vmovdqa       0x863d(%rip),%xmm9        # 351a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
   .byte  196,66,57,0,193                     // vpshufb       %xmm9,%xmm8,%xmm8
   .byte  196,66,97,0,201                     // vpshufb       %xmm9,%xmm3,%xmm9
   .byte  196,65,49,108,192                   // vpunpcklqdq   %xmm8,%xmm9,%xmm8
@@ -53161,28 +53420,28 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,28                              // ja            2c714 <_sk_store_a8_hsw_lowp+0x50>
+  .byte  119,28                              // ja            2cb98 <_sk_store_a8_hsw_lowp+0x50>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,146,0,0,0                 // lea           0x92(%rip),%r9        # 2c794 <_sk_store_a8_hsw_lowp+0xd0>
+  .byte  76,141,13,146,0,0,0                 // lea           0x92(%rip),%r9        # 2cc18 <_sk_store_a8_hsw_lowp+0xd0>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,67,121,20,4,16,0                // vpextrb       $0x0,%xmm8,(%r8,%rdx,1)
-  .byte  235,121                             // jmp           2c78d <_sk_store_a8_hsw_lowp+0xc9>
+  .byte  235,121                             // jmp           2cc11 <_sk_store_a8_hsw_lowp+0xc9>
   .byte  196,65,122,127,4,16                 // vmovdqu       %xmm8,(%r8,%rdx,1)
-  .byte  235,113                             // jmp           2c78d <_sk_store_a8_hsw_lowp+0xc9>
+  .byte  235,113                             // jmp           2cc11 <_sk_store_a8_hsw_lowp+0xc9>
   .byte  196,67,121,20,68,16,2,2             // vpextrb       $0x2,%xmm8,0x2(%r8,%rdx,1)
   .byte  196,67,121,21,4,16,0                // vpextrw       $0x0,%xmm8,(%r8,%rdx,1)
-  .byte  235,96                              // jmp           2c78d <_sk_store_a8_hsw_lowp+0xc9>
+  .byte  235,96                              // jmp           2cc11 <_sk_store_a8_hsw_lowp+0xc9>
   .byte  196,67,121,20,68,16,6,6             // vpextrb       $0x6,%xmm8,0x6(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,5,5             // vpextrb       $0x5,%xmm8,0x5(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,4,4             // vpextrb       $0x4,%xmm8,0x4(%r8,%rdx,1)
   .byte  196,65,121,126,4,16                 // vmovd         %xmm8,(%r8,%rdx,1)
-  .byte  235,64                              // jmp           2c78d <_sk_store_a8_hsw_lowp+0xc9>
+  .byte  235,64                              // jmp           2cc11 <_sk_store_a8_hsw_lowp+0xc9>
   .byte  196,67,121,20,68,16,10,10           // vpextrb       $0xa,%xmm8,0xa(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,9,9             // vpextrb       $0x9,%xmm8,0x9(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,8,8             // vpextrb       $0x8,%xmm8,0x8(%r8,%rdx,1)
-  .byte  235,32                              // jmp           2c787 <_sk_store_a8_hsw_lowp+0xc3>
+  .byte  235,32                              // jmp           2cc0b <_sk_store_a8_hsw_lowp+0xc3>
   .byte  196,67,121,20,68,16,14,14           // vpextrb       $0xe,%xmm8,0xe(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,13,13           // vpextrb       $0xd,%xmm8,0xd(%r8,%rdx,1)
   .byte  196,67,121,20,68,16,12,12           // vpextrb       $0xc,%xmm8,0xc(%r8,%rdx,1)
@@ -53191,7 +53450,7 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  119,255                             // ja            2c795 <_sk_store_a8_hsw_lowp+0xd1>
+  .byte  119,255                             // ja            2cc19 <_sk_store_a8_hsw_lowp+0xd1>
   .byte  255                                 // (bad)
   .byte  255,144,255,255,255,136             // callq         *-0x77000001(%rax)
   .byte  255                                 // (bad)
@@ -53226,6 +53485,100 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_a8_hsw_lowp
+.globl _sk_gather_a8_hsw_lowp
+FUNCTION(_sk_gather_a8_hsw_lowp)
+_sk_gather_a8_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,88,64,12                 // vpbroadcastd  0xc(%rax),%ymm8
+  .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
+  .byte  196,65,61,254,193                   // vpaddd        %ymm9,%ymm8,%ymm8
+  .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
+  .byte  196,193,124,95,194                  // vmaxps        %ymm10,%ymm0,%ymm0
+  .byte  196,193,116,95,202                  // vmaxps        %ymm10,%ymm1,%ymm1
+  .byte  196,65,116,93,216                   // vminps        %ymm8,%ymm1,%ymm11
+  .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
+  .byte  196,226,125,88,72,16                // vpbroadcastd  0x10(%rax),%ymm1
+  .byte  196,193,117,254,201                 // vpaddd        %ymm9,%ymm1,%ymm1
+  .byte  196,193,100,95,218                  // vmaxps        %ymm10,%ymm3,%ymm3
+  .byte  196,193,108,95,210                  // vmaxps        %ymm10,%ymm2,%ymm2
+  .byte  197,236,93,209                      // vminps        %ymm1,%ymm2,%ymm2
+  .byte  197,228,93,201                      // vminps        %ymm1,%ymm3,%ymm1
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  197,254,91,201                      // vcvttps2dq    %ymm1,%ymm1
+  .byte  197,254,91,210                      // vcvttps2dq    %ymm2,%ymm2
+  .byte  196,226,125,88,88,8                 // vpbroadcastd  0x8(%rax),%ymm3
+  .byte  196,226,101,64,210                  // vpmulld       %ymm2,%ymm3,%ymm2
+  .byte  196,226,101,64,217                  // vpmulld       %ymm1,%ymm3,%ymm3
+  .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
+  .byte  197,237,254,200                     // vpaddd        %ymm0,%ymm2,%ymm1
+  .byte  196,193,126,91,195                  // vcvttps2dq    %ymm11,%ymm0
+  .byte  197,229,254,192                     // vpaddd        %ymm0,%ymm3,%ymm0
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  197,249,110,208                     // vmovd         %eax,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,105,32,208,1                // vpinsrb       $0x1,%eax,%xmm2,%xmm2
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,201,1                // vextracti128  $0x1,%ymm1,%xmm1
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,105,32,208,2                // vpinsrb       $0x2,%eax,%xmm2,%xmm2
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,105,32,208,3                // vpinsrb       $0x3,%eax,%xmm2,%xmm2
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,105,32,208,4                // vpinsrb       $0x4,%eax,%xmm2,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,105,32,200,5                // vpinsrb       $0x5,%eax,%xmm2,%xmm1
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,6                // vpinsrb       $0x6,%eax,%xmm1,%xmm1
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,113,32,200,7                // vpinsrb       $0x7,%eax,%xmm1,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,8                // vpinsrb       $0x8,%eax,%xmm1,%xmm1
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,113,32,200,9                // vpinsrb       $0x9,%eax,%xmm1,%xmm1
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,192,1                // vextracti128  $0x1,%ymm0,%xmm0
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,10               // vpinsrb       $0xa,%eax,%xmm1,%xmm1
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,113,32,200,11               // vpinsrb       $0xb,%eax,%xmm1,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,12               // vpinsrb       $0xc,%eax,%xmm1,%xmm1
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,113,32,192,13               // vpinsrb       $0xd,%eax,%xmm1,%xmm0
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,121,32,192,14               // vpinsrb       $0xe,%eax,%xmm0,%xmm0
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,121,32,192,15               // vpinsrb       $0xf,%eax,%xmm0,%xmm0
+  .byte  196,226,125,48,216                  // vpmovzxbw     %xmm0,%ymm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  197,253,239,192                     // vpxor         %ymm0,%ymm0,%ymm0
+  .byte  197,245,239,201                     // vpxor         %ymm1,%ymm1,%ymm1
+  .byte  197,237,239,210                     // vpxor         %ymm2,%ymm2,%ymm2
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_g8_hsw_lowp
 .globl _sk_load_g8_hsw_lowp
 FUNCTION(_sk_load_g8_hsw_lowp)
@@ -53238,38 +53591,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2c80c <_sk_load_g8_hsw_lowp+0x3c>
+  .byte  119,37                              // ja            2ce37 <_sk_load_g8_hsw_lowp+0x3c>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,227,0,0,0                 // lea           0xe3(%rip),%r9        # 2c8d8 <_sk_load_g8_hsw_lowp+0x108>
+  .byte  76,141,13,224,0,0,0                 // lea           0xe0(%rip),%r9        # 2cf00 <_sk_load_g8_hsw_lowp+0x105>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,249,110,192                     // vmovd         %eax,%xmm0
-  .byte  233,175,0,0,0                       // jmpq          2c8bb <_sk_load_g8_hsw_lowp+0xeb>
+  .byte  233,175,0,0,0                       // jmpq          2cee6 <_sk_load_g8_hsw_lowp+0xeb>
   .byte  196,193,122,111,4,16                // vmovdqu       (%r8,%rdx,1),%xmm0
-  .byte  233,164,0,0,0                       // jmpq          2c8bb <_sk_load_g8_hsw_lowp+0xeb>
+  .byte  233,164,0,0,0                       // jmpq          2cee6 <_sk_load_g8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,2,2            // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm0,%xmm0
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,249,110,200                     // vmovd         %eax,%xmm1
   .byte  196,227,121,14,193,1                // vpblendw      $0x1,%xmm1,%xmm0,%xmm0
-  .byte  233,132,0,0,0                       // jmpq          2c8bb <_sk_load_g8_hsw_lowp+0xeb>
+  .byte  233,132,0,0,0                       // jmpq          2cee6 <_sk_load_g8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,6,6            // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,5,5            // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,4,4            // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,193,121,110,12,16               // vmovd         (%r8,%rdx,1),%xmm1
   .byte  196,227,121,2,193,1                 // vpblendd      $0x1,%xmm1,%xmm0,%xmm0
-  .byte  235,90                              // jmp           2c8bb <_sk_load_g8_hsw_lowp+0xeb>
+  .byte  235,90                              // jmp           2cee6 <_sk_load_g8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,10,10          // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,9,9            // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,8,8            // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,193,122,126,12,16               // vmovq         (%r8,%rdx,1),%xmm1
   .byte  196,227,113,2,192,12                // vpblendd      $0xc,%xmm0,%xmm1,%xmm0
-  .byte  235,48                              // jmp           2c8bb <_sk_load_g8_hsw_lowp+0xeb>
+  .byte  235,48                              // jmp           2cee6 <_sk_load_g8_hsw_lowp+0xeb>
   .byte  197,249,239,192                     // vpxor         %xmm0,%xmm0,%xmm0
   .byte  196,195,121,32,68,16,14,14          // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm0,%xmm0
   .byte  196,195,121,32,68,16,13,13          // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm0,%xmm0
@@ -53279,43 +53632,42 @@
   .byte  196,227,113,2,192,8                 // vpblendd      $0x8,%xmm0,%xmm1,%xmm0
   .byte  196,226,125,48,192                  // vpmovzxbw     %xmm0,%ymm0
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,29,73,106,0,0       // vpbroadcastw  0x6a49(%rip),%ymm3        # 33314 <_sk_srcover_rgba_8888_sse2_lowp+0x994>
+  .byte  196,226,125,121,29,230,122,0,0      // vpbroadcastw  0x7ae6(%rip),%ymm3        # 349dc <_sk_srcover_rgba_8888_sse2_lowp+0x9ac>
   .byte  197,253,111,200                     // vmovdqa       %ymm0,%ymm1
   .byte  197,253,111,208                     // vmovdqa       %ymm0,%ymm2
   .byte  255,224                             // jmpq          *%rax
-  .byte  15,31,0                             // nopl          (%rax)
-  .byte  38,255                              // es            (bad)
+  .byte  41,255                              // sub           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,75,255                          // decl          -0x1(%rbx)
+  .byte  255,78,255                          // decl          -0x1(%rsi)
+  .byte  255                                 // (bad)
+  .byte  255,66,255                          // incl          -0x1(%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  63                                  // (bad)
+  .byte  126,255                             // jle           2cf0d <_sk_load_g8_hsw_lowp+0x112>
+  .byte  255                                 // (bad)
+  .byte  255,118,255                         // pushq         -0x1(%rsi)
+  .byte  255                                 // (bad)
+  .byte  255,110,255                         // ljmp          *-0x1(%rsi)
+  .byte  255                                 // (bad)
+  .byte  255,98,255                          // jmpq          *-0x1(%rdx)
+  .byte  255                                 // (bad)
+  .byte  255,168,255,255,255,160             // ljmp          *-0x5f000001(%rax)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           2c8e5 <_sk_load_g8_hsw_lowp+0x115>
-  .byte  255                                 // (bad)
-  .byte  255,115,255                         // pushq         -0x1(%rbx)
-  .byte  255                                 // (bad)
-  .byte  255,107,255                         // ljmp          *-0x1(%rbx)
-  .byte  255                                 // (bad)
-  .byte  255,95,255                          // lcall         *-0x1(%rdi)
-  .byte  255                                 // (bad)
-  .byte  255,165,255,255,255,157             // jmpq          *-0x62000001(%rbp)
+  .byte  255,152,255,255,255,140             // lcall         *-0x73000001(%rax)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,149,255,255,255,137             // callq         *-0x76000001(%rbp)
+  .byte  255,210                             // callq         *%rdx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,207                             // dec           %edi
+  .byte  255,202                             // dec           %edx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,199                             // inc           %edi
+  .byte  255,194                             // inc           %edx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255                                 // (bad)
-  .byte  191,255,255,255,179                 // mov           $0xb3ffffff,%edi
-  .byte  255                                 // (bad)
+  .byte  255                                 // .byte         0xff
+  .byte  182,255                             // mov           $0xff,%dh
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
@@ -53331,38 +53683,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,37                              // ja            2c950 <_sk_load_g8_dst_hsw_lowp+0x3c>
+  .byte  119,37                              // ja            2cf78 <_sk_load_g8_dst_hsw_lowp+0x3c>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,227,0,0,0                 // lea           0xe3(%rip),%r9        # 2ca1c <_sk_load_g8_dst_hsw_lowp+0x108>
+  .byte  76,141,13,227,0,0,0                 // lea           0xe3(%rip),%r9        # 2d044 <_sk_load_g8_dst_hsw_lowp+0x108>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,249,110,224                     // vmovd         %eax,%xmm4
-  .byte  233,175,0,0,0                       // jmpq          2c9ff <_sk_load_g8_dst_hsw_lowp+0xeb>
+  .byte  233,175,0,0,0                       // jmpq          2d027 <_sk_load_g8_dst_hsw_lowp+0xeb>
   .byte  196,193,122,111,36,16               // vmovdqu       (%r8,%rdx,1),%xmm4
-  .byte  233,164,0,0,0                       // jmpq          2c9ff <_sk_load_g8_dst_hsw_lowp+0xeb>
+  .byte  233,164,0,0,0                       // jmpq          2d027 <_sk_load_g8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,2,2            // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm4,%xmm4
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,249,110,232                     // vmovd         %eax,%xmm5
   .byte  196,227,89,14,229,1                 // vpblendw      $0x1,%xmm5,%xmm4,%xmm4
-  .byte  233,132,0,0,0                       // jmpq          2c9ff <_sk_load_g8_dst_hsw_lowp+0xeb>
+  .byte  233,132,0,0,0                       // jmpq          2d027 <_sk_load_g8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,6,6            // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,5,5            // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,4,4            // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,193,121,110,44,16               // vmovd         (%r8,%rdx,1),%xmm5
   .byte  196,227,89,2,229,1                  // vpblendd      $0x1,%xmm5,%xmm4,%xmm4
-  .byte  235,90                              // jmp           2c9ff <_sk_load_g8_dst_hsw_lowp+0xeb>
+  .byte  235,90                              // jmp           2d027 <_sk_load_g8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,10,10          // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,9,9            // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,8,8            // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,193,122,126,44,16               // vmovq         (%r8,%rdx,1),%xmm5
   .byte  196,227,81,2,228,12                 // vpblendd      $0xc,%xmm4,%xmm5,%xmm4
-  .byte  235,48                              // jmp           2c9ff <_sk_load_g8_dst_hsw_lowp+0xeb>
+  .byte  235,48                              // jmp           2d027 <_sk_load_g8_dst_hsw_lowp+0xeb>
   .byte  197,217,239,228                     // vpxor         %xmm4,%xmm4,%xmm4
   .byte  196,195,89,32,100,16,14,14          // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm4,%xmm4
   .byte  196,195,89,32,100,16,13,13          // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm4,%xmm4
@@ -53372,7 +53724,7 @@
   .byte  196,227,81,2,228,8                  // vpblendd      $0x8,%xmm4,%xmm5,%xmm4
   .byte  196,226,125,48,228                  // vpmovzxbw     %xmm4,%ymm4
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  196,226,125,121,61,7,105,0,0        // vpbroadcastw  0x6907(%rip),%ymm7        # 33316 <_sk_srcover_rgba_8888_sse2_lowp+0x996>
+  .byte  196,226,125,121,61,167,121,0,0      // vpbroadcastw  0x79a7(%rip),%ymm7        # 349de <_sk_srcover_rgba_8888_sse2_lowp+0x9ae>
   .byte  197,253,111,236                     // vmovdqa       %ymm4,%ymm5
   .byte  197,253,111,244                     // vmovdqa       %ymm4,%ymm6
   .byte  255,224                             // jmpq          *%rax
@@ -53386,7 +53738,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           2ca29 <_sk_load_g8_dst_hsw_lowp+0x115>
+  .byte  123,255                             // jnp           2d051 <_sk_load_g8_dst_hsw_lowp+0x115>
   .byte  255                                 // (bad)
   .byte  255,115,255                         // pushq         -0x1(%rbx)
   .byte  255                                 // (bad)
@@ -53416,12 +53768,12 @@
 .globl _sk_luminance_to_alpha_hsw_lowp
 FUNCTION(_sk_luminance_to_alpha_hsw_lowp)
 _sk_luminance_to_alpha_hsw_lowp:
-  .byte  196,226,125,121,29,183,104,0,0      // vpbroadcastw  0x68b7(%rip),%ymm3        # 33318 <_sk_srcover_rgba_8888_sse2_lowp+0x998>
+  .byte  196,226,125,121,29,87,121,0,0       // vpbroadcastw  0x7957(%rip),%ymm3        # 349e0 <_sk_srcover_rgba_8888_sse2_lowp+0x9b0>
   .byte  197,253,213,195                     // vpmullw       %ymm3,%ymm0,%ymm0
-  .byte  196,226,125,121,29,172,104,0,0      // vpbroadcastw  0x68ac(%rip),%ymm3        # 3331a <_sk_srcover_rgba_8888_sse2_lowp+0x99a>
+  .byte  196,226,125,121,29,76,121,0,0       // vpbroadcastw  0x794c(%rip),%ymm3        # 349e2 <_sk_srcover_rgba_8888_sse2_lowp+0x9b2>
   .byte  197,245,213,203                     // vpmullw       %ymm3,%ymm1,%ymm1
   .byte  197,245,253,192                     // vpaddw        %ymm0,%ymm1,%ymm0
-  .byte  196,226,125,121,13,157,104,0,0      // vpbroadcastw  0x689d(%rip),%ymm1        # 3331c <_sk_srcover_rgba_8888_sse2_lowp+0x99c>
+  .byte  196,226,125,121,13,61,121,0,0       // vpbroadcastw  0x793d(%rip),%ymm1        # 349e4 <_sk_srcover_rgba_8888_sse2_lowp+0x9b4>
   .byte  197,237,213,201                     // vpmullw       %ymm1,%ymm2,%ymm1
   .byte  197,253,253,193                     // vpaddw        %ymm1,%ymm0,%ymm0
   .byte  197,229,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm3
@@ -53431,19 +53783,113 @@
   .byte  197,237,239,210                     // vpxor         %ymm2,%ymm2,%ymm2
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_g8_hsw_lowp
+.globl _sk_gather_g8_hsw_lowp
+FUNCTION(_sk_gather_g8_hsw_lowp)
+_sk_gather_g8_hsw_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,98,125,88,64,12                 // vpbroadcastd  0xc(%rax),%ymm8
+  .byte  196,65,53,118,201                   // vpcmpeqd      %ymm9,%ymm9,%ymm9
+  .byte  196,65,61,254,193                   // vpaddd        %ymm9,%ymm8,%ymm8
+  .byte  196,65,44,87,210                    // vxorps        %ymm10,%ymm10,%ymm10
+  .byte  196,193,124,95,194                  // vmaxps        %ymm10,%ymm0,%ymm0
+  .byte  196,193,116,95,202                  // vmaxps        %ymm10,%ymm1,%ymm1
+  .byte  196,65,116,93,216                   // vminps        %ymm8,%ymm1,%ymm11
+  .byte  196,193,124,93,192                  // vminps        %ymm8,%ymm0,%ymm0
+  .byte  196,226,125,88,72,16                // vpbroadcastd  0x10(%rax),%ymm1
+  .byte  196,193,117,254,201                 // vpaddd        %ymm9,%ymm1,%ymm1
+  .byte  196,193,100,95,218                  // vmaxps        %ymm10,%ymm3,%ymm3
+  .byte  196,193,108,95,210                  // vmaxps        %ymm10,%ymm2,%ymm2
+  .byte  197,236,93,209                      // vminps        %ymm1,%ymm2,%ymm2
+  .byte  197,228,93,201                      // vminps        %ymm1,%ymm3,%ymm1
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  197,254,91,201                      // vcvttps2dq    %ymm1,%ymm1
+  .byte  197,254,91,210                      // vcvttps2dq    %ymm2,%ymm2
+  .byte  196,226,125,88,88,8                 // vpbroadcastd  0x8(%rax),%ymm3
+  .byte  196,226,101,64,210                  // vpmulld       %ymm2,%ymm3,%ymm2
+  .byte  196,226,101,64,217                  // vpmulld       %ymm1,%ymm3,%ymm3
+  .byte  197,254,91,192                      // vcvttps2dq    %ymm0,%ymm0
+  .byte  197,237,254,200                     // vpaddd        %ymm0,%ymm2,%ymm1
+  .byte  196,193,126,91,195                  // vcvttps2dq    %ymm11,%ymm0
+  .byte  197,229,254,192                     // vpaddd        %ymm0,%ymm3,%ymm0
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  197,249,110,208                     // vmovd         %eax,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,105,32,208,1                // vpinsrb       $0x1,%eax,%xmm2,%xmm2
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,201,1                // vextracti128  $0x1,%ymm1,%xmm1
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,105,32,208,2                // vpinsrb       $0x2,%eax,%xmm2,%xmm2
+  .byte  196,193,249,126,201                 // vmovq         %xmm1,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,105,32,208,3                // vpinsrb       $0x3,%eax,%xmm2,%xmm2
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,105,32,208,4                // vpinsrb       $0x4,%eax,%xmm2,%xmm2
+  .byte  196,195,249,22,202,1                // vpextrq       $0x1,%xmm1,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,105,32,200,5                // vpinsrb       $0x5,%eax,%xmm2,%xmm1
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,6                // vpinsrb       $0x6,%eax,%xmm1,%xmm1
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,113,32,200,7                // vpinsrb       $0x7,%eax,%xmm1,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,8                // vpinsrb       $0x8,%eax,%xmm1,%xmm1
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,113,32,200,9                // vpinsrb       $0x9,%eax,%xmm1,%xmm1
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  196,227,125,57,192,1                // vextracti128  $0x1,%ymm0,%xmm0
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,10               // vpinsrb       $0xa,%eax,%xmm1,%xmm1
+  .byte  196,193,249,126,193                 // vmovq         %xmm0,%r9
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,113,32,200,11               // vpinsrb       $0xb,%eax,%xmm1,%xmm1
+  .byte  68,137,200                          // mov           %r9d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,113,32,200,12               // vpinsrb       $0xc,%eax,%xmm1,%xmm1
+  .byte  196,195,249,22,194,1                // vpextrq       $0x1,%xmm0,%r10
+  .byte  73,193,233,32                       // shr           $0x20,%r9
+  .byte  67,15,182,4,8                       // movzbl        (%r8,%r9,1),%eax
+  .byte  196,227,113,32,192,13               // vpinsrb       $0xd,%eax,%xmm1,%xmm0
+  .byte  68,137,208                          // mov           %r10d,%eax
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  196,227,121,32,192,14               // vpinsrb       $0xe,%eax,%xmm0,%xmm0
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  67,15,182,4,16                      // movzbl        (%r8,%r10,1),%eax
+  .byte  196,227,121,32,192,15               // vpinsrb       $0xf,%eax,%xmm0,%xmm0
+  .byte  196,226,125,48,192                  // vpmovzxbw     %xmm0,%ymm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  196,226,125,121,29,128,119,0,0      // vpbroadcastw  0x7780(%rip),%ymm3        # 349e6 <_sk_srcover_rgba_8888_sse2_lowp+0x9b6>
+  .byte  197,253,111,200                     // vmovdqa       %ymm0,%ymm1
+  .byte  197,253,111,208                     // vmovdqa       %ymm0,%ymm2
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_scale_1_float_hsw_lowp
 .globl _sk_scale_1_float_hsw_lowp
 FUNCTION(_sk_scale_1_float_hsw_lowp)
 _sk_scale_1_float_hsw_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,122,16,0                        // vmovss        (%rax),%xmm8
-  .byte  197,122,16,13,178,97,0,0            // vmovss        0x61b2(%rip),%xmm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  196,98,57,169,13,121,97,0,0         // vfmadd213ss   0x6179(%rip),%xmm8,%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,122,16,13,142,112,0,0           // vmovss        0x708e(%rip),%xmm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,57,169,13,85,112,0,0         // vfmadd213ss   0x7055(%rip),%xmm8,%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,122,44,193                  // vcvttss2si    %xmm9,%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
   .byte  196,66,125,121,192                  // vpbroadcastw  %xmm8,%ymm8
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,121,13,80,104,0,0        // vpbroadcastw  0x6850(%rip),%ymm9        # 3331e <_sk_srcover_rgba_8888_sse2_lowp+0x99e>
+  .byte  196,98,125,121,13,70,119,0,0        // vpbroadcastw  0x7746(%rip),%ymm9        # 349e8 <_sk_srcover_rgba_8888_sse2_lowp+0x9b8>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,189,213,201                     // vpmullw       %ymm1,%ymm8,%ymm1
@@ -53464,12 +53910,12 @@
 _sk_lerp_1_float_hsw_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  197,122,16,0                        // vmovss        (%rax),%xmm8
-  .byte  197,122,16,13,72,97,0,0             // vmovss        0x6148(%rip),%xmm9        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  196,98,57,169,13,15,97,0,0          // vfmadd213ss   0x610f(%rip),%xmm8,%xmm9        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  197,122,16,13,36,112,0,0            // vmovss        0x7024(%rip),%xmm9        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  196,98,57,169,13,235,111,0,0        // vfmadd213ss   0x6feb(%rip),%xmm8,%xmm9        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  196,193,122,44,193                  // vcvttss2si    %xmm9,%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
   .byte  196,66,125,121,192                  // vpbroadcastw  %xmm8,%ymm8
-  .byte  196,98,125,121,13,236,103,0,0       // vpbroadcastw  0x67ec(%rip),%ymm9        # 33320 <_sk_srcover_rgba_8888_sse2_lowp+0x9a0>
+  .byte  196,98,125,121,13,226,118,0,0       // vpbroadcastw  0x76e2(%rip),%ymm9        # 349ea <_sk_srcover_rgba_8888_sse2_lowp+0x9ba>
   .byte  196,65,53,249,208                   // vpsubw        %ymm8,%ymm9,%ymm10
   .byte  197,45,213,220                      // vpmullw       %ymm4,%ymm10,%ymm11
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
@@ -53506,38 +53952,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,38                              // ja            2cbd2 <_sk_scale_u8_hsw_lowp+0x3d>
+  .byte  119,38                              // ja            2d3a6 <_sk_scale_u8_hsw_lowp+0x3d>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,21,1,0,0                  // lea           0x115(%rip),%r9        # 2ccd0 <_sk_scale_u8_hsw_lowp+0x13b>
+  .byte  76,141,13,21,1,0,0                  // lea           0x115(%rip),%r9        # 2d4a4 <_sk_scale_u8_hsw_lowp+0x13b>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
-  .byte  233,179,0,0,0                       // jmpq          2cc85 <_sk_scale_u8_hsw_lowp+0xf0>
+  .byte  233,179,0,0,0                       // jmpq          2d459 <_sk_scale_u8_hsw_lowp+0xf0>
   .byte  196,65,122,111,4,16                 // vmovdqu       (%r8,%rdx,1),%xmm8
-  .byte  233,168,0,0,0                       // jmpq          2cc85 <_sk_scale_u8_hsw_lowp+0xf0>
+  .byte  233,168,0,0,0                       // jmpq          2d459 <_sk_scale_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,2,2              // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm8,%xmm8
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,121,110,200                     // vmovd         %eax,%xmm9
   .byte  196,67,57,14,193,1                  // vpblendw      $0x1,%xmm9,%xmm8,%xmm8
-  .byte  233,135,0,0,0                       // jmpq          2cc85 <_sk_scale_u8_hsw_lowp+0xf0>
+  .byte  233,135,0,0,0                       // jmpq          2d459 <_sk_scale_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,6,6              // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,5,5              // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,4,4              // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,65,121,110,12,16                // vmovd         (%r8,%rdx,1),%xmm9
   .byte  196,67,57,2,193,1                   // vpblendd      $0x1,%xmm9,%xmm8,%xmm8
-  .byte  235,92                              // jmp           2cc85 <_sk_scale_u8_hsw_lowp+0xf0>
+  .byte  235,92                              // jmp           2d459 <_sk_scale_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,10,10            // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,9,9              // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,8,8              // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,65,122,126,12,16                // vmovq         (%r8,%rdx,1),%xmm9
   .byte  196,67,49,2,192,12                  // vpblendd      $0xc,%xmm8,%xmm9,%xmm8
-  .byte  235,49                              // jmp           2cc85 <_sk_scale_u8_hsw_lowp+0xf0>
+  .byte  235,49                              // jmp           2d459 <_sk_scale_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,14,14            // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,13,13            // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm8,%xmm8
@@ -53547,7 +53993,7 @@
   .byte  196,67,49,2,192,8                   // vpblendd      $0x8,%xmm8,%xmm9,%xmm8
   .byte  196,66,125,48,192                   // vpmovzxbw     %xmm8,%ymm8
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,121,13,139,102,0,0       // vpbroadcastw  0x668b(%rip),%ymm9        # 33322 <_sk_srcover_rgba_8888_sse2_lowp+0x9a2>
+  .byte  196,98,125,121,13,129,117,0,0       // vpbroadcastw  0x7581(%rip),%ymm9        # 349ec <_sk_srcover_rgba_8888_sse2_lowp+0x9bc>
   .byte  196,193,125,253,193                 // vpaddw        %ymm9,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,189,213,201                     // vpmullw       %ymm1,%ymm8,%ymm1
@@ -53568,7 +54014,7 @@
   .byte  255,26                              // lcall         *(%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,75               // decl          0x4bffffff(%rip)        # 4c02ccdc <_sk_srcover_rgba_8888_sse2_lowp+0x4bffa35c>
+  .byte  255,13,255,255,255,75               // decl          0x4bffffff(%rip)        # 4c02d4b0 <_sk_srcover_rgba_8888_sse2_lowp+0x4bff9480>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,67,255                          // incl          -0x1(%rbx)
@@ -53607,38 +54053,38 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,38                              // ja            2cd49 <_sk_lerp_u8_hsw_lowp+0x3d>
+  .byte  119,38                              // ja            2d51d <_sk_lerp_u8_hsw_lowp+0x3d>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,58,1,0,0                  // lea           0x13a(%rip),%r9        # 2ce6c <_sk_lerp_u8_hsw_lowp+0x160>
+  .byte  76,141,13,58,1,0,0                  // lea           0x13a(%rip),%r9        # 2d640 <_sk_lerp_u8_hsw_lowp+0x160>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
-  .byte  233,179,0,0,0                       // jmpq          2cdfc <_sk_lerp_u8_hsw_lowp+0xf0>
+  .byte  233,179,0,0,0                       // jmpq          2d5d0 <_sk_lerp_u8_hsw_lowp+0xf0>
   .byte  196,65,122,111,4,16                 // vmovdqu       (%r8,%rdx,1),%xmm8
-  .byte  233,168,0,0,0                       // jmpq          2cdfc <_sk_lerp_u8_hsw_lowp+0xf0>
+  .byte  233,168,0,0,0                       // jmpq          2d5d0 <_sk_lerp_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,2,2              // vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm8,%xmm8
   .byte  65,15,183,4,16                      // movzwl        (%r8,%rdx,1),%eax
   .byte  197,121,110,200                     // vmovd         %eax,%xmm9
   .byte  196,67,57,14,193,1                  // vpblendw      $0x1,%xmm9,%xmm8,%xmm8
-  .byte  233,135,0,0,0                       // jmpq          2cdfc <_sk_lerp_u8_hsw_lowp+0xf0>
+  .byte  233,135,0,0,0                       // jmpq          2d5d0 <_sk_lerp_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,6,6              // vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,5,5              // vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,4,4              // vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,65,121,110,12,16                // vmovd         (%r8,%rdx,1),%xmm9
   .byte  196,67,57,2,193,1                   // vpblendd      $0x1,%xmm9,%xmm8,%xmm8
-  .byte  235,92                              // jmp           2cdfc <_sk_lerp_u8_hsw_lowp+0xf0>
+  .byte  235,92                              // jmp           2d5d0 <_sk_lerp_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,10,10            // vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,9,9              // vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,8,8              // vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,65,122,126,12,16                // vmovq         (%r8,%rdx,1),%xmm9
   .byte  196,67,49,2,192,12                  // vpblendd      $0xc,%xmm8,%xmm9,%xmm8
-  .byte  235,49                              // jmp           2cdfc <_sk_lerp_u8_hsw_lowp+0xf0>
+  .byte  235,49                              // jmp           2d5d0 <_sk_lerp_u8_hsw_lowp+0xf0>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,67,57,32,68,16,14,14            // vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm8,%xmm8
   .byte  196,67,57,32,68,16,13,13            // vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm8,%xmm8
@@ -53647,7 +54093,7 @@
   .byte  196,67,49,34,76,16,8,2              // vpinsrd       $0x2,0x8(%r8,%rdx,1),%xmm9,%xmm9
   .byte  196,67,49,2,192,8                   // vpblendd      $0x8,%xmm8,%xmm9,%xmm8
   .byte  196,66,125,48,192                   // vpmovzxbw     %xmm8,%ymm8
-  .byte  196,98,125,121,13,26,101,0,0        // vpbroadcastw  0x651a(%rip),%ymm9        # 33324 <_sk_srcover_rgba_8888_sse2_lowp+0x9a4>
+  .byte  196,98,125,121,13,16,116,0,0        // vpbroadcastw  0x7410(%rip),%ymm9        # 349ee <_sk_srcover_rgba_8888_sse2_lowp+0x9be>
   .byte  196,65,61,239,209                   // vpxor         %ymm9,%ymm8,%ymm10
   .byte  197,45,213,220                      // vpmullw       %ymm4,%ymm10,%ymm11
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
@@ -53679,7 +54125,7 @@
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  232,254,255,255,38                  // callq         2702ce77 <_sk_srcover_rgba_8888_sse2_lowp+0x26ffa4f7>
+  .byte  232,254,255,255,38                  // callq         2702d64b <_sk_srcover_rgba_8888_sse2_lowp+0x26ff961b>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,30                              // lcall         *(%rsi)
@@ -53700,7 +54146,7 @@
   .byte  255,52,255                          // pushq         (%rdi,%rdi,8)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  124,255                             // jl            2ce99 <_sk_lerp_u8_hsw_lowp+0x18d>
+  .byte  124,255                             // jl            2d66d <_sk_lerp_u8_hsw_lowp+0x18d>
   .byte  255                                 // (bad)
   .byte  255,116,255,255                     // pushq         -0x1(%rdi,%rdi,8)
   .byte  255,108,255,255                     // ljmp          *-0x1(%rdi,%rdi,8)
@@ -53721,24 +54167,24 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,38                              // ja            2cee8 <_sk_scale_565_hsw_lowp+0x40>
+  .byte  119,38                              // ja            2d6bc <_sk_scale_565_hsw_lowp+0x40>
   .byte  196,65,61,239,192                   // vpxor         %ymm8,%ymm8,%ymm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,255,1,0,0                 // lea           0x1ff(%rip),%r9        # 2d0d0 <_sk_scale_565_hsw_lowp+0x228>
+  .byte  76,141,13,255,1,0,0                 // lea           0x1ff(%rip),%r9        # 2d8a4 <_sk_scale_565_hsw_lowp+0x228>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
-  .byte  233,12,1,0,0                        // jmpq          2cff4 <_sk_scale_565_hsw_lowp+0x14c>
+  .byte  233,12,1,0,0                        // jmpq          2d7c8 <_sk_scale_565_hsw_lowp+0x14c>
   .byte  196,65,126,111,4,80                 // vmovdqu       (%r8,%rdx,2),%ymm8
-  .byte  233,1,1,0,0                         // jmpq          2cff4 <_sk_scale_565_hsw_lowp+0x14c>
+  .byte  233,1,1,0,0                         // jmpq          2d7c8 <_sk_scale_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,4,2             // vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,53,56,192,1                  // vinserti128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,121,110,12,80                // vmovd         (%r8,%rdx,2),%xmm9
   .byte  196,67,61,2,193,1                   // vpblendd      $0x1,%ymm9,%ymm8,%ymm8
-  .byte  233,221,0,0,0                       // jmpq          2cff4 <_sk_scale_565_hsw_lowp+0x14c>
+  .byte  233,221,0,0,0                       // jmpq          2d7c8 <_sk_scale_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,12,6            // vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,53,56,192,1                  // vinserti128   $0x1,%xmm8,%ymm9,%ymm8
@@ -53748,7 +54194,7 @@
   .byte  196,67,61,2,193,15                  // vpblendd      $0xf,%ymm9,%ymm8,%ymm8
   .byte  196,65,122,126,12,80                // vmovq         (%r8,%rdx,2),%xmm9
   .byte  196,67,61,2,193,3                   // vpblendd      $0x3,%ymm9,%ymm8,%ymm8
-  .byte  233,157,0,0,0                       // jmpq          2cff4 <_sk_scale_565_hsw_lowp+0x14c>
+  .byte  233,157,0,0,0                       // jmpq          2d7c8 <_sk_scale_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,20,2            // vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -53760,7 +54206,7 @@
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
   .byte  196,65,122,111,12,80                // vmovdqu       (%r8,%rdx,2),%xmm9
   .byte  196,67,53,2,192,240                 // vpblendd      $0xf0,%ymm8,%ymm9,%ymm8
-  .byte  235,84                              // jmp           2cff4 <_sk_scale_565_hsw_lowp+0x14c>
+  .byte  235,84                              // jmp           2d7c8 <_sk_scale_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,28,6            // vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -53774,13 +54220,13 @@
   .byte  196,65,122,111,20,80                // vmovdqu       (%r8,%rdx,2),%xmm10
   .byte  196,67,45,56,201,1                  // vinserti128   $0x1,%xmm9,%ymm10,%ymm9
   .byte  196,67,53,2,192,192                 // vpblendd      $0xc0,%ymm8,%ymm9,%ymm8
-  .byte  196,98,125,121,13,41,99,0,0         // vpbroadcastw  0x6329(%rip),%ymm9        # 33326 <_sk_srcover_rgba_8888_sse2_lowp+0x9a6>
+  .byte  196,98,125,121,13,31,114,0,0        // vpbroadcastw  0x721f(%rip),%ymm9        # 349f0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c0>
   .byte  196,193,45,113,208,8                // vpsrlw        $0x8,%ymm8,%ymm10
   .byte  196,65,45,219,201                   // vpand         %ymm9,%ymm10,%ymm9
   .byte  196,193,45,113,208,5                // vpsrlw        $0x5,%ymm8,%ymm10
-  .byte  196,98,125,121,29,17,99,0,0         // vpbroadcastw  0x6311(%rip),%ymm11        # 33328 <_sk_srcover_rgba_8888_sse2_lowp+0x9a8>
+  .byte  196,98,125,121,29,7,114,0,0         // vpbroadcastw  0x7207(%rip),%ymm11        # 349f2 <_sk_srcover_rgba_8888_sse2_lowp+0x9c2>
   .byte  196,65,45,219,211                   // vpand         %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,121,29,5,99,0,0          // vpbroadcastw  0x6305(%rip),%ymm11        # 3332a <_sk_srcover_rgba_8888_sse2_lowp+0x9aa>
+  .byte  196,98,125,121,29,251,113,0,0       // vpbroadcastw  0x71fb(%rip),%ymm11        # 349f4 <_sk_srcover_rgba_8888_sse2_lowp+0x9c4>
   .byte  196,65,61,219,219                   // vpand         %ymm11,%ymm8,%ymm11
   .byte  196,193,61,113,208,13               // vpsrlw        $0xd,%ymm8,%ymm8
   .byte  196,65,53,235,192                   // vpor          %ymm8,%ymm9,%ymm8
@@ -53790,7 +54236,7 @@
   .byte  196,193,45,113,243,3                // vpsllw        $0x3,%ymm11,%ymm10
   .byte  196,193,37,113,211,2                // vpsrlw        $0x2,%ymm11,%ymm11
   .byte  196,65,45,235,211                   // vpor          %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,121,29,204,98,0,0        // vpbroadcastw  0x62cc(%rip),%ymm11        # 3332c <_sk_srcover_rgba_8888_sse2_lowp+0x9ac>
+  .byte  196,98,125,121,29,194,113,0,0       // vpbroadcastw  0x71c2(%rip),%ymm11        # 349f6 <_sk_srcover_rgba_8888_sse2_lowp+0x9c6>
   .byte  196,65,101,239,227                  // vpxor         %ymm11,%ymm3,%ymm12
   .byte  196,65,69,239,219                   // vpxor         %ymm11,%ymm7,%ymm11
   .byte  196,65,37,101,220                   // vpcmpgtw      %ymm12,%ymm11,%ymm11
@@ -53800,7 +54246,7 @@
   .byte  196,66,21,62,232                    // vpmaxuw       %ymm8,%ymm13,%ymm13
   .byte  196,67,21,76,220,176                // vpblendvb     %ymm11,%ymm12,%ymm13,%ymm11
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
-  .byte  196,98,125,121,5,152,98,0,0         // vpbroadcastw  0x6298(%rip),%ymm8        # 3332e <_sk_srcover_rgba_8888_sse2_lowp+0x9ae>
+  .byte  196,98,125,121,5,142,113,0,0        // vpbroadcastw  0x718e(%rip),%ymm8        # 349f8 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
   .byte  196,193,125,253,192                 // vpaddw        %ymm8,%ymm0,%ymm0
   .byte  197,253,113,208,8                   // vpsrlw        $0x8,%ymm0,%ymm0
   .byte  197,181,213,201                     // vpmullw       %ymm1,%ymm9,%ymm1
@@ -53864,24 +54310,24 @@
   .byte  36,15                               // and           $0xf,%al
   .byte  254,200                             // dec           %al
   .byte  60,14                               // cmp           $0xe,%al
-  .byte  119,38                              // ja            2d14c <_sk_lerp_565_hsw_lowp+0x40>
+  .byte  119,38                              // ja            2d920 <_sk_lerp_565_hsw_lowp+0x40>
   .byte  196,65,61,239,192                   // vpxor         %ymm8,%ymm8,%ymm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,51,2,0,0                  // lea           0x233(%rip),%r9        # 2d368 <_sk_lerp_565_hsw_lowp+0x25c>
+  .byte  76,141,13,51,2,0,0                  // lea           0x233(%rip),%r9        # 2db3c <_sk_lerp_565_hsw_lowp+0x25c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  197,121,110,192                     // vmovd         %eax,%xmm8
-  .byte  233,12,1,0,0                        // jmpq          2d258 <_sk_lerp_565_hsw_lowp+0x14c>
+  .byte  233,12,1,0,0                        // jmpq          2da2c <_sk_lerp_565_hsw_lowp+0x14c>
   .byte  196,65,126,111,4,80                 // vmovdqu       (%r8,%rdx,2),%ymm8
-  .byte  233,1,1,0,0                         // jmpq          2d258 <_sk_lerp_565_hsw_lowp+0x14c>
+  .byte  233,1,1,0,0                         // jmpq          2da2c <_sk_lerp_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,4,2             // vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,53,56,192,1                  // vinserti128   $0x1,%xmm8,%ymm9,%ymm8
   .byte  196,65,121,110,12,80                // vmovd         (%r8,%rdx,2),%xmm9
   .byte  196,67,61,2,193,1                   // vpblendd      $0x1,%ymm9,%ymm8,%ymm8
-  .byte  233,221,0,0,0                       // jmpq          2d258 <_sk_lerp_565_hsw_lowp+0x14c>
+  .byte  233,221,0,0,0                       // jmpq          2da2c <_sk_lerp_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,12,6            // vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,53,56,192,1                  // vinserti128   $0x1,%xmm8,%ymm9,%ymm8
@@ -53891,7 +54337,7 @@
   .byte  196,67,61,2,193,15                  // vpblendd      $0xf,%ymm9,%ymm8,%ymm8
   .byte  196,65,122,126,12,80                // vmovq         (%r8,%rdx,2),%xmm9
   .byte  196,67,61,2,193,3                   // vpblendd      $0x3,%ymm9,%ymm8,%ymm8
-  .byte  233,157,0,0,0                       // jmpq          2d258 <_sk_lerp_565_hsw_lowp+0x14c>
+  .byte  233,157,0,0,0                       // jmpq          2da2c <_sk_lerp_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,20,2            // vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -53903,7 +54349,7 @@
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
   .byte  196,65,122,111,12,80                // vmovdqu       (%r8,%rdx,2),%xmm9
   .byte  196,67,53,2,192,240                 // vpblendd      $0xf0,%ymm8,%ymm9,%ymm8
-  .byte  235,84                              // jmp           2d258 <_sk_lerp_565_hsw_lowp+0x14c>
+  .byte  235,84                              // jmp           2da2c <_sk_lerp_565_hsw_lowp+0x14c>
   .byte  196,65,57,239,192                   // vpxor         %xmm8,%xmm8,%xmm8
   .byte  196,65,57,196,76,80,28,6            // vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm8,%xmm9
   .byte  196,67,61,56,193,1                  // vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -53917,13 +54363,13 @@
   .byte  196,65,122,111,20,80                // vmovdqu       (%r8,%rdx,2),%xmm10
   .byte  196,67,45,56,201,1                  // vinserti128   $0x1,%xmm9,%ymm10,%ymm9
   .byte  196,67,53,2,192,192                 // vpblendd      $0xc0,%ymm8,%ymm9,%ymm8
-  .byte  196,98,125,121,13,207,96,0,0        // vpbroadcastw  0x60cf(%rip),%ymm9        # 33330 <_sk_srcover_rgba_8888_sse2_lowp+0x9b0>
+  .byte  196,98,125,121,13,197,111,0,0       // vpbroadcastw  0x6fc5(%rip),%ymm9        # 349fa <_sk_srcover_rgba_8888_sse2_lowp+0x9ca>
   .byte  196,193,45,113,208,8                // vpsrlw        $0x8,%ymm8,%ymm10
   .byte  196,65,45,219,201                   // vpand         %ymm9,%ymm10,%ymm9
   .byte  196,193,45,113,208,5                // vpsrlw        $0x5,%ymm8,%ymm10
-  .byte  196,98,125,121,29,183,96,0,0        // vpbroadcastw  0x60b7(%rip),%ymm11        # 33332 <_sk_srcover_rgba_8888_sse2_lowp+0x9b2>
+  .byte  196,98,125,121,29,173,111,0,0       // vpbroadcastw  0x6fad(%rip),%ymm11        # 349fc <_sk_srcover_rgba_8888_sse2_lowp+0x9cc>
   .byte  196,65,45,219,211                   // vpand         %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,121,29,171,96,0,0        // vpbroadcastw  0x60ab(%rip),%ymm11        # 33334 <_sk_srcover_rgba_8888_sse2_lowp+0x9b4>
+  .byte  196,98,125,121,29,161,111,0,0       // vpbroadcastw  0x6fa1(%rip),%ymm11        # 349fe <_sk_srcover_rgba_8888_sse2_lowp+0x9ce>
   .byte  196,65,61,219,219                   // vpand         %ymm11,%ymm8,%ymm11
   .byte  196,193,61,113,208,13               // vpsrlw        $0xd,%ymm8,%ymm8
   .byte  196,65,53,235,192                   // vpor          %ymm8,%ymm9,%ymm8
@@ -53933,7 +54379,7 @@
   .byte  196,193,45,113,243,3                // vpsllw        $0x3,%ymm11,%ymm10
   .byte  196,193,37,113,211,2                // vpsrlw        $0x2,%ymm11,%ymm11
   .byte  196,65,45,235,211                   // vpor          %ymm11,%ymm10,%ymm10
-  .byte  196,98,125,121,29,114,96,0,0        // vpbroadcastw  0x6072(%rip),%ymm11        # 33336 <_sk_srcover_rgba_8888_sse2_lowp+0x9b6>
+  .byte  196,98,125,121,29,104,111,0,0       // vpbroadcastw  0x6f68(%rip),%ymm11        # 34a00 <_sk_srcover_rgba_8888_sse2_lowp+0x9d0>
   .byte  196,65,101,239,227                  // vpxor         %ymm11,%ymm3,%ymm12
   .byte  196,65,69,239,219                   // vpxor         %ymm11,%ymm7,%ymm11
   .byte  196,65,37,101,220                   // vpcmpgtw      %ymm12,%ymm11,%ymm11
@@ -53942,7 +54388,7 @@
   .byte  196,66,53,62,234                    // vpmaxuw       %ymm10,%ymm9,%ymm13
   .byte  196,66,21,62,232                    // vpmaxuw       %ymm8,%ymm13,%ymm13
   .byte  196,67,21,76,220,176                // vpblendvb     %ymm11,%ymm12,%ymm13,%ymm11
-  .byte  196,98,125,121,37,66,96,0,0         // vpbroadcastw  0x6042(%rip),%ymm12        # 33338 <_sk_srcover_rgba_8888_sse2_lowp+0x9b8>
+  .byte  196,98,125,121,37,56,111,0,0        // vpbroadcastw  0x6f38(%rip),%ymm12        # 34a02 <_sk_srcover_rgba_8888_sse2_lowp+0x9d2>
   .byte  196,65,61,239,236                   // vpxor         %ymm12,%ymm8,%ymm13
   .byte  197,21,213,236                      // vpmullw       %ymm4,%ymm13,%ymm13
   .byte  197,189,213,192                     // vpmullw       %ymm0,%ymm8,%ymm0
@@ -54023,9 +54469,9 @@
   .byte  65,254,202                          // dec           %r10b
   .byte  69,15,182,194                       // movzbl        %r10b,%r8d
   .byte  65,128,248,14                       // cmp           $0xe,%r8b
-  .byte  119,51                              // ja            2d400 <_sk_srcover_rgba_8888_hsw_lowp+0x5c>
+  .byte  119,51                              // ja            2dbd4 <_sk_srcover_rgba_8888_hsw_lowp+0x5c>
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
-  .byte  76,141,29,236,2,0,0                 // lea           0x2ec(%rip),%r11        # 2d6c4 <_sk_srcover_rgba_8888_hsw_lowp+0x320>
+  .byte  76,141,29,236,2,0,0                 // lea           0x2ec(%rip),%r11        # 2de98 <_sk_srcover_rgba_8888_hsw_lowp+0x320>
   .byte  75,99,4,131                         // movslq        (%r11,%r8,4),%rax
   .byte  76,1,216                            // add           %r11,%rax
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -54035,10 +54481,10 @@
   .byte  255,224                             // jmpq          *%rax
   .byte  196,193,121,110,41                  // vmovd         (%r9),%xmm5
   .byte  197,221,239,228                     // vpxor         %ymm4,%ymm4,%ymm4
-  .byte  233,22,1,0,0                        // jmpq          2d516 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  .byte  233,22,1,0,0                        // jmpq          2dcea <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   .byte  196,193,126,111,41                  // vmovdqu       (%r9),%ymm5
   .byte  196,193,126,111,97,32               // vmovdqu       0x20(%r9),%ymm4
-  .byte  233,6,1,0,0                         // jmpq          2d516 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  .byte  233,6,1,0,0                         // jmpq          2dcea <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   .byte  196,193,121,110,97,8                // vmovd         0x8(%r9),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,197,239,255                     // vpxor         %ymm7,%ymm7,%ymm7
@@ -54047,7 +54493,7 @@
   .byte  197,249,112,237,232                 // vpshufd       $0xe8,%xmm5,%xmm5
   .byte  196,227,93,2,237,3                  // vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   .byte  197,253,111,231                     // vmovdqa       %ymm7,%ymm4
-  .byte  233,216,0,0,0                       // jmpq          2d516 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  .byte  233,216,0,0,0                       // jmpq          2dcea <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   .byte  196,193,121,110,97,24               // vmovd         0x18(%r9),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -54063,7 +54509,7 @@
   .byte  196,193,122,111,41                  // vmovdqu       (%r9),%xmm5
   .byte  196,227,85,2,236,240                // vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   .byte  197,125,127,196                     // vmovdqa       %ymm8,%ymm4
-  .byte  233,129,0,0,0                       // jmpq          2d516 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  .byte  233,129,0,0,0                       // jmpq          2dcea <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   .byte  196,193,121,110,97,40               // vmovd         0x28(%r9),%xmm4
   .byte  196,226,121,89,228                  // vpbroadcastq  %xmm4,%xmm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -54073,7 +54519,7 @@
   .byte  196,193,121,110,105,32              // vmovd         0x20(%r9),%xmm5
   .byte  196,227,93,2,229,1                  // vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   .byte  196,193,126,111,41                  // vmovdqu       (%r9),%ymm5
-  .byte  235,76                              // jmp           2d516 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  .byte  235,76                              // jmp           2dcea <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   .byte  196,193,121,110,97,56               // vmovd         0x38(%r9),%xmm4
   .byte  196,226,125,89,228                  // vpbroadcastq  %xmm4,%ymm4
   .byte  197,213,239,237                     // vpxor         %ymm5,%ymm5,%ymm5
@@ -54093,9 +54539,9 @@
   .byte  196,227,69,14,236,170               // vpblendw      $0xaa,%ymm4,%ymm7,%ymm5
   .byte  196,227,77,14,228,170               // vpblendw      $0xaa,%ymm4,%ymm6,%ymm4
   .byte  196,226,93,43,237                   // vpackusdw     %ymm5,%ymm4,%ymm5
-  .byte  197,125,111,5,1,94,0,0              // vmovdqa       0x5e01(%rip),%ymm8        # 33340 <_sk_srcover_rgba_8888_sse2_lowp+0x9c0>
+  .byte  197,125,111,5,13,109,0,0            // vmovdqa       0x6d0d(%rip),%ymm8        # 34a20 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
   .byte  196,193,85,219,224                  // vpand         %ymm8,%ymm5,%ymm4
-  .byte  196,98,125,121,13,19,94,0,0         // vpbroadcastw  0x5e13(%rip),%ymm9        # 33360 <_sk_srcover_rgba_8888_sse2_lowp+0x9e0>
+  .byte  196,98,125,121,13,31,109,0,0        // vpbroadcastw  0x6d1f(%rip),%ymm9        # 34a40 <_sk_srcover_rgba_8888_sse2_lowp+0xa10>
   .byte  197,213,113,213,8                   // vpsrlw        $0x8,%ymm5,%ymm5
   .byte  197,197,114,215,16                  // vpsrld        $0x10,%ymm7,%ymm7
   .byte  197,205,114,214,16                  // vpsrld        $0x10,%ymm6,%ymm6
@@ -54134,19 +54580,19 @@
   .byte  196,193,53,114,243,16               // vpslld        $0x10,%ymm11,%ymm9
   .byte  196,65,53,235,202                   // vpor          %ymm10,%ymm9,%ymm9
   .byte  65,128,250,14                       // cmp           $0xe,%r10b
-  .byte  119,26                              // ja            2d627 <_sk_srcover_rgba_8888_hsw_lowp+0x283>
-  .byte  76,141,21,236,0,0,0                 // lea           0xec(%rip),%r10        # 2d700 <_sk_srcover_rgba_8888_hsw_lowp+0x35c>
+  .byte  119,26                              // ja            2ddfb <_sk_srcover_rgba_8888_hsw_lowp+0x283>
+  .byte  76,141,21,236,0,0,0                 // lea           0xec(%rip),%r10        # 2ded4 <_sk_srcover_rgba_8888_hsw_lowp+0x35c>
   .byte  75,99,4,130                         // movslq        (%r10,%r8,4),%rax
   .byte  76,1,208                            // add           %r10,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  196,65,121,126,1                    // vmovd         %xmm8,(%r9)
-  .byte  233,151,0,0,0                       // jmpq          2d6be <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  .byte  233,151,0,0,0                       // jmpq          2de92 <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   .byte  196,65,126,127,1                    // vmovdqu       %ymm8,(%r9)
   .byte  196,65,126,127,73,32                // vmovdqu       %ymm9,0x20(%r9)
-  .byte  233,135,0,0,0                       // jmpq          2d6be <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  .byte  233,135,0,0,0                       // jmpq          2de92 <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   .byte  196,67,121,22,65,8,2                // vpextrd       $0x2,%xmm8,0x8(%r9)
   .byte  196,65,121,214,1                    // vmovq         %xmm8,(%r9)
-  .byte  235,121                             // jmp           2d6be <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  .byte  235,121                             // jmp           2de92 <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,67,121,22,73,24,2               // vpextrd       $0x2,%xmm9,0x18(%r9)
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
@@ -54154,12 +54600,12 @@
   .byte  196,67,125,57,193,1                 // vextracti128  $0x1,%ymm8,%xmm9
   .byte  196,65,121,126,73,16                // vmovd         %xmm9,0x10(%r9)
   .byte  196,65,122,127,1                    // vmovdqu       %xmm8,(%r9)
-  .byte  235,76                              // jmp           2d6be <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  .byte  235,76                              // jmp           2de92 <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   .byte  196,67,121,22,73,40,2               // vpextrd       $0x2,%xmm9,0x28(%r9)
   .byte  196,67,121,22,73,36,1               // vpextrd       $0x1,%xmm9,0x24(%r9)
   .byte  196,65,121,126,73,32                // vmovd         %xmm9,0x20(%r9)
   .byte  196,65,126,127,1                    // vmovdqu       %ymm8,(%r9)
-  .byte  235,49                              // jmp           2d6be <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  .byte  235,49                              // jmp           2de92 <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
   .byte  196,67,121,22,81,56,2               // vpextrd       $0x2,%xmm10,0x38(%r9)
   .byte  196,67,125,57,202,1                 // vextracti128  $0x1,%ymm9,%xmm10
@@ -54207,7 +54653,7 @@
   .byte  255,6                               // incl          (%rsi)
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,62               // lcall         *0x3effffff(%rip)        # 3f02d704 <_sk_srcover_rgba_8888_sse2_lowp+0x3effad84>
+  .byte  255,29,255,255,255,62               // lcall         *0x3effffff(%rip)        # 3f02ded8 <_sk_srcover_rgba_8888_sse2_lowp+0x3eff9ea8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,55                              // pushq         (%rdi)
@@ -54225,7 +54671,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  121,255                             // jns           2d725 <_sk_srcover_rgba_8888_hsw_lowp+0x381>
+  .byte  121,255                             // jns           2def9 <_sk_srcover_rgba_8888_hsw_lowp+0x381>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -54258,13 +54704,13 @@
   .byte  73,137,246                          // mov           %rsi,%r14
   .byte  72,137,77,192                       // mov           %rcx,-0x40(%rbp)
   .byte  72,57,203                           // cmp           %rcx,%rbx
-  .byte  15,131,131,0,0,0                    // jae           2d7f2 <_sk_start_pipeline_sse41_lowp+0xb6>
+  .byte  15,131,131,0,0,0                    // jae           2dfc6 <_sk_start_pipeline_sse41_lowp+0xb6>
   .byte  72,139,69,208                       // mov           -0x30(%rbp),%rax
   .byte  72,141,64,8                         // lea           0x8(%rax),%rax
   .byte  72,137,69,200                       // mov           %rax,-0x38(%rbp)
   .byte  76,57,125,200                       // cmp           %r15,-0x38(%rbp)
   .byte  72,139,85,208                       // mov           -0x30(%rbp),%rdx
-  .byte  119,59                              // ja            2d7c0 <_sk_start_pipeline_sse41_lowp+0x84>
+  .byte  119,59                              // ja            2df94 <_sk_start_pipeline_sse41_lowp+0x84>
   .byte  76,139,101,208                      // mov           -0x30(%rbp),%r12
   .byte  49,255                              // xor           %edi,%edi
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -54283,10 +54729,10 @@
   .byte  73,131,196,16                       // add           $0x10,%r12
   .byte  77,57,252                           // cmp           %r15,%r12
   .byte  73,137,212                          // mov           %rdx,%r12
-  .byte  118,201                             // jbe           2d789 <_sk_start_pipeline_sse41_lowp+0x4d>
+  .byte  118,201                             // jbe           2df5d <_sk_start_pipeline_sse41_lowp+0x4d>
   .byte  76,137,255                          // mov           %r15,%rdi
   .byte  72,41,215                           // sub           %rdx,%rdi
-  .byte  116,33                              // je            2d7e9 <_sk_start_pipeline_sse41_lowp+0xad>
+  .byte  116,33                              // je            2dfbd <_sk_start_pipeline_sse41_lowp+0xad>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -54300,7 +54746,7 @@
   .byte  65,255,213                          // callq         *%r13
   .byte  72,255,195                          // inc           %rbx
   .byte  72,59,93,192                        // cmp           -0x40(%rbp),%rbx
-  .byte  117,137                             // jne           2d77b <_sk_start_pipeline_sse41_lowp+0x3f>
+  .byte  117,137                             // jne           2df4f <_sk_start_pipeline_sse41_lowp+0x3f>
   .byte  72,131,196,24                       // add           $0x18,%rsp
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
@@ -54331,11 +54777,51 @@
   .byte  102,15,110,209                      // movd          %ecx,%xmm2
   .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,88,21,144,91,0,0                 // addps         0x5b90(%rip),%xmm2        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,21,156,106,0,0                // addps         0x6a9c(%rip),%xmm2        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_translate_sse41_lowp
+.globl _sk_matrix_translate_sse41_lowp
+FUNCTION(_sk_matrix_translate_sse41_lowp)
+_sk_matrix_translate_sse41_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
+  .byte  243,68,15,16,72,4                   // movss         0x4(%rax),%xmm9
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  65,15,88,200                        // addps         %xmm8,%xmm1
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  65,15,88,209                        // addps         %xmm9,%xmm2
+  .byte  65,15,88,217                        // addps         %xmm9,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_matrix_scale_translate_sse41_lowp
+.globl _sk_matrix_scale_translate_sse41_lowp
+FUNCTION(_sk_matrix_scale_translate_sse41_lowp)
+_sk_matrix_scale_translate_sse41_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
+  .byte  243,68,15,16,72,4                   // movss         0x4(%rax),%xmm9
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  243,68,15,16,80,8                   // movss         0x8(%rax),%xmm10
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
+  .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
+  .byte  65,15,88,194                        // addps         %xmm10,%xmm0
+  .byte  65,15,88,202                        // addps         %xmm10,%xmm1
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  65,15,89,217                        // mulps         %xmm9,%xmm3
+  .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
+  .byte  65,15,88,208                        // addps         %xmm8,%xmm2
+  .byte  65,15,88,216                        // addps         %xmm8,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_matrix_2x3_sse41_lowp
 .globl _sk_matrix_2x3_sse41_lowp
 FUNCTION(_sk_matrix_2x3_sse41_lowp)
@@ -54377,6 +54863,70 @@
   .byte  65,15,40,218                        // movaps        %xmm10,%xmm3
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_perspective_sse41_lowp
+.globl _sk_matrix_perspective_sse41_lowp
+FUNCTION(_sk_matrix_perspective_sse41_lowp)
+_sk_matrix_perspective_sse41_lowp:
+  .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,15,16,8                         // movss         (%rax),%xmm1
+  .byte  243,68,15,16,80,4                   // movss         0x4(%rax),%xmm10
+  .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  243,68,15,16,72,8                   // movss         0x8(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  68,15,40,219                        // movaps        %xmm3,%xmm11
+  .byte  69,15,89,218                        // mulps         %xmm10,%xmm11
+  .byte  68,15,89,210                        // mulps         %xmm2,%xmm10
+  .byte  69,15,88,209                        // addps         %xmm9,%xmm10
+  .byte  69,15,88,217                        // addps         %xmm9,%xmm11
+  .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
+  .byte  68,15,89,201                        // mulps         %xmm1,%xmm9
+  .byte  69,15,88,202                        // addps         %xmm10,%xmm9
+  .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
+  .byte  65,15,88,203                        // addps         %xmm11,%xmm1
+  .byte  243,68,15,16,96,12                  // movss         0xc(%rax),%xmm12
+  .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
+  .byte  243,68,15,16,80,16                  // movss         0x10(%rax),%xmm10
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  243,68,15,16,104,20                 // movss         0x14(%rax),%xmm13
+  .byte  69,15,198,237,0                     // shufps        $0x0,%xmm13,%xmm13
+  .byte  68,15,40,219                        // movaps        %xmm3,%xmm11
+  .byte  69,15,89,218                        // mulps         %xmm10,%xmm11
+  .byte  68,15,89,210                        // mulps         %xmm2,%xmm10
+  .byte  69,15,88,213                        // addps         %xmm13,%xmm10
+  .byte  69,15,88,221                        // addps         %xmm13,%xmm11
+  .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
+  .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
+  .byte  69,15,88,213                        // addps         %xmm13,%xmm10
+  .byte  69,15,89,224                        // mulps         %xmm8,%xmm12
+  .byte  69,15,88,220                        // addps         %xmm12,%xmm11
+  .byte  243,68,15,16,96,24                  // movss         0x18(%rax),%xmm12
+  .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
+  .byte  243,68,15,16,104,28                 // movss         0x1c(%rax),%xmm13
+  .byte  69,15,198,237,0                     // shufps        $0x0,%xmm13,%xmm13
+  .byte  243,68,15,16,112,32                 // movss         0x20(%rax),%xmm14
+  .byte  69,15,198,246,0                     // shufps        $0x0,%xmm14,%xmm14
+  .byte  65,15,89,221                        // mulps         %xmm13,%xmm3
+  .byte  68,15,89,234                        // mulps         %xmm2,%xmm13
+  .byte  69,15,88,238                        // addps         %xmm14,%xmm13
+  .byte  65,15,88,222                        // addps         %xmm14,%xmm3
+  .byte  65,15,89,196                        // mulps         %xmm12,%xmm0
+  .byte  65,15,88,197                        // addps         %xmm13,%xmm0
+  .byte  69,15,89,224                        // mulps         %xmm8,%xmm12
+  .byte  68,15,88,227                        // addps         %xmm3,%xmm12
+  .byte  15,83,192                           // rcpps         %xmm0,%xmm0
+  .byte  65,15,83,212                        // rcpps         %xmm12,%xmm2
+  .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
+  .byte  15,89,202                           // mulps         %xmm2,%xmm1
+  .byte  68,15,89,208                        // mulps         %xmm0,%xmm10
+  .byte  68,15,89,218                        // mulps         %xmm2,%xmm11
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
+  .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
+  .byte  65,15,40,219                        // movaps        %xmm11,%xmm3
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_uniform_color_sse41_lowp
 .globl _sk_uniform_color_sse41_lowp
 FUNCTION(_sk_uniform_color_sse41_lowp)
@@ -54404,7 +54954,7 @@
 FUNCTION(_sk_black_color_sse41_lowp)
 _sk_black_color_sse41_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,77,90,0,0                  // movaps        0x5a4d(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,225,103,0,0                // movaps        0x67e1(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -54415,7 +54965,7 @@
 FUNCTION(_sk_white_color_sse41_lowp)
 _sk_white_color_sse41_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,5,57,90,0,0                   // movaps        0x5a39(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,5,205,103,0,0                 // movaps        0x67cd(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -54426,10 +54976,10 @@
 FUNCTION(_sk_set_rgb_sse41_lowp)
 _sk_set_rgb_sse41_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  243,15,16,21,16,83,0,0              // movss         0x5310(%rip),%xmm2        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,15,16,21,116,96,0,0             // movss         0x6074(%rip),%xmm2        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  243,15,16,0                         // movss         (%rax),%xmm0
   .byte  243,15,89,194                       // mulss         %xmm2,%xmm0
-  .byte  243,68,15,16,5,207,82,0,0           // movss         0x52cf(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,16,5,51,96,0,0            // movss         0x6033(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,88,192                    // addss         %xmm8,%xmm0
   .byte  243,68,15,44,192                    // cvttss2si     %xmm0,%r8d
   .byte  102,65,15,110,192                   // movd          %r8d,%xmm0
@@ -54456,7 +55006,7 @@
 FUNCTION(_sk_premul_sse41_lowp)
 _sk_premul_sse41_lowp:
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,68,15,111,5,171,89,0,0          // movdqa        0x59ab(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,63,103,0,0          // movdqa        0x673f(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
@@ -54504,7 +55054,7 @@
 .globl _sk_invert_sse41_lowp
 FUNCTION(_sk_invert_sse41_lowp)
 _sk_invert_sse41_lowp:
-  .byte  102,68,15,111,5,73,89,0,0           // movdqa        0x5949(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,221,102,0,0         // movdqa        0x66dd(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,200                   // psubw         %xmm0,%xmm9
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -54536,7 +55086,7 @@
 _sk_srcatop_sse41_lowp:
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,15,111,29,237,88,0,0            // movdqa        0x58ed(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,29,129,102,0,0           // movdqa        0x6681(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,111,203                   // movdqa        %xmm3,%xmm9
   .byte  102,69,15,249,200                   // psubw         %xmm8,%xmm9
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
@@ -54568,7 +55118,7 @@
 _sk_dstatop_sse41_lowp:
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
   .byte  102,68,15,213,195                   // pmullw        %xmm3,%xmm8
-  .byte  102,68,15,111,13,109,88,0,0         // movdqa        0x586d(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,1,102,0,0          // movdqa        0x6601(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,253,193                   // paddw         %xmm9,%xmm8
   .byte  102,68,15,111,213                   // movdqa        %xmm5,%xmm10
   .byte  102,68,15,213,211                   // pmullw        %xmm3,%xmm10
@@ -54597,7 +55147,7 @@
 FUNCTION(_sk_srcin_sse41_lowp)
 _sk_srcin_sse41_lowp:
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,68,15,111,5,248,87,0,0          // movdqa        0x57f8(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,140,101,0,0         // movdqa        0x658c(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,207                      // pmullw        %xmm7,%xmm1
@@ -54618,7 +55168,7 @@
 _sk_dstin_sse41_lowp:
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,68,15,111,5,175,87,0,0          // movdqa        0x57af(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,67,101,0,0          // movdqa        0x6543(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
@@ -54639,7 +55189,7 @@
 .globl _sk_srcout_sse41_lowp
 FUNCTION(_sk_srcout_sse41_lowp)
 _sk_srcout_sse41_lowp:
-  .byte  102,68,15,111,5,102,87,0,0          // movdqa        0x5766(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,250,100,0,0         // movdqa        0x64fa(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,207                   // psubw         %xmm7,%xmm9
   .byte  102,65,15,213,193                   // pmullw        %xmm9,%xmm0
@@ -54662,7 +55212,7 @@
 FUNCTION(_sk_dstout_sse41_lowp)
 _sk_dstout_sse41_lowp:
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,68,15,111,5,15,87,0,0           // movdqa        0x570f(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,163,100,0,0         // movdqa        0x64a3(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  102,15,249,216                      // psubw         %xmm0,%xmm3
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
@@ -54687,7 +55237,7 @@
 .globl _sk_srcover_sse41_lowp
 FUNCTION(_sk_srcover_sse41_lowp)
 _sk_srcover_sse41_lowp:
-  .byte  102,68,15,111,13,181,86,0,0         // movdqa        0x56b5(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,73,100,0,0         // movdqa        0x6449(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,68,15,249,195                   // psubw         %xmm3,%xmm8
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -54716,7 +55266,7 @@
 .globl _sk_dstover_sse41_lowp
 FUNCTION(_sk_dstover_sse41_lowp)
 _sk_dstover_sse41_lowp:
-  .byte  102,68,15,111,5,59,86,0,0           // movdqa        0x563b(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,207,99,0,0          // movdqa        0x63cf(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,207                   // psubw         %xmm7,%xmm9
   .byte  102,65,15,213,193                   // pmullw        %xmm9,%xmm0
@@ -54743,7 +55293,7 @@
 FUNCTION(_sk_modulate_sse41_lowp)
 _sk_modulate_sse41_lowp:
   .byte  102,15,213,196                      // pmullw        %xmm4,%xmm0
-  .byte  102,68,15,111,5,212,85,0,0          // movdqa        0x55d4(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,104,99,0,0          // movdqa        0x6368(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,205                      // pmullw        %xmm5,%xmm1
@@ -54762,7 +55312,7 @@
 .globl _sk_multiply_sse41_lowp
 FUNCTION(_sk_multiply_sse41_lowp)
 _sk_multiply_sse41_lowp:
-  .byte  102,68,15,111,13,147,85,0,0         // movdqa        0x5593(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,39,99,0,0          // movdqa        0x6327(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,68,15,249,195                   // psubw         %xmm3,%xmm8
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -54805,7 +55355,7 @@
 FUNCTION(_sk_plus__sse41_lowp)
 _sk_plus__sse41_lowp:
   .byte  102,15,253,196                      // paddw         %xmm4,%xmm0
-  .byte  102,68,15,111,5,216,84,0,0          // movdqa        0x54d8(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,108,98,0,0          // movdqa        0x626c(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,56,58,192                 // pminuw        %xmm8,%xmm0
   .byte  102,15,253,205                      // paddw         %xmm5,%xmm1
   .byte  102,65,15,56,58,200                 // pminuw        %xmm8,%xmm1
@@ -54825,7 +55375,7 @@
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,204                      // pmullw        %xmm4,%xmm1
-  .byte  102,68,15,111,21,146,84,0,0         // movdqa        0x5492(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,21,38,98,0,0          // movdqa        0x6226(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,202                   // paddw         %xmm10,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -54857,7 +55407,7 @@
 FUNCTION(_sk_xor__sse41_lowp)
 _sk_xor__sse41_lowp:
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
-  .byte  102,68,15,111,13,16,84,0,0          // movdqa        0x5410(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,164,97,0,0         // movdqa        0x61a4(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,209                   // movdqa        %xmm9,%xmm10
   .byte  102,68,15,249,215                   // psubw         %xmm7,%xmm10
   .byte  102,65,15,213,194                   // pmullw        %xmm10,%xmm0
@@ -54900,7 +55450,7 @@
   .byte  102,68,15,213,207                   // pmullw        %xmm7,%xmm9
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
   .byte  102,65,15,56,62,201                 // pmaxuw        %xmm9,%xmm1
-  .byte  102,68,15,111,13,87,83,0,0          // movdqa        0x5357(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,235,96,0,0         // movdqa        0x60eb(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,201                   // paddw         %xmm9,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -54944,7 +55494,7 @@
   .byte  102,68,15,213,207                   // pmullw        %xmm7,%xmm9
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
   .byte  102,65,15,56,58,201                 // pminuw        %xmm9,%xmm1
-  .byte  102,68,15,111,13,154,82,0,0         // movdqa        0x529a(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,46,96,0,0          // movdqa        0x602e(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,201                   // paddw         %xmm9,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -54988,10 +55538,10 @@
   .byte  102,68,15,213,207                   // pmullw        %xmm7,%xmm9
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
   .byte  102,65,15,56,58,201                 // pminuw        %xmm9,%xmm1
-  .byte  102,68,15,111,13,221,81,0,0         // movdqa        0x51dd(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,113,95,0,0         // movdqa        0x5f71(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,201                   // paddw         %xmm9,%xmm1
   .byte  102,15,113,209,7                    // psrlw         $0x7,%xmm1
-  .byte  102,68,15,111,21,42,89,0,0          // movdqa        0x592a(%rip),%xmm10        # 33ad0 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
+  .byte  102,68,15,111,21,190,102,0,0        // movdqa        0x66be(%rip),%xmm10        # 351b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
   .byte  102,65,15,219,202                   // pand          %xmm10,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
   .byte  102,68,15,111,221                   // movdqa        %xmm5,%xmm11
@@ -55033,10 +55583,10 @@
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,204                      // pmullw        %xmm4,%xmm1
-  .byte  102,68,15,111,13,25,81,0,0          // movdqa        0x5119(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,173,94,0,0         // movdqa        0x5ead(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,201                   // paddw         %xmm9,%xmm1
   .byte  102,15,113,209,7                    // psrlw         $0x7,%xmm1
-  .byte  102,68,15,111,21,102,88,0,0         // movdqa        0x5866(%rip),%xmm10        # 33ad0 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
+  .byte  102,68,15,111,21,250,101,0,0        // movdqa        0x65fa(%rip),%xmm10        # 351b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
   .byte  102,65,15,219,202                   // pand          %xmm10,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
@@ -55072,14 +55622,14 @@
   .byte  102,15,111,245                      // movdqa        %xmm5,%xmm6
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
-  .byte  102,68,15,111,29,116,80,0,0         // movdqa        0x5074(%rip),%xmm11        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,29,8,94,0,0           // movdqa        0x5e08(%rip),%xmm11        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,249,220                   // psubw         %xmm12,%xmm11
   .byte  102,65,15,111,227                   // movdqa        %xmm11,%xmm4
   .byte  102,65,15,213,224                   // pmullw        %xmm8,%xmm4
   .byte  102,68,15,111,251                   // movdqa        %xmm3,%xmm15
   .byte  102,69,15,249,248                   // psubw         %xmm8,%xmm15
   .byte  102,69,15,253,192                   // paddw         %xmm8,%xmm8
-  .byte  102,68,15,111,13,189,87,0,0         // movdqa        0x57bd(%rip),%xmm9        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,13,81,101,0,0         // movdqa        0x6551(%rip),%xmm9        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,68,15,111,243                   // movdqa        %xmm3,%xmm14
   .byte  102,69,15,239,241                   // pxor          %xmm9,%xmm14
   .byte  102,65,15,111,196                   // movdqa        %xmm12,%xmm0
@@ -55126,7 +55676,7 @@
   .byte  102,68,15,111,232                   // movdqa        %xmm0,%xmm13
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,65,15,56,16,210                 // pblendvb      %xmm0,%xmm10,%xmm2
-  .byte  102,68,15,111,13,101,79,0,0         // movdqa        0x4f65(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,249,92,0,0         // movdqa        0x5cf9(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,249,195                      // psubw         %xmm3,%xmm0
   .byte  102,15,111,248                      // movdqa        %xmm0,%xmm7
@@ -55165,7 +55715,7 @@
 _sk_overlay_sse41_lowp:
   .byte  102,68,15,111,231                   // movdqa        %xmm7,%xmm12
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
-  .byte  102,68,15,111,29,200,78,0,0         // movdqa        0x4ec8(%rip),%xmm11        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,29,92,92,0,0          // movdqa        0x5c5c(%rip),%xmm11        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,249,220                   // psubw         %xmm12,%xmm11
   .byte  102,65,15,111,195                   // movdqa        %xmm11,%xmm0
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
@@ -55177,7 +55727,7 @@
   .byte  102,68,15,249,252                   // psubw         %xmm4,%xmm15
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,253,192                      // paddw         %xmm0,%xmm0
-  .byte  102,68,15,111,13,249,85,0,0         // movdqa        0x55f9(%rip),%xmm9        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,13,141,99,0,0         // movdqa        0x638d(%rip),%xmm9        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,69,15,111,212                   // movdqa        %xmm12,%xmm10
   .byte  102,68,15,213,211                   // pmullw        %xmm3,%xmm10
   .byte  102,69,15,213,254                   // pmullw        %xmm14,%xmm15
@@ -55223,7 +55773,7 @@
   .byte  102,15,253,210                      // paddw         %xmm2,%xmm2
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,65,15,56,16,210                 // pblendvb      %xmm0,%xmm10,%xmm2
-  .byte  102,68,15,111,13,171,77,0,0         // movdqa        0x4dab(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,63,91,0,0          // movdqa        0x5b3f(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,249,195                      // psubw         %xmm3,%xmm0
   .byte  102,15,111,248                      // movdqa        %xmm0,%xmm7
@@ -55266,49 +55816,49 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            2e687 <_sk_load_8888_sse41_lowp+0x44>
+  .byte  119,41                              // ja            2efd3 <_sk_load_8888_sse41_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,239,0,0,0                 // lea           0xef(%rip),%r9        # 2e75c <_sk_load_8888_sse41_lowp+0x119>
+  .byte  76,141,13,239,0,0,0                 // lea           0xef(%rip),%r9        # 2f0a8 <_sk_load_8888_sse41_lowp+0x119>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,20,144                // movd          (%r8,%rdx,4),%xmm2
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,89                              // jmp           2e6e0 <_sk_load_8888_sse41_lowp+0x9d>
+  .byte  235,89                              // jmp           2f02c <_sk_load_8888_sse41_lowp+0x9d>
   .byte  243,65,15,111,20,144                // movdqu        (%r8,%rdx,4),%xmm2
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,74                              // jmp           2e6e0 <_sk_load_8888_sse41_lowp+0x9d>
+  .byte  235,74                              // jmp           2f02c <_sk_load_8888_sse41_lowp+0x9d>
   .byte  102,65,15,110,68,144,8              // movd          0x8(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  243,65,15,126,20,144                // movq          (%r8,%rdx,4),%xmm2
   .byte  102,65,15,58,14,208,240             // pblendw       $0xf0,%xmm8,%xmm2
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
-  .byte  235,37                              // jmp           2e6e0 <_sk_load_8888_sse41_lowp+0x9d>
+  .byte  235,37                              // jmp           2f02c <_sk_load_8888_sse41_lowp+0x9d>
   .byte  102,65,15,110,68,144,24             // movd          0x18(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,69,15,58,34,68,144,20,1         // pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   .byte  102,69,15,58,34,68,144,16,0         // pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   .byte  243,65,15,111,20,144                // movdqu        (%r8,%rdx,4),%xmm2
-  .byte  102,15,111,5,8,84,0,0               // movdqa        0x5408(%rip),%xmm0        # 33af0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
+  .byte  102,15,111,5,156,97,0,0             // movdqa        0x619c(%rip),%xmm0        # 351d0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,15,56,0,200                     // pshufb        %xmm0,%xmm1
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  102,15,56,0,216                     // pshufb        %xmm0,%xmm3
   .byte  102,15,108,203                      // punpcklqdq    %xmm3,%xmm1
-  .byte  102,68,15,111,13,104,76,0,0         // movdqa        0x4c68(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,252,89,0,0         // movdqa        0x59fc(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
-  .byte  102,68,15,111,21,225,83,0,0         // movdqa        0x53e1(%rip),%xmm10        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,21,117,97,0,0         // movdqa        0x6175(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
   .byte  102,65,15,56,0,218                  // pshufb        %xmm10,%xmm3
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
   .byte  102,65,15,108,219                   // punpcklqdq    %xmm11,%xmm3
-  .byte  102,68,15,111,21,206,83,0,0         // movdqa        0x53ce(%rip),%xmm10        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,21,98,97,0,0          // movdqa        0x6162(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,210                  // pshufb        %xmm10,%xmm2
   .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
   .byte  102,65,15,108,208                   // punpcklqdq    %xmm8,%xmm2
@@ -55324,7 +55874,7 @@
   .byte  58,255                              // cmp           %bh,%bh
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  126,255                             // jle           2e769 <_sk_load_8888_sse41_lowp+0x126>
+  .byte  126,255                             // jle           2f0b5 <_sk_load_8888_sse41_lowp+0x126>
   .byte  255                                 // (bad)
   .byte  255,117,255                         // pushq         -0x1(%rbp)
   .byte  255                                 // (bad)
@@ -55346,49 +55896,49 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            2e7bc <_sk_load_8888_dst_sse41_lowp+0x44>
+  .byte  119,41                              // ja            2f108 <_sk_load_8888_dst_sse41_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2e894 <_sk_load_8888_dst_sse41_lowp+0x11c>
+  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2f1e0 <_sk_load_8888_dst_sse41_lowp+0x11c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,52,144                // movd          (%r8,%rdx,4),%xmm6
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,89                              // jmp           2e815 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  .byte  235,89                              // jmp           2f161 <_sk_load_8888_dst_sse41_lowp+0x9d>
   .byte  243,65,15,111,52,144                // movdqu        (%r8,%rdx,4),%xmm6
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,74                              // jmp           2e815 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  .byte  235,74                              // jmp           2f161 <_sk_load_8888_dst_sse41_lowp+0x9d>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  243,65,15,126,52,144                // movq          (%r8,%rdx,4),%xmm6
   .byte  102,65,15,58,14,240,240             // pblendw       $0xf0,%xmm8,%xmm6
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
-  .byte  235,37                              // jmp           2e815 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  .byte  235,37                              // jmp           2f161 <_sk_load_8888_dst_sse41_lowp+0x9d>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,69,15,58,34,68,144,20,1         // pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   .byte  102,69,15,58,34,68,144,16,0         // pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   .byte  243,65,15,111,52,144                // movdqu        (%r8,%rdx,4),%xmm6
-  .byte  102,15,111,37,211,82,0,0            // movdqa        0x52d3(%rip),%xmm4        # 33af0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
+  .byte  102,15,111,37,103,96,0,0            // movdqa        0x6067(%rip),%xmm4        # 351d0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
   .byte  102,15,111,238                      // movdqa        %xmm6,%xmm5
   .byte  102,15,56,0,236                     // pshufb        %xmm4,%xmm5
   .byte  102,65,15,111,248                   // movdqa        %xmm8,%xmm7
   .byte  102,15,56,0,252                     // pshufb        %xmm4,%xmm7
   .byte  102,15,108,239                      // punpcklqdq    %xmm7,%xmm5
-  .byte  102,68,15,111,13,51,75,0,0          // movdqa        0x4b33(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,199,88,0,0         // movdqa        0x58c7(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,229                      // movdqa        %xmm5,%xmm4
   .byte  102,65,15,219,225                   // pand          %xmm9,%xmm4
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
-  .byte  102,68,15,111,21,172,82,0,0         // movdqa        0x52ac(%rip),%xmm10        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,21,64,96,0,0          // movdqa        0x6040(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,254                      // movdqa        %xmm6,%xmm7
   .byte  102,65,15,56,0,250                  // pshufb        %xmm10,%xmm7
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
   .byte  102,65,15,108,251                   // punpcklqdq    %xmm11,%xmm7
-  .byte  102,68,15,111,21,153,82,0,0         // movdqa        0x5299(%rip),%xmm10        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,21,45,96,0,0          // movdqa        0x602d(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,242                  // pshufb        %xmm10,%xmm6
   .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
   .byte  102,65,15,108,240                   // punpcklqdq    %xmm8,%xmm6
@@ -55404,7 +55954,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           2e8a1 <_sk_load_8888_dst_sse41_lowp+0x129>
+  .byte  123,255                             // jnp           2f1ed <_sk_load_8888_dst_sse41_lowp+0x129>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -55441,20 +55991,20 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,27                              // ja            2e936 <_sk_store_8888_sse41_lowp+0x86>
+  .byte  119,27                              // ja            2f282 <_sk_store_8888_sse41_lowp+0x86>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,87,0,0,0                  // lea           0x57(%rip),%r9        # 2e97c <_sk_store_8888_sse41_lowp+0xcc>
+  .byte  76,141,13,87,0,0,0                  // lea           0x57(%rip),%r9        # 2f2c8 <_sk_store_8888_sse41_lowp+0xcc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,126,4,144                 // movd          %xmm8,(%r8,%rdx,4)
-  .byte  235,63                              // jmp           2e975 <_sk_store_8888_sse41_lowp+0xc5>
+  .byte  235,63                              // jmp           2f2c1 <_sk_store_8888_sse41_lowp+0xc5>
   .byte  243,69,15,127,4,144                 // movdqu        %xmm8,(%r8,%rdx,4)
   .byte  243,69,15,127,76,144,16             // movdqu        %xmm9,0x10(%r8,%rdx,4)
-  .byte  235,48                              // jmp           2e975 <_sk_store_8888_sse41_lowp+0xc5>
+  .byte  235,48                              // jmp           2f2c1 <_sk_store_8888_sse41_lowp+0xc5>
   .byte  102,69,15,58,22,68,144,8,2          // pextrd        $0x2,%xmm8,0x8(%r8,%rdx,4)
   .byte  102,69,15,214,4,144                 // movq          %xmm8,(%r8,%rdx,4)
-  .byte  235,31                              // jmp           2e975 <_sk_store_8888_sse41_lowp+0xc5>
+  .byte  235,31                              // jmp           2f2c1 <_sk_store_8888_sse41_lowp+0xc5>
   .byte  102,69,15,58,22,76,144,24,2         // pextrd        $0x2,%xmm9,0x18(%r8,%rdx,4)
   .byte  102,69,15,58,22,76,144,20,1         // pextrd        $0x1,%xmm9,0x14(%r8,%rdx,4)
   .byte  102,69,15,126,76,144,16             // movd          %xmm9,0x10(%r8,%rdx,4)
@@ -55498,49 +56048,49 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            2e9dc <_sk_load_bgra_sse41_lowp+0x44>
+  .byte  119,41                              // ja            2f328 <_sk_load_bgra_sse41_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2eab4 <_sk_load_bgra_sse41_lowp+0x11c>
+  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2f400 <_sk_load_bgra_sse41_lowp+0x11c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,4,144                 // movd          (%r8,%rdx,4),%xmm0
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,89                              // jmp           2ea35 <_sk_load_bgra_sse41_lowp+0x9d>
+  .byte  235,89                              // jmp           2f381 <_sk_load_bgra_sse41_lowp+0x9d>
   .byte  243,65,15,111,4,144                 // movdqu        (%r8,%rdx,4),%xmm0
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,74                              // jmp           2ea35 <_sk_load_bgra_sse41_lowp+0x9d>
+  .byte  235,74                              // jmp           2f381 <_sk_load_bgra_sse41_lowp+0x9d>
   .byte  102,65,15,110,68,144,8              // movd          0x8(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  243,65,15,126,4,144                 // movq          (%r8,%rdx,4),%xmm0
   .byte  102,65,15,58,14,192,240             // pblendw       $0xf0,%xmm8,%xmm0
   .byte  102,68,15,111,193                   // movdqa        %xmm1,%xmm8
-  .byte  235,37                              // jmp           2ea35 <_sk_load_bgra_sse41_lowp+0x9d>
+  .byte  235,37                              // jmp           2f381 <_sk_load_bgra_sse41_lowp+0x9d>
   .byte  102,65,15,110,68,144,24             // movd          0x18(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,69,15,58,34,68,144,20,1         // pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   .byte  102,69,15,58,34,68,144,16,0         // pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   .byte  243,65,15,111,4,144                 // movdqu        (%r8,%rdx,4),%xmm0
-  .byte  102,15,111,21,179,80,0,0            // movdqa        0x50b3(%rip),%xmm2        # 33af0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
+  .byte  102,15,111,21,71,94,0,0             // movdqa        0x5e47(%rip),%xmm2        # 351d0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,56,0,202                     // pshufb        %xmm2,%xmm1
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  102,15,56,0,218                     // pshufb        %xmm2,%xmm3
   .byte  102,15,108,203                      // punpcklqdq    %xmm3,%xmm1
-  .byte  102,68,15,111,13,19,73,0,0          // movdqa        0x4913(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,167,86,0,0         // movdqa        0x56a7(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,65,15,219,209                   // pand          %xmm9,%xmm2
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
-  .byte  102,68,15,111,21,140,80,0,0         // movdqa        0x508c(%rip),%xmm10        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,21,32,94,0,0          // movdqa        0x5e20(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
   .byte  102,65,15,56,0,218                  // pshufb        %xmm10,%xmm3
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
   .byte  102,65,15,108,219                   // punpcklqdq    %xmm11,%xmm3
-  .byte  102,68,15,111,21,121,80,0,0         // movdqa        0x5079(%rip),%xmm10        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,21,13,94,0,0          // movdqa        0x5e0d(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,194                  // pshufb        %xmm10,%xmm0
   .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
   .byte  102,65,15,108,192                   // punpcklqdq    %xmm8,%xmm0
@@ -55556,7 +56106,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           2eac1 <_sk_load_bgra_sse41_lowp+0x129>
+  .byte  123,255                             // jnp           2f40d <_sk_load_bgra_sse41_lowp+0x129>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -55578,49 +56128,49 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            2eb14 <_sk_load_bgra_dst_sse41_lowp+0x44>
+  .byte  119,41                              // ja            2f460 <_sk_load_bgra_dst_sse41_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2ebec <_sk_load_bgra_dst_sse41_lowp+0x11c>
+  .byte  76,141,13,242,0,0,0                 // lea           0xf2(%rip),%r9        # 2f538 <_sk_load_bgra_dst_sse41_lowp+0x11c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,36,144                // movd          (%r8,%rdx,4),%xmm4
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,89                              // jmp           2eb6d <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  .byte  235,89                              // jmp           2f4b9 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   .byte  243,65,15,111,36,144                // movdqu        (%r8,%rdx,4),%xmm4
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,74                              // jmp           2eb6d <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  .byte  235,74                              // jmp           2f4b9 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  243,65,15,126,36,144                // movq          (%r8,%rdx,4),%xmm4
   .byte  102,65,15,58,14,224,240             // pblendw       $0xf0,%xmm8,%xmm4
   .byte  102,68,15,111,197                   // movdqa        %xmm5,%xmm8
-  .byte  235,37                              // jmp           2eb6d <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  .byte  235,37                              // jmp           2f4b9 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,69,15,58,34,68,144,20,1         // pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   .byte  102,69,15,58,34,68,144,16,0         // pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   .byte  243,65,15,111,36,144                // movdqu        (%r8,%rdx,4),%xmm4
-  .byte  102,15,111,53,123,79,0,0            // movdqa        0x4f7b(%rip),%xmm6        # 33af0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
+  .byte  102,15,111,53,15,93,0,0             // movdqa        0x5d0f(%rip),%xmm6        # 351d0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,56,0,238                     // pshufb        %xmm6,%xmm5
   .byte  102,65,15,111,248                   // movdqa        %xmm8,%xmm7
   .byte  102,15,56,0,254                     // pshufb        %xmm6,%xmm7
   .byte  102,15,108,239                      // punpcklqdq    %xmm7,%xmm5
-  .byte  102,68,15,111,13,219,71,0,0         // movdqa        0x47db(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,111,85,0,0         // movdqa        0x556f(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,245                      // movdqa        %xmm5,%xmm6
   .byte  102,65,15,219,241                   // pand          %xmm9,%xmm6
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
-  .byte  102,68,15,111,21,84,79,0,0          // movdqa        0x4f54(%rip),%xmm10        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,21,232,92,0,0         // movdqa        0x5ce8(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,252                      // movdqa        %xmm4,%xmm7
   .byte  102,65,15,56,0,250                  // pshufb        %xmm10,%xmm7
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
   .byte  102,65,15,108,251                   // punpcklqdq    %xmm11,%xmm7
-  .byte  102,68,15,111,21,65,79,0,0          // movdqa        0x4f41(%rip),%xmm10        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,21,213,92,0,0         // movdqa        0x5cd5(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,226                  // pshufb        %xmm10,%xmm4
   .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
   .byte  102,65,15,108,224                   // punpcklqdq    %xmm8,%xmm4
@@ -55636,7 +56186,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           2ebf9 <_sk_load_bgra_dst_sse41_lowp+0x129>
+  .byte  123,255                             // jnp           2f545 <_sk_load_bgra_dst_sse41_lowp+0x129>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -55673,20 +56223,20 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,27                              // ja            2ec8e <_sk_store_bgra_sse41_lowp+0x86>
+  .byte  119,27                              // ja            2f5da <_sk_store_bgra_sse41_lowp+0x86>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,87,0,0,0                  // lea           0x57(%rip),%r9        # 2ecd4 <_sk_store_bgra_sse41_lowp+0xcc>
+  .byte  76,141,13,87,0,0,0                  // lea           0x57(%rip),%r9        # 2f620 <_sk_store_bgra_sse41_lowp+0xcc>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,126,4,144                 // movd          %xmm8,(%r8,%rdx,4)
-  .byte  235,63                              // jmp           2eccd <_sk_store_bgra_sse41_lowp+0xc5>
+  .byte  235,63                              // jmp           2f619 <_sk_store_bgra_sse41_lowp+0xc5>
   .byte  243,69,15,127,4,144                 // movdqu        %xmm8,(%r8,%rdx,4)
   .byte  243,69,15,127,76,144,16             // movdqu        %xmm9,0x10(%r8,%rdx,4)
-  .byte  235,48                              // jmp           2eccd <_sk_store_bgra_sse41_lowp+0xc5>
+  .byte  235,48                              // jmp           2f619 <_sk_store_bgra_sse41_lowp+0xc5>
   .byte  102,69,15,58,22,68,144,8,2          // pextrd        $0x2,%xmm8,0x8(%r8,%rdx,4)
   .byte  102,69,15,214,4,144                 // movq          %xmm8,(%r8,%rdx,4)
-  .byte  235,31                              // jmp           2eccd <_sk_store_bgra_sse41_lowp+0xc5>
+  .byte  235,31                              // jmp           2f619 <_sk_store_bgra_sse41_lowp+0xc5>
   .byte  102,69,15,58,22,76,144,24,2         // pextrd        $0x2,%xmm9,0x18(%r8,%rdx,4)
   .byte  102,69,15,58,22,76,144,20,1         // pextrd        $0x1,%xmm9,0x14(%r8,%rdx,4)
   .byte  102,69,15,126,76,144,16             // movd          %xmm9,0x10(%r8,%rdx,4)
@@ -55789,17 +56339,17 @@
   .byte  102,65,15,196,202,5                 // pinsrw        $0x5,%r10d,%xmm1
   .byte  102,65,15,196,201,6                 // pinsrw        $0x6,%r9d,%xmm1
   .byte  102,15,196,200,7                    // pinsrw        $0x7,%eax,%xmm1
-  .byte  102,68,15,111,13,68,69,0,0          // movdqa        0x4544(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,216,82,0,0         // movdqa        0x52d8(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
-  .byte  102,68,15,111,21,189,76,0,0         // movdqa        0x4cbd(%rip),%xmm10        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,21,81,90,0,0          // movdqa        0x5a51(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
   .byte  102,65,15,56,0,218                  // pshufb        %xmm10,%xmm3
   .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
   .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
   .byte  102,65,15,108,219                   // punpcklqdq    %xmm11,%xmm3
-  .byte  102,68,15,111,21,170,76,0,0         // movdqa        0x4caa(%rip),%xmm10        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,21,62,90,0,0          // movdqa        0x5a3e(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,210                  // pshufb        %xmm10,%xmm2
   .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
   .byte  102,65,15,108,208                   // punpcklqdq    %xmm8,%xmm2
@@ -55812,6 +56362,101 @@
   .byte  93                                  // pop           %rbp
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_bgra_sse41_lowp
+.globl _sk_gather_bgra_sse41_lowp
+FUNCTION(_sk_gather_bgra_sse41_lowp)
+_sk_gather_bgra_sse41_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  243,68,15,91,195                    // cvttps2dq     %xmm3,%xmm8
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,15,56,64,211                    // pmulld        %xmm3,%xmm2
+  .byte  102,68,15,56,64,195                 // pmulld        %xmm3,%xmm8
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,200                   // paddd         %xmm8,%xmm1
+  .byte  102,72,15,58,22,200,1               // pextrq        $0x1,%xmm1,%rax
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,73,15,126,202                   // movq          %xmm1,%r10
+  .byte  69,137,211                          // mov           %r10d,%r11d
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  102,72,15,58,22,195,1               // pextrq        $0x1,%xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  67,139,44,184                       // mov           (%r8,%r15,4),%ebp
+  .byte  71,139,52,176                       // mov           (%r8,%r14,4),%r14d
+  .byte  65,139,28,152                       // mov           (%r8,%rbx,4),%ebx
+  .byte  71,139,28,152                       // mov           (%r8,%r11,4),%r11d
+  .byte  71,139,20,144                       // mov           (%r8,%r10,4),%r10d
+  .byte  71,139,12,136                       // mov           (%r8,%r9,4),%r9d
+  .byte  65,139,4,128                        // mov           (%r8,%rax,4),%eax
+  .byte  102,69,15,110,195                   // movd          %r11d,%xmm8
+  .byte  102,69,15,58,34,194,1               // pinsrd        $0x1,%r10d,%xmm8
+  .byte  102,69,15,58,34,193,2               // pinsrd        $0x2,%r9d,%xmm8
+  .byte  102,68,15,58,34,192,3               // pinsrd        $0x3,%eax,%xmm8
+  .byte  102,67,15,110,12,160                // movd          (%r8,%r12,4),%xmm1
+  .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
+  .byte  102,15,58,34,197,1                  // pinsrd        $0x1,%ebp,%xmm0
+  .byte  102,65,15,58,34,198,2               // pinsrd        $0x2,%r14d,%xmm0
+  .byte  102,15,58,34,195,3                  // pinsrd        $0x3,%ebx,%xmm0
+  .byte  102,15,196,205,1                    // pinsrw        $0x1,%ebp,%xmm1
+  .byte  102,65,15,196,206,2                 // pinsrw        $0x2,%r14d,%xmm1
+  .byte  102,15,196,203,3                    // pinsrw        $0x3,%ebx,%xmm1
+  .byte  102,65,15,196,203,4                 // pinsrw        $0x4,%r11d,%xmm1
+  .byte  102,65,15,196,202,5                 // pinsrw        $0x5,%r10d,%xmm1
+  .byte  102,65,15,196,201,6                 // pinsrw        $0x6,%r9d,%xmm1
+  .byte  102,15,196,200,7                    // pinsrw        $0x7,%eax,%xmm1
+  .byte  102,68,15,111,13,64,81,0,0          // movdqa        0x5140(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
+  .byte  102,65,15,219,209                   // pand          %xmm9,%xmm2
+  .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
+  .byte  102,68,15,111,21,185,88,0,0         // movdqa        0x58b9(%rip),%xmm10        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
+  .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
+  .byte  102,65,15,56,0,218                  // pshufb        %xmm10,%xmm3
+  .byte  102,69,15,111,216                   // movdqa        %xmm8,%xmm11
+  .byte  102,69,15,56,0,218                  // pshufb        %xmm10,%xmm11
+  .byte  102,65,15,108,219                   // punpcklqdq    %xmm11,%xmm3
+  .byte  102,68,15,111,21,166,88,0,0         // movdqa        0x58a6(%rip),%xmm10        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,65,15,56,0,194                  // pshufb        %xmm10,%xmm0
+  .byte  102,69,15,56,0,194                  // pshufb        %xmm10,%xmm8
+  .byte  102,65,15,108,192                   // punpcklqdq    %xmm8,%xmm0
+  .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_565_sse41_lowp
 .globl _sk_load_565_sse41_lowp
 FUNCTION(_sk_load_565_sse41_lowp)
@@ -55825,23 +56470,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2eec4 <_sk_load_565_sse41_lowp+0x3c>
+  .byte  119,34                              // ja            2f9a8 <_sk_load_565_sse41_lowp+0x3c>
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,196,0,0,0                 // lea           0xc4(%rip),%r9        # 2ef74 <_sk_load_565_sse41_lowp+0xec>
+  .byte  76,141,13,196,0,0,0                 // lea           0xc4(%rip),%r9        # 2fa58 <_sk_load_565_sse41_lowp+0xec>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,192                      // movd          %eax,%xmm0
-  .byte  235,74                              // jmp           2ef0e <_sk_load_565_sse41_lowp+0x86>
+  .byte  235,74                              // jmp           2f9f2 <_sk_load_565_sse41_lowp+0x86>
   .byte  243,65,15,111,4,80                  // movdqu        (%r8,%rdx,2),%xmm0
-  .byte  235,66                              // jmp           2ef0e <_sk_load_565_sse41_lowp+0x86>
+  .byte  235,66                              // jmp           2f9f2 <_sk_load_565_sse41_lowp+0x86>
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  102,65,15,196,76,80,4,2             // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm1
   .byte  102,65,15,110,4,80                  // movd          (%r8,%rdx,2),%xmm0
   .byte  102,15,58,14,193,252                // pblendw       $0xfc,%xmm1,%xmm0
-  .byte  235,40                              // jmp           2ef0e <_sk_load_565_sse41_lowp+0x86>
+  .byte  235,40                              // jmp           2f9f2 <_sk_load_565_sse41_lowp+0x86>
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  102,65,15,196,76,80,12,6            // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm1
   .byte  102,65,15,196,76,80,10,5            // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm1
@@ -55850,11 +56495,11 @@
   .byte  102,15,58,14,193,240                // pblendw       $0xf0,%xmm1,%xmm0
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
   .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
-  .byte  102,15,219,29,1,76,0,0              // pand          0x4c01(%rip),%xmm3        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,29,253,87,0,0            // pand          0x57fd(%rip),%xmm3        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
-  .byte  102,15,219,13,0,76,0,0              // pand          0x4c00(%rip),%xmm1        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,15,111,21,8,76,0,0              // movdqa        0x4c08(%rip),%xmm2        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,15,219,13,252,87,0,0            // pand          0x57fc(%rip),%xmm1        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,21,4,88,0,0              // movdqa        0x5804(%rip),%xmm2        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,15,219,208                      // pand          %xmm0,%xmm2
   .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
   .byte  102,15,235,195                      // por           %xmm3,%xmm0
@@ -55867,7 +56512,7 @@
   .byte  102,15,113,210,2                    // psrlw         $0x2,%xmm2
   .byte  102,15,235,211                      // por           %xmm3,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,254,67,0,0                 // movaps        0x43fe(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,250,79,0,0                 // movaps        0x4ffa(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  255,224                             // jmpq          *%rax
   .byte  69,255                              // rex.RB        (bad)
   .byte  255                                 // (bad)
@@ -55878,7 +56523,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  126,255                             // jle           2ef89 <_sk_load_565_sse41_lowp+0x101>
+  .byte  126,255                             // jle           2fa6d <_sk_load_565_sse41_lowp+0x101>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -55897,23 +56542,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2efcc <_sk_load_565_dst_sse41_lowp+0x3c>
+  .byte  119,34                              // ja            2fab0 <_sk_load_565_dst_sse41_lowp+0x3c>
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,196,0,0,0                 // lea           0xc4(%rip),%r9        # 2f07c <_sk_load_565_dst_sse41_lowp+0xec>
+  .byte  76,141,13,196,0,0,0                 // lea           0xc4(%rip),%r9        # 2fb60 <_sk_load_565_dst_sse41_lowp+0xec>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,224                      // movd          %eax,%xmm4
-  .byte  235,74                              // jmp           2f016 <_sk_load_565_dst_sse41_lowp+0x86>
+  .byte  235,74                              // jmp           2fafa <_sk_load_565_dst_sse41_lowp+0x86>
   .byte  243,65,15,111,36,80                 // movdqu        (%r8,%rdx,2),%xmm4
-  .byte  235,66                              // jmp           2f016 <_sk_load_565_dst_sse41_lowp+0x86>
+  .byte  235,66                              // jmp           2fafa <_sk_load_565_dst_sse41_lowp+0x86>
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  102,65,15,196,108,80,4,2            // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm5
   .byte  102,65,15,110,36,80                 // movd          (%r8,%rdx,2),%xmm4
   .byte  102,15,58,14,229,252                // pblendw       $0xfc,%xmm5,%xmm4
-  .byte  235,40                              // jmp           2f016 <_sk_load_565_dst_sse41_lowp+0x86>
+  .byte  235,40                              // jmp           2fafa <_sk_load_565_dst_sse41_lowp+0x86>
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  102,65,15,196,108,80,12,6           // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm5
   .byte  102,65,15,196,108,80,10,5           // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm5
@@ -55922,11 +56567,11 @@
   .byte  102,15,58,14,229,240                // pblendw       $0xf0,%xmm5,%xmm4
   .byte  102,15,111,252                      // movdqa        %xmm4,%xmm7
   .byte  102,15,113,215,8                    // psrlw         $0x8,%xmm7
-  .byte  102,15,219,61,249,74,0,0            // pand          0x4af9(%rip),%xmm7        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,61,245,86,0,0            // pand          0x56f5(%rip),%xmm7        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,113,213,5                    // psrlw         $0x5,%xmm5
-  .byte  102,15,219,45,248,74,0,0            // pand          0x4af8(%rip),%xmm5        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,15,111,53,0,75,0,0              // movdqa        0x4b00(%rip),%xmm6        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,15,219,45,244,86,0,0            // pand          0x56f4(%rip),%xmm5        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,53,252,86,0,0            // movdqa        0x56fc(%rip),%xmm6        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,15,219,244                      // pand          %xmm4,%xmm6
   .byte  102,15,113,212,13                   // psrlw         $0xd,%xmm4
   .byte  102,15,235,231                      // por           %xmm7,%xmm4
@@ -55939,7 +56584,7 @@
   .byte  102,15,113,214,2                    // psrlw         $0x2,%xmm6
   .byte  102,15,235,247                      // por           %xmm7,%xmm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,246,66,0,0                 // movaps        0x42f6(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,61,242,78,0,0                 // movaps        0x4ef2(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  255,224                             // jmpq          *%rax
   .byte  69,255                              // rex.RB        (bad)
   .byte  255                                 // (bad)
@@ -55950,7 +56595,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  126,255                             // jle           2f091 <_sk_load_565_dst_sse41_lowp+0x101>
+  .byte  126,255                             // jle           2fb75 <_sk_load_565_dst_sse41_lowp+0x101>
   .byte  255                                 // (bad)
   .byte  255,114,255                         // pushq         -0x1(%rdx)
   .byte  255                                 // (bad)
@@ -55967,10 +56612,10 @@
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
   .byte  102,65,15,113,240,8                 // psllw         $0x8,%xmm8
-  .byte  102,68,15,219,5,148,74,0,0          // pand          0x4a94(%rip),%xmm8        # 33b50 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
+  .byte  102,68,15,219,5,144,86,0,0          // pand          0x5690(%rip),%xmm8        # 35230 <_sk_srcover_rgba_8888_sse2_lowp+0x1200>
   .byte  102,68,15,111,201                   // movdqa        %xmm1,%xmm9
   .byte  102,65,15,113,241,3                 // psllw         $0x3,%xmm9
-  .byte  102,68,15,219,13,144,74,0,0         // pand          0x4a90(%rip),%xmm9        # 33b60 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,219,13,140,86,0,0         // pand          0x568c(%rip),%xmm9        # 35240 <_sk_srcover_rgba_8888_sse2_lowp+0x1210>
   .byte  102,69,15,235,200                   // por           %xmm8,%xmm9
   .byte  102,68,15,111,194                   // movdqa        %xmm2,%xmm8
   .byte  102,65,15,113,208,3                 // psrlw         $0x3,%xmm8
@@ -55979,19 +56624,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,29                              // ja            2f10c <_sk_store_565_sse41_lowp+0x74>
+  .byte  119,29                              // ja            2fbf0 <_sk_store_565_sse41_lowp+0x74>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,83,0,0,0                  // lea           0x53(%rip),%r9        # 2f14c <_sk_store_565_sse41_lowp+0xb4>
+  .byte  76,141,13,83,0,0,0                  // lea           0x53(%rip),%r9        # 2fc30 <_sk_store_565_sse41_lowp+0xb4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,58,21,4,80,0              // pextrw        $0x0,%xmm8,(%r8,%rdx,2)
-  .byte  235,58                              // jmp           2f146 <_sk_store_565_sse41_lowp+0xae>
+  .byte  235,58                              // jmp           2fc2a <_sk_store_565_sse41_lowp+0xae>
   .byte  243,69,15,127,4,80                  // movdqu        %xmm8,(%r8,%rdx,2)
-  .byte  235,50                              // jmp           2f146 <_sk_store_565_sse41_lowp+0xae>
+  .byte  235,50                              // jmp           2fc2a <_sk_store_565_sse41_lowp+0xae>
   .byte  102,69,15,58,21,68,80,4,2           // pextrw        $0x2,%xmm8,0x4(%r8,%rdx,2)
   .byte  102,69,15,126,4,80                  // movd          %xmm8,(%r8,%rdx,2)
-  .byte  235,33                              // jmp           2f146 <_sk_store_565_sse41_lowp+0xae>
+  .byte  235,33                              // jmp           2fc2a <_sk_store_565_sse41_lowp+0xae>
   .byte  102,69,15,58,21,68,80,12,6          // pextrw        $0x6,%xmm8,0xc(%r8,%rdx,2)
   .byte  102,69,15,58,21,68,80,10,5          // pextrw        $0x5,%xmm8,0xa(%r8,%rdx,2)
   .byte  102,69,15,58,21,68,80,8,4           // pextrw        $0x4,%xmm8,0x8(%r8,%rdx,2)
@@ -56011,7 +56656,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  235,255                             // jmp           2f15d <_sk_store_565_sse41_lowp+0xc5>
+  .byte  235,255                             // jmp           2fc41 <_sk_store_565_sse41_lowp+0xc5>
   .byte  255                                 // (bad)
   .byte  255,226                             // jmpq          *%rdx
   .byte  255                                 // (bad)
@@ -56021,6 +56666,98 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_565_sse41_lowp
+.globl _sk_gather_565_sse41_lowp
+FUNCTION(_sk_gather_565_sse41_lowp)
+_sk_gather_565_sse41_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  243,68,15,91,195                    // cvttps2dq     %xmm3,%xmm8
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,15,56,64,211                    // pmulld        %xmm3,%xmm2
+  .byte  102,68,15,56,64,195                 // pmulld        %xmm3,%xmm8
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,200                   // paddd         %xmm8,%xmm1
+  .byte  102,72,15,58,22,200,1               // pextrq        $0x1,%xmm1,%rax
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,73,15,126,202                   // movq          %xmm1,%r10
+  .byte  69,137,211                          // mov           %r10d,%r11d
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  102,72,15,58,22,195,1               // pextrq        $0x1,%xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  71,15,183,60,120                    // movzwl        (%r8,%r15,2),%r15d
+  .byte  67,15,183,44,96                     // movzwl        (%r8,%r12,2),%ebp
+  .byte  102,15,110,197                      // movd          %ebp,%xmm0
+  .byte  102,65,15,196,199,1                 // pinsrw        $0x1,%r15d,%xmm0
+  .byte  67,15,183,44,112                    // movzwl        (%r8,%r14,2),%ebp
+  .byte  102,15,196,197,2                    // pinsrw        $0x2,%ebp,%xmm0
+  .byte  65,15,183,28,88                     // movzwl        (%r8,%rbx,2),%ebx
+  .byte  102,15,196,195,3                    // pinsrw        $0x3,%ebx,%xmm0
+  .byte  67,15,183,44,88                     // movzwl        (%r8,%r11,2),%ebp
+  .byte  102,15,196,197,4                    // pinsrw        $0x4,%ebp,%xmm0
+  .byte  67,15,183,44,80                     // movzwl        (%r8,%r10,2),%ebp
+  .byte  102,15,196,197,5                    // pinsrw        $0x5,%ebp,%xmm0
+  .byte  67,15,183,44,72                     // movzwl        (%r8,%r9,2),%ebp
+  .byte  102,15,196,197,6                    // pinsrw        $0x6,%ebp,%xmm0
+  .byte  65,15,183,4,64                      // movzwl        (%r8,%rax,2),%eax
+  .byte  102,15,196,192,7                    // pinsrw        $0x7,%eax,%xmm0
+  .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
+  .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
+  .byte  102,15,219,29,154,84,0,0            // pand          0x549a(%rip),%xmm3        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
+  .byte  102,15,219,13,153,84,0,0            // pand          0x5499(%rip),%xmm1        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,21,161,84,0,0            // movdqa        0x54a1(%rip),%xmm2        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
+  .byte  102,15,219,208                      // pand          %xmm0,%xmm2
+  .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
+  .byte  102,15,235,195                      // por           %xmm3,%xmm0
+  .byte  102,15,111,217                      // movdqa        %xmm1,%xmm3
+  .byte  102,15,113,243,2                    // psllw         $0x2,%xmm3
+  .byte  102,15,113,209,4                    // psrlw         $0x4,%xmm1
+  .byte  102,15,235,203                      // por           %xmm3,%xmm1
+  .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
+  .byte  102,15,113,243,3                    // psllw         $0x3,%xmm3
+  .byte  102,15,113,210,2                    // psrlw         $0x2,%xmm2
+  .byte  102,15,235,211                      // por           %xmm3,%xmm2
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  15,40,29,151,76,0,0                 // movaps        0x4c97(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_a8_sse41_lowp
 .globl _sk_load_a8_sse41_lowp
 FUNCTION(_sk_load_a8_sse41_lowp)
@@ -56033,18 +56770,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2f1a1 <_sk_load_a8_sse41_lowp+0x39>
+  .byte  119,34                              // ja            2fdfc <_sk_load_a8_sse41_lowp+0x39>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,143,0,0,0                 // lea           0x8f(%rip),%r9        # 2f21c <_sk_load_a8_sse41_lowp+0xb4>
+  .byte  76,141,13,144,0,0,0                 // lea           0x90(%rip),%r9        # 2fe78 <_sk_load_a8_sse41_lowp+0xb5>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,216                      // movd          %eax,%xmm3
-  .byte  235,100                             // jmp           2f205 <_sk_load_a8_sse41_lowp+0x9d>
+  .byte  235,100                             // jmp           2fe60 <_sk_load_a8_sse41_lowp+0x9d>
   .byte  102,65,15,56,48,28,16               // pmovzxbw      (%r8,%rdx,1),%xmm3
-  .byte  235,91                              // jmp           2f205 <_sk_load_a8_sse41_lowp+0x9d>
+  .byte  235,91                              // jmp           2fe60 <_sk_load_a8_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,192,2                    // pinsrw        $0x2,%eax,%xmm0
@@ -56052,7 +56789,7 @@
   .byte  102,15,110,200                      // movd          %eax,%xmm1
   .byte  102,15,56,48,217                    // pmovzxbw      %xmm1,%xmm3
   .byte  102,15,58,14,216,252                // pblendw       $0xfc,%xmm0,%xmm3
-  .byte  235,54                              // jmp           2f205 <_sk_load_a8_sse41_lowp+0x9d>
+  .byte  235,54                              // jmp           2fe60 <_sk_load_a8_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,192,6                    // pinsrw        $0x6,%eax,%xmm0
@@ -56063,28 +56800,28 @@
   .byte  102,65,15,110,12,16                 // movd          (%r8,%rdx,1),%xmm1
   .byte  102,15,56,48,217                    // pmovzxbw      %xmm1,%xmm3
   .byte  102,15,58,14,216,240                // pblendw       $0xf0,%xmm0,%xmm3
-  .byte  102,15,219,29,99,65,0,0             // pand          0x4163(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,29,232,75,0,0            // pand          0x4be8(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
   .byte  255,224                             // jmpq          *%rax
-  .byte  122,255                             // jp            2f21d <_sk_load_a8_sse41_lowp+0xb5>
+  .byte  144                                 // nop
+  .byte  121,255                             // jns           2fe79 <_sk_load_a8_sse41_lowp+0xb6>
   .byte  255                                 // (bad)
-  .byte  255,157,255,255,255,142             // lcall         *-0x71000001(%rbp)
+  .byte  255,156,255,255,255,141,255         // lcall         *-0x720001(%rdi,%rdi,8)
+  .byte  255                                 // (bad)
+  .byte  255,215                             // callq         *%rdi
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255                                 // (bad)
-  .byte  216,255                             // fdivr         %st(7),%st
-  .byte  255                                 // (bad)
-  .byte  255,205                             // dec           %ebp
+  .byte  255,204                             // dec           %esp
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,194                             // inc           %edx
+  .byte  255,193                             // inc           %ecx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
-  .byte  179,255                             // mov           $0xff,%bl
+  .byte  178,255                             // mov           $0xff,%dl
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
@@ -56100,18 +56837,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2f271 <_sk_load_a8_dst_sse41_lowp+0x39>
+  .byte  119,34                              // ja            2fecd <_sk_load_a8_dst_sse41_lowp+0x39>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,143,0,0,0                 // lea           0x8f(%rip),%r9        # 2f2ec <_sk_load_a8_dst_sse41_lowp+0xb4>
+  .byte  76,141,13,143,0,0,0                 // lea           0x8f(%rip),%r9        # 2ff48 <_sk_load_a8_dst_sse41_lowp+0xb4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,248                      // movd          %eax,%xmm7
-  .byte  235,100                             // jmp           2f2d5 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  .byte  235,100                             // jmp           2ff31 <_sk_load_a8_dst_sse41_lowp+0x9d>
   .byte  102,65,15,56,48,60,16               // pmovzxbw      (%r8,%rdx,1),%xmm7
-  .byte  235,91                              // jmp           2f2d5 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  .byte  235,91                              // jmp           2ff31 <_sk_load_a8_dst_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,224,2                    // pinsrw        $0x2,%eax,%xmm4
@@ -56119,7 +56856,7 @@
   .byte  102,15,110,232                      // movd          %eax,%xmm5
   .byte  102,15,56,48,253                    // pmovzxbw      %xmm5,%xmm7
   .byte  102,15,58,14,252,252                // pblendw       $0xfc,%xmm4,%xmm7
-  .byte  235,54                              // jmp           2f2d5 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  .byte  235,54                              // jmp           2ff31 <_sk_load_a8_dst_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,224,6                    // pinsrw        $0x6,%eax,%xmm4
@@ -56130,13 +56867,13 @@
   .byte  102,65,15,110,44,16                 // movd          (%r8,%rdx,1),%xmm5
   .byte  102,15,56,48,253                    // pmovzxbw      %xmm5,%xmm7
   .byte  102,15,58,14,252,240                // pblendw       $0xf0,%xmm4,%xmm7
-  .byte  102,15,219,61,147,64,0,0            // pand          0x4093(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,61,23,75,0,0             // pand          0x4b17(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  15,87,246                           // xorps         %xmm6,%xmm6
   .byte  255,224                             // jmpq          *%rax
-  .byte  122,255                             // jp            2f2ed <_sk_load_a8_dst_sse41_lowp+0xb5>
+  .byte  122,255                             // jp            2ff49 <_sk_load_a8_dst_sse41_lowp+0xb5>
   .byte  255                                 // (bad)
   .byte  255,157,255,255,255,142             // lcall         *-0x71000001(%rbp)
   .byte  255                                 // (bad)
@@ -56167,28 +56904,28 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,29                              // ja            2f33c <_sk_store_a8_sse41_lowp+0x34>
+  .byte  119,29                              // ja            2ff98 <_sk_store_a8_sse41_lowp+0x34>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,131,0,0,0                 // lea           0x83(%rip),%r9        # 2f3ac <_sk_store_a8_sse41_lowp+0xa4>
+  .byte  76,141,13,131,0,0,0                 // lea           0x83(%rip),%r9        # 30008 <_sk_store_a8_sse41_lowp+0xa4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,58,20,28,16,0             // pextrb        $0x0,%xmm3,(%r8,%rdx,1)
-  .byte  235,105                             // jmp           2f3a5 <_sk_store_a8_sse41_lowp+0x9d>
+  .byte  235,105                             // jmp           30001 <_sk_store_a8_sse41_lowp+0x9d>
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
-  .byte  102,68,15,56,0,5,117,71,0,0         // pshufb        0x4775(%rip),%xmm8        # 33ac0 <_sk_srcover_rgba_8888_sse2_lowp+0x1140>
+  .byte  102,68,15,56,0,5,249,81,0,0         // pshufb        0x51f9(%rip),%xmm8        # 351a0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
   .byte  102,69,15,214,4,16                  // movq          %xmm8,(%r8,%rdx,1)
-  .byte  235,82                              // jmp           2f3a5 <_sk_store_a8_sse41_lowp+0x9d>
+  .byte  235,82                              // jmp           30001 <_sk_store_a8_sse41_lowp+0x9d>
   .byte  102,65,15,58,20,92,16,2,4           // pextrb        $0x4,%xmm3,0x2(%r8,%rdx,1)
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
-  .byte  102,68,15,56,0,5,21,64,0,0          // pshufb        0x4015(%rip),%xmm8        # 33380 <_sk_srcover_rgba_8888_sse2_lowp+0xa00>
+  .byte  102,68,15,56,0,5,153,74,0,0         // pshufb        0x4a99(%rip),%xmm8        # 34a60 <_sk_srcover_rgba_8888_sse2_lowp+0xa30>
   .byte  102,69,15,58,21,4,16,0              // pextrw        $0x0,%xmm8,(%r8,%rdx,1)
-  .byte  235,48                              // jmp           2f3a5 <_sk_store_a8_sse41_lowp+0x9d>
+  .byte  235,48                              // jmp           30001 <_sk_store_a8_sse41_lowp+0x9d>
   .byte  102,65,15,58,20,92,16,6,12          // pextrb        $0xc,%xmm3,0x6(%r8,%rdx,1)
   .byte  102,65,15,58,20,92,16,5,10          // pextrb        $0xa,%xmm3,0x5(%r8,%rdx,1)
   .byte  102,65,15,58,20,92,16,4,8           // pextrb        $0x8,%xmm3,0x4(%r8,%rdx,1)
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
-  .byte  102,68,15,56,0,5,241,63,0,0         // pshufb        0x3ff1(%rip),%xmm8        # 33390 <_sk_srcover_rgba_8888_sse2_lowp+0xa10>
+  .byte  102,68,15,56,0,5,117,74,0,0         // pshufb        0x4a75(%rip),%xmm8        # 34a70 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
   .byte  102,69,15,126,4,16                  // movd          %xmm8,(%r8,%rdx,1)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
@@ -56212,6 +56949,83 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_a8_sse41_lowp
+.globl _sk_gather_a8_sse41_lowp
+FUNCTION(_sk_gather_a8_sse41_lowp)
+_sk_gather_a8_sse41_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  243,68,15,91,195                    // cvttps2dq     %xmm3,%xmm8
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,15,56,64,211                    // pmulld        %xmm3,%xmm2
+  .byte  102,68,15,56,64,195                 // pmulld        %xmm3,%xmm8
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,200                   // paddd         %xmm8,%xmm1
+  .byte  102,72,15,58,22,200,1               // pextrq        $0x1,%xmm1,%rax
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,73,15,126,202                   // movq          %xmm1,%r10
+  .byte  69,137,211                          // mov           %r10d,%r11d
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  102,72,15,58,22,195,1               // pextrq        $0x1,%xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  71,15,182,60,56                     // movzbl        (%r8,%r15,1),%r15d
+  .byte  67,15,182,44,32                     // movzbl        (%r8,%r12,1),%ebp
+  .byte  102,15,110,197                      // movd          %ebp,%xmm0
+  .byte  102,65,15,58,32,199,1               // pinsrb        $0x1,%r15d,%xmm0
+  .byte  67,15,182,44,48                     // movzbl        (%r8,%r14,1),%ebp
+  .byte  102,15,58,32,197,2                  // pinsrb        $0x2,%ebp,%xmm0
+  .byte  65,15,182,28,24                     // movzbl        (%r8,%rbx,1),%ebx
+  .byte  102,15,58,32,195,3                  // pinsrb        $0x3,%ebx,%xmm0
+  .byte  67,15,182,44,24                     // movzbl        (%r8,%r11,1),%ebp
+  .byte  102,15,58,32,197,4                  // pinsrb        $0x4,%ebp,%xmm0
+  .byte  67,15,182,44,16                     // movzbl        (%r8,%r10,1),%ebp
+  .byte  102,15,58,32,197,5                  // pinsrb        $0x5,%ebp,%xmm0
+  .byte  67,15,182,44,8                      // movzbl        (%r8,%r9,1),%ebp
+  .byte  102,15,58,32,197,6                  // pinsrb        $0x6,%ebp,%xmm0
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  102,15,58,32,192,7                  // pinsrb        $0x7,%eax,%xmm0
+  .byte  102,15,56,48,216                    // pmovzxbw      %xmm0,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
+  .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
+  .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_g8_sse41_lowp
 .globl _sk_load_g8_sse41_lowp
 FUNCTION(_sk_load_g8_sse41_lowp)
@@ -56224,18 +57038,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2f401 <_sk_load_g8_sse41_lowp+0x39>
+  .byte  119,34                              // ja            3018a <_sk_load_g8_sse41_lowp+0x39>
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 2f480 <_sk_load_g8_sse41_lowp+0xb8>
+  .byte  76,141,13,150,0,0,0                 // lea           0x96(%rip),%r9        # 3020c <_sk_load_g8_sse41_lowp+0xbb>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,192                      // movd          %eax,%xmm0
-  .byte  235,100                             // jmp           2f465 <_sk_load_g8_sse41_lowp+0x9d>
+  .byte  235,100                             // jmp           301ee <_sk_load_g8_sse41_lowp+0x9d>
   .byte  102,65,15,56,48,4,16                // pmovzxbw      (%r8,%rdx,1),%xmm0
-  .byte  235,91                              // jmp           2f465 <_sk_load_g8_sse41_lowp+0x9d>
+  .byte  235,91                              // jmp           301ee <_sk_load_g8_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  102,15,196,200,2                    // pinsrw        $0x2,%eax,%xmm1
@@ -56243,7 +57057,7 @@
   .byte  102,15,110,192                      // movd          %eax,%xmm0
   .byte  102,15,56,48,192                    // pmovzxbw      %xmm0,%xmm0
   .byte  102,15,58,14,193,252                // pblendw       $0xfc,%xmm1,%xmm0
-  .byte  235,54                              // jmp           2f465 <_sk_load_g8_sse41_lowp+0x9d>
+  .byte  235,54                              // jmp           301ee <_sk_load_g8_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
   .byte  102,15,196,200,6                    // pinsrw        $0x6,%eax,%xmm1
@@ -56254,25 +57068,26 @@
   .byte  102,65,15,110,4,16                  // movd          (%r8,%rdx,1),%xmm0
   .byte  102,15,56,48,192                    // pmovzxbw      %xmm0,%xmm0
   .byte  102,15,58,14,193,240                // pblendw       $0xf0,%xmm1,%xmm0
-  .byte  102,15,219,5,3,63,0,0               // pand          0x3f03(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,5,90,72,0,0              // pand          0x485a(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,250,62,0,0                 // movaps        0x3efa(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,81,72,0,0                  // movaps        0x4851(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
-  .byte  118,255                             // jbe           2f481 <_sk_load_g8_sse41_lowp+0xb9>
+  .byte  15,31,0                             // nopl          (%rax)
+  .byte  115,255                             // jae           3020d <_sk_load_g8_sse41_lowp+0xbc>
   .byte  255                                 // (bad)
-  .byte  255,153,255,255,255,138             // lcall         *-0x75000001(%rcx)
+  .byte  255,150,255,255,255,135             // callq         *-0x78000001(%rsi)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,212                             // callq         *%rsp
+  .byte  255,209                             // callq         *%rcx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,201                             // dec           %ecx
+  .byte  255,198                             // inc           %esi
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  190,255,255,255,175                 // mov           $0xafffffff,%esi
+  .byte  187,255,255,255,172                 // mov           $0xacffffff,%ebx
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -56289,18 +57104,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            2f4d5 <_sk_load_g8_dst_sse41_lowp+0x39>
+  .byte  119,34                              // ja            30261 <_sk_load_g8_dst_sse41_lowp+0x39>
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 2f554 <_sk_load_g8_dst_sse41_lowp+0xb8>
+  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 302e0 <_sk_load_g8_dst_sse41_lowp+0xb8>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,224                      // movd          %eax,%xmm4
-  .byte  235,100                             // jmp           2f539 <_sk_load_g8_dst_sse41_lowp+0x9d>
+  .byte  235,100                             // jmp           302c5 <_sk_load_g8_dst_sse41_lowp+0x9d>
   .byte  102,65,15,56,48,36,16               // pmovzxbw      (%r8,%rdx,1),%xmm4
-  .byte  235,91                              // jmp           2f539 <_sk_load_g8_dst_sse41_lowp+0x9d>
+  .byte  235,91                              // jmp           302c5 <_sk_load_g8_dst_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  102,15,196,232,2                    // pinsrw        $0x2,%eax,%xmm5
@@ -56308,7 +57123,7 @@
   .byte  102,15,110,224                      // movd          %eax,%xmm4
   .byte  102,15,56,48,228                    // pmovzxbw      %xmm4,%xmm4
   .byte  102,15,58,14,229,252                // pblendw       $0xfc,%xmm5,%xmm4
-  .byte  235,54                              // jmp           2f539 <_sk_load_g8_dst_sse41_lowp+0x9d>
+  .byte  235,54                              // jmp           302c5 <_sk_load_g8_dst_sse41_lowp+0x9d>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,237                      // pxor          %xmm5,%xmm5
   .byte  102,15,196,232,6                    // pinsrw        $0x6,%eax,%xmm5
@@ -56319,13 +57134,13 @@
   .byte  102,65,15,110,36,16                 // movd          (%r8,%rdx,1),%xmm4
   .byte  102,15,56,48,228                    // pmovzxbw      %xmm4,%xmm4
   .byte  102,15,58,14,229,240                // pblendw       $0xf0,%xmm5,%xmm4
-  .byte  102,15,219,37,47,62,0,0             // pand          0x3e2f(%rip),%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,37,131,71,0,0            // pand          0x4783(%rip),%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,38,62,0,0                  // movaps        0x3e26(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,61,122,71,0,0                 // movaps        0x477a(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  255,224                             // jmpq          *%rax
-  .byte  118,255                             // jbe           2f555 <_sk_load_g8_dst_sse41_lowp+0xb9>
+  .byte  118,255                             // jbe           302e1 <_sk_load_g8_dst_sse41_lowp+0xb9>
   .byte  255                                 // (bad)
   .byte  255,153,255,255,255,138             // lcall         *-0x75000001(%rcx)
   .byte  255                                 // (bad)
@@ -56347,10 +57162,10 @@
 FUNCTION(_sk_luminance_to_alpha_sse41_lowp)
 _sk_luminance_to_alpha_sse41_lowp:
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
-  .byte  102,15,213,5,244,69,0,0             // pmullw        0x45f4(%rip),%xmm0        # 33b70 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
-  .byte  102,15,213,13,252,69,0,0            // pmullw        0x45fc(%rip),%xmm1        # 33b80 <_sk_srcover_rgba_8888_sse2_lowp+0x1200>
+  .byte  102,15,213,5,72,79,0,0              // pmullw        0x4f48(%rip),%xmm0        # 35250 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  .byte  102,15,213,13,80,79,0,0             // pmullw        0x4f50(%rip),%xmm1        # 35260 <_sk_srcover_rgba_8888_sse2_lowp+0x1230>
   .byte  102,15,253,200                      // paddw         %xmm0,%xmm1
-  .byte  102,15,213,29,0,70,0,0              // pmullw        0x4600(%rip),%xmm3        # 33b90 <_sk_srcover_rgba_8888_sse2_lowp+0x1210>
+  .byte  102,15,213,29,84,79,0,0             // pmullw        0x4f54(%rip),%xmm3        # 35270 <_sk_srcover_rgba_8888_sse2_lowp+0x1240>
   .byte  102,15,253,217                      // paddw         %xmm1,%xmm3
   .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -56359,20 +57174,97 @@
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_g8_sse41_lowp
+.globl _sk_gather_g8_sse41_lowp
+FUNCTION(_sk_gather_g8_sse41_lowp)
+_sk_gather_g8_sse41_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  76,139,0                            // mov           (%rax),%r8
+  .byte  243,68,15,91,195                    // cvttps2dq     %xmm3,%xmm8
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,15,56,64,211                    // pmulld        %xmm3,%xmm2
+  .byte  102,68,15,56,64,195                 // pmulld        %xmm3,%xmm8
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,200                   // paddd         %xmm8,%xmm1
+  .byte  102,72,15,58,22,200,1               // pextrq        $0x1,%xmm1,%rax
+  .byte  65,137,193                          // mov           %eax,%r9d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,73,15,126,202                   // movq          %xmm1,%r10
+  .byte  69,137,211                          // mov           %r10d,%r11d
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  102,72,15,58,22,195,1               // pextrq        $0x1,%xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  71,15,182,60,56                     // movzbl        (%r8,%r15,1),%r15d
+  .byte  67,15,182,44,32                     // movzbl        (%r8,%r12,1),%ebp
+  .byte  102,15,110,197                      // movd          %ebp,%xmm0
+  .byte  102,65,15,58,32,199,1               // pinsrb        $0x1,%r15d,%xmm0
+  .byte  67,15,182,44,48                     // movzbl        (%r8,%r14,1),%ebp
+  .byte  102,15,58,32,197,2                  // pinsrb        $0x2,%ebp,%xmm0
+  .byte  65,15,182,28,24                     // movzbl        (%r8,%rbx,1),%ebx
+  .byte  102,15,58,32,195,3                  // pinsrb        $0x3,%ebx,%xmm0
+  .byte  67,15,182,44,24                     // movzbl        (%r8,%r11,1),%ebp
+  .byte  102,15,58,32,197,4                  // pinsrb        $0x4,%ebp,%xmm0
+  .byte  67,15,182,44,16                     // movzbl        (%r8,%r10,1),%ebp
+  .byte  102,15,58,32,197,5                  // pinsrb        $0x5,%ebp,%xmm0
+  .byte  67,15,182,44,8                      // movzbl        (%r8,%r9,1),%ebp
+  .byte  102,15,58,32,197,6                  // pinsrb        $0x6,%ebp,%xmm0
+  .byte  65,15,182,4,0                       // movzbl        (%r8,%rax,1),%eax
+  .byte  102,15,58,32,192,7                  // pinsrb        $0x7,%eax,%xmm0
+  .byte  102,15,56,48,192                    // pmovzxbw      %xmm0,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  15,40,29,253,69,0,0                 // movaps        0x45fd(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_scale_1_float_sse41_lowp
 .globl _sk_scale_1_float_sse41_lowp
 FUNCTION(_sk_scale_1_float_sse41_lowp)
 _sk_scale_1_float_sse41_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
-  .byte  243,68,15,89,5,163,54,0,0           // mulss         0x36a3(%rip),%xmm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  243,68,15,88,5,106,54,0,0           // addss         0x366a(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,5,151,62,0,0           // mulss         0x3e97(%rip),%xmm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,68,15,88,5,94,62,0,0            // addss         0x3e5e(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,44,192                    // cvttss2si     %xmm8,%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  242,69,15,112,192,0                 // pshuflw       $0x0,%xmm8,%xmm8
   .byte  102,69,15,112,192,80                // pshufd        $0x50,%xmm8,%xmm8
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
-  .byte  102,68,15,111,13,138,61,0,0         // movdqa        0x3d8a(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,174,69,0,0         // movdqa        0x45ae(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -56393,13 +57285,13 @@
 _sk_lerp_1_float_sse41_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
-  .byte  243,68,15,89,5,43,54,0,0            // mulss         0x362b(%rip),%xmm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  243,68,15,88,5,242,53,0,0           // addss         0x35f2(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,5,31,62,0,0            // mulss         0x3e1f(%rip),%xmm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,68,15,88,5,230,61,0,0           // addss         0x3de6(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,44,192                    // cvttss2si     %xmm8,%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  242,69,15,112,192,0                 // pshuflw       $0x0,%xmm8,%xmm8
   .byte  102,69,15,112,192,80                // pshufd        $0x50,%xmm8,%xmm8
-  .byte  102,68,15,111,13,23,61,0,0          // movdqa        0x3d17(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,59,69,0,0          // movdqa        0x453b(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -56439,18 +57331,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,36                              // ja            2f710 <_sk_scale_u8_sse41_lowp+0x3b>
+  .byte  119,36                              // ja            305cc <_sk_scale_u8_sse41_lowp+0x3b>
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,221,0,0,0                 // lea           0xdd(%rip),%r9        # 2f7d8 <_sk_scale_u8_sse41_lowp+0x103>
+  .byte  76,141,13,221,0,0,0                 // lea           0xdd(%rip),%r9        # 30694 <_sk_scale_u8_sse41_lowp+0x103>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
-  .byte  235,111                             // jmp           2f77f <_sk_scale_u8_sse41_lowp+0xaa>
+  .byte  235,111                             // jmp           3063b <_sk_scale_u8_sse41_lowp+0xaa>
   .byte  102,69,15,56,48,4,16                // pmovzxbw      (%r8,%rdx,1),%xmm8
-  .byte  235,102                             // jmp           2f77f <_sk_scale_u8_sse41_lowp+0xaa>
+  .byte  235,102                             // jmp           3063b <_sk_scale_u8_sse41_lowp+0xaa>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,68,15,196,200,2                 // pinsrw        $0x2,%eax,%xmm9
@@ -56458,7 +57350,7 @@
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  102,69,15,56,48,192                 // pmovzxbw      %xmm8,%xmm8
   .byte  102,69,15,58,14,193,252             // pblendw       $0xfc,%xmm9,%xmm8
-  .byte  235,60                              // jmp           2f77f <_sk_scale_u8_sse41_lowp+0xaa>
+  .byte  235,60                              // jmp           3063b <_sk_scale_u8_sse41_lowp+0xaa>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,68,15,196,200,6                 // pinsrw        $0x6,%eax,%xmm9
@@ -56469,9 +57361,9 @@
   .byte  102,69,15,110,4,16                  // movd          (%r8,%rdx,1),%xmm8
   .byte  102,69,15,56,48,192                 // pmovzxbw      %xmm8,%xmm8
   .byte  102,69,15,58,14,193,240             // pblendw       $0xf0,%xmm9,%xmm8
-  .byte  102,68,15,219,5,232,59,0,0          // pand          0x3be8(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,219,5,12,68,0,0           // pand          0x440c(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
-  .byte  102,68,15,111,13,218,59,0,0         // movdqa        0x3bda(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,254,67,0,0         // movdqa        0x43fe(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -56496,7 +57388,7 @@
   .byte  255,148,255,255,255,136,255         // callq         *-0x770001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  124,255                             // jl            2f7ed <_sk_scale_u8_sse41_lowp+0x118>
+  .byte  124,255                             // jl            306a9 <_sk_scale_u8_sse41_lowp+0x118>
   .byte  255                                 // (bad)
   .byte  255,107,255                         // ljmp          *-0x1(%rbx)
   .byte  255                                 // (bad)
@@ -56514,18 +57406,18 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,36                              // ja            2f82f <_sk_lerp_u8_sse41_lowp+0x3b>
+  .byte  119,36                              // ja            306eb <_sk_lerp_u8_sse41_lowp+0x3b>
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,30,1,0,0                  // lea           0x11e(%rip),%r9        # 2f938 <_sk_lerp_u8_sse41_lowp+0x144>
+  .byte  76,141,13,30,1,0,0                  // lea           0x11e(%rip),%r9        # 307f4 <_sk_lerp_u8_sse41_lowp+0x144>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
-  .byte  235,111                             // jmp           2f89e <_sk_lerp_u8_sse41_lowp+0xaa>
+  .byte  235,111                             // jmp           3075a <_sk_lerp_u8_sse41_lowp+0xaa>
   .byte  102,69,15,56,48,4,16                // pmovzxbw      (%r8,%rdx,1),%xmm8
-  .byte  235,102                             // jmp           2f89e <_sk_lerp_u8_sse41_lowp+0xaa>
+  .byte  235,102                             // jmp           3075a <_sk_lerp_u8_sse41_lowp+0xaa>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,68,15,196,200,2                 // pinsrw        $0x2,%eax,%xmm9
@@ -56533,7 +57425,7 @@
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  102,69,15,56,48,192                 // pmovzxbw      %xmm8,%xmm8
   .byte  102,69,15,58,14,193,252             // pblendw       $0xfc,%xmm9,%xmm8
-  .byte  235,60                              // jmp           2f89e <_sk_lerp_u8_sse41_lowp+0xaa>
+  .byte  235,60                              // jmp           3075a <_sk_lerp_u8_sse41_lowp+0xaa>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,69,15,239,201                   // pxor          %xmm9,%xmm9
   .byte  102,68,15,196,200,6                 // pinsrw        $0x6,%eax,%xmm9
@@ -56544,8 +57436,8 @@
   .byte  102,69,15,110,4,16                  // movd          (%r8,%rdx,1),%xmm8
   .byte  102,69,15,56,48,192                 // pmovzxbw      %xmm8,%xmm8
   .byte  102,69,15,58,14,193,240             // pblendw       $0xf0,%xmm9,%xmm8
-  .byte  102,68,15,219,5,201,58,0,0          // pand          0x3ac9(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
-  .byte  102,68,15,111,21,192,58,0,0         // movdqa        0x3ac0(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,219,5,237,66,0,0          // pand          0x42ed(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,68,15,111,21,228,66,0,0         // movdqa        0x42e4(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,69,15,239,202                   // pxor          %xmm10,%xmm9
   .byte  102,69,15,111,217                   // movdqa        %xmm9,%xmm11
@@ -56575,7 +57467,7 @@
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  255,224                             // jmpq          *%rax
   .byte  144                                 // nop
-  .byte  235,254                             // jmp           2f938 <_sk_lerp_u8_sse41_lowp+0x144>
+  .byte  235,254                             // jmp           307f4 <_sk_lerp_u8_sse41_lowp+0x144>
   .byte  255                                 // (bad)
   .byte  255,17                              // callq         *(%rcx)
   .byte  255                                 // (bad)
@@ -56609,23 +57501,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,35                              // ja            2f996 <_sk_scale_565_sse41_lowp+0x42>
+  .byte  119,35                              // ja            30852 <_sk_scale_565_sse41_lowp+0x42>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,111,1,0,0                 // lea           0x16f(%rip),%r9        # 2faf0 <_sk_scale_565_sse41_lowp+0x19c>
+  .byte  76,141,13,111,1,0,0                 // lea           0x16f(%rip),%r9        # 309ac <_sk_scale_565_sse41_lowp+0x19c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,68,15,110,200                   // movd          %eax,%xmm9
-  .byte  235,76                              // jmp           2f9e2 <_sk_scale_565_sse41_lowp+0x8e>
+  .byte  235,76                              // jmp           3089e <_sk_scale_565_sse41_lowp+0x8e>
   .byte  243,69,15,111,12,80                 // movdqu        (%r8,%rdx,2),%xmm9
-  .byte  235,68                              // jmp           2f9e2 <_sk_scale_565_sse41_lowp+0x8e>
+  .byte  235,68                              // jmp           3089e <_sk_scale_565_sse41_lowp+0x8e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,4,2             // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   .byte  102,69,15,110,12,80                 // movd          (%r8,%rdx,2),%xmm9
   .byte  102,68,15,58,14,200,252             // pblendw       $0xfc,%xmm0,%xmm9
-  .byte  235,41                              // jmp           2f9e2 <_sk_scale_565_sse41_lowp+0x8e>
+  .byte  235,41                              // jmp           3089e <_sk_scale_565_sse41_lowp+0x8e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,12,6            // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   .byte  102,65,15,196,68,80,10,5            // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -56634,11 +57526,11 @@
   .byte  102,68,15,58,14,200,240             // pblendw       $0xf0,%xmm0,%xmm9
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
-  .byte  102,15,219,5,44,65,0,0              // pand          0x412c(%rip),%xmm0        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,5,80,73,0,0              // pand          0x4950(%rip),%xmm0        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,69,15,111,209                   // movdqa        %xmm9,%xmm10
   .byte  102,65,15,113,210,5                 // psrlw         $0x5,%xmm10
-  .byte  102,68,15,219,21,40,65,0,0          // pand          0x4128(%rip),%xmm10        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,68,15,111,29,47,65,0,0          // movdqa        0x412f(%rip),%xmm11        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,68,15,219,21,76,73,0,0          // pand          0x494c(%rip),%xmm10        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,111,29,83,73,0,0          // movdqa        0x4953(%rip),%xmm11        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,69,15,219,217                   // pand          %xmm9,%xmm11
   .byte  102,65,15,113,209,13                // psrlw         $0xd,%xmm9
   .byte  102,68,15,235,200                   // por           %xmm0,%xmm9
@@ -56650,7 +57542,7 @@
   .byte  102,15,113,240,3                    // psllw         $0x3,%xmm0
   .byte  102,65,15,113,211,2                 // psrlw         $0x2,%xmm11
   .byte  102,68,15,235,216                   // por           %xmm0,%xmm11
-  .byte  102,15,111,5,141,64,0,0             // movdqa        0x408d(%rip),%xmm0        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,15,111,5,177,72,0,0             // movdqa        0x48b1(%rip),%xmm0        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,68,15,111,227                   // movdqa        %xmm3,%xmm12
   .byte  102,68,15,239,224                   // pxor          %xmm0,%xmm12
   .byte  102,15,239,199                      // pxor          %xmm7,%xmm0
@@ -56663,7 +57555,7 @@
   .byte  102,69,15,56,62,225                 // pmaxuw        %xmm9,%xmm12
   .byte  102,69,15,56,16,229                 // pblendvb      %xmm0,%xmm13,%xmm12
   .byte  102,69,15,213,200                   // pmullw        %xmm8,%xmm9
-  .byte  102,15,111,5,213,56,0,0             // movdqa        0x38d5(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,5,249,64,0,0             // movdqa        0x40f9(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,253,200                   // paddw         %xmm0,%xmm9
   .byte  102,65,15,113,209,8                 // psrlw         $0x8,%xmm9
   .byte  102,68,15,213,209                   // pmullw        %xmm1,%xmm10
@@ -56717,23 +57609,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,35                              // ja            2fb4e <_sk_lerp_565_sse41_lowp+0x42>
+  .byte  119,35                              // ja            30a0a <_sk_lerp_565_sse41_lowp+0x42>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,187,1,0,0                 // lea           0x1bb(%rip),%r9        # 2fcf4 <_sk_lerp_565_sse41_lowp+0x1e8>
+  .byte  76,141,13,187,1,0,0                 // lea           0x1bb(%rip),%r9        # 30bb0 <_sk_lerp_565_sse41_lowp+0x1e8>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,68,15,110,200                   // movd          %eax,%xmm9
-  .byte  235,76                              // jmp           2fb9a <_sk_lerp_565_sse41_lowp+0x8e>
+  .byte  235,76                              // jmp           30a56 <_sk_lerp_565_sse41_lowp+0x8e>
   .byte  243,69,15,111,12,80                 // movdqu        (%r8,%rdx,2),%xmm9
-  .byte  235,68                              // jmp           2fb9a <_sk_lerp_565_sse41_lowp+0x8e>
+  .byte  235,68                              // jmp           30a56 <_sk_lerp_565_sse41_lowp+0x8e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,4,2             // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   .byte  102,69,15,110,12,80                 // movd          (%r8,%rdx,2),%xmm9
   .byte  102,68,15,58,14,200,252             // pblendw       $0xfc,%xmm0,%xmm9
-  .byte  235,41                              // jmp           2fb9a <_sk_lerp_565_sse41_lowp+0x8e>
+  .byte  235,41                              // jmp           30a56 <_sk_lerp_565_sse41_lowp+0x8e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,12,6            // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   .byte  102,65,15,196,68,80,10,5            // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -56742,11 +57634,11 @@
   .byte  102,68,15,58,14,200,240             // pblendw       $0xf0,%xmm0,%xmm9
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
-  .byte  102,15,219,5,116,63,0,0             // pand          0x3f74(%rip),%xmm0        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,5,152,71,0,0             // pand          0x4798(%rip),%xmm0        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,69,15,111,209                   // movdqa        %xmm9,%xmm10
   .byte  102,65,15,113,210,5                 // psrlw         $0x5,%xmm10
-  .byte  102,68,15,219,21,112,63,0,0         // pand          0x3f70(%rip),%xmm10        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,68,15,111,29,119,63,0,0         // movdqa        0x3f77(%rip),%xmm11        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,68,15,219,21,148,71,0,0         // pand          0x4794(%rip),%xmm10        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,111,29,155,71,0,0         // movdqa        0x479b(%rip),%xmm11        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,69,15,219,217                   // pand          %xmm9,%xmm11
   .byte  102,65,15,113,209,13                // psrlw         $0xd,%xmm9
   .byte  102,68,15,235,200                   // por           %xmm0,%xmm9
@@ -56758,7 +57650,7 @@
   .byte  102,15,113,240,3                    // psllw         $0x3,%xmm0
   .byte  102,65,15,113,211,2                 // psrlw         $0x2,%xmm11
   .byte  102,68,15,235,216                   // por           %xmm0,%xmm11
-  .byte  102,15,111,5,213,62,0,0             // movdqa        0x3ed5(%rip),%xmm0        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,15,111,5,249,70,0,0             // movdqa        0x46f9(%rip),%xmm0        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,68,15,111,227                   // movdqa        %xmm3,%xmm12
   .byte  102,68,15,239,224                   // pxor          %xmm0,%xmm12
   .byte  102,15,239,199                      // pxor          %xmm7,%xmm0
@@ -56770,7 +57662,7 @@
   .byte  102,69,15,56,62,227                 // pmaxuw        %xmm11,%xmm12
   .byte  102,69,15,56,62,225                 // pmaxuw        %xmm9,%xmm12
   .byte  102,69,15,56,16,229                 // pblendvb      %xmm0,%xmm13,%xmm12
-  .byte  102,68,15,111,45,33,55,0,0          // movdqa        0x3721(%rip),%xmm13        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,45,69,63,0,0          // movdqa        0x3f45(%rip),%xmm13        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,111,193                   // movdqa        %xmm9,%xmm0
   .byte  102,65,15,239,197                   // pxor          %xmm13,%xmm0
   .byte  102,15,213,196                      // pmullw        %xmm4,%xmm0
@@ -56834,48 +57726,48 @@
   .byte  65,254,202                          // dec           %r10b
   .byte  69,15,182,202                       // movzbl        %r10b,%r9d
   .byte  65,128,249,6                        // cmp           $0x6,%r9b
-  .byte  119,38                              // ja            2fd5b <_sk_srcover_rgba_8888_sse41_lowp+0x4b>
+  .byte  119,38                              // ja            30c17 <_sk_srcover_rgba_8888_sse41_lowp+0x4b>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  76,141,29,23,2,0,0                  // lea           0x217(%rip),%r11        # 2ff58 <_sk_srcover_rgba_8888_sse41_lowp+0x248>
+  .byte  76,141,29,23,2,0,0                  // lea           0x217(%rip),%r11        # 30e14 <_sk_srcover_rgba_8888_sse41_lowp+0x248>
   .byte  75,99,4,139                         // movslq        (%r11,%r9,4),%rax
   .byte  76,1,216                            // add           %r11,%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,110,52,144                // movd          (%r8,%rdx,4),%xmm6
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,89                              // jmp           2fdb4 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  .byte  235,89                              // jmp           30c70 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   .byte  243,65,15,111,52,144                // movdqu        (%r8,%rdx,4),%xmm6
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,74                              // jmp           2fdb4 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  .byte  235,74                              // jmp           30c70 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  243,65,15,126,52,144                // movq          (%r8,%rdx,4),%xmm6
   .byte  102,65,15,58,14,240,240             // pblendw       $0xf0,%xmm8,%xmm6
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
-  .byte  235,37                              // jmp           2fdb4 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  .byte  235,37                              // jmp           30c70 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,69,15,58,34,68,144,20,1         // pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   .byte  102,69,15,58,34,68,144,16,0         // pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   .byte  243,65,15,111,52,144                // movdqu        (%r8,%rdx,4),%xmm6
-  .byte  102,15,111,37,52,61,0,0             // movdqa        0x3d34(%rip),%xmm4        # 33af0 <_sk_srcover_rgba_8888_sse2_lowp+0x1170>
+  .byte  102,15,111,37,88,69,0,0             // movdqa        0x4558(%rip),%xmm4        # 351d0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
   .byte  102,15,111,238                      // movdqa        %xmm6,%xmm5
   .byte  102,15,56,0,236                     // pshufb        %xmm4,%xmm5
   .byte  102,65,15,111,248                   // movdqa        %xmm8,%xmm7
   .byte  102,15,56,0,252                     // pshufb        %xmm4,%xmm7
   .byte  102,15,108,239                      // punpcklqdq    %xmm7,%xmm5
-  .byte  102,68,15,111,37,148,53,0,0         // movdqa        0x3594(%rip),%xmm12        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,37,184,61,0,0         // movdqa        0x3db8(%rip),%xmm12        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,229                      // movdqa        %xmm5,%xmm4
   .byte  102,65,15,219,228                   // pand          %xmm12,%xmm4
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
-  .byte  102,68,15,111,13,13,61,0,0          // movdqa        0x3d0d(%rip),%xmm9        # 33b00 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
+  .byte  102,68,15,111,13,49,69,0,0          // movdqa        0x4531(%rip),%xmm9        # 351e0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
   .byte  102,15,111,254                      // movdqa        %xmm6,%xmm7
   .byte  102,65,15,56,0,249                  // pshufb        %xmm9,%xmm7
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
   .byte  102,69,15,56,0,209                  // pshufb        %xmm9,%xmm10
   .byte  102,65,15,108,250                   // punpcklqdq    %xmm10,%xmm7
-  .byte  102,68,15,111,13,250,60,0,0         // movdqa        0x3cfa(%rip),%xmm9        # 33b10 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
+  .byte  102,68,15,111,13,30,69,0,0          // movdqa        0x451e(%rip),%xmm9        # 351f0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
   .byte  102,65,15,56,0,241                  // pshufb        %xmm9,%xmm6
   .byte  102,69,15,56,0,193                  // pshufb        %xmm9,%xmm8
   .byte  102,65,15,108,240                   // punpcklqdq    %xmm8,%xmm6
@@ -56917,19 +57809,19 @@
   .byte  102,15,97,195                       // punpcklwd     %xmm3,%xmm0
   .byte  102,65,15,235,196                   // por           %xmm12,%xmm0
   .byte  65,128,250,6                        // cmp           $0x6,%r10b
-  .byte  119,24                              // ja            2fefe <_sk_srcover_rgba_8888_sse41_lowp+0x1ee>
-  .byte  76,141,21,135,0,0,0                 // lea           0x87(%rip),%r10        # 2ff74 <_sk_srcover_rgba_8888_sse41_lowp+0x264>
+  .byte  119,24                              // ja            30dba <_sk_srcover_rgba_8888_sse41_lowp+0x1ee>
+  .byte  76,141,21,135,0,0,0                 // lea           0x87(%rip),%r10        # 30e30 <_sk_srcover_rgba_8888_sse41_lowp+0x264>
   .byte  75,99,4,138                         // movslq        (%r10,%r9,4),%rax
   .byte  76,1,208                            // add           %r10,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,65,15,126,4,144                 // movd          %xmm0,(%r8,%rdx,4)
-  .byte  235,63                              // jmp           2ff3d <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  .byte  235,63                              // jmp           30df9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   .byte  243,65,15,127,4,144                 // movdqu        %xmm0,(%r8,%rdx,4)
   .byte  243,65,15,127,76,144,16             // movdqu        %xmm1,0x10(%r8,%rdx,4)
-  .byte  235,48                              // jmp           2ff3d <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  .byte  235,48                              // jmp           30df9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   .byte  102,65,15,58,22,68,144,8,2          // pextrd        $0x2,%xmm0,0x8(%r8,%rdx,4)
   .byte  102,65,15,214,4,144                 // movq          %xmm0,(%r8,%rdx,4)
-  .byte  235,31                              // jmp           2ff3d <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  .byte  235,31                              // jmp           30df9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   .byte  102,65,15,58,22,76,144,24,2         // pextrd        $0x2,%xmm1,0x18(%r8,%rdx,4)
   .byte  102,65,15,58,22,76,144,20,1         // pextrd        $0x1,%xmm1,0x14(%r8,%rdx,4)
   .byte  102,65,15,126,76,144,16             // movd          %xmm1,0x10(%r8,%rdx,4)
@@ -56994,13 +57886,13 @@
   .byte  73,137,246                          // mov           %rsi,%r14
   .byte  72,137,77,192                       // mov           %rcx,-0x40(%rbp)
   .byte  72,57,203                           // cmp           %rcx,%rbx
-  .byte  15,131,131,0,0,0                    // jae           30046 <_sk_start_pipeline_sse2_lowp+0xb6>
+  .byte  15,131,131,0,0,0                    // jae           30f02 <_sk_start_pipeline_sse2_lowp+0xb6>
   .byte  72,139,69,208                       // mov           -0x30(%rbp),%rax
   .byte  72,141,64,8                         // lea           0x8(%rax),%rax
   .byte  72,137,69,200                       // mov           %rax,-0x38(%rbp)
   .byte  76,57,125,200                       // cmp           %r15,-0x38(%rbp)
   .byte  72,139,85,208                       // mov           -0x30(%rbp),%rdx
-  .byte  119,59                              // ja            30014 <_sk_start_pipeline_sse2_lowp+0x84>
+  .byte  119,59                              // ja            30ed0 <_sk_start_pipeline_sse2_lowp+0x84>
   .byte  76,139,101,208                      // mov           -0x30(%rbp),%r12
   .byte  49,255                              // xor           %edi,%edi
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -57019,10 +57911,10 @@
   .byte  73,131,196,16                       // add           $0x10,%r12
   .byte  77,57,252                           // cmp           %r15,%r12
   .byte  73,137,212                          // mov           %rdx,%r12
-  .byte  118,201                             // jbe           2ffdd <_sk_start_pipeline_sse2_lowp+0x4d>
+  .byte  118,201                             // jbe           30e99 <_sk_start_pipeline_sse2_lowp+0x4d>
   .byte  76,137,255                          // mov           %r15,%rdi
   .byte  72,41,215                           // sub           %rdx,%rdi
-  .byte  116,33                              // je            3003d <_sk_start_pipeline_sse2_lowp+0xad>
+  .byte  116,33                              // je            30ef9 <_sk_start_pipeline_sse2_lowp+0xad>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -57036,7 +57928,7 @@
   .byte  65,255,213                          // callq         *%r13
   .byte  72,255,195                          // inc           %rbx
   .byte  72,59,93,192                        // cmp           -0x40(%rbp),%rbx
-  .byte  117,137                             // jne           2ffcf <_sk_start_pipeline_sse2_lowp+0x3f>
+  .byte  117,137                             // jne           30e8b <_sk_start_pipeline_sse2_lowp+0x3f>
   .byte  72,131,196,24                       // add           $0x18,%rsp
   .byte  91                                  // pop           %rbx
   .byte  65,92                               // pop           %r12
@@ -57067,11 +57959,51 @@
   .byte  102,15,110,209                      // movd          %ecx,%xmm2
   .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,88,21,60,51,0,0                  // addps         0x333c(%rip),%xmm2        # 333c0 <_sk_srcover_rgba_8888_sse2_lowp+0xa40>
+  .byte  15,88,21,96,59,0,0                  // addps         0x3b60(%rip),%xmm2        # 34aa0 <_sk_srcover_rgba_8888_sse2_lowp+0xa70>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_translate_sse2_lowp
+.globl _sk_matrix_translate_sse2_lowp
+FUNCTION(_sk_matrix_translate_sse2_lowp)
+_sk_matrix_translate_sse2_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
+  .byte  243,68,15,16,72,4                   // movss         0x4(%rax),%xmm9
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  65,15,88,192                        // addps         %xmm8,%xmm0
+  .byte  65,15,88,200                        // addps         %xmm8,%xmm1
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  65,15,88,209                        // addps         %xmm9,%xmm2
+  .byte  65,15,88,217                        // addps         %xmm9,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
+HIDDEN _sk_matrix_scale_translate_sse2_lowp
+.globl _sk_matrix_scale_translate_sse2_lowp
+FUNCTION(_sk_matrix_scale_translate_sse2_lowp)
+_sk_matrix_scale_translate_sse2_lowp:
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
+  .byte  243,68,15,16,72,4                   // movss         0x4(%rax),%xmm9
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  243,68,15,16,80,8                   // movss         0x8(%rax),%xmm10
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
+  .byte  65,15,89,192                        // mulps         %xmm8,%xmm0
+  .byte  65,15,88,194                        // addps         %xmm10,%xmm0
+  .byte  65,15,88,202                        // addps         %xmm10,%xmm1
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  65,15,89,217                        // mulps         %xmm9,%xmm3
+  .byte  65,15,89,209                        // mulps         %xmm9,%xmm2
+  .byte  65,15,88,208                        // addps         %xmm8,%xmm2
+  .byte  65,15,88,216                        // addps         %xmm8,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_matrix_2x3_sse2_lowp
 .globl _sk_matrix_2x3_sse2_lowp
 FUNCTION(_sk_matrix_2x3_sse2_lowp)
@@ -57113,6 +58045,70 @@
   .byte  65,15,40,218                        // movaps        %xmm10,%xmm3
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_matrix_perspective_sse2_lowp
+.globl _sk_matrix_perspective_sse2_lowp
+FUNCTION(_sk_matrix_perspective_sse2_lowp)
+_sk_matrix_perspective_sse2_lowp:
+  .byte  68,15,40,193                        // movaps        %xmm1,%xmm8
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,15,16,8                         // movss         (%rax),%xmm1
+  .byte  243,68,15,16,80,4                   // movss         0x4(%rax),%xmm10
+  .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  243,68,15,16,72,8                   // movss         0x8(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  68,15,40,219                        // movaps        %xmm3,%xmm11
+  .byte  69,15,89,218                        // mulps         %xmm10,%xmm11
+  .byte  68,15,89,210                        // mulps         %xmm2,%xmm10
+  .byte  69,15,88,209                        // addps         %xmm9,%xmm10
+  .byte  69,15,88,217                        // addps         %xmm9,%xmm11
+  .byte  68,15,40,200                        // movaps        %xmm0,%xmm9
+  .byte  68,15,89,201                        // mulps         %xmm1,%xmm9
+  .byte  69,15,88,202                        // addps         %xmm10,%xmm9
+  .byte  65,15,89,200                        // mulps         %xmm8,%xmm1
+  .byte  65,15,88,203                        // addps         %xmm11,%xmm1
+  .byte  243,68,15,16,96,12                  // movss         0xc(%rax),%xmm12
+  .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
+  .byte  243,68,15,16,80,16                  // movss         0x10(%rax),%xmm10
+  .byte  69,15,198,210,0                     // shufps        $0x0,%xmm10,%xmm10
+  .byte  243,68,15,16,104,20                 // movss         0x14(%rax),%xmm13
+  .byte  69,15,198,237,0                     // shufps        $0x0,%xmm13,%xmm13
+  .byte  68,15,40,219                        // movaps        %xmm3,%xmm11
+  .byte  69,15,89,218                        // mulps         %xmm10,%xmm11
+  .byte  68,15,89,210                        // mulps         %xmm2,%xmm10
+  .byte  69,15,88,213                        // addps         %xmm13,%xmm10
+  .byte  69,15,88,221                        // addps         %xmm13,%xmm11
+  .byte  68,15,40,232                        // movaps        %xmm0,%xmm13
+  .byte  69,15,89,236                        // mulps         %xmm12,%xmm13
+  .byte  69,15,88,213                        // addps         %xmm13,%xmm10
+  .byte  69,15,89,224                        // mulps         %xmm8,%xmm12
+  .byte  69,15,88,220                        // addps         %xmm12,%xmm11
+  .byte  243,68,15,16,96,24                  // movss         0x18(%rax),%xmm12
+  .byte  69,15,198,228,0                     // shufps        $0x0,%xmm12,%xmm12
+  .byte  243,68,15,16,104,28                 // movss         0x1c(%rax),%xmm13
+  .byte  69,15,198,237,0                     // shufps        $0x0,%xmm13,%xmm13
+  .byte  243,68,15,16,112,32                 // movss         0x20(%rax),%xmm14
+  .byte  69,15,198,246,0                     // shufps        $0x0,%xmm14,%xmm14
+  .byte  65,15,89,221                        // mulps         %xmm13,%xmm3
+  .byte  68,15,89,234                        // mulps         %xmm2,%xmm13
+  .byte  69,15,88,238                        // addps         %xmm14,%xmm13
+  .byte  65,15,88,222                        // addps         %xmm14,%xmm3
+  .byte  65,15,89,196                        // mulps         %xmm12,%xmm0
+  .byte  65,15,88,197                        // addps         %xmm13,%xmm0
+  .byte  69,15,89,224                        // mulps         %xmm8,%xmm12
+  .byte  68,15,88,227                        // addps         %xmm3,%xmm12
+  .byte  15,83,192                           // rcpps         %xmm0,%xmm0
+  .byte  65,15,83,212                        // rcpps         %xmm12,%xmm2
+  .byte  68,15,89,200                        // mulps         %xmm0,%xmm9
+  .byte  15,89,202                           // mulps         %xmm2,%xmm1
+  .byte  68,15,89,208                        // mulps         %xmm0,%xmm10
+  .byte  68,15,89,218                        // mulps         %xmm2,%xmm11
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  65,15,40,193                        // movaps        %xmm9,%xmm0
+  .byte  65,15,40,210                        // movaps        %xmm10,%xmm2
+  .byte  65,15,40,219                        // movaps        %xmm11,%xmm3
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_uniform_color_sse2_lowp
 .globl _sk_uniform_color_sse2_lowp
 FUNCTION(_sk_uniform_color_sse2_lowp)
@@ -57140,7 +58136,7 @@
 FUNCTION(_sk_black_color_sse2_lowp)
 _sk_black_color_sse2_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,249,49,0,0                 // movaps        0x31f9(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,165,56,0,0                 // movaps        0x38a5(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -57151,7 +58147,7 @@
 FUNCTION(_sk_white_color_sse2_lowp)
 _sk_white_color_sse2_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,5,229,49,0,0                  // movaps        0x31e5(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,5,145,56,0,0                  // movaps        0x3891(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -57162,10 +58158,10 @@
 FUNCTION(_sk_set_rgb_sse2_lowp)
 _sk_set_rgb_sse2_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  243,15,16,21,188,42,0,0             // movss         0x2abc(%rip),%xmm2        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,15,16,21,56,49,0,0              // movss         0x3138(%rip),%xmm2        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   .byte  243,15,16,0                         // movss         (%rax),%xmm0
   .byte  243,15,89,194                       // mulss         %xmm2,%xmm0
-  .byte  243,68,15,16,5,123,42,0,0           // movss         0x2a7b(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,16,5,247,48,0,0           // movss         0x30f7(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,88,192                    // addss         %xmm8,%xmm0
   .byte  243,68,15,44,192                    // cvttss2si     %xmm0,%r8d
   .byte  102,65,15,110,192                   // movd          %r8d,%xmm0
@@ -57192,7 +58188,7 @@
 FUNCTION(_sk_premul_sse2_lowp)
 _sk_premul_sse2_lowp:
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,68,15,111,5,87,49,0,0           // movdqa        0x3157(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,3,56,0,0            // movdqa        0x3803(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
@@ -57240,7 +58236,7 @@
 .globl _sk_invert_sse2_lowp
 FUNCTION(_sk_invert_sse2_lowp)
 _sk_invert_sse2_lowp:
-  .byte  102,68,15,111,5,245,48,0,0          // movdqa        0x30f5(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,161,55,0,0          // movdqa        0x37a1(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,200                   // psubw         %xmm0,%xmm9
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -57272,7 +58268,7 @@
 _sk_srcatop_sse2_lowp:
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,15,111,29,153,48,0,0            // movdqa        0x3099(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,29,69,55,0,0             // movdqa        0x3745(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,111,203                   // movdqa        %xmm3,%xmm9
   .byte  102,69,15,249,200                   // psubw         %xmm8,%xmm9
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
@@ -57304,7 +58300,7 @@
 _sk_dstatop_sse2_lowp:
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
   .byte  102,68,15,213,195                   // pmullw        %xmm3,%xmm8
-  .byte  102,68,15,111,13,25,48,0,0          // movdqa        0x3019(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,197,54,0,0         // movdqa        0x36c5(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,253,193                   // paddw         %xmm9,%xmm8
   .byte  102,68,15,111,213                   // movdqa        %xmm5,%xmm10
   .byte  102,68,15,213,211                   // pmullw        %xmm3,%xmm10
@@ -57333,7 +58329,7 @@
 FUNCTION(_sk_srcin_sse2_lowp)
 _sk_srcin_sse2_lowp:
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,68,15,111,5,164,47,0,0          // movdqa        0x2fa4(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,80,54,0,0           // movdqa        0x3650(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,207                      // pmullw        %xmm7,%xmm1
@@ -57354,7 +58350,7 @@
 _sk_dstin_sse2_lowp:
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,68,15,111,5,91,47,0,0           // movdqa        0x2f5b(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,7,54,0,0            // movdqa        0x3607(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
@@ -57375,7 +58371,7 @@
 .globl _sk_srcout_sse2_lowp
 FUNCTION(_sk_srcout_sse2_lowp)
 _sk_srcout_sse2_lowp:
-  .byte  102,68,15,111,5,18,47,0,0           // movdqa        0x2f12(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,190,53,0,0          // movdqa        0x35be(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,207                   // psubw         %xmm7,%xmm9
   .byte  102,65,15,213,193                   // pmullw        %xmm9,%xmm0
@@ -57398,7 +58394,7 @@
 FUNCTION(_sk_dstout_sse2_lowp)
 _sk_dstout_sse2_lowp:
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  102,68,15,111,5,187,46,0,0          // movdqa        0x2ebb(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,103,53,0,0          // movdqa        0x3567(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  102,15,249,216                      // psubw         %xmm0,%xmm3
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
@@ -57423,7 +58419,7 @@
 .globl _sk_srcover_sse2_lowp
 FUNCTION(_sk_srcover_sse2_lowp)
 _sk_srcover_sse2_lowp:
-  .byte  102,68,15,111,13,97,46,0,0          // movdqa        0x2e61(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,13,53,0,0          // movdqa        0x350d(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,68,15,249,195                   // psubw         %xmm3,%xmm8
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -57452,7 +58448,7 @@
 .globl _sk_dstover_sse2_lowp
 FUNCTION(_sk_dstover_sse2_lowp)
 _sk_dstover_sse2_lowp:
-  .byte  102,68,15,111,5,231,45,0,0          // movdqa        0x2de7(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,147,52,0,0          // movdqa        0x3493(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,68,15,249,207                   // psubw         %xmm7,%xmm9
   .byte  102,65,15,213,193                   // pmullw        %xmm9,%xmm0
@@ -57479,7 +58475,7 @@
 FUNCTION(_sk_modulate_sse2_lowp)
 _sk_modulate_sse2_lowp:
   .byte  102,15,213,196                      // pmullw        %xmm4,%xmm0
-  .byte  102,68,15,111,5,128,45,0,0          // movdqa        0x2d80(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,44,52,0,0           // movdqa        0x342c(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,192                   // paddw         %xmm8,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,205                      // pmullw        %xmm5,%xmm1
@@ -57498,7 +58494,7 @@
 .globl _sk_multiply_sse2_lowp
 FUNCTION(_sk_multiply_sse2_lowp)
 _sk_multiply_sse2_lowp:
-  .byte  102,68,15,111,13,63,45,0,0          // movdqa        0x2d3f(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,235,51,0,0         // movdqa        0x33eb(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,193                   // movdqa        %xmm9,%xmm8
   .byte  102,68,15,249,195                   // psubw         %xmm3,%xmm8
   .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
@@ -57543,12 +58539,12 @@
   .byte  102,68,15,111,193                   // movdqa        %xmm1,%xmm8
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,253,204                      // paddw         %xmm4,%xmm1
-  .byte  102,68,15,111,21,171,52,0,0         // movdqa        0x34ab(%rip),%xmm10        # 33ba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  .byte  102,68,15,111,21,87,59,0,0          // movdqa        0x3b57(%rip),%xmm10        # 35280 <_sk_srcover_rgba_8888_sse2_lowp+0x1250>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,217,194                   // psubusw       %xmm10,%xmm0
   .byte  102,69,15,239,228                   // pxor          %xmm12,%xmm12
   .byte  102,65,15,117,196                   // pcmpeqw       %xmm12,%xmm0
-  .byte  102,68,15,111,29,95,44,0,0          // movdqa        0x2c5f(%rip),%xmm11        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,29,11,51,0,0          // movdqa        0x330b(%rip),%xmm11        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
   .byte  102,65,15,223,195                   // pandn         %xmm11,%xmm0
   .byte  102,15,235,193                      // por           %xmm1,%xmm0
@@ -57587,7 +58583,7 @@
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,204                      // pmullw        %xmm4,%xmm1
-  .byte  102,68,15,111,21,191,43,0,0         // movdqa        0x2bbf(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,21,107,50,0,0         // movdqa        0x326b(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,202                   // paddw         %xmm10,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -57619,7 +58615,7 @@
 FUNCTION(_sk_xor__sse2_lowp)
 _sk_xor__sse2_lowp:
   .byte  102,68,15,111,195                   // movdqa        %xmm3,%xmm8
-  .byte  102,68,15,111,13,61,43,0,0          // movdqa        0x2b3d(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,233,49,0,0         // movdqa        0x31e9(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,209                   // movdqa        %xmm9,%xmm10
   .byte  102,68,15,249,215                   // psubw         %xmm7,%xmm10
   .byte  102,65,15,213,194                   // pmullw        %xmm10,%xmm0
@@ -57661,7 +58657,7 @@
   .byte  102,65,15,253,194                   // paddw         %xmm10,%xmm0
   .byte  102,68,15,213,215                   // pmullw        %xmm7,%xmm10
   .byte  102,68,15,213,227                   // pmullw        %xmm3,%xmm12
-  .byte  102,68,15,111,13,248,49,0,0         // movdqa        0x31f8(%rip),%xmm9        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,13,164,56,0,0         // movdqa        0x38a4(%rip),%xmm9        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,69,15,111,218                   // movdqa        %xmm10,%xmm11
   .byte  102,69,15,239,217                   // pxor          %xmm9,%xmm11
   .byte  102,65,15,111,204                   // movdqa        %xmm12,%xmm1
@@ -57670,7 +58666,7 @@
   .byte  102,68,15,219,225                   // pand          %xmm1,%xmm12
   .byte  102,65,15,223,202                   // pandn         %xmm10,%xmm1
   .byte  102,65,15,235,204                   // por           %xmm12,%xmm1
-  .byte  102,68,15,111,21,87,42,0,0          // movdqa        0x2a57(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,21,3,49,0,0           // movdqa        0x3103(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,202                   // paddw         %xmm10,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -57726,7 +58722,7 @@
   .byte  102,65,15,253,196                   // paddw         %xmm12,%xmm0
   .byte  102,68,15,213,231                   // pmullw        %xmm7,%xmm12
   .byte  102,68,15,213,211                   // pmullw        %xmm3,%xmm10
-  .byte  102,68,15,111,13,207,48,0,0         // movdqa        0x30cf(%rip),%xmm9        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,13,123,55,0,0         // movdqa        0x377b(%rip),%xmm9        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,69,15,111,220                   // movdqa        %xmm12,%xmm11
   .byte  102,69,15,239,217                   // pxor          %xmm9,%xmm11
   .byte  102,65,15,111,202                   // movdqa        %xmm10,%xmm1
@@ -57735,7 +58731,7 @@
   .byte  102,68,15,219,225                   // pand          %xmm1,%xmm12
   .byte  102,65,15,223,202                   // pandn         %xmm10,%xmm1
   .byte  102,65,15,235,204                   // por           %xmm12,%xmm1
-  .byte  102,68,15,111,21,46,41,0,0          // movdqa        0x292e(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,21,218,47,0,0         // movdqa        0x2fda(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,202                   // paddw         %xmm10,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -57791,7 +58787,7 @@
   .byte  102,65,15,253,196                   // paddw         %xmm12,%xmm0
   .byte  102,68,15,213,231                   // pmullw        %xmm7,%xmm12
   .byte  102,68,15,213,211                   // pmullw        %xmm3,%xmm10
-  .byte  102,68,15,111,13,166,47,0,0         // movdqa        0x2fa6(%rip),%xmm9        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,13,82,54,0,0          // movdqa        0x3652(%rip),%xmm9        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,69,15,111,220                   // movdqa        %xmm12,%xmm11
   .byte  102,69,15,239,217                   // pxor          %xmm9,%xmm11
   .byte  102,65,15,111,202                   // movdqa        %xmm10,%xmm1
@@ -57800,10 +58796,10 @@
   .byte  102,68,15,219,225                   // pand          %xmm1,%xmm12
   .byte  102,65,15,223,202                   // pandn         %xmm10,%xmm1
   .byte  102,65,15,235,204                   // por           %xmm12,%xmm1
-  .byte  102,68,15,111,21,5,40,0,0           // movdqa        0x2805(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,21,177,46,0,0         // movdqa        0x2eb1(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,202                   // paddw         %xmm10,%xmm1
   .byte  102,15,113,209,7                    // psrlw         $0x7,%xmm1
-  .byte  102,68,15,111,29,82,47,0,0          // movdqa        0x2f52(%rip),%xmm11        # 33ad0 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
+  .byte  102,68,15,111,29,254,53,0,0         // movdqa        0x35fe(%rip),%xmm11        # 351b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
   .byte  102,65,15,219,203                   // pand          %xmm11,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
   .byte  102,68,15,111,229                   // movdqa        %xmm5,%xmm12
@@ -57858,10 +58854,10 @@
   .byte  102,15,111,196                      // movdqa        %xmm4,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,204                      // pmullw        %xmm4,%xmm1
-  .byte  102,68,15,111,13,2,39,0,0           // movdqa        0x2702(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,174,45,0,0         // movdqa        0x2dae(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,201                   // paddw         %xmm9,%xmm1
   .byte  102,15,113,209,7                    // psrlw         $0x7,%xmm1
-  .byte  102,68,15,111,21,79,46,0,0          // movdqa        0x2e4f(%rip),%xmm10        # 33ad0 <_sk_srcover_rgba_8888_sse2_lowp+0x1150>
+  .byte  102,68,15,111,21,251,52,0,0         // movdqa        0x34fb(%rip),%xmm10        # 351b0 <_sk_srcover_rgba_8888_sse2_lowp+0x1180>
   .byte  102,65,15,219,202                   // pand          %xmm10,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
@@ -57897,7 +58893,7 @@
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,68,15,111,194                   // movdqa        %xmm2,%xmm8
   .byte  102,68,15,111,200                   // movdqa        %xmm0,%xmm9
-  .byte  102,15,111,21,94,38,0,0             // movdqa        0x265e(%rip),%xmm2        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,21,10,45,0,0             // movdqa        0x2d0a(%rip),%xmm2        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,111,226                   // movdqa        %xmm2,%xmm12
   .byte  102,68,15,249,231                   // psubw         %xmm7,%xmm12
   .byte  102,65,15,111,196                   // movdqa        %xmm12,%xmm0
@@ -57910,7 +58906,7 @@
   .byte  102,68,15,111,243                   // movdqa        %xmm3,%xmm14
   .byte  102,69,15,249,241                   // psubw         %xmm9,%xmm14
   .byte  102,69,15,253,201                   // paddw         %xmm9,%xmm9
-  .byte  102,15,111,21,138,45,0,0            // movdqa        0x2d8a(%rip),%xmm2        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,15,111,21,54,52,0,0             // movdqa        0x3436(%rip),%xmm2        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,15,111,199                      // movdqa        %xmm7,%xmm0
   .byte  102,68,15,111,215                   // movdqa        %xmm7,%xmm10
   .byte  102,15,111,231                      // movdqa        %xmm7,%xmm4
@@ -57929,7 +58925,7 @@
   .byte  102,68,15,219,232                   // pand          %xmm0,%xmm13
   .byte  102,65,15,223,193                   // pandn         %xmm9,%xmm0
   .byte  102,65,15,235,197                   // por           %xmm13,%xmm0
-  .byte  102,15,253,5,189,37,0,0             // paddw         0x25bd(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,253,5,105,44,0,0             // paddw         0x2c69(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,199                   // paddw         %xmm15,%xmm0
   .byte  102,69,15,111,204                   // movdqa        %xmm12,%xmm9
   .byte  102,68,15,213,201                   // pmullw        %xmm1,%xmm9
@@ -57952,7 +58948,7 @@
   .byte  102,65,15,219,249                   // pand          %xmm9,%xmm7
   .byte  102,68,15,223,201                   // pandn         %xmm1,%xmm9
   .byte  102,68,15,235,207                   // por           %xmm7,%xmm9
-  .byte  102,68,15,253,13,72,37,0,0          // paddw         0x2548(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,253,13,244,43,0,0         // paddw         0x2bf4(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,253,205                   // paddw         %xmm13,%xmm9
   .byte  102,69,15,213,224                   // pmullw        %xmm8,%xmm12
   .byte  102,69,15,111,251                   // movdqa        %xmm11,%xmm15
@@ -57974,7 +58970,7 @@
   .byte  102,68,15,219,210                   // pand          %xmm2,%xmm10
   .byte  102,65,15,223,208                   // pandn         %xmm8,%xmm2
   .byte  102,65,15,235,210                   // por           %xmm10,%xmm2
-  .byte  102,68,15,111,5,216,36,0,0          // movdqa        0x24d8(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,132,43,0,0          // movdqa        0x2b84(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,208                   // paddw         %xmm8,%xmm2
   .byte  102,65,15,253,215                   // paddw         %xmm15,%xmm2
   .byte  102,68,15,213,220                   // pmullw        %xmm4,%xmm11
@@ -58002,7 +58998,7 @@
   .byte  102,15,127,124,36,232               // movdqa        %xmm7,-0x18(%rsp)
   .byte  102,68,15,111,193                   // movdqa        %xmm1,%xmm8
   .byte  102,68,15,111,248                   // movdqa        %xmm0,%xmm15
-  .byte  102,15,111,53,105,36,0,0            // movdqa        0x2469(%rip),%xmm6        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,53,21,43,0,0             // movdqa        0x2b15(%rip),%xmm6        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,111,222                   // movdqa        %xmm6,%xmm11
   .byte  102,69,15,249,222                   // psubw         %xmm14,%xmm11
   .byte  102,65,15,111,203                   // movdqa        %xmm11,%xmm1
@@ -58024,7 +59020,7 @@
   .byte  102,15,253,201                      // paddw         %xmm1,%xmm1
   .byte  102,69,15,111,225                   // movdqa        %xmm9,%xmm12
   .byte  102,68,15,249,225                   // psubw         %xmm1,%xmm12
-  .byte  102,15,111,13,109,43,0,0            // movdqa        0x2b6d(%rip),%xmm1        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,15,111,13,25,50,0,0             // movdqa        0x3219(%rip),%xmm1        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,69,15,111,238                   // movdqa        %xmm14,%xmm13
   .byte  102,68,15,239,233                   // pxor          %xmm1,%xmm13
   .byte  102,15,239,249                      // pxor          %xmm1,%xmm7
@@ -58051,7 +59047,7 @@
   .byte  102,68,15,213,197                   // pmullw        %xmm5,%xmm8
   .byte  102,15,111,229                      // movdqa        %xmm5,%xmm4
   .byte  102,15,253,228                      // paddw         %xmm4,%xmm4
-  .byte  102,15,111,61,236,42,0,0            // movdqa        0x2aec(%rip),%xmm7        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,15,111,61,152,49,0,0            // movdqa        0x3198(%rip),%xmm7        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,15,239,231                      // pxor          %xmm7,%xmm4
   .byte  102,65,15,101,229                   // pcmpgtw       %xmm13,%xmm4
   .byte  102,69,15,253,192                   // paddw         %xmm8,%xmm8
@@ -58111,27 +59107,27 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            3110a <_sk_load_8888_sse2_lowp+0x44>
+  .byte  119,41                              // ja            3213e <_sk_load_8888_sse2_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,24,1,0,0                  // lea           0x118(%rip),%r9        # 31208 <_sk_load_8888_sse2_lowp+0x142>
+  .byte  76,141,13,24,1,0,0                  // lea           0x118(%rip),%r9        # 3223c <_sk_load_8888_sse2_lowp+0x142>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  255,224                             // jmpq          *%rax
   .byte  243,65,15,16,28,144                 // movss         (%r8,%rdx,4),%xmm3
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,102                             // jmp           31170 <_sk_load_8888_sse2_lowp+0xaa>
+  .byte  235,102                             // jmp           321a4 <_sk_load_8888_sse2_lowp+0xaa>
   .byte  102,65,15,16,28,144                 // movupd        (%r8,%rdx,4),%xmm3
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,87                              // jmp           31170 <_sk_load_8888_sse2_lowp+0xaa>
+  .byte  235,87                              // jmp           321a4 <_sk_load_8888_sse2_lowp+0xaa>
   .byte  102,65,15,110,68,144,8              // movd          0x8(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,69,15,18,4,144                  // movlpd        (%r8,%rdx,4),%xmm8
   .byte  102,65,15,40,216                    // movapd        %xmm8,%xmm3
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
-  .byte  235,52                              // jmp           31170 <_sk_load_8888_sse2_lowp+0xaa>
+  .byte  235,52                              // jmp           321a4 <_sk_load_8888_sse2_lowp+0xaa>
   .byte  102,65,15,110,68,144,24             // movd          0x18(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  243,65,15,16,68,144,20              // movss         0x14(%r8,%rdx,4),%xmm0
@@ -58148,7 +59144,7 @@
   .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
   .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
   .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
-  .byte  102,68,15,111,13,214,33,0,0         // movdqa        0x21d6(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,130,40,0,0         // movdqa        0x2882(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
@@ -58204,27 +59200,27 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            31268 <_sk_load_8888_dst_sse2_lowp+0x44>
+  .byte  119,41                              // ja            3229c <_sk_load_8888_dst_sse2_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 31364 <_sk_load_8888_dst_sse2_lowp+0x140>
+  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 32398 <_sk_load_8888_dst_sse2_lowp+0x140>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  255,224                             // jmpq          *%rax
   .byte  243,65,15,16,60,144                 // movss         (%r8,%rdx,4),%xmm7
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,102                             // jmp           312ce <_sk_load_8888_dst_sse2_lowp+0xaa>
+  .byte  235,102                             // jmp           32302 <_sk_load_8888_dst_sse2_lowp+0xaa>
   .byte  102,65,15,16,60,144                 // movupd        (%r8,%rdx,4),%xmm7
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,87                              // jmp           312ce <_sk_load_8888_dst_sse2_lowp+0xaa>
+  .byte  235,87                              // jmp           32302 <_sk_load_8888_dst_sse2_lowp+0xaa>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,69,15,18,4,144                  // movlpd        (%r8,%rdx,4),%xmm8
   .byte  102,65,15,40,248                    // movapd        %xmm8,%xmm7
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
-  .byte  235,52                              // jmp           312ce <_sk_load_8888_dst_sse2_lowp+0xaa>
+  .byte  235,52                              // jmp           32302 <_sk_load_8888_dst_sse2_lowp+0xaa>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  243,65,15,16,100,144,20             // movss         0x14(%r8,%rdx,4),%xmm4
@@ -58241,7 +59237,7 @@
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
   .byte  102,15,114,228,16                   // psrad         $0x10,%xmm4
   .byte  102,15,107,236                      // packssdw      %xmm4,%xmm5
-  .byte  102,68,15,111,13,120,32,0,0         // movdqa        0x2078(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,36,39,0,0          // movdqa        0x2724(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,229                      // movdqa        %xmm5,%xmm4
   .byte  102,65,15,219,225                   // pand          %xmm9,%xmm4
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
@@ -58309,21 +59305,21 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,27                              // ja            31403 <_sk_store_8888_sse2_lowp+0x83>
+  .byte  119,27                              // ja            32437 <_sk_store_8888_sse2_lowp+0x83>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,98,0,0,0                  // lea           0x62(%rip),%r9        # 31454 <_sk_store_8888_sse2_lowp+0xd4>
+  .byte  76,141,13,98,0,0,0                  // lea           0x62(%rip),%r9        # 32488 <_sk_store_8888_sse2_lowp+0xd4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,126,12,144                // movd          %xmm9,(%r8,%rdx,4)
-  .byte  235,75                              // jmp           3144e <_sk_store_8888_sse2_lowp+0xce>
+  .byte  235,75                              // jmp           32482 <_sk_store_8888_sse2_lowp+0xce>
   .byte  243,69,15,127,12,144                // movdqu        %xmm9,(%r8,%rdx,4)
   .byte  243,69,15,127,68,144,16             // movdqu        %xmm8,0x10(%r8,%rdx,4)
-  .byte  235,60                              // jmp           3144e <_sk_store_8888_sse2_lowp+0xce>
+  .byte  235,60                              // jmp           32482 <_sk_store_8888_sse2_lowp+0xce>
   .byte  102,69,15,112,193,78                // pshufd        $0x4e,%xmm9,%xmm8
   .byte  102,69,15,126,68,144,8              // movd          %xmm8,0x8(%r8,%rdx,4)
   .byte  102,69,15,214,12,144                // movq          %xmm9,(%r8,%rdx,4)
-  .byte  235,39                              // jmp           3144e <_sk_store_8888_sse2_lowp+0xce>
+  .byte  235,39                              // jmp           32482 <_sk_store_8888_sse2_lowp+0xce>
   .byte  102,69,15,112,208,78                // pshufd        $0x4e,%xmm8,%xmm10
   .byte  102,69,15,126,84,144,24             // movd          %xmm10,0x18(%r8,%rdx,4)
   .byte  102,69,15,112,208,229               // pshufd        $0xe5,%xmm8,%xmm10
@@ -58368,27 +59364,27 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            314b4 <_sk_load_bgra_sse2_lowp+0x44>
+  .byte  119,41                              // ja            324e8 <_sk_load_bgra_sse2_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 315b0 <_sk_load_bgra_sse2_lowp+0x140>
+  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 325e4 <_sk_load_bgra_sse2_lowp+0x140>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  255,224                             // jmpq          *%rax
   .byte  243,65,15,16,28,144                 // movss         (%r8,%rdx,4),%xmm3
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,102                             // jmp           3151a <_sk_load_bgra_sse2_lowp+0xaa>
+  .byte  235,102                             // jmp           3254e <_sk_load_bgra_sse2_lowp+0xaa>
   .byte  102,65,15,16,28,144                 // movupd        (%r8,%rdx,4),%xmm3
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,87                              // jmp           3151a <_sk_load_bgra_sse2_lowp+0xaa>
+  .byte  235,87                              // jmp           3254e <_sk_load_bgra_sse2_lowp+0xaa>
   .byte  102,65,15,110,68,144,8              // movd          0x8(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,69,15,18,4,144                  // movlpd        (%r8,%rdx,4),%xmm8
   .byte  102,65,15,40,216                    // movapd        %xmm8,%xmm3
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
-  .byte  235,52                              // jmp           3151a <_sk_load_bgra_sse2_lowp+0xaa>
+  .byte  235,52                              // jmp           3254e <_sk_load_bgra_sse2_lowp+0xaa>
   .byte  102,65,15,110,68,144,24             // movd          0x18(%r8,%rdx,4),%xmm0
   .byte  102,68,15,112,192,69                // pshufd        $0x45,%xmm0,%xmm8
   .byte  243,65,15,16,68,144,20              // movss         0x14(%r8,%rdx,4),%xmm0
@@ -58405,7 +59401,7 @@
   .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
   .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
   .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
-  .byte  102,68,15,111,13,44,30,0,0          // movdqa        0x1e2c(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,216,36,0,0         // movdqa        0x24d8(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,65,15,219,209                   // pand          %xmm9,%xmm2
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
@@ -58458,27 +59454,27 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,41                              // ja            31610 <_sk_load_bgra_dst_sse2_lowp+0x44>
+  .byte  119,41                              // ja            32644 <_sk_load_bgra_dst_sse2_lowp+0x44>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 3170c <_sk_load_bgra_dst_sse2_lowp+0x140>
+  .byte  76,141,13,22,1,0,0                  // lea           0x116(%rip),%r9        # 32740 <_sk_load_bgra_dst_sse2_lowp+0x140>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  255,224                             // jmpq          *%rax
   .byte  243,65,15,16,60,144                 // movss         (%r8,%rdx,4),%xmm7
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,102                             // jmp           31676 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  .byte  235,102                             // jmp           326aa <_sk_load_bgra_dst_sse2_lowp+0xaa>
   .byte  102,65,15,16,60,144                 // movupd        (%r8,%rdx,4),%xmm7
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,87                              // jmp           31676 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  .byte  235,87                              // jmp           326aa <_sk_load_bgra_dst_sse2_lowp+0xaa>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,69,15,18,4,144                  // movlpd        (%r8,%rdx,4),%xmm8
   .byte  102,65,15,40,248                    // movapd        %xmm8,%xmm7
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
-  .byte  235,52                              // jmp           31676 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  .byte  235,52                              // jmp           326aa <_sk_load_bgra_dst_sse2_lowp+0xaa>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  243,65,15,16,100,144,20             // movss         0x14(%r8,%rdx,4),%xmm4
@@ -58495,7 +59491,7 @@
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
   .byte  102,15,114,228,16                   // psrad         $0x10,%xmm4
   .byte  102,15,107,236                      // packssdw      %xmm4,%xmm5
-  .byte  102,68,15,111,13,208,28,0,0         // movdqa        0x1cd0(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,124,35,0,0         // movdqa        0x237c(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,245                      // movdqa        %xmm5,%xmm6
   .byte  102,65,15,219,241                   // pand          %xmm9,%xmm6
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
@@ -58563,21 +59559,21 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,27                              // ja            317ab <_sk_store_bgra_sse2_lowp+0x83>
+  .byte  119,27                              // ja            327df <_sk_store_bgra_sse2_lowp+0x83>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,98,0,0,0                  // lea           0x62(%rip),%r9        # 317fc <_sk_store_bgra_sse2_lowp+0xd4>
+  .byte  76,141,13,98,0,0,0                  // lea           0x62(%rip),%r9        # 32830 <_sk_store_bgra_sse2_lowp+0xd4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,126,12,144                // movd          %xmm9,(%r8,%rdx,4)
-  .byte  235,75                              // jmp           317f6 <_sk_store_bgra_sse2_lowp+0xce>
+  .byte  235,75                              // jmp           3282a <_sk_store_bgra_sse2_lowp+0xce>
   .byte  243,69,15,127,12,144                // movdqu        %xmm9,(%r8,%rdx,4)
   .byte  243,69,15,127,68,144,16             // movdqu        %xmm8,0x10(%r8,%rdx,4)
-  .byte  235,60                              // jmp           317f6 <_sk_store_bgra_sse2_lowp+0xce>
+  .byte  235,60                              // jmp           3282a <_sk_store_bgra_sse2_lowp+0xce>
   .byte  102,69,15,112,193,78                // pshufd        $0x4e,%xmm9,%xmm8
   .byte  102,69,15,126,68,144,8              // movd          %xmm8,0x8(%r8,%rdx,4)
   .byte  102,69,15,214,12,144                // movq          %xmm9,(%r8,%rdx,4)
-  .byte  235,39                              // jmp           317f6 <_sk_store_bgra_sse2_lowp+0xce>
+  .byte  235,39                              // jmp           3282a <_sk_store_bgra_sse2_lowp+0xce>
   .byte  102,69,15,112,208,78                // pshufd        $0x4e,%xmm8,%xmm10
   .byte  102,69,15,126,84,144,24             // movd          %xmm10,0x18(%r8,%rdx,4)
   .byte  102,69,15,112,208,229               // pshufd        $0xe5,%xmm8,%xmm10
@@ -58691,7 +59687,7 @@
   .byte  102,15,114,241,16                   // pslld         $0x10,%xmm1
   .byte  102,15,114,225,16                   // psrad         $0x10,%xmm1
   .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
-  .byte  102,68,15,111,13,236,25,0,0         // movdqa        0x19ec(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,152,32,0,0         // movdqa        0x2098(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
@@ -58719,6 +59715,116 @@
   .byte  65,95                               // pop           %r15
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_bgra_sse2_lowp
+.globl _sk_gather_bgra_sse2_lowp
+FUNCTION(_sk_gather_bgra_sse2_lowp)
+_sk_gather_bgra_sse2_lowp:
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  243,68,15,91,195                    // cvttps2dq     %xmm3,%xmm8
+  .byte  243,68,15,91,202                    // cvttps2dq     %xmm2,%xmm9
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,69,15,112,209,245               // pshufd        $0xf5,%xmm9,%xmm10
+  .byte  102,68,15,244,211                   // pmuludq       %xmm3,%xmm10
+  .byte  102,68,15,111,219                   // movdqa        %xmm3,%xmm11
+  .byte  102,69,15,112,224,245               // pshufd        $0xf5,%xmm8,%xmm12
+  .byte  102,68,15,244,227                   // pmuludq       %xmm3,%xmm12
+  .byte  102,65,15,244,217                   // pmuludq       %xmm9,%xmm3
+  .byte  72,139,0                            // mov           (%rax),%rax
+  .byte  102,15,112,219,232                  // pshufd        $0xe8,%xmm3,%xmm3
+  .byte  102,65,15,112,210,232               // pshufd        $0xe8,%xmm10,%xmm2
+  .byte  102,15,98,218                       // punpckldq     %xmm2,%xmm3
+  .byte  102,69,15,244,216                   // pmuludq       %xmm8,%xmm11
+  .byte  102,69,15,112,195,232               // pshufd        $0xe8,%xmm11,%xmm8
+  .byte  102,65,15,112,212,232               // pshufd        $0xe8,%xmm12,%xmm2
+  .byte  102,68,15,98,194                    // punpckldq     %xmm2,%xmm8
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,200                   // paddd         %xmm8,%xmm1
+  .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
+  .byte  102,73,15,126,208                   // movq          %xmm2,%r8
+  .byte  69,137,193                          // mov           %r8d,%r9d
+  .byte  73,193,232,32                       // shr           $0x20,%r8
+  .byte  102,73,15,126,202                   // movq          %xmm1,%r10
+  .byte  69,137,211                          // mov           %r10d,%r11d
+  .byte  73,193,234,32                       // shr           $0x20,%r10
+  .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
+  .byte  102,72,15,126,203                   // movq          %xmm1,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  102,66,15,110,4,184                 // movd          (%rax,%r15,4),%xmm0
+  .byte  102,66,15,110,28,160                // movd          (%rax,%r12,4),%xmm3
+  .byte  102,15,98,216                       // punpckldq     %xmm0,%xmm3
+  .byte  102,15,110,4,152                    // movd          (%rax,%rbx,4),%xmm0
+  .byte  102,66,15,110,12,176                // movd          (%rax,%r14,4),%xmm1
+  .byte  102,15,98,200                       // punpckldq     %xmm0,%xmm1
+  .byte  102,15,108,217                      // punpcklqdq    %xmm1,%xmm3
+  .byte  102,66,15,110,4,144                 // movd          (%rax,%r10,4),%xmm0
+  .byte  102,70,15,110,4,152                 // movd          (%rax,%r11,4),%xmm8
+  .byte  102,68,15,98,192                    // punpckldq     %xmm0,%xmm8
+  .byte  102,66,15,110,4,128                 // movd          (%rax,%r8,4),%xmm0
+  .byte  102,66,15,110,12,136                // movd          (%rax,%r9,4),%xmm1
+  .byte  102,15,98,200                       // punpckldq     %xmm0,%xmm1
+  .byte  102,68,15,108,193                   // punpcklqdq    %xmm1,%xmm8
+  .byte  102,65,15,111,192                   // movdqa        %xmm8,%xmm0
+  .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
+  .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
+  .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
+  .byte  102,15,114,241,16                   // pslld         $0x10,%xmm1
+  .byte  102,15,114,225,16                   // psrad         $0x10,%xmm1
+  .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
+  .byte  102,68,15,111,13,185,30,0,0         // movdqa        0x1eb9(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
+  .byte  102,65,15,219,209                   // pand          %xmm9,%xmm2
+  .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
+  .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
+  .byte  102,15,114,208,16                   // psrld         $0x10,%xmm0
+  .byte  102,69,15,111,208                   // movdqa        %xmm8,%xmm10
+  .byte  102,65,15,114,210,16                // psrld         $0x10,%xmm10
+  .byte  102,65,15,114,208,24                // psrld         $0x18,%xmm8
+  .byte  102,65,15,114,240,16                // pslld         $0x10,%xmm8
+  .byte  102,65,15,114,224,16                // psrad         $0x10,%xmm8
+  .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
+  .byte  102,15,114,243,16                   // pslld         $0x10,%xmm3
+  .byte  102,15,114,227,16                   // psrad         $0x10,%xmm3
+  .byte  102,65,15,107,216                   // packssdw      %xmm8,%xmm3
+  .byte  102,65,15,114,242,16                // pslld         $0x10,%xmm10
+  .byte  102,65,15,114,226,16                // psrad         $0x10,%xmm10
+  .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
+  .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
+  .byte  102,65,15,107,194                   // packssdw      %xmm10,%xmm0
+  .byte  102,65,15,219,193                   // pand          %xmm9,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_565_sse2_lowp
 .globl _sk_load_565_sse2_lowp
 FUNCTION(_sk_load_565_sse2_lowp)
@@ -58732,23 +59838,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            31a33 <_sk_load_565_sse2_lowp+0x3c>
+  .byte  119,34                              // ja            32c46 <_sk_load_565_sse2_lowp+0x3c>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,189,0,0,0                 // lea           0xbd(%rip),%r9        # 31adc <_sk_load_565_sse2_lowp+0xe5>
+  .byte  76,141,13,190,0,0,0                 // lea           0xbe(%rip),%r9        # 32cf0 <_sk_load_565_sse2_lowp+0xe6>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,192                      // movd          %eax,%xmm0
-  .byte  235,66                              // jmp           31a75 <_sk_load_565_sse2_lowp+0x7e>
+  .byte  235,66                              // jmp           32c88 <_sk_load_565_sse2_lowp+0x7e>
   .byte  243,65,15,111,4,80                  // movdqu        (%r8,%rdx,2),%xmm0
-  .byte  235,58                              // jmp           31a75 <_sk_load_565_sse2_lowp+0x7e>
+  .byte  235,58                              // jmp           32c88 <_sk_load_565_sse2_lowp+0x7e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,4,2             // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   .byte  243,65,15,16,12,80                  // movss         (%r8,%rdx,2),%xmm1
   .byte  243,15,16,193                       // movss         %xmm1,%xmm0
-  .byte  235,34                              // jmp           31a75 <_sk_load_565_sse2_lowp+0x7e>
+  .byte  235,34                              // jmp           32c88 <_sk_load_565_sse2_lowp+0x7e>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,65,15,196,68,80,12,6            // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   .byte  102,65,15,196,68,80,10,5            // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -58756,11 +59862,11 @@
   .byte  102,65,15,18,4,80                   // movlpd        (%r8,%rdx,2),%xmm0
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
   .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
-  .byte  102,15,219,29,154,32,0,0            // pand          0x209a(%rip),%xmm3        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,29,103,37,0,0            // pand          0x2567(%rip),%xmm3        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
-  .byte  102,15,219,13,153,32,0,0            // pand          0x2099(%rip),%xmm1        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,15,111,21,161,32,0,0            // movdqa        0x20a1(%rip),%xmm2        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,15,219,13,102,37,0,0            // pand          0x2566(%rip),%xmm1        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,21,110,37,0,0            // movdqa        0x256e(%rip),%xmm2        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,15,219,208                      // pand          %xmm0,%xmm2
   .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
   .byte  102,15,235,195                      // por           %xmm3,%xmm0
@@ -58773,19 +59879,19 @@
   .byte  102,15,113,210,2                    // psrlw         $0x2,%xmm2
   .byte  102,15,235,211                      // por           %xmm3,%xmm2
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,151,24,0,0                 // movaps        0x1897(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,100,29,0,0                 // movaps        0x1d64(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  255,224                             // jmpq          *%rax
-  .byte  144                                 // nop
-  .byte  76,255                              // rex.WR        (bad)
+  .byte  102,144                             // xchg          %ax,%ax
+  .byte  75,255                              // rex.WXB       (bad)
   .byte  255                                 // (bad)
-  .byte  255,107,255                         // ljmp          *-0x1(%rbx)
+  .byte  255,106,255                         // ljmp          *-0x1(%rdx)
   .byte  255                                 // (bad)
-  .byte  255,95,255                          // lcall         *-0x1(%rdi)
+  .byte  255,94,255                          // lcall         *-0x1(%rsi)
   .byte  255                                 // (bad)
-  .byte  255,147,255,255,255,139             // callq         *-0x74000001(%rbx)
+  .byte  255,146,255,255,255,138             // callq         *-0x75000001(%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,131,255,255,255,119             // incl          0x77ffffff(%rbx)
+  .byte  255,130,255,255,255,118             // incl          0x76ffffff(%rdx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
@@ -58803,23 +59909,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            31b34 <_sk_load_565_dst_sse2_lowp+0x3c>
+  .byte  119,34                              // ja            32d48 <_sk_load_565_dst_sse2_lowp+0x3c>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,188,0,0,0                 // lea           0xbc(%rip),%r9        # 31bdc <_sk_load_565_dst_sse2_lowp+0xe4>
+  .byte  76,141,13,188,0,0,0                 // lea           0xbc(%rip),%r9        # 32df0 <_sk_load_565_dst_sse2_lowp+0xe4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,224                      // movd          %eax,%xmm4
-  .byte  235,66                              // jmp           31b76 <_sk_load_565_dst_sse2_lowp+0x7e>
+  .byte  235,66                              // jmp           32d8a <_sk_load_565_dst_sse2_lowp+0x7e>
   .byte  243,65,15,111,36,80                 // movdqu        (%r8,%rdx,2),%xmm4
-  .byte  235,58                              // jmp           31b76 <_sk_load_565_dst_sse2_lowp+0x7e>
+  .byte  235,58                              // jmp           32d8a <_sk_load_565_dst_sse2_lowp+0x7e>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,65,15,196,100,80,4,2            // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm4
   .byte  243,65,15,16,44,80                  // movss         (%r8,%rdx,2),%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,34                              // jmp           31b76 <_sk_load_565_dst_sse2_lowp+0x7e>
+  .byte  235,34                              // jmp           32d8a <_sk_load_565_dst_sse2_lowp+0x7e>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,65,15,196,100,80,12,6           // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm4
   .byte  102,65,15,196,100,80,10,5           // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm4
@@ -58827,11 +59933,11 @@
   .byte  102,65,15,18,36,80                  // movlpd        (%r8,%rdx,2),%xmm4
   .byte  102,15,111,252                      // movdqa        %xmm4,%xmm7
   .byte  102,15,113,215,8                    // psrlw         $0x8,%xmm7
-  .byte  102,15,219,61,153,31,0,0            // pand          0x1f99(%rip),%xmm7        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,15,219,61,101,36,0,0            // pand          0x2465(%rip),%xmm7        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,113,213,5                    // psrlw         $0x5,%xmm5
-  .byte  102,15,219,45,152,31,0,0            // pand          0x1f98(%rip),%xmm5        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,15,111,53,160,31,0,0            // movdqa        0x1fa0(%rip),%xmm6        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,15,219,45,100,36,0,0            // pand          0x2464(%rip),%xmm5        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,53,108,36,0,0            // movdqa        0x246c(%rip),%xmm6        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,15,219,244                      // pand          %xmm4,%xmm6
   .byte  102,15,113,212,13                   // psrlw         $0xd,%xmm4
   .byte  102,15,235,231                      // por           %xmm7,%xmm4
@@ -58844,7 +59950,7 @@
   .byte  102,15,113,214,2                    // psrlw         $0x2,%xmm6
   .byte  102,15,235,247                      // por           %xmm7,%xmm6
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,150,23,0,0                 // movaps        0x1796(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,61,98,28,0,0                  // movaps        0x1c62(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  255,224                             // jmpq          *%rax
   .byte  77,255                              // rex.WRB       (bad)
   .byte  255                                 // (bad)
@@ -58868,10 +59974,10 @@
   .byte  76,3,0                              // add           (%rax),%r8
   .byte  102,68,15,111,192                   // movdqa        %xmm0,%xmm8
   .byte  102,65,15,113,240,8                 // psllw         $0x8,%xmm8
-  .byte  102,68,15,219,5,52,31,0,0           // pand          0x1f34(%rip),%xmm8        # 33b50 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
+  .byte  102,68,15,219,5,0,36,0,0            // pand          0x2400(%rip),%xmm8        # 35230 <_sk_srcover_rgba_8888_sse2_lowp+0x1200>
   .byte  102,68,15,111,201                   // movdqa        %xmm1,%xmm9
   .byte  102,65,15,113,241,3                 // psllw         $0x3,%xmm9
-  .byte  102,68,15,219,13,48,31,0,0          // pand          0x1f30(%rip),%xmm9        # 33b60 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,219,13,252,35,0,0         // pand          0x23fc(%rip),%xmm9        # 35240 <_sk_srcover_rgba_8888_sse2_lowp+0x1210>
   .byte  102,69,15,235,200                   // por           %xmm8,%xmm9
   .byte  102,68,15,111,194                   // movdqa        %xmm2,%xmm8
   .byte  102,65,15,113,208,3                 // psrlw         $0x3,%xmm8
@@ -58880,21 +59986,21 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,31                              // ja            31c6e <_sk_store_565_sse2_lowp+0x76>
+  .byte  119,31                              // ja            32e82 <_sk_store_565_sse2_lowp+0x76>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,95,0,0,0                  // lea           0x5f(%rip),%r9        # 31cb8 <_sk_store_565_sse2_lowp+0xc0>
+  .byte  76,141,13,95,0,0,0                  // lea           0x5f(%rip),%r9        # 32ecc <_sk_store_565_sse2_lowp+0xc0>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,68,15,126,192                   // movd          %xmm8,%eax
   .byte  102,65,137,4,80                     // mov           %ax,(%r8,%rdx,2)
-  .byte  235,70                              // jmp           31cb4 <_sk_store_565_sse2_lowp+0xbc>
+  .byte  235,70                              // jmp           32ec8 <_sk_store_565_sse2_lowp+0xbc>
   .byte  243,69,15,127,4,80                  // movdqu        %xmm8,(%r8,%rdx,2)
-  .byte  235,62                              // jmp           31cb4 <_sk_store_565_sse2_lowp+0xbc>
+  .byte  235,62                              // jmp           32ec8 <_sk_store_565_sse2_lowp+0xbc>
   .byte  102,65,15,197,192,2                 // pextrw        $0x2,%xmm8,%eax
   .byte  102,65,137,68,80,4                  // mov           %ax,0x4(%r8,%rdx,2)
   .byte  102,69,15,126,4,80                  // movd          %xmm8,(%r8,%rdx,2)
-  .byte  235,42                              // jmp           31cb4 <_sk_store_565_sse2_lowp+0xbc>
+  .byte  235,42                              // jmp           32ec8 <_sk_store_565_sse2_lowp+0xbc>
   .byte  102,65,15,197,192,6                 // pextrw        $0x6,%xmm8,%eax
   .byte  102,65,137,68,80,12                 // mov           %ax,0xc(%r8,%rdx,2)
   .byte  102,65,15,197,192,5                 // pextrw        $0x5,%xmm8,%eax
@@ -58926,6 +60032,118 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_565_sse2_lowp
+.globl _sk_gather_565_sse2_lowp
+FUNCTION(_sk_gather_565_sse2_lowp)
+_sk_gather_565_sse2_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,64,12                  // movss         0xc(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,118,201                   // pcmpeqd       %xmm9,%xmm9
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  69,15,87,210                        // xorps         %xmm10,%xmm10
+  .byte  65,15,95,202                        // maxps         %xmm10,%xmm1
+  .byte  65,15,95,194                        // maxps         %xmm10,%xmm0
+  .byte  65,15,93,192                        // minps         %xmm8,%xmm0
+  .byte  65,15,93,200                        // minps         %xmm8,%xmm1
+  .byte  243,68,15,16,64,16                  // movss         0x10(%rax),%xmm8
+  .byte  69,15,198,192,0                     // shufps        $0x0,%xmm8,%xmm8
+  .byte  102,69,15,254,193                   // paddd         %xmm9,%xmm8
+  .byte  65,15,95,210                        // maxps         %xmm10,%xmm2
+  .byte  65,15,95,218                        // maxps         %xmm10,%xmm3
+  .byte  65,15,93,216                        // minps         %xmm8,%xmm3
+  .byte  65,15,93,208                        // minps         %xmm8,%xmm2
+  .byte  243,68,15,91,194                    // cvttps2dq     %xmm2,%xmm8
+  .byte  243,68,15,91,203                    // cvttps2dq     %xmm3,%xmm9
+  .byte  102,15,110,80,8                     // movd          0x8(%rax),%xmm2
+  .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
+  .byte  102,69,15,112,209,245               // pshufd        $0xf5,%xmm9,%xmm10
+  .byte  102,68,15,244,210                   // pmuludq       %xmm2,%xmm10
+  .byte  102,68,15,111,218                   // movdqa        %xmm2,%xmm11
+  .byte  102,69,15,112,224,245               // pshufd        $0xf5,%xmm8,%xmm12
+  .byte  102,68,15,244,226                   // pmuludq       %xmm2,%xmm12
+  .byte  102,65,15,244,209                   // pmuludq       %xmm9,%xmm2
+  .byte  76,139,16                           // mov           (%rax),%r10
+  .byte  102,15,112,210,232                  // pshufd        $0xe8,%xmm2,%xmm2
+  .byte  102,65,15,112,218,232               // pshufd        $0xe8,%xmm10,%xmm3
+  .byte  102,15,98,211                       // punpckldq     %xmm3,%xmm2
+  .byte  102,69,15,244,216                   // pmuludq       %xmm8,%xmm11
+  .byte  102,69,15,112,195,232               // pshufd        $0xe8,%xmm11,%xmm8
+  .byte  102,65,15,112,220,232               // pshufd        $0xe8,%xmm12,%xmm3
+  .byte  102,68,15,98,195                    // punpckldq     %xmm3,%xmm8
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,15,254,202                      // paddd         %xmm2,%xmm1
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,65,15,254,192                   // paddd         %xmm8,%xmm0
+  .byte  102,15,112,208,78                   // pshufd        $0x4e,%xmm0,%xmm2
+  .byte  102,73,15,126,208                   // movq          %xmm2,%r8
+  .byte  69,137,193                          // mov           %r8d,%r9d
+  .byte  73,193,232,32                       // shr           $0x20,%r8
+  .byte  102,72,15,126,192                   // movq          %xmm0,%rax
+  .byte  65,137,195                          // mov           %eax,%r11d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,15,112,193,78                   // pshufd        $0x4e,%xmm1,%xmm0
+  .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,73,15,126,207                   // movq          %xmm1,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  67,15,183,44,122                    // movzwl        (%r10,%r15,2),%ebp
+  .byte  102,15,110,197                      // movd          %ebp,%xmm0
+  .byte  67,15,183,44,98                     // movzwl        (%r10,%r12,2),%ebp
+  .byte  102,15,110,205                      // movd          %ebp,%xmm1
+  .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
+  .byte  65,15,183,28,90                     // movzwl        (%r10,%rbx,2),%ebx
+  .byte  102,15,110,195                      // movd          %ebx,%xmm0
+  .byte  67,15,183,28,114                    // movzwl        (%r10,%r14,2),%ebx
+  .byte  102,15,110,211                      // movd          %ebx,%xmm2
+  .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
+  .byte  102,15,98,202                       // punpckldq     %xmm2,%xmm1
+  .byte  65,15,183,4,66                      // movzwl        (%r10,%rax,2),%eax
+  .byte  102,15,110,208                      // movd          %eax,%xmm2
+  .byte  67,15,183,4,90                      // movzwl        (%r10,%r11,2),%eax
+  .byte  102,15,110,192                      // movd          %eax,%xmm0
+  .byte  102,15,97,194                       // punpcklwd     %xmm2,%xmm0
+  .byte  67,15,183,4,66                      // movzwl        (%r10,%r8,2),%eax
+  .byte  102,15,110,208                      // movd          %eax,%xmm2
+  .byte  67,15,183,4,74                      // movzwl        (%r10,%r9,2),%eax
+  .byte  102,15,110,216                      // movd          %eax,%xmm3
+  .byte  102,15,97,218                       // punpcklwd     %xmm2,%xmm3
+  .byte  102,15,98,195                       // punpckldq     %xmm3,%xmm0
+  .byte  102,15,108,193                      // punpcklqdq    %xmm1,%xmm0
+  .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
+  .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
+  .byte  102,15,219,29,169,33,0,0            // pand          0x21a9(%rip),%xmm3        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
+  .byte  102,15,219,13,168,33,0,0            // pand          0x21a8(%rip),%xmm1        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,15,111,21,176,33,0,0            // movdqa        0x21b0(%rip),%xmm2        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
+  .byte  102,15,219,208                      // pand          %xmm0,%xmm2
+  .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
+  .byte  102,15,235,195                      // por           %xmm3,%xmm0
+  .byte  102,15,111,217                      // movdqa        %xmm1,%xmm3
+  .byte  102,15,113,243,2                    // psllw         $0x2,%xmm3
+  .byte  102,15,113,209,4                    // psrlw         $0x4,%xmm1
+  .byte  102,15,235,203                      // por           %xmm3,%xmm1
+  .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
+  .byte  102,15,113,243,3                    // psllw         $0x3,%xmm3
+  .byte  102,15,113,210,2                    // psrlw         $0x2,%xmm2
+  .byte  102,15,235,211                      // por           %xmm3,%xmm2
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  15,40,29,166,25,0,0                 // movaps        0x19a6(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_a8_sse2_lowp
 .globl _sk_load_a8_sse2_lowp
 FUNCTION(_sk_load_a8_sse2_lowp)
@@ -58938,19 +60156,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            31d0d <_sk_load_a8_sse2_lowp+0x39>
+  .byte  119,34                              // ja            330ed <_sk_load_a8_sse2_lowp+0x39>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,139,0,0,0                 // lea           0x8b(%rip),%r9        # 31d84 <_sk_load_a8_sse2_lowp+0xb0>
+  .byte  76,141,13,139,0,0,0                 // lea           0x8b(%rip),%r9        # 33164 <_sk_load_a8_sse2_lowp+0xb0>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,216                      // movd          %eax,%xmm3
-  .byte  235,97                              // jmp           31d6e <_sk_load_a8_sse2_lowp+0x9a>
+  .byte  235,97                              // jmp           3314e <_sk_load_a8_sse2_lowp+0x9a>
   .byte  243,65,15,126,28,16                 // movq          (%r8,%rdx,1),%xmm3
   .byte  102,15,96,216                       // punpcklbw     %xmm0,%xmm3
-  .byte  235,85                              // jmp           31d6e <_sk_load_a8_sse2_lowp+0x9a>
+  .byte  235,85                              // jmp           3314e <_sk_load_a8_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,216,2                    // pinsrw        $0x2,%eax,%xmm3
@@ -58958,7 +60176,7 @@
   .byte  102,15,110,192                      // movd          %eax,%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  243,15,16,216                       // movss         %xmm0,%xmm3
-  .byte  235,51                              // jmp           31d6e <_sk_load_a8_sse2_lowp+0x9a>
+  .byte  235,51                              // jmp           3314e <_sk_load_a8_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,216,6                    // pinsrw        $0x6,%eax,%xmm3
@@ -58969,14 +60187,14 @@
   .byte  102,65,15,110,4,16                  // movd          (%r8,%rdx,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  242,15,16,216                       // movsd         %xmm0,%xmm3
-  .byte  102,15,219,29,250,21,0,0            // pand          0x15fa(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,29,250,24,0,0            // pand          0x18fa(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
   .byte  255,224                             // jmpq          *%rax
   .byte  144                                 // nop
-  .byte  126,255                             // jle           31d85 <_sk_load_a8_sse2_lowp+0xb1>
+  .byte  126,255                             // jle           33165 <_sk_load_a8_sse2_lowp+0xb1>
   .byte  255                                 // (bad)
   .byte  255,164,255,255,255,149,255         // jmpq          *-0x6a0001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
@@ -59006,19 +60224,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            31dd9 <_sk_load_a8_dst_sse2_lowp+0x39>
+  .byte  119,34                              // ja            331b9 <_sk_load_a8_dst_sse2_lowp+0x39>
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,139,0,0,0                 // lea           0x8b(%rip),%r9        # 31e50 <_sk_load_a8_dst_sse2_lowp+0xb0>
+  .byte  76,141,13,139,0,0,0                 // lea           0x8b(%rip),%r9        # 33230 <_sk_load_a8_dst_sse2_lowp+0xb0>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,248                      // movd          %eax,%xmm7
-  .byte  235,97                              // jmp           31e3a <_sk_load_a8_dst_sse2_lowp+0x9a>
+  .byte  235,97                              // jmp           3321a <_sk_load_a8_dst_sse2_lowp+0x9a>
   .byte  243,65,15,126,60,16                 // movq          (%r8,%rdx,1),%xmm7
   .byte  102,15,96,248                       // punpcklbw     %xmm0,%xmm7
-  .byte  235,85                              // jmp           31e3a <_sk_load_a8_dst_sse2_lowp+0x9a>
+  .byte  235,85                              // jmp           3321a <_sk_load_a8_dst_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  102,15,196,248,2                    // pinsrw        $0x2,%eax,%xmm7
@@ -59026,7 +60244,7 @@
   .byte  102,15,110,224                      // movd          %eax,%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  243,15,16,252                       // movss         %xmm4,%xmm7
-  .byte  235,51                              // jmp           31e3a <_sk_load_a8_dst_sse2_lowp+0x9a>
+  .byte  235,51                              // jmp           3321a <_sk_load_a8_dst_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  102,15,196,248,6                    // pinsrw        $0x6,%eax,%xmm7
@@ -59037,14 +60255,14 @@
   .byte  102,65,15,110,36,16                 // movd          (%r8,%rdx,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  242,15,16,252                       // movsd         %xmm4,%xmm7
-  .byte  102,15,219,61,46,21,0,0             // pand          0x152e(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,61,46,24,0,0             // pand          0x182e(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,87,237                           // xorps         %xmm5,%xmm5
   .byte  15,87,246                           // xorps         %xmm6,%xmm6
   .byte  255,224                             // jmpq          *%rax
   .byte  144                                 // nop
-  .byte  126,255                             // jle           31e51 <_sk_load_a8_dst_sse2_lowp+0xb1>
+  .byte  126,255                             // jle           33231 <_sk_load_a8_dst_sse2_lowp+0xb1>
   .byte  255                                 // (bad)
   .byte  255,164,255,255,255,149,255         // jmpq          *-0x6a0001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
@@ -59074,42 +60292,42 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,29                              // ja            31ea0 <_sk_store_a8_sse2_lowp+0x34>
+  .byte  119,29                              // ja            33280 <_sk_store_a8_sse2_lowp+0x34>
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 31f20 <_sk_store_a8_sse2_lowp+0xb4>
+  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 33300 <_sk_store_a8_sse2_lowp+0xb4>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,15,126,216                      // movd          %xmm3,%eax
   .byte  65,136,4,16                         // mov           %al,(%r8,%rdx,1)
-  .byte  235,123                             // jmp           31f1b <_sk_store_a8_sse2_lowp+0xaf>
-  .byte  102,68,15,111,5,199,20,0,0          // movdqa        0x14c7(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  235,123                             // jmp           332fb <_sk_store_a8_sse2_lowp+0xaf>
+  .byte  102,68,15,111,5,199,23,0,0          // movdqa        0x17c7(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,219,195                   // pand          %xmm3,%xmm8
   .byte  102,69,15,103,192                   // packuswb      %xmm8,%xmm8
   .byte  102,69,15,214,4,16                  // movq          %xmm8,(%r8,%rdx,1)
-  .byte  235,96                              // jmp           31f1b <_sk_store_a8_sse2_lowp+0xaf>
+  .byte  235,96                              // jmp           332fb <_sk_store_a8_sse2_lowp+0xaf>
   .byte  102,15,197,195,2                    // pextrw        $0x2,%xmm3,%eax
   .byte  65,136,68,16,2                      // mov           %al,0x2(%r8,%rdx,1)
-  .byte  102,68,15,111,5,162,20,0,0          // movdqa        0x14a2(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,162,23,0,0          // movdqa        0x17a2(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,219,195                   // pand          %xmm3,%xmm8
   .byte  102,69,15,103,192                   // packuswb      %xmm8,%xmm8
   .byte  102,68,15,126,192                   // movd          %xmm8,%eax
   .byte  102,65,137,4,16                     // mov           %ax,(%r8,%rdx,1)
-  .byte  235,55                              // jmp           31f1b <_sk_store_a8_sse2_lowp+0xaf>
+  .byte  235,55                              // jmp           332fb <_sk_store_a8_sse2_lowp+0xaf>
   .byte  102,15,197,195,6                    // pextrw        $0x6,%xmm3,%eax
   .byte  65,136,68,16,6                      // mov           %al,0x6(%r8,%rdx,1)
   .byte  102,15,197,195,5                    // pextrw        $0x5,%xmm3,%eax
   .byte  65,136,68,16,5                      // mov           %al,0x5(%r8,%rdx,1)
   .byte  102,15,197,195,4                    // pextrw        $0x4,%xmm3,%eax
   .byte  65,136,68,16,4                      // mov           %al,0x4(%r8,%rdx,1)
-  .byte  102,68,15,111,5,101,20,0,0          // movdqa        0x1465(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,101,23,0,0          // movdqa        0x1765(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,68,15,219,195                   // pand          %xmm3,%xmm8
   .byte  102,69,15,103,192                   // packuswb      %xmm8,%xmm8
   .byte  102,69,15,126,4,16                  // movd          %xmm8,(%r8,%rdx,1)
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  144                                 // nop
-  .byte  118,255                             // jbe           31f21 <_sk_store_a8_sse2_lowp+0xb5>
+  .byte  118,255                             // jbe           33301 <_sk_store_a8_sse2_lowp+0xb5>
   .byte  255                                 // (bad)
   .byte  255,165,255,255,255,155             // jmpq          *-0x64000001(%rbp)
   .byte  255                                 // (bad)
@@ -59128,6 +60346,100 @@
   .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
+HIDDEN _sk_gather_a8_sse2_lowp
+.globl _sk_gather_a8_sse2_lowp
+FUNCTION(_sk_gather_a8_sse2_lowp)
+_sk_gather_a8_sse2_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,72,12                  // movss         0xc(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  102,69,15,118,210                   // pcmpeqd       %xmm10,%xmm10
+  .byte  102,69,15,254,202                   // paddd         %xmm10,%xmm9
+  .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
+  .byte  65,15,95,192                        // maxps         %xmm8,%xmm0
+  .byte  65,15,95,200                        // maxps         %xmm8,%xmm1
+  .byte  65,15,93,201                        // minps         %xmm9,%xmm1
+  .byte  65,15,93,193                        // minps         %xmm9,%xmm0
+  .byte  243,68,15,16,72,16                  // movss         0x10(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  102,69,15,254,202                   // paddd         %xmm10,%xmm9
+  .byte  65,15,95,216                        // maxps         %xmm8,%xmm3
+  .byte  65,15,95,208                        // maxps         %xmm8,%xmm2
+  .byte  65,15,93,209                        // minps         %xmm9,%xmm2
+  .byte  65,15,93,217                        // minps         %xmm9,%xmm3
+  .byte  243,68,15,91,203                    // cvttps2dq     %xmm3,%xmm9
+  .byte  243,68,15,91,210                    // cvttps2dq     %xmm2,%xmm10
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,69,15,112,218,245               // pshufd        $0xf5,%xmm10,%xmm11
+  .byte  102,68,15,244,219                   // pmuludq       %xmm3,%xmm11
+  .byte  102,68,15,111,227                   // movdqa        %xmm3,%xmm12
+  .byte  102,69,15,112,233,245               // pshufd        $0xf5,%xmm9,%xmm13
+  .byte  102,68,15,244,235                   // pmuludq       %xmm3,%xmm13
+  .byte  102,65,15,244,218                   // pmuludq       %xmm10,%xmm3
+  .byte  102,15,112,219,232                  // pshufd        $0xe8,%xmm3,%xmm3
+  .byte  102,65,15,112,211,232               // pshufd        $0xe8,%xmm11,%xmm2
+  .byte  102,15,98,218                       // punpckldq     %xmm2,%xmm3
+  .byte  76,139,16                           // mov           (%rax),%r10
+  .byte  102,69,15,244,225                   // pmuludq       %xmm9,%xmm12
+  .byte  102,69,15,112,204,232               // pshufd        $0xe8,%xmm12,%xmm9
+  .byte  102,65,15,112,213,232               // pshufd        $0xe8,%xmm13,%xmm2
+  .byte  102,68,15,98,202                    // punpckldq     %xmm2,%xmm9
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,201                   // paddd         %xmm9,%xmm1
+  .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
+  .byte  102,73,15,126,208                   // movq          %xmm2,%r8
+  .byte  69,137,193                          // mov           %r8d,%r9d
+  .byte  73,193,232,32                       // shr           $0x20,%r8
+  .byte  102,72,15,126,200                   // movq          %xmm1,%rax
+  .byte  65,137,195                          // mov           %eax,%r11d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,15,112,192,78                   // pshufd        $0x4e,%xmm0,%xmm0
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  71,15,182,36,34                     // movzbl        (%r10,%r12,1),%r12d
+  .byte  67,15,182,44,58                     // movzbl        (%r10,%r15,1),%ebp
+  .byte  193,229,8                           // shl           $0x8,%ebp
+  .byte  68,9,229                            // or            %r12d,%ebp
+  .byte  71,15,182,52,50                     // movzbl        (%r10,%r14,1),%r14d
+  .byte  65,15,182,28,26                     // movzbl        (%r10,%rbx,1),%ebx
+  .byte  193,227,8                           // shl           $0x8,%ebx
+  .byte  68,9,243                            // or            %r14d,%ebx
+  .byte  102,15,110,219                      // movd          %ebx,%xmm3
+  .byte  102,15,196,221,1                    // pinsrw        $0x1,%ebp,%xmm3
+  .byte  67,15,182,44,26                     // movzbl        (%r10,%r11,1),%ebp
+  .byte  65,15,182,4,2                       // movzbl        (%r10,%rax,1),%eax
+  .byte  193,224,8                           // shl           $0x8,%eax
+  .byte  9,232                               // or            %ebp,%eax
+  .byte  102,15,196,216,2                    // pinsrw        $0x2,%eax,%xmm3
+  .byte  67,15,182,4,10                      // movzbl        (%r10,%r9,1),%eax
+  .byte  67,15,182,44,2                      // movzbl        (%r10,%r8,1),%ebp
+  .byte  193,229,8                           // shl           $0x8,%ebp
+  .byte  9,197                               // or            %eax,%ebp
+  .byte  102,15,196,221,3                    // pinsrw        $0x3,%ebp,%xmm3
+  .byte  102,65,15,96,216                    // punpcklbw     %xmm8,%xmm3
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
+  .byte  102,15,239,201                      // pxor          %xmm1,%xmm1
+  .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_load_g8_sse2_lowp
 .globl _sk_load_g8_sse2_lowp
 FUNCTION(_sk_load_g8_sse2_lowp)
@@ -59140,19 +60452,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            31f75 <_sk_load_g8_sse2_lowp+0x39>
+  .byte  119,34                              // ja            334bd <_sk_load_g8_sse2_lowp+0x39>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 31ff4 <_sk_load_g8_sse2_lowp+0xb8>
+  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 3353c <_sk_load_g8_sse2_lowp+0xb8>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,192                      // movd          %eax,%xmm0
-  .byte  235,97                              // jmp           31fd6 <_sk_load_g8_sse2_lowp+0x9a>
+  .byte  235,97                              // jmp           3351e <_sk_load_g8_sse2_lowp+0x9a>
   .byte  243,65,15,126,4,16                  // movq          (%r8,%rdx,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
-  .byte  235,85                              // jmp           31fd6 <_sk_load_g8_sse2_lowp+0x9a>
+  .byte  235,85                              // jmp           3351e <_sk_load_g8_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,192,2                    // pinsrw        $0x2,%eax,%xmm0
@@ -59160,7 +60472,7 @@
   .byte  102,15,110,200                      // movd          %eax,%xmm1
   .byte  102,15,96,200                       // punpcklbw     %xmm0,%xmm1
   .byte  243,15,16,193                       // movss         %xmm1,%xmm0
-  .byte  235,51                              // jmp           31fd6 <_sk_load_g8_sse2_lowp+0x9a>
+  .byte  235,51                              // jmp           3351e <_sk_load_g8_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,192,6                    // pinsrw        $0x6,%eax,%xmm0
@@ -59171,14 +60483,14 @@
   .byte  102,65,15,110,12,16                 // movd          (%r8,%rdx,1),%xmm1
   .byte  102,15,96,200                       // punpcklbw     %xmm0,%xmm1
   .byte  242,15,16,193                       // movsd         %xmm1,%xmm0
-  .byte  102,15,219,5,146,19,0,0             // pand          0x1392(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,5,42,21,0,0              // pand          0x152a(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,29,137,19,0,0                 // movaps        0x1389(%rip),%xmm3        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,29,33,21,0,0                  // movaps        0x1521(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  118,255                             // jbe           31ff5 <_sk_load_g8_sse2_lowp+0xb9>
+  .byte  118,255                             // jbe           3353d <_sk_load_g8_sse2_lowp+0xb9>
   .byte  255                                 // (bad)
   .byte  255,156,255,255,255,141,255         // lcall         *-0x720001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
@@ -59206,19 +60518,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            32049 <_sk_load_g8_dst_sse2_lowp+0x39>
+  .byte  119,34                              // ja            33591 <_sk_load_g8_dst_sse2_lowp+0x39>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 320c8 <_sk_load_g8_dst_sse2_lowp+0xb8>
+  .byte  76,141,13,147,0,0,0                 // lea           0x93(%rip),%r9        # 33610 <_sk_load_g8_dst_sse2_lowp+0xb8>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,15,110,224                      // movd          %eax,%xmm4
-  .byte  235,97                              // jmp           320aa <_sk_load_g8_dst_sse2_lowp+0x9a>
+  .byte  235,97                              // jmp           335f2 <_sk_load_g8_dst_sse2_lowp+0x9a>
   .byte  243,65,15,126,36,16                 // movq          (%r8,%rdx,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
-  .byte  235,85                              // jmp           320aa <_sk_load_g8_dst_sse2_lowp+0x9a>
+  .byte  235,85                              // jmp           335f2 <_sk_load_g8_dst_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,224,2                    // pinsrw        $0x2,%eax,%xmm4
@@ -59226,7 +60538,7 @@
   .byte  102,15,110,232                      // movd          %eax,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,51                              // jmp           320aa <_sk_load_g8_dst_sse2_lowp+0x9a>
+  .byte  235,51                              // jmp           335f2 <_sk_load_g8_dst_sse2_lowp+0x9a>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,224,6                    // pinsrw        $0x6,%eax,%xmm4
@@ -59237,14 +60549,14 @@
   .byte  102,65,15,110,44,16                 // movd          (%r8,%rdx,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,229                       // movsd         %xmm5,%xmm4
-  .byte  102,15,219,37,190,18,0,0            // pand          0x12be(%rip),%xmm4        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,219,37,86,20,0,0             // pand          0x1456(%rip),%xmm4        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  72,173                              // lods          %ds:(%rsi),%rax
-  .byte  15,40,61,181,18,0,0                 // movaps        0x12b5(%rip),%xmm7        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  15,40,61,77,20,0,0                  // movaps        0x144d(%rip),%xmm7        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  118,255                             // jbe           320c9 <_sk_load_g8_dst_sse2_lowp+0xb9>
+  .byte  118,255                             // jbe           33611 <_sk_load_g8_dst_sse2_lowp+0xb9>
   .byte  255                                 // (bad)
   .byte  255,156,255,255,255,141,255         // lcall         *-0x720001(%rdi,%rdi,8)
   .byte  255                                 // (bad)
@@ -59265,10 +60577,10 @@
 FUNCTION(_sk_luminance_to_alpha_sse2_lowp)
 _sk_luminance_to_alpha_sse2_lowp:
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
-  .byte  102,15,213,5,128,26,0,0             // pmullw        0x1a80(%rip),%xmm0        # 33b70 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
-  .byte  102,15,213,13,136,26,0,0            // pmullw        0x1a88(%rip),%xmm1        # 33b80 <_sk_srcover_rgba_8888_sse2_lowp+0x1200>
+  .byte  102,15,213,5,24,28,0,0              // pmullw        0x1c18(%rip),%xmm0        # 35250 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  .byte  102,15,213,13,32,28,0,0             // pmullw        0x1c20(%rip),%xmm1        # 35260 <_sk_srcover_rgba_8888_sse2_lowp+0x1230>
   .byte  102,15,253,200                      // paddw         %xmm0,%xmm1
-  .byte  102,15,213,29,140,26,0,0            // pmullw        0x1a8c(%rip),%xmm3        # 33b90 <_sk_srcover_rgba_8888_sse2_lowp+0x1210>
+  .byte  102,15,213,29,36,28,0,0             // pmullw        0x1c24(%rip),%xmm3        # 35270 <_sk_srcover_rgba_8888_sse2_lowp+0x1240>
   .byte  102,15,253,217                      // paddw         %xmm1,%xmm3
   .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
   .byte  72,173                              // lods          %ds:(%rsi),%rax
@@ -59277,20 +60589,114 @@
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  255,224                             // jmpq          *%rax
 
+HIDDEN _sk_gather_g8_sse2_lowp
+.globl _sk_gather_g8_sse2_lowp
+FUNCTION(_sk_gather_g8_sse2_lowp)
+_sk_gather_g8_sse2_lowp:
+  .byte  85                                  // push          %rbp
+  .byte  65,87                               // push          %r15
+  .byte  65,86                               // push          %r14
+  .byte  65,84                               // push          %r12
+  .byte  83                                  // push          %rbx
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  243,68,15,16,72,12                  // movss         0xc(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  102,69,15,118,210                   // pcmpeqd       %xmm10,%xmm10
+  .byte  102,69,15,254,202                   // paddd         %xmm10,%xmm9
+  .byte  69,15,87,192                        // xorps         %xmm8,%xmm8
+  .byte  65,15,95,192                        // maxps         %xmm8,%xmm0
+  .byte  65,15,95,200                        // maxps         %xmm8,%xmm1
+  .byte  65,15,93,201                        // minps         %xmm9,%xmm1
+  .byte  65,15,93,193                        // minps         %xmm9,%xmm0
+  .byte  243,68,15,16,72,16                  // movss         0x10(%rax),%xmm9
+  .byte  69,15,198,201,0                     // shufps        $0x0,%xmm9,%xmm9
+  .byte  102,69,15,254,202                   // paddd         %xmm10,%xmm9
+  .byte  65,15,95,216                        // maxps         %xmm8,%xmm3
+  .byte  65,15,95,208                        // maxps         %xmm8,%xmm2
+  .byte  65,15,93,209                        // minps         %xmm9,%xmm2
+  .byte  65,15,93,217                        // minps         %xmm9,%xmm3
+  .byte  243,68,15,91,203                    // cvttps2dq     %xmm3,%xmm9
+  .byte  243,68,15,91,210                    // cvttps2dq     %xmm2,%xmm10
+  .byte  102,15,110,88,8                     // movd          0x8(%rax),%xmm3
+  .byte  102,15,112,219,0                    // pshufd        $0x0,%xmm3,%xmm3
+  .byte  102,69,15,112,218,245               // pshufd        $0xf5,%xmm10,%xmm11
+  .byte  102,68,15,244,219                   // pmuludq       %xmm3,%xmm11
+  .byte  102,68,15,111,227                   // movdqa        %xmm3,%xmm12
+  .byte  102,69,15,112,233,245               // pshufd        $0xf5,%xmm9,%xmm13
+  .byte  102,68,15,244,235                   // pmuludq       %xmm3,%xmm13
+  .byte  102,65,15,244,218                   // pmuludq       %xmm10,%xmm3
+  .byte  102,15,112,219,232                  // pshufd        $0xe8,%xmm3,%xmm3
+  .byte  102,65,15,112,211,232               // pshufd        $0xe8,%xmm11,%xmm2
+  .byte  102,15,98,218                       // punpckldq     %xmm2,%xmm3
+  .byte  76,139,16                           // mov           (%rax),%r10
+  .byte  102,69,15,244,225                   // pmuludq       %xmm9,%xmm12
+  .byte  102,69,15,112,204,232               // pshufd        $0xe8,%xmm12,%xmm9
+  .byte  102,65,15,112,213,232               // pshufd        $0xe8,%xmm13,%xmm2
+  .byte  102,68,15,98,202                    // punpckldq     %xmm2,%xmm9
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,65,15,254,201                   // paddd         %xmm9,%xmm1
+  .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
+  .byte  102,73,15,126,208                   // movq          %xmm2,%r8
+  .byte  69,137,193                          // mov           %r8d,%r9d
+  .byte  73,193,232,32                       // shr           $0x20,%r8
+  .byte  102,72,15,126,200                   // movq          %xmm1,%rax
+  .byte  65,137,195                          // mov           %eax,%r11d
+  .byte  72,193,232,32                       // shr           $0x20,%rax
+  .byte  102,72,15,126,195                   // movq          %xmm0,%rbx
+  .byte  65,137,222                          // mov           %ebx,%r14d
+  .byte  72,193,235,32                       // shr           $0x20,%rbx
+  .byte  102,15,112,192,78                   // pshufd        $0x4e,%xmm0,%xmm0
+  .byte  102,73,15,126,199                   // movq          %xmm0,%r15
+  .byte  69,137,252                          // mov           %r15d,%r12d
+  .byte  73,193,239,32                       // shr           $0x20,%r15
+  .byte  71,15,182,36,34                     // movzbl        (%r10,%r12,1),%r12d
+  .byte  67,15,182,44,58                     // movzbl        (%r10,%r15,1),%ebp
+  .byte  193,229,8                           // shl           $0x8,%ebp
+  .byte  68,9,229                            // or            %r12d,%ebp
+  .byte  71,15,182,52,50                     // movzbl        (%r10,%r14,1),%r14d
+  .byte  65,15,182,28,26                     // movzbl        (%r10,%rbx,1),%ebx
+  .byte  193,227,8                           // shl           $0x8,%ebx
+  .byte  68,9,243                            // or            %r14d,%ebx
+  .byte  102,15,110,195                      // movd          %ebx,%xmm0
+  .byte  102,15,196,197,1                    // pinsrw        $0x1,%ebp,%xmm0
+  .byte  67,15,182,44,26                     // movzbl        (%r10,%r11,1),%ebp
+  .byte  65,15,182,4,2                       // movzbl        (%r10,%rax,1),%eax
+  .byte  193,224,8                           // shl           $0x8,%eax
+  .byte  9,232                               // or            %ebp,%eax
+  .byte  102,15,196,192,2                    // pinsrw        $0x2,%eax,%xmm0
+  .byte  67,15,182,4,10                      // movzbl        (%r10,%r9,1),%eax
+  .byte  67,15,182,44,2                      // movzbl        (%r10,%r8,1),%ebp
+  .byte  193,229,8                           // shl           $0x8,%ebp
+  .byte  9,197                               // or            %eax,%ebp
+  .byte  102,15,196,197,3                    // pinsrw        $0x3,%ebp,%xmm0
+  .byte  102,65,15,96,192                    // punpcklbw     %xmm8,%xmm0
+  .byte  72,173                              // lods          %ds:(%rsi),%rax
+  .byte  15,40,29,146,18,0,0                 // movaps        0x1292(%rip),%xmm3        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
+  .byte  91                                  // pop           %rbx
+  .byte  65,92                               // pop           %r12
+  .byte  65,94                               // pop           %r14
+  .byte  65,95                               // pop           %r15
+  .byte  93                                  // pop           %rbp
+  .byte  255,224                             // jmpq          *%rax
+
 HIDDEN _sk_scale_1_float_sse2_lowp
 .globl _sk_scale_1_float_sse2_lowp
 FUNCTION(_sk_scale_1_float_sse2_lowp)
 _sk_scale_1_float_sse2_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
-  .byte  243,68,15,89,5,47,11,0,0            // mulss         0xb2f(%rip),%xmm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  243,68,15,88,5,246,10,0,0           // addss         0xaf6(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,5,44,11,0,0            // mulss         0xb2c(%rip),%xmm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,68,15,88,5,243,10,0,0           // addss         0xaf3(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,44,192                    // cvttss2si     %xmm8,%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  242,69,15,112,192,0                 // pshuflw       $0x0,%xmm8,%xmm8
   .byte  102,69,15,112,192,80                // pshufd        $0x50,%xmm8,%xmm8
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
-  .byte  102,68,15,111,13,22,18,0,0          // movdqa        0x1216(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,67,18,0,0          // movdqa        0x1243(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -59311,13 +60717,13 @@
 _sk_lerp_1_float_sse2_lowp:
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  243,68,15,16,0                      // movss         (%rax),%xmm8
-  .byte  243,68,15,89,5,183,10,0,0           // mulss         0xab7(%rip),%xmm8        # 32c5c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  .byte  243,68,15,88,5,126,10,0,0           // addss         0xa7e(%rip),%xmm8        # 32c2c <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  .byte  243,68,15,89,5,180,10,0,0           // mulss         0xab4(%rip),%xmm8        # 3430c <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  .byte  243,68,15,88,5,123,10,0,0           // addss         0xa7b(%rip),%xmm8        # 342dc <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   .byte  243,65,15,44,192                    // cvttss2si     %xmm8,%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
   .byte  242,69,15,112,192,0                 // pshuflw       $0x0,%xmm8,%xmm8
   .byte  102,69,15,112,192,80                // pshufd        $0x50,%xmm8,%xmm8
-  .byte  102,68,15,111,13,163,17,0,0         // movdqa        0x11a3(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,208,17,0,0         // movdqa        0x11d0(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -59357,19 +60763,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,36                              // ja            32284 <_sk_scale_u8_sse2_lowp+0x3b>
+  .byte  119,36                              // ja            33937 <_sk_scale_u8_sse2_lowp+0x3b>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,221,0,0,0                 // lea           0xdd(%rip),%r9        # 3234c <_sk_scale_u8_sse2_lowp+0x103>
+  .byte  76,141,13,218,0,0,0                 // lea           0xda(%rip),%r9        # 339fc <_sk_scale_u8_sse2_lowp+0x100>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
-  .byte  235,109                             // jmp           322f1 <_sk_scale_u8_sse2_lowp+0xa8>
+  .byte  235,109                             // jmp           339a4 <_sk_scale_u8_sse2_lowp+0xa8>
   .byte  243,69,15,126,4,16                  // movq          (%r8,%rdx,1),%xmm8
   .byte  102,68,15,96,192                    // punpcklbw     %xmm0,%xmm8
-  .byte  235,96                              // jmp           322f1 <_sk_scale_u8_sse2_lowp+0xa8>
+  .byte  235,96                              // jmp           339a4 <_sk_scale_u8_sse2_lowp+0xa8>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,68,15,196,192,2                 // pinsrw        $0x2,%eax,%xmm8
@@ -59377,7 +60783,7 @@
   .byte  102,68,15,110,200                   // movd          %eax,%xmm9
   .byte  102,68,15,96,200                    // punpcklbw     %xmm0,%xmm9
   .byte  243,69,15,16,193                    // movss         %xmm9,%xmm8
-  .byte  235,57                              // jmp           322f1 <_sk_scale_u8_sse2_lowp+0xa8>
+  .byte  235,57                              // jmp           339a4 <_sk_scale_u8_sse2_lowp+0xa8>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,68,15,196,192,6                 // pinsrw        $0x6,%eax,%xmm8
@@ -59388,9 +60794,9 @@
   .byte  102,69,15,110,12,16                 // movd          (%r8,%rdx,1),%xmm9
   .byte  102,68,15,96,200                    // punpcklbw     %xmm0,%xmm9
   .byte  242,69,15,16,193                    // movsd         %xmm9,%xmm8
-  .byte  102,68,15,219,5,118,16,0,0          // pand          0x1076(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,219,5,163,16,0,0          // pand          0x10a3(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,213,192                   // pmullw        %xmm8,%xmm0
-  .byte  102,68,15,111,13,104,16,0,0         // movdqa        0x1068(%rip),%xmm9        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,13,149,16,0,0         // movdqa        0x1095(%rip),%xmm9        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,253,193                   // paddw         %xmm9,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,65,15,213,200                   // pmullw        %xmm8,%xmm1
@@ -59405,20 +60811,19 @@
   .byte  72,173                              // lods          %ds:(%rsi),%rax
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  255,224                             // jmpq          *%rax
-  .byte  15,31,0                             // nopl          (%rax)
-  .byte  44,255                              // sub           $0xff,%al
-  .byte  255                                 // (bad)
-  .byte  255,86,255                          // callq         *-0x1(%rsi)
-  .byte  255                                 // (bad)
-  .byte  255,69,255                          // incl          -0x1(%rbp)
-  .byte  255                                 // (bad)
-  .byte  255,149,255,255,255,137             // callq         *-0x76000001(%rbp)
+  .byte  47                                  // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
+  .byte  255,89,255                          // lcall         *-0x1(%rcx)
   .byte  255                                 // (bad)
-  .byte  125,255                             // jge           32361 <_sk_scale_u8_sse2_lowp+0x118>
+  .byte  255,72,255                          // decl          -0x1(%rax)
   .byte  255                                 // (bad)
-  .byte  255,108,255,255                     // ljmp          *-0x1(%rdi,%rdi,8)
+  .byte  255,152,255,255,255,140             // lcall         *-0x73000001(%rax)
+  .byte  255                                 // (bad)
+  .byte  255                                 // (bad)
+  .byte  255,128,255,255,255,111             // incl          0x6fffffff(%rax)
+  .byte  255                                 // (bad)
+  .byte  255                                 // (bad)
   .byte  255                                 // .byte         0xff
 
 HIDDEN _sk_lerp_u8_sse2_lowp
@@ -59433,19 +60838,19 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,36                              // ja            323a3 <_sk_lerp_u8_sse2_lowp+0x3b>
+  .byte  119,36                              // ja            33a53 <_sk_lerp_u8_sse2_lowp+0x3b>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,30,1,0,0                  // lea           0x11e(%rip),%r9        # 324ac <_sk_lerp_u8_sse2_lowp+0x144>
+  .byte  76,141,13,30,1,0,0                  // lea           0x11e(%rip),%r9        # 33b5c <_sk_lerp_u8_sse2_lowp+0x144>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,182,4,16                      // movzbl        (%r8,%rdx,1),%eax
   .byte  102,68,15,110,192                   // movd          %eax,%xmm8
-  .byte  235,109                             // jmp           32410 <_sk_lerp_u8_sse2_lowp+0xa8>
+  .byte  235,109                             // jmp           33ac0 <_sk_lerp_u8_sse2_lowp+0xa8>
   .byte  243,69,15,126,4,16                  // movq          (%r8,%rdx,1),%xmm8
   .byte  102,68,15,96,192                    // punpcklbw     %xmm0,%xmm8
-  .byte  235,96                              // jmp           32410 <_sk_lerp_u8_sse2_lowp+0xa8>
+  .byte  235,96                              // jmp           33ac0 <_sk_lerp_u8_sse2_lowp+0xa8>
   .byte  65,15,182,68,16,2                   // movzbl        0x2(%r8,%rdx,1),%eax
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,68,15,196,192,2                 // pinsrw        $0x2,%eax,%xmm8
@@ -59453,7 +60858,7 @@
   .byte  102,68,15,110,200                   // movd          %eax,%xmm9
   .byte  102,68,15,96,200                    // punpcklbw     %xmm0,%xmm9
   .byte  243,69,15,16,193                    // movss         %xmm9,%xmm8
-  .byte  235,57                              // jmp           32410 <_sk_lerp_u8_sse2_lowp+0xa8>
+  .byte  235,57                              // jmp           33ac0 <_sk_lerp_u8_sse2_lowp+0xa8>
   .byte  65,15,182,68,16,6                   // movzbl        0x6(%r8,%rdx,1),%eax
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
   .byte  102,68,15,196,192,6                 // pinsrw        $0x6,%eax,%xmm8
@@ -59464,8 +60869,8 @@
   .byte  102,69,15,110,12,16                 // movd          (%r8,%rdx,1),%xmm9
   .byte  102,68,15,96,200                    // punpcklbw     %xmm0,%xmm9
   .byte  242,69,15,16,193                    // movsd         %xmm9,%xmm8
-  .byte  102,68,15,219,5,87,15,0,0           // pand          0xf57(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
-  .byte  102,68,15,111,21,78,15,0,0          // movdqa        0xf4e(%rip),%xmm10        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,219,5,135,15,0,0          // pand          0xf87(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
+  .byte  102,68,15,111,21,126,15,0,0         // movdqa        0xf7e(%rip),%xmm10        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,69,15,111,200                   // movdqa        %xmm8,%xmm9
   .byte  102,69,15,239,202                   // pxor          %xmm10,%xmm9
   .byte  102,69,15,111,217                   // movdqa        %xmm9,%xmm11
@@ -59495,9 +60900,9 @@
   .byte  102,65,15,111,216                   // movdqa        %xmm8,%xmm3
   .byte  255,224                             // jmpq          *%rax
   .byte  15,31,0                             // nopl          (%rax)
-  .byte  235,254                             // jmp           324ac <_sk_lerp_u8_sse2_lowp+0x144>
+  .byte  235,254                             // jmp           33b5c <_sk_lerp_u8_sse2_lowp+0x144>
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,4                // callq         *0x4ffffff(%rip)        # 50324b4 <_sk_srcover_rgba_8888_sse2_lowp+0x4fffb34>
+  .byte  255,21,255,255,255,4                // callq         *0x4ffffff(%rip)        # 5033b64 <_sk_srcover_rgba_8888_sse2_lowp+0x4fffb34>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,84,255,255                      // callq         *-0x1(%rdi,%rdi,8)
@@ -59527,23 +60932,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            32513 <_sk_scale_565_sse2_lowp+0x4b>
+  .byte  119,34                              // ja            33bc3 <_sk_scale_565_sse2_lowp+0x4b>
   .byte  102,15,239,246                      // pxor          %xmm6,%xmm6
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,229,1,0,0                 // lea           0x1e5(%rip),%r9        # 326e4 <_sk_scale_565_sse2_lowp+0x21c>
+  .byte  76,141,13,229,1,0,0                 // lea           0x1e5(%rip),%r9        # 33d94 <_sk_scale_565_sse2_lowp+0x21c>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,240                      // movd          %eax,%xmm6
-  .byte  235,67                              // jmp           32556 <_sk_scale_565_sse2_lowp+0x8e>
+  .byte  235,67                              // jmp           33c06 <_sk_scale_565_sse2_lowp+0x8e>
   .byte  243,65,15,111,52,80                 // movdqu        (%r8,%rdx,2),%xmm6
-  .byte  235,59                              // jmp           32556 <_sk_scale_565_sse2_lowp+0x8e>
+  .byte  235,59                              // jmp           33c06 <_sk_scale_565_sse2_lowp+0x8e>
   .byte  102,15,239,246                      // pxor          %xmm6,%xmm6
   .byte  102,65,15,196,116,80,4,2            // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm6
   .byte  243,69,15,16,12,80                  // movss         (%r8,%rdx,2),%xmm9
   .byte  243,65,15,16,241                    // movss         %xmm9,%xmm6
-  .byte  235,34                              // jmp           32556 <_sk_scale_565_sse2_lowp+0x8e>
+  .byte  235,34                              // jmp           33c06 <_sk_scale_565_sse2_lowp+0x8e>
   .byte  102,15,239,246                      // pxor          %xmm6,%xmm6
   .byte  102,65,15,196,116,80,12,6           // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm6
   .byte  102,65,15,196,116,80,10,5           // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm6
@@ -59551,11 +60956,11 @@
   .byte  102,65,15,18,52,80                  // movlpd        (%r8,%rdx,2),%xmm6
   .byte  102,68,15,111,222                   // movdqa        %xmm6,%xmm11
   .byte  102,65,15,113,211,8                 // psrlw         $0x8,%xmm11
-  .byte  102,68,15,219,29,182,21,0,0         // pand          0x15b6(%rip),%xmm11        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,68,15,219,29,230,21,0,0         // pand          0x15e6(%rip),%xmm11        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,68,15,111,206                   // movdqa        %xmm6,%xmm9
   .byte  102,65,15,113,209,5                 // psrlw         $0x5,%xmm9
-  .byte  102,68,15,219,13,178,21,0,0         // pand          0x15b2(%rip),%xmm9        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,68,15,111,21,185,21,0,0         // movdqa        0x15b9(%rip),%xmm10        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,68,15,219,13,226,21,0,0         // pand          0x15e2(%rip),%xmm9        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,111,21,233,21,0,0         // movdqa        0x15e9(%rip),%xmm10        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,68,15,219,214                   // pand          %xmm6,%xmm10
   .byte  102,15,113,214,13                   // psrlw         $0xd,%xmm6
   .byte  102,65,15,235,243                   // por           %xmm11,%xmm6
@@ -59567,7 +60972,7 @@
   .byte  102,65,15,113,243,3                 // psllw         $0x3,%xmm11
   .byte  102,65,15,113,210,2                 // psrlw         $0x2,%xmm10
   .byte  102,69,15,235,211                   // por           %xmm11,%xmm10
-  .byte  102,68,15,111,37,21,21,0,0          // movdqa        0x1515(%rip),%xmm12        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,37,69,21,0,0          // movdqa        0x1545(%rip),%xmm12        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,68,15,111,235                   // movdqa        %xmm3,%xmm13
   .byte  102,69,15,239,236                   // pxor          %xmm12,%xmm13
   .byte  102,68,15,111,223                   // movdqa        %xmm7,%xmm11
@@ -59608,7 +61013,7 @@
   .byte  102,68,15,213,201                   // pmullw        %xmm1,%xmm9
   .byte  102,68,15,213,210                   // pmullw        %xmm2,%xmm10
   .byte  102,68,15,213,219                   // pmullw        %xmm3,%xmm11
-  .byte  102,15,111,5,215,12,0,0             // movdqa        0xcd7(%rip),%xmm0        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,15,111,5,7,13,0,0               // movdqa        0xd07(%rip),%xmm0        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,253,240                      // paddw         %xmm0,%xmm6
   .byte  102,68,15,253,200                   // paddw         %xmm0,%xmm9
   .byte  102,68,15,253,208                   // paddw         %xmm0,%xmm10
@@ -59655,23 +61060,23 @@
   .byte  36,7                                // and           $0x7,%al
   .byte  254,200                             // dec           %al
   .byte  60,6                                // cmp           $0x6,%al
-  .byte  119,34                              // ja            3274b <_sk_lerp_565_sse2_lowp+0x4b>
+  .byte  119,34                              // ja            33dfb <_sk_lerp_565_sse2_lowp+0x4b>
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  15,182,192                          // movzbl        %al,%eax
-  .byte  76,141,13,45,2,0,0                  // lea           0x22d(%rip),%r9        # 32964 <_sk_lerp_565_sse2_lowp+0x264>
+  .byte  76,141,13,45,2,0,0                  // lea           0x22d(%rip),%r9        # 34014 <_sk_lerp_565_sse2_lowp+0x264>
   .byte  73,99,4,129                         // movslq        (%r9,%rax,4),%rax
   .byte  76,1,200                            // add           %r9,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  65,15,183,4,80                      // movzwl        (%r8,%rdx,2),%eax
   .byte  102,15,110,208                      // movd          %eax,%xmm2
-  .byte  235,67                              // jmp           3278e <_sk_lerp_565_sse2_lowp+0x8e>
+  .byte  235,67                              // jmp           33e3e <_sk_lerp_565_sse2_lowp+0x8e>
   .byte  243,65,15,111,20,80                 // movdqu        (%r8,%rdx,2),%xmm2
-  .byte  235,59                              // jmp           3278e <_sk_lerp_565_sse2_lowp+0x8e>
+  .byte  235,59                              // jmp           33e3e <_sk_lerp_565_sse2_lowp+0x8e>
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  102,65,15,196,84,80,4,2             // pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm2
   .byte  243,69,15,16,12,80                  // movss         (%r8,%rdx,2),%xmm9
   .byte  243,65,15,16,209                    // movss         %xmm9,%xmm2
-  .byte  235,34                              // jmp           3278e <_sk_lerp_565_sse2_lowp+0x8e>
+  .byte  235,34                              // jmp           33e3e <_sk_lerp_565_sse2_lowp+0x8e>
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  102,65,15,196,84,80,12,6            // pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm2
   .byte  102,65,15,196,84,80,10,5            // pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm2
@@ -59679,11 +61084,11 @@
   .byte  102,65,15,18,20,80                  // movlpd        (%r8,%rdx,2),%xmm2
   .byte  102,68,15,111,218                   // movdqa        %xmm2,%xmm11
   .byte  102,65,15,113,211,8                 // psrlw         $0x8,%xmm11
-  .byte  102,68,15,219,29,126,19,0,0         // pand          0x137e(%rip),%xmm11        # 33b20 <_sk_srcover_rgba_8888_sse2_lowp+0x11a0>
+  .byte  102,68,15,219,29,174,19,0,0         // pand          0x13ae(%rip),%xmm11        # 35200 <_sk_srcover_rgba_8888_sse2_lowp+0x11d0>
   .byte  102,68,15,111,202                   // movdqa        %xmm2,%xmm9
   .byte  102,65,15,113,209,5                 // psrlw         $0x5,%xmm9
-  .byte  102,68,15,219,13,122,19,0,0         // pand          0x137a(%rip),%xmm9        # 33b30 <_sk_srcover_rgba_8888_sse2_lowp+0x11b0>
-  .byte  102,68,15,111,21,129,19,0,0         // movdqa        0x1381(%rip),%xmm10        # 33b40 <_sk_srcover_rgba_8888_sse2_lowp+0x11c0>
+  .byte  102,68,15,219,13,170,19,0,0         // pand          0x13aa(%rip),%xmm9        # 35210 <_sk_srcover_rgba_8888_sse2_lowp+0x11e0>
+  .byte  102,68,15,111,21,177,19,0,0         // movdqa        0x13b1(%rip),%xmm10        # 35220 <_sk_srcover_rgba_8888_sse2_lowp+0x11f0>
   .byte  102,68,15,219,210                   // pand          %xmm2,%xmm10
   .byte  102,15,113,210,13                   // psrlw         $0xd,%xmm2
   .byte  102,65,15,235,211                   // por           %xmm11,%xmm2
@@ -59695,7 +61100,7 @@
   .byte  102,65,15,113,243,3                 // psllw         $0x3,%xmm11
   .byte  102,65,15,113,210,2                 // psrlw         $0x2,%xmm10
   .byte  102,69,15,235,211                   // por           %xmm11,%xmm10
-  .byte  102,68,15,111,37,221,18,0,0         // movdqa        0x12dd(%rip),%xmm12        # 33ae0 <_sk_srcover_rgba_8888_sse2_lowp+0x1160>
+  .byte  102,68,15,111,37,13,19,0,0          // movdqa        0x130d(%rip),%xmm12        # 351c0 <_sk_srcover_rgba_8888_sse2_lowp+0x1190>
   .byte  102,68,15,111,235                   // movdqa        %xmm3,%xmm13
   .byte  102,69,15,239,236                   // pxor          %xmm12,%xmm13
   .byte  102,68,15,111,223                   // movdqa        %xmm7,%xmm11
@@ -59734,7 +61139,7 @@
   .byte  102,69,15,235,222                   // por           %xmm14,%xmm11
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
   .byte  102,15,213,209                      // pmullw        %xmm1,%xmm2
-  .byte  102,68,15,111,5,169,10,0,0          // movdqa        0xaa9(%rip),%xmm8        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,5,217,10,0,0          // movdqa        0xad9(%rip),%xmm8        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,65,15,239,192                   // pxor          %xmm8,%xmm0
   .byte  102,15,213,196                      // pmullw        %xmm4,%xmm0
   .byte  102,65,15,253,208                   // paddw         %xmm8,%xmm2
@@ -59803,26 +61208,26 @@
   .byte  65,254,202                          // dec           %r10b
   .byte  69,15,182,202                       // movzbl        %r10b,%r9d
   .byte  65,128,249,6                        // cmp           $0x6,%r9b
-  .byte  119,38                              // ja            329cb <_sk_srcover_rgba_8888_sse2_lowp+0x4b>
+  .byte  119,38                              // ja            3407b <_sk_srcover_rgba_8888_sse2_lowp+0x4b>
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  76,141,29,67,2,0,0                  // lea           0x243(%rip),%r11        # 32bf4 <_sk_srcover_rgba_8888_sse2_lowp+0x274>
+  .byte  76,141,29,67,2,0,0                  // lea           0x243(%rip),%r11        # 342a4 <_sk_srcover_rgba_8888_sse2_lowp+0x274>
   .byte  75,99,4,139                         // movslq        (%r11,%r9,4),%rax
   .byte  76,1,216                            // add           %r11,%rax
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  255,224                             // jmpq          *%rax
   .byte  243,65,15,16,60,144                 // movss         (%r8,%rdx,4),%xmm7
   .byte  102,69,15,239,192                   // pxor          %xmm8,%xmm8
-  .byte  235,102                             // jmp           32a31 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  .byte  235,102                             // jmp           340e1 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   .byte  102,65,15,16,60,144                 // movupd        (%r8,%rdx,4),%xmm7
   .byte  243,69,15,111,68,144,16             // movdqu        0x10(%r8,%rdx,4),%xmm8
-  .byte  235,87                              // jmp           32a31 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  .byte  235,87                              // jmp           340e1 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   .byte  102,65,15,110,100,144,8             // movd          0x8(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,69,15,18,4,144                  // movlpd        (%r8,%rdx,4),%xmm8
   .byte  102,65,15,40,248                    // movapd        %xmm8,%xmm7
   .byte  102,68,15,111,196                   // movdqa        %xmm4,%xmm8
-  .byte  235,52                              // jmp           32a31 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  .byte  235,52                              // jmp           340e1 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   .byte  102,65,15,110,100,144,24            // movd          0x18(%r8,%rdx,4),%xmm4
   .byte  102,68,15,112,196,69                // pshufd        $0x45,%xmm4,%xmm8
   .byte  243,65,15,16,100,144,20             // movss         0x14(%r8,%rdx,4),%xmm4
@@ -59839,7 +61244,7 @@
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
   .byte  102,15,114,228,16                   // psrad         $0x10,%xmm4
   .byte  102,15,107,236                      // packssdw      %xmm4,%xmm5
-  .byte  102,68,15,111,37,21,9,0,0           // movdqa        0x915(%rip),%xmm12        # 33370 <_sk_srcover_rgba_8888_sse2_lowp+0x9f0>
+  .byte  102,68,15,111,37,69,9,0,0           // movdqa        0x945(%rip),%xmm12        # 34a50 <_sk_srcover_rgba_8888_sse2_lowp+0xa20>
   .byte  102,15,111,229                      // movdqa        %xmm5,%xmm4
   .byte  102,65,15,219,228                   // pand          %xmm12,%xmm4
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
@@ -59897,20 +61302,20 @@
   .byte  102,15,105,193                      // punpckhwd     %xmm1,%xmm0
   .byte  102,15,235,194                      // por           %xmm2,%xmm0
   .byte  65,128,250,6                        // cmp           $0x6,%r10b
-  .byte  119,24                              // ja            32b93 <_sk_srcover_rgba_8888_sse2_lowp+0x213>
-  .byte  76,141,21,142,0,0,0                 // lea           0x8e(%rip),%r10        # 32c10 <_sk_srcover_rgba_8888_sse2_lowp+0x290>
+  .byte  119,24                              // ja            34243 <_sk_srcover_rgba_8888_sse2_lowp+0x213>
+  .byte  76,141,21,142,0,0,0                 // lea           0x8e(%rip),%r10        # 342c0 <_sk_srcover_rgba_8888_sse2_lowp+0x290>
   .byte  75,99,4,138                         // movslq        (%r10,%r9,4),%rax
   .byte  76,1,208                            // add           %r10,%rax
   .byte  255,224                             // jmpq          *%rax
   .byte  102,69,15,126,36,144                // movd          %xmm12,(%r8,%rdx,4)
-  .byte  235,73                              // jmp           32bdc <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  .byte  235,73                              // jmp           3428c <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   .byte  243,69,15,127,36,144                // movdqu        %xmm12,(%r8,%rdx,4)
   .byte  243,65,15,127,68,144,16             // movdqu        %xmm0,0x10(%r8,%rdx,4)
-  .byte  235,58                              // jmp           32bdc <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  .byte  235,58                              // jmp           3428c <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   .byte  102,65,15,112,196,78                // pshufd        $0x4e,%xmm12,%xmm0
   .byte  102,65,15,126,68,144,8              // movd          %xmm0,0x8(%r8,%rdx,4)
   .byte  102,69,15,214,36,144                // movq          %xmm12,(%r8,%rdx,4)
-  .byte  235,37                              // jmp           32bdc <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  .byte  235,37                              // jmp           3428c <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
   .byte  102,65,15,126,76,144,24             // movd          %xmm1,0x18(%r8,%rdx,4)
   .byte  102,15,112,200,229                  // pshufd        $0xe5,%xmm0,%xmm1
@@ -59943,7 +61348,7 @@
   .byte  254                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  123,255                             // jnp           32c11 <_sk_srcover_rgba_8888_sse2_lowp+0x291>
+  .byte  123,255                             // jnp           342c1 <_sk_srcover_rgba_8888_sse2_lowp+0x291>
   .byte  255                                 // (bad)
   .byte  255,159,255,255,255,146             // lcall         *-0x6d000001(%rdi)
   .byte  255                                 // (bad)
@@ -59976,14 +61381,14 @@
   .byte  252                                 // cld
   .byte  190,0,0,128,191                     // mov           $0xbf800000,%esi
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        32c90 <.literal4+0x64>
+  .byte  224,64                              // loopne        34340 <.literal4+0x64>
   .byte  154                                 // (bad)
   .byte  153                                 // cltd
   .byte  153                                 // cltd
   .byte  62,61,10,23,63,174                  // ds            cmp $0xae3f170a,%eax
-  .byte  71,225,61                           // rex.RXB       loope 32c99 <.literal4+0x6d>
+  .byte  71,225,61                           // rex.RXB       loope 34349 <.literal4+0x6d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            32ca3 <.literal4+0x77>
+  .byte  127,67                              // jg            34353 <.literal4+0x77>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,127,145,131                     // cmpb          $0x83,-0x6f(%rdi)
   .byte  158                                 // sahf
@@ -60020,7 +61425,7 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        32cbd <.literal4+0x91>
+  .byte  224,7                               // loopne        3436d <.literal4+0x91>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -60031,8 +61436,8 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,52,255                            // add           %dh,(%rdi,%rdi,8)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            32cd0 <.literal4+0xa4>
-  .byte  119,115                             // ja            32d45 <.literal4+0x119>
+  .byte  127,0                               // jg            34380 <.literal4+0xa4>
+  .byte  119,115                             // ja            343f5 <.literal4+0x119>
   .byte  248                                 // clc
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,249,68,180                   // mov           $0xb444f93f,%edi
@@ -60055,12 +61460,12 @@
   .byte  190,80,128,3,62                     // mov           $0x3e038050,%esi
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           32d5b <.literal4+0x12f>
+  .byte  118,63                              // jbe           3440b <.literal4+0x12f>
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            32d6a <.literal4+0x13e>
+  .byte  124,66                              // jl            3441a <.literal4+0x13e>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  137,136,136,55,0,15                 // mov           %ecx,0xf003788(%rax)
@@ -60070,9 +61475,9 @@
   .byte  137,136,136,59,15,0                 // mov           %ecx,0xf3b88(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  137,136,136,61,0,0                  // mov           %ecx,0x3d88(%rax)
-  .byte  112,65                              // jo            32d8d <.literal4+0x161>
+  .byte  112,65                              // jo            3443d <.literal4+0x161>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            32d97 <.literal4+0x16b>
+  .byte  127,71                              // jg            34447 <.literal4+0x16b>
   .byte  89                                  // pop           %rcx
   .byte  23                                  // (bad)
   .byte  55                                  // (bad)
@@ -60100,14 +61505,14 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,0,0                             // addb          $0x0,(%rax)
   .byte  0,191,0,0,192,191                   // add           %bh,-0x40400000(%rdi)
-  .byte  114,28                              // jb            32db6 <.literal4+0x18a>
+  .byte  114,28                              // jb            34466 <.literal4+0x18a>
   .byte  199                                 // (bad)
   .byte  62,85                               // ds            push %rbp
   .byte  85                                  // push          %rbp
   .byte  149                                 // xchg          %eax,%ebp
   .byte  191,0,0,192,63                      // mov           $0x3fc00000,%edi
   .byte  57,142,99,61,114,249                // cmp           %ecx,-0x68dc29d(%rsi)
-  .byte  127,63                              // jg            32deb <_sk_srcover_rgba_8888_sse2_lowp+0x46b>
+  .byte  127,63                              // jg            3449b <_sk_srcover_rgba_8888_sse2_lowp+0x46b>
   .byte  3,0                                 // add           (%rax),%eax
   .byte  0,0                                 // add           %al,(%rax)
   .byte  109                                 // insl          (%dx),%es:(%rdi)
@@ -60122,7 +61527,7 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            32dc5 <.literal4+0x199>
+  .byte  127,0                               // jg            34475 <.literal4+0x199>
   .byte  0,128,56,0,64,254                   // add           %al,-0x1bfffc8(%rax)
   .byte  255                                 // .byte         0xff
 
@@ -60174,16 +61579,16 @@
   .byte  0,1                                 // add           %al,(%rcx)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a032e48 <_sk_srcover_rgba_8888_sse2_lowp+0xa0004c8>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a0344e8 <_sk_srcover_rgba_8888_sse2_lowp+0xa0004b8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12032e50 <_sk_srcover_rgba_8888_sse2_lowp+0x120004d0>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 120344f0 <_sk_srcover_rgba_8888_sse2_lowp+0x120004c0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a032e58 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0004d8>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a0344f8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0004c8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3032e60 <_sk_srcover_rgba_8888_sse2_lowp+0x30004e0>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3034500 <_sk_srcover_rgba_8888_sse2_lowp+0x30004d0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60226,16 +61631,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a032ea8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000528>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034548 <_sk_srcover_rgba_8888_sse2_lowp+0xa000518>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12032eb0 <_sk_srcover_rgba_8888_sse2_lowp+0x12000530>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12034550 <_sk_srcover_rgba_8888_sse2_lowp+0x12000520>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a032eb8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000538>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a034558 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000528>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3032ec0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000540>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3034560 <_sk_srcover_rgba_8888_sse2_lowp+0x3000530>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60278,16 +61683,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a032f08 <_sk_srcover_rgba_8888_sse2_lowp+0xa000588>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a0345a8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000578>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12032f10 <_sk_srcover_rgba_8888_sse2_lowp+0x12000590>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 120345b0 <_sk_srcover_rgba_8888_sse2_lowp+0x12000580>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a032f18 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000598>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a0345b8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000588>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3032f20 <_sk_srcover_rgba_8888_sse2_lowp+0x30005a0>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 30345c0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000590>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60330,16 +61735,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a032f68 <_sk_srcover_rgba_8888_sse2_lowp+0xa0005e8>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034608 <_sk_srcover_rgba_8888_sse2_lowp+0xa0005d8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12032f70 <_sk_srcover_rgba_8888_sse2_lowp+0x120005f0>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12034610 <_sk_srcover_rgba_8888_sse2_lowp+0x120005e0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a032f78 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0005f8>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a034618 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0005e8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3032f80 <_sk_srcover_rgba_8888_sse2_lowp+0x3000600>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3034620 <_sk_srcover_rgba_8888_sse2_lowp+0x30005f0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60382,16 +61787,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a032fc8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000648>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034668 <_sk_srcover_rgba_8888_sse2_lowp+0xa000638>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12032fd0 <_sk_srcover_rgba_8888_sse2_lowp+0x12000650>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12034670 <_sk_srcover_rgba_8888_sse2_lowp+0x12000640>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a032fd8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000658>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a034678 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000648>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3032fe0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000660>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3034680 <_sk_srcover_rgba_8888_sse2_lowp+0x3000650>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60434,16 +61839,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a033028 <_sk_srcover_rgba_8888_sse2_lowp+0xa0006a8>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a0346c8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000698>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12033030 <_sk_srcover_rgba_8888_sse2_lowp+0x120006b0>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 120346d0 <_sk_srcover_rgba_8888_sse2_lowp+0x120006a0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a033038 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0006b8>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a0346d8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0006a8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3033040 <_sk_srcover_rgba_8888_sse2_lowp+0x30006c0>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 30346e0 <_sk_srcover_rgba_8888_sse2_lowp+0x30006b0>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60486,16 +61891,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a033088 <_sk_srcover_rgba_8888_sse2_lowp+0xa000708>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034728 <_sk_srcover_rgba_8888_sse2_lowp+0xa0006f8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12033090 <_sk_srcover_rgba_8888_sse2_lowp+0x12000710>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12034730 <_sk_srcover_rgba_8888_sse2_lowp+0x12000700>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a033098 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000718>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a034738 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000708>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 30330a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000720>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3034740 <_sk_srcover_rgba_8888_sse2_lowp+0x3000710>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60538,16 +61943,16 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a0330e8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000768>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034788 <_sk_srcover_rgba_8888_sse2_lowp+0xa000758>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 120330f0 <_sk_srcover_rgba_8888_sse2_lowp+0x12000770>
+  .byte  255,13,255,255,255,17               // decl          0x11ffffff(%rip)        # 12034790 <_sk_srcover_rgba_8888_sse2_lowp+0x12000760>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a0330f8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000778>
+  .byte  255,21,255,255,255,25               // callq         *0x19ffffff(%rip)        # 1a034798 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000768>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 3033100 <_sk_srcover_rgba_8888_sse2_lowp+0x3000780>
+  .byte  255,29,255,255,255,2                // lcall         *0x2ffffff(%rip)        # 30347a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000770>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -60810,6 +62215,22 @@
   .byte  255,0                               // incl          (%rax)
   .byte  255,0                               // incl          (%rax)
   .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
+  .byte  255,0                               // incl          (%rax)
   .byte  248                                 // clc
   .byte  0,63                                // add           %bh,(%rdi)
   .byte  0,31                                // add           %bl,(%rdi)
@@ -60820,10 +62241,16 @@
   .byte  0,255                               // add           %bh,%bh
   .byte  0,0                                 // add           %al,(%rax)
   .byte  248                                 // clc
-  .byte  224,255                             // loopne        33313 <.const+0x533>
-  .byte  255,0                               // incl          (%rax)
-  .byte  255,0                               // incl          (%rax)
-  .byte  54,0,183,0,19,0,255                 // add           %dh,%ss:-0xffed00(%rdi)
+  .byte  224,255                             // loopne        349d3 <.const+0x553>
+  .byte  248                                 // clc
+  .byte  0,63                                // add           %bh,(%rdi)
+  .byte  0,31                                // add           %bl,(%rdi)
+  .byte  0,255                               // add           %bh,%bh
+  .byte  0,255                               // add           %bh,%bh
+  .byte  0,255                               // add           %bh,%bh
+  .byte  0,54                                // add           %dh,(%rsi)
+  .byte  0,183,0,19,0,255                    // add           %dh,-0xffed00(%rdi)
+  .byte  0,255                               // add           %bh,%bh
   .byte  0,255                               // add           %bh,%bh
   .byte  0,255                               // add           %bh,%bh
   .byte  0,255                               // add           %bh,%bh
@@ -60840,6 +62267,17 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
+  .byte  0,0                                 // add           %al,(%rax)
   .byte  255,0                               // incl          (%rax)
   .byte  255,0                               // incl          (%rax)
   .byte  255,0                               // incl          (%rax)
@@ -60950,11 +62388,11 @@
   .byte  0,128,191,0,0,128                   // add           %al,-0x7fffff41(%rax)
   .byte  191,0,0,224,64                      // mov           $0x40e00000,%edi
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        33488 <.literal16+0x118>
+  .byte  224,64                              // loopne        34b68 <.literal16+0x118>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        3348c <.literal16+0x11c>
+  .byte  224,64                              // loopne        34b6c <.literal16+0x11c>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,64                              // loopne        33490 <.literal16+0x120>
+  .byte  224,64                              // loopne        34b70 <.literal16+0x120>
   .byte  154                                 // (bad)
   .byte  153                                 // cltd
   .byte  153                                 // cltd
@@ -60974,13 +62412,13 @@
   .byte  10,23                               // or            (%rdi),%dl
   .byte  63                                  // (bad)
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 334b1 <.literal16+0x141>
+  .byte  71,225,61                           // rex.RXB       loope 34b91 <.literal16+0x141>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 334b5 <.literal16+0x145>
+  .byte  71,225,61                           // rex.RXB       loope 34b95 <.literal16+0x145>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 334b9 <.literal16+0x149>
+  .byte  71,225,61                           // rex.RXB       loope 34b99 <.literal16+0x149>
   .byte  174                                 // scas          %es:(%rdi),%al
-  .byte  71,225,61                           // rex.RXB       loope 334bd <.literal16+0x14d>
+  .byte  71,225,61                           // rex.RXB       loope 34b9d <.literal16+0x14d>
   .byte  255,0                               // incl          (%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  255,0                               // incl          (%rax)
@@ -60991,10 +62429,10 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  1,255                               // add           %edi,%edi
   .byte  255                                 // (bad)
-  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a033498 <_sk_srcover_rgba_8888_sse2_lowp+0xa000b18>
+  .byte  255,5,255,255,255,9                 // incl          0x9ffffff(%rip)        # a034b78 <_sk_srcover_rgba_8888_sse2_lowp+0xa000b48>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 30334a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000b20>
+  .byte  255,13,255,255,255,2                // decl          0x2ffffff(%rip)        # 3034b80 <_sk_srcover_rgba_8888_sse2_lowp+0x3000b50>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255,6                               // incl          (%rsi)
@@ -61009,11 +62447,11 @@
   .byte  255,0                               // incl          (%rax)
   .byte  0,127,67                            // add           %bh,0x43(%rdi)
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            334fb <.literal16+0x18b>
+  .byte  127,67                              // jg            34bdb <.literal16+0x18b>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            334ff <.literal16+0x18f>
+  .byte  127,67                              // jg            34bdf <.literal16+0x18f>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  127,67                              // jg            33503 <.literal16+0x193>
+  .byte  127,67                              // jg            34be3 <.literal16+0x193>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,127,0,0                         // cmpb          $0x0,0x0(%rdi)
   .byte  128,127,0,0                         // cmpb          $0x0,0x0(%rdi)
@@ -61169,13 +62607,13 @@
   .byte  132,55                              // test          %dh,(%rdi)
   .byte  8,33                                // or            %ah,(%rcx)
   .byte  132,55                              // test          %dh,(%rdi)
-  .byte  224,7                               // loopne        33619 <.literal16+0x2a9>
+  .byte  224,7                               // loopne        34cf9 <.literal16+0x2a9>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        3361d <.literal16+0x2ad>
+  .byte  224,7                               // loopne        34cfd <.literal16+0x2ad>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        33621 <.literal16+0x2b1>
+  .byte  224,7                               // loopne        34d01 <.literal16+0x2b1>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  224,7                               // loopne        33625 <.literal16+0x2b5>
+  .byte  224,7                               // loopne        34d05 <.literal16+0x2b5>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  33,8                                // and           %ecx,(%rax)
   .byte  2,58                                // add           (%rdx),%bh
@@ -61229,17 +62667,17 @@
   .byte  0,0                                 // add           %al,(%rax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            33694 <.literal16+0x324>
+  .byte  127,0                               // jg            34d74 <.literal16+0x324>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            33698 <.literal16+0x328>
+  .byte  127,0                               // jg            34d78 <.literal16+0x328>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            3369c <.literal16+0x32c>
+  .byte  127,0                               // jg            34d7c <.literal16+0x32c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            336a0 <.literal16+0x330>
-  .byte  119,115                             // ja            33715 <.literal16+0x3a5>
+  .byte  127,0                               // jg            34d80 <.literal16+0x330>
+  .byte  119,115                             // ja            34df5 <.literal16+0x3a5>
   .byte  248                                 // clc
   .byte  194,119,115                         // retq          $0x7377
   .byte  248                                 // clc
@@ -61250,7 +62688,7 @@
   .byte  194,117,191                         // retq          $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
-  .byte  117,191                             // jne           33679 <.literal16+0x309>
+  .byte  117,191                             // jne           34d59 <.literal16+0x309>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // (bad)
   .byte  249                                 // stc
@@ -61262,7 +62700,7 @@
   .byte  249                                 // stc
   .byte  68,180,62                           // rex.R         mov $0x3e,%spl
   .byte  163,233,220,63,163,233,220,63,163   // movabs        %eax,0xa33fdce9a33fdce9
-  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a676ba <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe9a34d3a>
+  .byte  233,220,63,163,233                  // jmpq          ffffffffe9a68d9a <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe9a34d6a>
   .byte  220,63                              // fdivrl        (%rdi)
   .byte  81                                  // push          %rcx
   .byte  140,242                             // mov           %?,%edx
@@ -61352,16 +62790,16 @@
   .byte  128,3,62                            // addb          $0x3e,(%rbx)
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           33803 <.literal16+0x493>
+  .byte  118,63                              // jbe           34ee3 <.literal16+0x493>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           33807 <.literal16+0x497>
+  .byte  118,63                              // jbe           34ee7 <.literal16+0x497>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           3380b <.literal16+0x49b>
+  .byte  118,63                              // jbe           34eeb <.literal16+0x49b>
   .byte  31                                  // (bad)
   .byte  215                                 // xlat          %ds:(%rbx)
-  .byte  118,63                              // jbe           3380f <.literal16+0x49f>
+  .byte  118,63                              // jbe           34eef <.literal16+0x49f>
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%rax)
@@ -61382,13 +62820,13 @@
   .byte  65,0,0                              // add           %al,(%r8)
   .byte  248                                 // clc
   .byte  65,0,0                              // add           %al,(%r8)
-  .byte  124,66                              // jl            33846 <.literal16+0x4d6>
+  .byte  124,66                              // jl            34f26 <.literal16+0x4d6>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            3384a <.literal16+0x4da>
+  .byte  124,66                              // jl            34f2a <.literal16+0x4da>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            3384e <.literal16+0x4de>
+  .byte  124,66                              // jl            34f2e <.literal16+0x4de>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  124,66                              // jl            33852 <.literal16+0x4e2>
+  .byte  124,66                              // jl            34f32 <.literal16+0x4e2>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,240                               // add           %dh,%al
@@ -61434,25 +62872,25 @@
   .byte  136,136,61,137,136,136              // mov           %cl,-0x777776c3(%rax)
   .byte  61,137,136,136,61                   // cmp           $0x3d888889,%eax
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            338d5 <.literal16+0x565>
+  .byte  112,65                              // jo            34fb5 <.literal16+0x565>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            338d9 <.literal16+0x569>
+  .byte  112,65                              // jo            34fb9 <.literal16+0x569>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            338dd <.literal16+0x56d>
+  .byte  112,65                              // jo            34fbd <.literal16+0x56d>
   .byte  0,0                                 // add           %al,(%rax)
-  .byte  112,65                              // jo            338e1 <.literal16+0x571>
+  .byte  112,65                              // jo            34fc1 <.literal16+0x571>
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%rax)
   .byte  0,0                                 // add           %al,(%rax)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            338b3 <.literal16+0x543>
+  .byte  127,0                               // jg            34f93 <.literal16+0x543>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            338b7 <.literal16+0x547>
+  .byte  127,0                               // jg            34f97 <.literal16+0x547>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            338bb <.literal16+0x54b>
+  .byte  127,0                               // jg            34f9b <.literal16+0x54b>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            338bf <.literal16+0x54f>
+  .byte  127,0                               // jg            34f9f <.literal16+0x54f>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  4,0                                 // add           $0x0,%al
   .byte  0,0                                 // add           %al,(%rax)
@@ -61478,13 +62916,13 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,255                             // jg            338f8 <.literal16+0x588>
+  .byte  127,255                             // jg            34fd8 <.literal16+0x588>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,255                             // jg            338fc <.literal16+0x58c>
+  .byte  127,255                             // jg            34fdc <.literal16+0x58c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            33901 <.literal16+0x591>
+  .byte  127,0                               // jg            34fe1 <.literal16+0x591>
   .byte  0,128,56,0,0,128                    // add           %al,-0x7fffffc8(%rax)
   .byte  56,0                                // cmp           %al,(%rax)
   .byte  0,128,56,0,0,128                    // add           %al,-0x7fffffc8(%rax)
@@ -61498,13 +62936,13 @@
   .byte  64,254                              // rex           (bad)
   .byte  255,0                               // incl          (%rax)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            3396b <.literal16+0x5fb>
+  .byte  127,71                              // jg            3504b <.literal16+0x5fb>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            3396f <.literal16+0x5ff>
+  .byte  127,71                              // jg            3504f <.literal16+0x5ff>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            33973 <.literal16+0x603>
+  .byte  127,71                              // jg            35053 <.literal16+0x603>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            33977 <.literal16+0x607>
+  .byte  127,71                              // jg            35057 <.literal16+0x607>
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,208                              // ds            (bad)
@@ -61534,10 +62972,10 @@
   .byte  61,152,221,147,61                   // cmp           $0x3d93dd98,%eax
   .byte  152                                 // cwtl
   .byte  221,147,61,45,16,17                 // fstl          0x11102d3d(%rbx)
-  .byte  192,45,16,17,192,45,16              // shrb          $0x10,0x2dc01110(%rip)        # 2dc34a7a <_sk_srcover_rgba_8888_sse2_lowp+0x2dc020fa>
+  .byte  192,45,16,17,192,45,16              // shrb          $0x10,0x2dc01110(%rip)        # 2dc3615a <_sk_srcover_rgba_8888_sse2_lowp+0x2dc0212a>
   .byte  17,192                              // adc           %eax,%eax
   .byte  45,16,17,192,18                     // sub           $0x12c01110,%eax
-  .byte  120,57                              // js            339ac <.literal16+0x63c>
+  .byte  120,57                              // js            3508c <.literal16+0x63c>
   .byte  64,18,120,57                        // adc           0x39(%rax),%dil
   .byte  64,18,120,57                        // adc           0x39(%rax),%dil
   .byte  64,18,120,57                        // adc           0x39(%rax),%dil
@@ -61602,13 +63040,13 @@
   .byte  192,191,0,0,192,191,0               // sarb          $0x0,-0x40400000(%rdi)
   .byte  0,192                               // add           %al,%al
   .byte  191,0,0,192,191                     // mov           $0xbfc00000,%edi
-  .byte  114,28                              // jb            33a5e <.literal16+0x6ee>
+  .byte  114,28                              // jb            3513e <.literal16+0x6ee>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         33a62 <.literal16+0x6f2>
+  .byte  62,114,28                           // jb,pt         35142 <.literal16+0x6f2>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         33a66 <.literal16+0x6f6>
+  .byte  62,114,28                           // jb,pt         35146 <.literal16+0x6f6>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         33a6a <.literal16+0x6fa>
+  .byte  62,114,28                           // jb,pt         3514a <.literal16+0x6fa>
   .byte  199                                 // (bad)
   .byte  62,85                               // ds            push %rbp
   .byte  85                                  // push          %rbp
@@ -61627,15 +63065,15 @@
   .byte  0,192                               // add           %al,%al
   .byte  63                                  // (bad)
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%rsi)
-  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d66c8b5 <_sk_srcover_rgba_8888_sse2_lowp+0x3d639f35>
+  .byte  99,61,57,142,99,61                  // movslq        0x3d638e39(%rip),%edi        # 3d66df95 <_sk_srcover_rgba_8888_sse2_lowp+0x3d639f65>
   .byte  57,142,99,61,114,249                // cmp           %ecx,-0x68dc29d(%rsi)
-  .byte  127,63                              // jg            33ac3 <.literal16+0x753>
-  .byte  114,249                             // jb            33a7f <.literal16+0x70f>
-  .byte  127,63                              // jg            33ac7 <.literal16+0x757>
-  .byte  114,249                             // jb            33a83 <.literal16+0x713>
-  .byte  127,63                              // jg            33acb <.literal16+0x75b>
-  .byte  114,249                             // jb            33a87 <.literal16+0x717>
-  .byte  127,63                              // jg            33acf <.literal16+0x75f>
+  .byte  127,63                              // jg            351a3 <.literal16+0x753>
+  .byte  114,249                             // jb            3515f <.literal16+0x70f>
+  .byte  127,63                              // jg            351a7 <.literal16+0x757>
+  .byte  114,249                             // jb            35163 <.literal16+0x713>
+  .byte  127,63                              // jg            351ab <.literal16+0x75b>
+  .byte  114,249                             // jb            35167 <.literal16+0x717>
+  .byte  127,63                              // jg            351af <.literal16+0x75f>
   .byte  3,0                                 // add           (%rax),%eax
   .byte  0,0                                 // add           %al,(%rax)
   .byte  3,0                                 // add           (%rax),%eax
@@ -61745,14 +63183,14 @@
   .byte  0,248                               // add           %bh,%al
   .byte  0,248                               // add           %bh,%al
   .byte  0,248                               // add           %bh,%al
-  .byte  224,255                             // loopne        33b61 <.literal16+0x7f1>
-  .byte  224,255                             // loopne        33b63 <.literal16+0x7f3>
-  .byte  224,255                             // loopne        33b65 <.literal16+0x7f5>
-  .byte  224,255                             // loopne        33b67 <.literal16+0x7f7>
-  .byte  224,255                             // loopne        33b69 <.literal16+0x7f9>
-  .byte  224,255                             // loopne        33b6b <.literal16+0x7fb>
-  .byte  224,255                             // loopne        33b6d <.literal16+0x7fd>
-  .byte  224,255                             // loopne        33b6f <.literal16+0x7ff>
+  .byte  224,255                             // loopne        35241 <.literal16+0x7f1>
+  .byte  224,255                             // loopne        35243 <.literal16+0x7f3>
+  .byte  224,255                             // loopne        35245 <.literal16+0x7f5>
+  .byte  224,255                             // loopne        35247 <.literal16+0x7f7>
+  .byte  224,255                             // loopne        35249 <.literal16+0x7f9>
+  .byte  224,255                             // loopne        3524b <.literal16+0x7fb>
+  .byte  224,255                             // loopne        3524d <.literal16+0x7fd>
+  .byte  224,255                             // loopne        3524f <.literal16+0x7ff>
   .byte  54,0,54                             // add           %dh,%ss:(%rsi)
   .byte  0,54                                // add           %dh,(%rsi)
   .byte  0,54                                // add           %dh,(%rsi)
@@ -61794,9 +63232,9 @@
   .byte  4,0                                 // add           $0x0,%al
   .byte  0,0                                 // add           %al,(%rax)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            33bcb <.literal8+0x1b>
+  .byte  127,0                               // jg            352ab <.literal8+0x1b>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            33bcf <.literal8+0x1f>
+  .byte  127,0                               // jg            352af <.literal8+0x1f>
   .byte  0,0                                 // add           %al,(%rax)
   .byte  128,0,0                             // addb          $0x0,(%rax)
   .byte  0                                   // .byte         0x0
@@ -61908,7 +63346,7 @@
   .byte  102,15,110,73,4                     // movd          0x4(%ecx),%xmm1
   .byte  102,15,112,201,0                    // pshufd        $0x0,%xmm1,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,88,136,119,207,0,0               // addps         0xcf77(%eax),%xmm1
+  .byte  15,88,136,151,216,0,0               // addps         0xd897(%eax),%xmm1
   .byte  141,114,8                           // lea           0x8(%edx),%esi
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
   .byte  15,41,81,64                         // movaps        %xmm2,0x40(%ecx)
@@ -61916,7 +63354,7 @@
   .byte  15,41,81,32                         // movaps        %xmm2,0x20(%ecx)
   .byte  15,41,81,16                         // movaps        %xmm2,0x10(%ecx)
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,144,135,207,0,0               // movaps        0xcf87(%eax),%xmm2
+  .byte  15,40,144,167,216,0,0               // movaps        0xd8a7(%eax),%xmm2
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
   .byte  86                                  // push          %esi
   .byte  81                                  // push          %ecx
@@ -61943,15 +63381,15 @@
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  102,15,110,32                       // movd          (%eax),%xmm4
   .byte  102,15,112,228,0                    // pshufd        $0x0,%xmm4,%xmm4
-  .byte  102,15,254,161,32,207,0,0           // paddd         0xcf20(%ecx),%xmm4
+  .byte  102,15,254,161,64,216,0,0           // paddd         0xd840(%ecx),%xmm4
   .byte  102,15,110,104,4                    // movd          0x4(%eax),%xmm5
   .byte  102,15,112,253,0                    // pshufd        $0x0,%xmm5,%xmm7
   .byte  102,15,239,252                      // pxor          %xmm4,%xmm7
-  .byte  102,15,111,177,48,207,0,0           // movdqa        0xcf30(%ecx),%xmm6
+  .byte  102,15,111,177,80,216,0,0           // movdqa        0xd850(%ecx),%xmm6
   .byte  102,15,111,207                      // movdqa        %xmm7,%xmm1
   .byte  102,15,219,206                      // pand          %xmm6,%xmm1
   .byte  102,15,114,241,5                    // pslld         $0x5,%xmm1
-  .byte  102,15,111,129,80,207,0,0           // movdqa        0xcf50(%ecx),%xmm0
+  .byte  102,15,111,129,112,216,0,0          // movdqa        0xd870(%ecx),%xmm0
   .byte  102,15,111,239                      // movdqa        %xmm7,%xmm5
   .byte  102,15,219,232                      // pand          %xmm0,%xmm5
   .byte  102,15,114,245,2                    // pslld         $0x2,%xmm5
@@ -61961,7 +63399,7 @@
   .byte  102,15,219,196                      // pand          %xmm4,%xmm0
   .byte  102,15,254,192                      // paddd         %xmm0,%xmm0
   .byte  102,15,235,198                      // por           %xmm6,%xmm0
-  .byte  102,15,111,137,64,207,0,0           // movdqa        0xcf40(%ecx),%xmm1
+  .byte  102,15,111,137,96,216,0,0           // movdqa        0xd860(%ecx),%xmm1
   .byte  102,15,219,249                      // pand          %xmm1,%xmm7
   .byte  102,15,219,225                      // pand          %xmm1,%xmm4
   .byte  102,15,114,212,2                    // psrld         $0x2,%xmm4
@@ -61972,8 +63410,8 @@
   .byte  102,15,114,215,1                    // psrld         $0x1,%xmm7
   .byte  102,15,235,231                      // por           %xmm7,%xmm4
   .byte  15,91,196                           // cvtdq2ps      %xmm4,%xmm0
-  .byte  15,89,129,96,207,0,0                // mulps         0xcf60(%ecx),%xmm0
-  .byte  15,88,129,112,207,0,0               // addps         0xcf70(%ecx),%xmm0
+  .byte  15,89,129,128,216,0,0               // mulps         0xd880(%ecx),%xmm0
+  .byte  15,88,129,144,216,0,0               // addps         0xd890(%ecx),%xmm0
   .byte  243,15,16,14                        // movss         (%esi),%xmm1
   .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
@@ -62041,7 +63479,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  141,81,4                            // lea           0x4(%ecx),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,152,212,205,0,0               // movaps        0xcdd4(%eax),%xmm3
+  .byte  15,40,152,244,214,0,0               // movaps        0xd6f4(%eax),%xmm3
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -62064,7 +63502,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  141,81,4                            // lea           0x4(%ecx),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,128,164,205,0,0               // movaps        0xcda4(%eax),%xmm0
+  .byte  15,40,128,196,214,0,0               // movaps        0xd6c4(%eax),%xmm0
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -62154,7 +63592,7 @@
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  15,40,97,64                         // movaps        0x40(%ecx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
-  .byte  15,40,170,245,204,0,0               // movaps        0xccf5(%edx),%xmm5
+  .byte  15,40,170,21,214,0,0                // movaps        0xd615(%edx),%xmm5
   .byte  15,92,235                           // subps         %xmm3,%xmm5
   .byte  15,40,113,16                        // movaps        0x10(%ecx),%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
@@ -62193,7 +63631,7 @@
   .byte  15,40,105,64                        // movaps        0x40(%ecx),%xmm5
   .byte  15,40,113,16                        // movaps        0x10(%ecx),%xmm6
   .byte  15,89,243                           // mulps         %xmm3,%xmm6
-  .byte  15,40,162,150,204,0,0               // movaps        0xcc96(%edx),%xmm4
+  .byte  15,40,162,182,213,0,0               // movaps        0xd5b6(%edx),%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  15,88,198                           // addps         %xmm6,%xmm0
@@ -62277,7 +63715,7 @@
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
-  .byte  15,40,160,213,203,0,0               // movaps        0xcbd5(%eax),%xmm4
+  .byte  15,40,160,245,212,0,0               // movaps        0xd4f5(%eax),%xmm4
   .byte  15,92,98,64                         // subps         0x40(%edx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
@@ -62303,7 +63741,7 @@
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
-  .byte  15,40,160,157,203,0,0               // movaps        0xcb9d(%eax),%xmm4
+  .byte  15,40,160,189,212,0,0               // movaps        0xd4bd(%eax),%xmm4
   .byte  15,92,227                           // subps         %xmm3,%xmm4
   .byte  15,40,66,16                         // movaps        0x10(%edx),%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
@@ -62333,7 +63771,7 @@
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
-  .byte  15,40,160,86,203,0,0                // movaps        0xcb56(%eax),%xmm4
+  .byte  15,40,160,118,212,0,0               // movaps        0xd476(%eax),%xmm4
   .byte  15,92,227                           // subps         %xmm3,%xmm4
   .byte  15,40,106,16                        // movaps        0x10(%edx),%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
@@ -62367,7 +63805,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
   .byte  15,40,106,64                        // movaps        0x40(%edx),%xmm5
-  .byte  15,40,160,6,203,0,0                 // movaps        0xcb06(%eax),%xmm4
+  .byte  15,40,160,38,212,0,0                // movaps        0xd426(%eax),%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  15,88,66,16                         // addps         0x10(%edx),%xmm0
@@ -62423,7 +63861,7 @@
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  15,40,80,64                         // movaps        0x40(%eax),%xmm2
   .byte  15,41,85,200                        // movaps        %xmm2,-0x38(%ebp)
-  .byte  15,40,129,135,202,0,0               // movaps        0xca87(%ecx),%xmm0
+  .byte  15,40,129,167,211,0,0               // movaps        0xd3a7(%ecx),%xmm0
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,92,242                           // subps         %xmm2,%xmm6
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
@@ -62483,7 +63921,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
   .byte  15,88,66,16                         // addps         0x10(%edx),%xmm0
-  .byte  15,40,160,225,201,0,0               // movaps        0xc9e1(%eax),%xmm4
+  .byte  15,40,160,1,211,0,0                 // movaps        0xd301(%eax),%xmm4
   .byte  15,93,196                           // minps         %xmm4,%xmm0
   .byte  15,88,74,32                         // addps         0x20(%edx),%xmm1
   .byte  15,93,204                           // minps         %xmm4,%xmm1
@@ -62555,7 +63993,7 @@
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  15,40,105,64                        // movaps        0x40(%ecx),%xmm5
-  .byte  15,40,154,51,201,0,0                // movaps        0xc933(%edx),%xmm3
+  .byte  15,40,154,83,210,0,0                // movaps        0xd253(%edx),%xmm3
   .byte  15,40,243                           // movaps        %xmm3,%xmm6
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
@@ -62617,7 +64055,7 @@
   .byte  15,89,203                           // mulps         %xmm3,%xmm1
   .byte  15,95,209                           // maxps         %xmm1,%xmm2
   .byte  15,92,234                           // subps         %xmm2,%xmm5
-  .byte  15,40,138,203,200,0,0               // movaps        0xc8cb(%edx),%xmm1
+  .byte  15,40,138,235,209,0,0               // movaps        0xd1eb(%edx),%xmm1
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -62666,7 +64104,7 @@
   .byte  15,89,203                           // mulps         %xmm3,%xmm1
   .byte  15,93,209                           // minps         %xmm1,%xmm2
   .byte  15,92,234                           // subps         %xmm2,%xmm5
-  .byte  15,40,138,75,200,0,0                // movaps        0xc84b(%edx),%xmm1
+  .byte  15,40,138,107,209,0,0               // movaps        0xd16b(%edx),%xmm1
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -62719,7 +64157,7 @@
   .byte  15,93,214                           // minps         %xmm6,%xmm2
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  15,92,226                           // subps         %xmm2,%xmm4
-  .byte  15,40,146,200,199,0,0               // movaps        0xc7c8(%edx),%xmm2
+  .byte  15,40,146,232,208,0,0               // movaps        0xd0e8(%edx),%xmm2
   .byte  15,92,211                           // subps         %xmm3,%xmm2
   .byte  15,89,213                           // mulps         %xmm5,%xmm2
   .byte  15,88,218                           // addps         %xmm2,%xmm3
@@ -62763,7 +64201,7 @@
   .byte  15,89,202                           // mulps         %xmm2,%xmm1
   .byte  15,88,201                           // addps         %xmm1,%xmm1
   .byte  15,92,233                           // subps         %xmm1,%xmm5
-  .byte  15,40,138,66,199,0,0                // movaps        0xc742(%edx),%xmm1
+  .byte  15,40,138,98,208,0,0                // movaps        0xd062(%edx),%xmm1
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,89,73,64                         // mulps         0x40(%ecx),%xmm1
   .byte  15,88,217                           // addps         %xmm1,%xmm3
@@ -62804,7 +64242,7 @@
   .byte  15,93,198                           // minps         %xmm6,%xmm0
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  15,92,232                           // subps         %xmm0,%xmm5
-  .byte  15,40,177,195,198,0,0               // movaps        0xc6c3(%ecx),%xmm6
+  .byte  15,40,177,227,207,0,0               // movaps        0xcfe3(%ecx),%xmm6
   .byte  15,40,198                           // movaps        %xmm6,%xmm0
   .byte  15,92,193                           // subps         %xmm1,%xmm0
   .byte  15,41,69,200                        // movaps        %xmm0,-0x38(%ebp)
@@ -62921,7 +64359,7 @@
   .byte  89                                  // pop           %ecx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  15,40,72,16                         // movaps        0x10(%eax),%xmm1
-  .byte  15,40,145,58,197,0,0                // movaps        0xc53a(%ecx),%xmm2
+  .byte  15,40,145,90,206,0,0                // movaps        0xce5a(%ecx),%xmm2
   .byte  15,40,226                           // movaps        %xmm2,%xmm4
   .byte  15,92,214                           // subps         %xmm6,%xmm2
   .byte  15,40,250                           // movaps        %xmm2,%xmm7
@@ -63038,7 +64476,7 @@
   .byte  89                                  // pop           %ecx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  15,40,72,64                         // movaps        0x40(%eax),%xmm1
-  .byte  15,40,145,217,195,0,0               // movaps        0xc3d9(%ecx),%xmm2
+  .byte  15,40,145,249,204,0,0               // movaps        0xccf9(%ecx),%xmm2
   .byte  15,40,234                           // movaps        %xmm2,%xmm5
   .byte  15,92,233                           // subps         %xmm1,%xmm5
   .byte  15,40,225                           // movaps        %xmm1,%xmm4
@@ -63148,7 +64586,7 @@
   .byte  89                                  // pop           %ecx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  15,40,80,64                         // movaps        0x40(%eax),%xmm2
-  .byte  15,40,161,142,194,0,0               // movaps        0xc28e(%ecx),%xmm4
+  .byte  15,40,161,174,203,0,0               // movaps        0xcbae(%ecx),%xmm4
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  15,92,242                           // subps         %xmm2,%xmm6
   .byte  15,40,206                           // movaps        %xmm6,%xmm1
@@ -63276,17 +64714,17 @@
   .byte  232,0,0,0,0                         // call          f52 <_sk_softlight_sse2+0x52>
   .byte  89                                  // pop           %ecx
   .byte  15,40,194                           // movaps        %xmm2,%xmm0
-  .byte  15,40,137,110,193,0,0               // movaps        0xc16e(%ecx),%xmm1
+  .byte  15,40,137,142,202,0,0               // movaps        0xca8e(%ecx),%xmm1
   .byte  15,41,141,104,255,255,255           // movaps        %xmm1,-0x98(%ebp)
   .byte  15,88,193                           // addps         %xmm1,%xmm0
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
-  .byte  15,40,185,254,192,0,0               // movaps        0xc0fe(%ecx),%xmm7
+  .byte  15,40,185,30,202,0,0                // movaps        0xca1e(%ecx),%xmm7
   .byte  15,41,125,232                       // movaps        %xmm7,-0x18(%ebp)
   .byte  15,92,250                           // subps         %xmm2,%xmm7
   .byte  15,82,242                           // rsqrtps       %xmm2,%xmm6
   .byte  15,83,206                           // rcpps         %xmm6,%xmm1
   .byte  15,92,202                           // subps         %xmm2,%xmm1
-  .byte  15,40,161,126,193,0,0               // movaps        0xc17e(%ecx),%xmm4
+  .byte  15,40,161,158,202,0,0               // movaps        0xca9e(%ecx),%xmm4
   .byte  15,41,101,152                       // movaps        %xmm4,-0x68(%ebp)
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  15,88,208                           // addps         %xmm0,%xmm2
@@ -63492,7 +64930,7 @@
   .byte  15,40,226                           // movaps        %xmm2,%xmm4
   .byte  15,40,72,64                         // movaps        0x40(%eax),%xmm1
   .byte  15,41,77,216                        // movaps        %xmm1,-0x28(%ebp)
-  .byte  15,40,169,61,190,0,0                // movaps        0xbe3d(%ecx),%xmm5
+  .byte  15,40,169,93,199,0,0                // movaps        0xc75d(%ecx),%xmm5
   .byte  15,92,233                           // subps         %xmm1,%xmm5
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
   .byte  15,41,133,72,255,255,255            // movaps        %xmm0,-0xb8(%ebp)
@@ -63525,17 +64963,17 @@
   .byte  15,89,233                           // mulps         %xmm1,%xmm5
   .byte  15,41,173,120,255,255,255           // movaps        %xmm5,-0x88(%ebp)
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,153,205,190,0,0               // movaps        0xbecd(%ecx),%xmm3
+  .byte  15,40,153,237,199,0,0               // movaps        0xc7ed(%ecx),%xmm3
   .byte  15,41,93,232                        // movaps        %xmm3,-0x18(%ebp)
   .byte  15,40,117,152                       // movaps        -0x68(%ebp),%xmm6
   .byte  15,40,206                           // movaps        %xmm6,%xmm1
   .byte  15,89,203                           // mulps         %xmm3,%xmm1
-  .byte  15,40,153,221,190,0,0               // movaps        0xbedd(%ecx),%xmm3
+  .byte  15,40,153,253,199,0,0               // movaps        0xc7fd(%ecx),%xmm3
   .byte  15,40,109,136                       // movaps        -0x78(%ebp),%xmm5
   .byte  15,40,253                           // movaps        %xmm5,%xmm7
   .byte  15,89,251                           // mulps         %xmm3,%xmm7
   .byte  15,88,249                           // addps         %xmm1,%xmm7
-  .byte  15,40,137,61,190,0,0                // movaps        0xbe3d(%ecx),%xmm1
+  .byte  15,40,137,93,199,0,0                // movaps        0xc75d(%ecx),%xmm1
   .byte  15,92,77,184                        // subps         -0x48(%ebp),%xmm1
   .byte  15,89,241                           // mulps         %xmm1,%xmm6
   .byte  15,41,117,152                       // movaps        %xmm6,-0x68(%ebp)
@@ -63545,7 +64983,7 @@
   .byte  15,40,77,200                        // movaps        -0x38(%ebp),%xmm1
   .byte  15,89,241                           // mulps         %xmm1,%xmm6
   .byte  15,41,117,168                       // movaps        %xmm6,-0x58(%ebp)
-  .byte  15,40,169,237,190,0,0               // movaps        0xbeed(%ecx),%xmm5
+  .byte  15,40,169,13,200,0,0                // movaps        0xc80d(%ecx),%xmm5
   .byte  15,89,205                           // mulps         %xmm5,%xmm1
   .byte  15,88,207                           // addps         %xmm7,%xmm1
   .byte  15,94,212                           // divps         %xmm4,%xmm2
@@ -63738,17 +65176,17 @@
   .byte  15,89,253                           // mulps         %xmm5,%xmm7
   .byte  232,0,0,0,0                         // call          1567 <_sk_saturation_sse2+0xb0>
   .byte  89                                  // pop           %ecx
-  .byte  15,40,153,121,187,0,0               // movaps        0xbb79(%ecx),%xmm3
+  .byte  15,40,153,153,196,0,0               // movaps        0xc499(%ecx),%xmm3
   .byte  15,41,157,88,255,255,255            // movaps        %xmm3,-0xa8(%ebp)
   .byte  15,40,77,184                        // movaps        -0x48(%ebp),%xmm1
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,40,153,137,187,0,0               // movaps        0xbb89(%ecx),%xmm3
+  .byte  15,40,153,169,196,0,0               // movaps        0xc4a9(%ecx),%xmm3
   .byte  15,41,93,232                        // movaps        %xmm3,-0x18(%ebp)
   .byte  15,40,238                           // movaps        %xmm6,%xmm5
   .byte  15,89,235                           // mulps         %xmm3,%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  15,40,129,233,186,0,0               // movaps        0xbae9(%ecx),%xmm0
+  .byte  15,40,129,9,196,0,0                 // movaps        0xc409(%ecx),%xmm0
   .byte  15,41,133,104,255,255,255           // movaps        %xmm0,-0x98(%ebp)
   .byte  15,92,69,216                        // subps         -0x28(%ebp),%xmm0
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
@@ -63760,7 +65198,7 @@
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  15,41,141,120,255,255,255           // movaps        %xmm1,-0x88(%ebp)
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
-  .byte  15,40,129,153,187,0,0               // movaps        0xbb99(%ecx),%xmm0
+  .byte  15,40,129,185,196,0,0               // movaps        0xc4b9(%ecx),%xmm0
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  15,88,205                           // addps         %xmm5,%xmm1
   .byte  15,40,109,168                       // movaps        -0x58(%ebp),%xmm5
@@ -63919,16 +65357,16 @@
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  15,40,80,16                         // movaps        0x10(%eax),%xmm2
   .byte  15,40,72,32                         // movaps        0x20(%eax),%xmm1
-  .byte  15,40,177,30,185,0,0                // movaps        0xb91e(%ecx),%xmm6
+  .byte  15,40,177,62,194,0,0                // movaps        0xc23e(%ecx),%xmm6
   .byte  15,41,117,216                       // movaps        %xmm6,-0x28(%ebp)
   .byte  15,40,194                           // movaps        %xmm2,%xmm0
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
-  .byte  15,40,153,46,185,0,0                // movaps        0xb92e(%ecx),%xmm3
+  .byte  15,40,153,78,194,0,0                // movaps        0xc24e(%ecx),%xmm3
   .byte  15,41,93,184                        // movaps        %xmm3,-0x48(%ebp)
   .byte  15,40,241                           // movaps        %xmm1,%xmm6
   .byte  15,89,243                           // mulps         %xmm3,%xmm6
   .byte  15,88,240                           // addps         %xmm0,%xmm6
-  .byte  15,40,129,142,184,0,0               // movaps        0xb88e(%ecx),%xmm0
+  .byte  15,40,129,174,193,0,0               // movaps        0xc1ae(%ecx),%xmm0
   .byte  15,41,69,168                        // movaps        %xmm0,-0x58(%ebp)
   .byte  15,92,69,200                        // subps         -0x38(%ebp),%xmm0
   .byte  15,89,208                           // mulps         %xmm0,%xmm2
@@ -63938,7 +65376,7 @@
   .byte  15,40,80,48                         // movaps        0x30(%eax),%xmm2
   .byte  15,89,194                           // mulps         %xmm2,%xmm0
   .byte  15,41,133,72,255,255,255            // movaps        %xmm0,-0xb8(%ebp)
-  .byte  15,40,137,62,185,0,0                // movaps        0xb93e(%ecx),%xmm1
+  .byte  15,40,137,94,194,0,0                // movaps        0xc25e(%ecx),%xmm1
   .byte  15,89,209                           // mulps         %xmm1,%xmm2
   .byte  15,88,214                           // addps         %xmm6,%xmm2
   .byte  15,40,88,64                         // movaps        0x40(%eax),%xmm3
@@ -64097,18 +65535,18 @@
   .byte  232,0,0,0,0                         // call          1a25 <_sk_luminosity_sse2+0x11>
   .byte  89                                  // pop           %ecx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
-  .byte  15,40,161,187,182,0,0               // movaps        0xb6bb(%ecx),%xmm4
+  .byte  15,40,161,219,191,0,0               // movaps        0xbfdb(%ecx),%xmm4
   .byte  15,41,101,232                       // movaps        %xmm4,-0x18(%ebp)
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
-  .byte  15,40,161,203,182,0,0               // movaps        0xb6cb(%ecx),%xmm4
+  .byte  15,40,161,235,191,0,0               // movaps        0xbfeb(%ecx),%xmm4
   .byte  15,41,101,216                       // movaps        %xmm4,-0x28(%ebp)
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
   .byte  15,88,234                           // addps         %xmm2,%xmm5
   .byte  15,40,80,64                         // movaps        0x40(%eax),%xmm2
   .byte  15,41,85,184                        // movaps        %xmm2,-0x48(%ebp)
-  .byte  15,40,161,43,182,0,0                // movaps        0xb62b(%ecx),%xmm4
+  .byte  15,40,161,75,191,0,0                // movaps        0xbf4b(%ecx),%xmm4
   .byte  15,40,252                           // movaps        %xmm4,%xmm7
   .byte  15,92,250                           // subps         %xmm2,%xmm7
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
@@ -64118,7 +65556,7 @@
   .byte  15,89,254                           // mulps         %xmm6,%xmm7
   .byte  15,41,189,104,255,255,255           // movaps        %xmm7,-0x98(%ebp)
   .byte  15,40,214                           // movaps        %xmm6,%xmm2
-  .byte  15,40,137,219,182,0,0               // movaps        0xb6db(%ecx),%xmm1
+  .byte  15,40,137,251,191,0,0               // movaps        0xbffb(%ecx),%xmm1
   .byte  15,89,209                           // mulps         %xmm1,%xmm2
   .byte  15,88,213                           // addps         %xmm5,%xmm2
   .byte  15,40,104,16                        // movaps        0x10(%eax),%xmm5
@@ -64294,7 +65732,7 @@
   .byte  139,49                              // mov           (%ecx),%esi
   .byte  15,133,224,0,0,0                    // jne           1d92 <_sk_srcover_rgba_8888_sse2+0x119>
   .byte  243,15,111,20,178                   // movdqu        (%edx,%esi,4),%xmm2
-  .byte  102,15,111,167,128,180,0,0          // movdqa        0xb480(%edi),%xmm4
+  .byte  102,15,111,167,160,189,0,0          // movdqa        0xbda0(%edi),%xmm4
   .byte  102,15,111,234                      // movdqa        %xmm2,%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  102,15,111,242                      // movdqa        %xmm2,%xmm6
@@ -64305,9 +65743,9 @@
   .byte  102,15,219,252                      // pand          %xmm4,%xmm7
   .byte  15,91,229                           // cvtdq2ps      %xmm5,%xmm4
   .byte  15,41,97,16                         // movaps        %xmm4,0x10(%ecx)
-  .byte  15,40,175,192,179,0,0               // movaps        0xb3c0(%edi),%xmm5
+  .byte  15,40,175,224,188,0,0               // movaps        0xbce0(%edi),%xmm5
   .byte  15,92,235                           // subps         %xmm3,%xmm5
-  .byte  15,40,143,144,180,0,0               // movaps        0xb490(%edi),%xmm1
+  .byte  15,40,143,176,189,0,0               // movaps        0xbdb0(%edi),%xmm1
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,88,224                           // addps         %xmm0,%xmm4
@@ -64421,7 +65859,7 @@
   .byte  232,0,0,0,0                         // call          1e3c <_sk_clamp_1_sse2+0xb>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,40,160,20,178,0,0                // movaps        0xb214(%eax),%xmm4
+  .byte  15,40,160,52,187,0,0                // movaps        0xbb34(%eax),%xmm4
   .byte  15,93,196                           // minps         %xmm4,%xmm0
   .byte  15,93,204                           // minps         %xmm4,%xmm1
   .byte  15,93,212                           // minps         %xmm4,%xmm2
@@ -64445,7 +65883,7 @@
   .byte  232,0,0,0,0                         // call          1e6f <_sk_clamp_a_sse2+0xb>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,93,152,225,177,0,0               // minps         0xb1e1(%eax),%xmm3
+  .byte  15,93,152,1,187,0,0                 // minps         0xbb01(%eax),%xmm3
   .byte  15,93,195                           // minps         %xmm3,%xmm0
   .byte  15,93,203                           // minps         %xmm3,%xmm1
   .byte  15,93,211                           // minps         %xmm3,%xmm2
@@ -64470,7 +65908,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,85,8                            // mov           0x8(%ebp),%edx
   .byte  15,40,98,64                         // movaps        0x40(%edx),%xmm4
-  .byte  15,93,160,177,177,0,0               // minps         0xb1b1(%eax),%xmm4
+  .byte  15,93,160,209,186,0,0               // minps         0xbad1(%eax),%xmm4
   .byte  15,41,98,64                         // movaps        %xmm4,0x40(%edx)
   .byte  15,40,106,16                        // movaps        0x10(%edx),%xmm5
   .byte  15,40,114,32                        // movaps        0x20(%edx),%xmm6
@@ -64544,7 +65982,7 @@
   .byte  232,0,0,0,0                         // call          1f4a <_sk_invert_sse2+0xb>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,40,160,6,177,0,0                 // movaps        0xb106(%eax),%xmm4
+  .byte  15,40,160,38,186,0,0                // movaps        0xba26(%eax),%xmm4
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,92,232                           // subps         %xmm0,%xmm5
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
@@ -64666,10 +66104,10 @@
   .byte  232,0,0,0,0                         // call          2048 <_sk_unpremul_sse2+0xb>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,40,160,8,176,0,0                 // movaps        0xb008(%eax),%xmm4
+  .byte  15,40,160,40,185,0,0                // movaps        0xb928(%eax),%xmm4
   .byte  15,94,227                           // divps         %xmm3,%xmm4
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
-  .byte  15,194,168,232,176,0,0,1            // cmpltps       0xb0e8(%eax),%xmm5
+  .byte  15,194,168,8,186,0,0,1              // cmpltps       0xba08(%eax),%xmm5
   .byte  15,84,236                           // andps         %xmm4,%xmm5
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
   .byte  15,89,205                           // mulps         %xmm5,%xmm1
@@ -64698,19 +66136,19 @@
   .byte  88                                  // pop           %eax
   .byte  15,40,202                           // movaps        %xmm2,%xmm1
   .byte  15,89,201                           // mulps         %xmm1,%xmm1
-  .byte  15,40,160,74,176,0,0                // movaps        0xb04a(%eax),%xmm4
+  .byte  15,40,160,106,185,0,0               // movaps        0xb96a(%eax),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
-  .byte  15,40,176,186,176,0,0               // movaps        0xb0ba(%eax),%xmm6
+  .byte  15,40,176,218,185,0,0               // movaps        0xb9da(%eax),%xmm6
   .byte  15,41,117,216                       // movaps        %xmm6,-0x28(%ebp)
   .byte  15,88,198                           // addps         %xmm6,%xmm0
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,176,170,176,0,0               // movaps        0xb0aa(%eax),%xmm6
+  .byte  15,40,176,202,185,0,0               // movaps        0xb9ca(%eax),%xmm6
   .byte  15,40,202                           // movaps        %xmm2,%xmm1
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
-  .byte  15,40,184,202,176,0,0               // movaps        0xb0ca(%eax),%xmm7
+  .byte  15,40,184,234,185,0,0               // movaps        0xb9ea(%eax),%xmm7
   .byte  15,41,125,232                       // movaps        %xmm7,-0x18(%ebp)
   .byte  15,88,199                           // addps         %xmm7,%xmm0
-  .byte  15,40,184,218,176,0,0               // movaps        0xb0da(%eax),%xmm7
+  .byte  15,40,184,250,185,0,0               // movaps        0xb9fa(%eax),%xmm7
   .byte  15,194,215,1                        // cmpltps       %xmm7,%xmm2
   .byte  15,84,202                           // andps         %xmm2,%xmm1
   .byte  15,85,208                           // andnps        %xmm0,%xmm2
@@ -64770,18 +66208,18 @@
   .byte  15,40,72,16                         // movaps        0x10(%eax),%xmm1
   .byte  15,40,233                           // movaps        %xmm1,%xmm5
   .byte  15,89,237                           // mulps         %xmm5,%xmm5
-  .byte  15,40,161,111,175,0,0               // movaps        0xaf6f(%ecx),%xmm4
+  .byte  15,40,161,143,184,0,0               // movaps        0xb88f(%ecx),%xmm4
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
-  .byte  15,40,177,223,175,0,0               // movaps        0xafdf(%ecx),%xmm6
+  .byte  15,40,177,255,184,0,0               // movaps        0xb8ff(%ecx),%xmm6
   .byte  15,88,198                           // addps         %xmm6,%xmm0
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
-  .byte  15,40,169,207,175,0,0               // movaps        0xafcf(%ecx),%xmm5
+  .byte  15,40,169,239,184,0,0               // movaps        0xb8ef(%ecx),%xmm5
   .byte  15,40,209                           // movaps        %xmm1,%xmm2
   .byte  15,89,213                           // mulps         %xmm5,%xmm2
-  .byte  15,40,185,239,175,0,0               // movaps        0xafef(%ecx),%xmm7
+  .byte  15,40,185,15,185,0,0                // movaps        0xb90f(%ecx),%xmm7
   .byte  15,88,199                           // addps         %xmm7,%xmm0
-  .byte  15,40,153,255,175,0,0               // movaps        0xafff(%ecx),%xmm3
+  .byte  15,40,153,31,185,0,0                // movaps        0xb91f(%ecx),%xmm3
   .byte  15,194,203,1                        // cmpltps       %xmm3,%xmm1
   .byte  15,84,209                           // andps         %xmm1,%xmm2
   .byte  15,85,200                           // andnps        %xmm0,%xmm1
@@ -64843,25 +66281,25 @@
   .byte  232,0,0,0,0                         // call          225c <_sk_to_srgb_sse2+0x18>
   .byte  88                                  // pop           %eax
   .byte  15,82,218                           // rsqrtps       %xmm2,%xmm3
-  .byte  15,40,160,52,175,0,0                // movaps        0xaf34(%eax),%xmm4
+  .byte  15,40,160,84,184,0,0                // movaps        0xb854(%eax),%xmm4
   .byte  15,40,235                           // movaps        %xmm3,%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
-  .byte  15,40,128,68,175,0,0                // movaps        0xaf44(%eax),%xmm0
+  .byte  15,40,128,100,184,0,0               // movaps        0xb864(%eax),%xmm0
   .byte  15,41,69,232                        // movaps        %xmm0,-0x18(%ebp)
   .byte  15,88,232                           // addps         %xmm0,%xmm5
   .byte  15,89,235                           // mulps         %xmm3,%xmm5
-  .byte  15,40,128,84,175,0,0                // movaps        0xaf54(%eax),%xmm0
+  .byte  15,40,128,116,184,0,0               // movaps        0xb874(%eax),%xmm0
   .byte  15,41,69,200                        // movaps        %xmm0,-0x38(%ebp)
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  15,40,128,100,175,0,0               // movaps        0xaf64(%eax),%xmm0
+  .byte  15,40,128,132,184,0,0               // movaps        0xb884(%eax),%xmm0
   .byte  15,41,69,216                        // movaps        %xmm0,-0x28(%ebp)
   .byte  15,88,216                           // addps         %xmm0,%xmm3
   .byte  15,83,195                           // rcpps         %xmm3,%xmm0
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
-  .byte  15,40,168,36,175,0,0                // movaps        0xaf24(%eax),%xmm5
+  .byte  15,40,168,68,184,0,0                // movaps        0xb844(%eax),%xmm5
   .byte  15,40,202                           // movaps        %xmm2,%xmm1
   .byte  15,89,205                           // mulps         %xmm5,%xmm1
-  .byte  15,40,152,116,175,0,0               // movaps        0xaf74(%eax),%xmm3
+  .byte  15,40,152,148,184,0,0               // movaps        0xb894(%eax),%xmm3
   .byte  15,194,211,1                        // cmpltps       %xmm3,%xmm2
   .byte  15,84,202                           // andps         %xmm2,%xmm1
   .byte  15,85,208                           // andnps        %xmm0,%xmm2
@@ -64932,14 +66370,14 @@
   .byte  15,40,195                           // movaps        %xmm3,%xmm0
   .byte  15,92,193                           // subps         %xmm1,%xmm0
   .byte  15,41,69,216                        // movaps        %xmm0,-0x28(%ebp)
-  .byte  15,40,136,253,172,0,0               // movaps        0xacfd(%eax),%xmm1
+  .byte  15,40,136,29,182,0,0                // movaps        0xb61d(%eax),%xmm1
   .byte  15,94,200                           // divps         %xmm0,%xmm1
   .byte  15,40,199                           // movaps        %xmm7,%xmm0
   .byte  15,92,194                           // subps         %xmm2,%xmm0
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
   .byte  15,40,239                           // movaps        %xmm7,%xmm5
   .byte  15,194,234,1                        // cmpltps       %xmm2,%xmm5
-  .byte  15,84,168,141,174,0,0               // andps         0xae8d(%eax),%xmm5
+  .byte  15,84,168,173,183,0,0               // andps         0xb7ad(%eax),%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
   .byte  15,40,195                           // movaps        %xmm3,%xmm0
   .byte  15,194,196,0                        // cmpeqps       %xmm4,%xmm0
@@ -64949,9 +66387,9 @@
   .byte  15,92,231                           // subps         %xmm7,%xmm4
   .byte  15,89,209                           // mulps         %xmm1,%xmm2
   .byte  15,89,225                           // mulps         %xmm1,%xmm4
-  .byte  15,40,136,157,174,0,0               // movaps        0xae9d(%eax),%xmm1
+  .byte  15,40,136,189,183,0,0               // movaps        0xb7bd(%eax),%xmm1
   .byte  15,88,209                           // addps         %xmm1,%xmm2
-  .byte  15,88,160,173,174,0,0               // addps         0xaead(%eax),%xmm4
+  .byte  15,88,160,205,183,0,0               // addps         0xb7cd(%eax),%xmm4
   .byte  15,84,214                           // andps         %xmm6,%xmm2
   .byte  15,85,244                           // andnps        %xmm4,%xmm6
   .byte  15,86,242                           // orps          %xmm2,%xmm6
@@ -64963,7 +66401,7 @@
   .byte  15,194,226,4                        // cmpneqps      %xmm2,%xmm4
   .byte  15,92,203                           // subps         %xmm3,%xmm1
   .byte  15,88,218                           // addps         %xmm2,%xmm3
-  .byte  15,40,168,237,172,0,0               // movaps        0xaced(%eax),%xmm5
+  .byte  15,40,168,13,182,0,0                // movaps        0xb60d(%eax),%xmm5
   .byte  15,92,202                           // subps         %xmm2,%xmm1
   .byte  15,40,211                           // movaps        %xmm3,%xmm2
   .byte  15,89,213                           // mulps         %xmm5,%xmm2
@@ -64973,7 +66411,7 @@
   .byte  15,86,233                           // orps          %xmm1,%xmm5
   .byte  15,40,77,216                        // movaps        -0x28(%ebp),%xmm1
   .byte  15,94,205                           // divps         %xmm5,%xmm1
-  .byte  15,89,128,189,174,0,0               // mulps         0xaebd(%eax),%xmm0
+  .byte  15,89,128,221,183,0,0               // mulps         0xb7dd(%eax),%xmm0
   .byte  15,84,196                           // andps         %xmm4,%xmm0
   .byte  15,84,204                           // andps         %xmm4,%xmm1
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
@@ -64998,7 +66436,7 @@
   .byte  15,41,69,184                        // movaps        %xmm0,-0x48(%ebp)
   .byte  232,0,0,0,0                         // call          244e <_sk_hsl_to_rgb_sse2+0x19>
   .byte  88                                  // pop           %eax
-  .byte  15,40,160,242,171,0,0               // movaps        0xabf2(%eax),%xmm4
+  .byte  15,40,160,18,181,0,0                // movaps        0xb512(%eax),%xmm4
   .byte  15,40,220                           // movaps        %xmm4,%xmm3
   .byte  15,41,165,104,255,255,255           // movaps        %xmm4,-0x98(%ebp)
   .byte  15,194,218,2                        // cmpleps       %xmm2,%xmm3
@@ -65011,13 +66449,13 @@
   .byte  15,84,243                           // andps         %xmm3,%xmm6
   .byte  15,85,217                           // andnps        %xmm1,%xmm3
   .byte  15,86,222                           // orps          %xmm6,%xmm3
-  .byte  15,40,168,210,173,0,0               // movaps        0xadd2(%eax),%xmm5
+  .byte  15,40,168,242,182,0,0               // movaps        0xb6f2(%eax),%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
   .byte  243,15,91,205                       // cvttps2dq     %xmm5,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,194,193,1                        // cmpltps       %xmm1,%xmm0
-  .byte  15,40,176,2,172,0,0                 // movaps        0xac02(%eax),%xmm6
+  .byte  15,40,176,34,181,0,0                // movaps        0xb522(%eax),%xmm6
   .byte  15,41,117,136                       // movaps        %xmm6,-0x78(%ebp)
   .byte  15,84,198                           // andps         %xmm6,%xmm0
   .byte  15,92,200                           // subps         %xmm0,%xmm1
@@ -65027,19 +66465,19 @@
   .byte  15,41,93,216                        // movaps        %xmm3,-0x28(%ebp)
   .byte  15,88,210                           // addps         %xmm2,%xmm2
   .byte  15,92,211                           // subps         %xmm3,%xmm2
-  .byte  15,40,184,226,173,0,0               // movaps        0xade2(%eax),%xmm7
+  .byte  15,40,184,2,183,0,0                 // movaps        0xb702(%eax),%xmm7
   .byte  15,41,189,88,255,255,255            // movaps        %xmm7,-0xa8(%ebp)
   .byte  15,194,253,2                        // cmpleps       %xmm5,%xmm7
   .byte  15,92,218                           // subps         %xmm2,%xmm3
   .byte  15,41,93,200                        // movaps        %xmm3,-0x38(%ebp)
   .byte  15,40,204                           // movaps        %xmm4,%xmm1
   .byte  15,194,205,2                        // cmpleps       %xmm5,%xmm1
-  .byte  15,40,160,194,173,0,0               // movaps        0xadc2(%eax),%xmm4
+  .byte  15,40,160,226,182,0,0               // movaps        0xb6e2(%eax),%xmm4
   .byte  15,41,165,120,255,255,255           // movaps        %xmm4,-0x88(%ebp)
   .byte  15,194,229,2                        // cmpleps       %xmm5,%xmm4
-  .byte  15,40,176,146,173,0,0               // movaps        0xad92(%eax),%xmm6
+  .byte  15,40,176,178,182,0,0               // movaps        0xb6b2(%eax),%xmm6
   .byte  15,89,238                           // mulps         %xmm6,%xmm5
-  .byte  15,40,128,178,173,0,0               // movaps        0xadb2(%eax),%xmm0
+  .byte  15,40,128,210,182,0,0               // movaps        0xb6d2(%eax),%xmm0
   .byte  15,41,69,168                        // movaps        %xmm0,-0x58(%ebp)
   .byte  15,92,197                           // subps         %xmm5,%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
@@ -65065,18 +66503,18 @@
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,203                           // movaps        %xmm3,%xmm1
   .byte  15,194,200,1                        // cmpltps       %xmm0,%xmm1
-  .byte  15,84,136,2,172,0,0                 // andps         0xac02(%eax),%xmm1
+  .byte  15,84,136,34,181,0,0                // andps         0xb522(%eax),%xmm1
   .byte  15,92,193                           // subps         %xmm1,%xmm0
   .byte  15,40,203                           // movaps        %xmm3,%xmm1
   .byte  15,92,200                           // subps         %xmm0,%xmm1
-  .byte  15,40,128,226,173,0,0               // movaps        0xade2(%eax),%xmm0
+  .byte  15,40,128,2,183,0,0                 // movaps        0xb702(%eax),%xmm0
   .byte  15,194,193,2                        // cmpleps       %xmm1,%xmm0
-  .byte  15,40,152,242,171,0,0               // movaps        0xabf2(%eax),%xmm3
+  .byte  15,40,152,18,181,0,0                // movaps        0xb512(%eax),%xmm3
   .byte  15,194,217,2                        // cmpleps       %xmm1,%xmm3
-  .byte  15,40,160,194,173,0,0               // movaps        0xadc2(%eax),%xmm4
+  .byte  15,40,160,226,182,0,0               // movaps        0xb6e2(%eax),%xmm4
   .byte  15,194,225,2                        // cmpleps       %xmm1,%xmm4
   .byte  15,89,206                           // mulps         %xmm6,%xmm1
-  .byte  15,40,168,178,173,0,0               // movaps        0xadb2(%eax),%xmm5
+  .byte  15,40,168,210,182,0,0               // movaps        0xb6d2(%eax),%xmm5
   .byte  15,92,233                           // subps         %xmm1,%xmm5
   .byte  15,89,239                           // mulps         %xmm7,%xmm5
   .byte  15,88,234                           // addps         %xmm2,%xmm5
@@ -65096,7 +66534,7 @@
   .byte  15,40,125,232                       // movaps        -0x18(%ebp),%xmm7
   .byte  15,85,252                           // andnps        %xmm4,%xmm7
   .byte  15,40,93,184                        // movaps        -0x48(%ebp),%xmm3
-  .byte  15,88,152,242,173,0,0               // addps         0xadf2(%eax),%xmm3
+  .byte  15,88,152,18,183,0,0                // addps         0xb712(%eax),%xmm3
   .byte  243,15,91,195                       // cvttps2dq     %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
   .byte  15,40,203                           // movaps        %xmm3,%xmm1
@@ -65198,9 +66636,9 @@
   .byte  102,15,110,36,62                    // movd          (%esi,%edi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,162,72,170,0,0           // pand          0xaa48(%edx),%xmm4
+  .byte  102,15,219,162,104,179,0,0          // pand          0xb368(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,136,171,0,0               // mulps         0xab88(%edx),%xmm4
+  .byte  15,89,162,168,180,0,0               // mulps         0xb4a8(%edx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
@@ -65266,17 +66704,17 @@
   .byte  15,133,157,0,0,0                    // jne           2840 <_sk_scale_565_sse2+0xd8>
   .byte  243,15,126,52,126                   // movq          (%esi,%edi,2),%xmm6
   .byte  102,15,97,240                       // punpcklwd     %xmm0,%xmm6
-  .byte  102,15,111,162,222,170,0,0          // movdqa        0xaade(%edx),%xmm4
+  .byte  102,15,111,162,254,179,0,0          // movdqa        0xb3fe(%edx),%xmm4
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
   .byte  15,91,212                           // cvtdq2ps      %xmm4,%xmm2
-  .byte  15,89,146,238,170,0,0               // mulps         0xaaee(%edx),%xmm2
-  .byte  102,15,111,170,254,170,0,0          // movdqa        0xaafe(%edx),%xmm5
+  .byte  15,89,146,14,180,0,0                // mulps         0xb40e(%edx),%xmm2
+  .byte  102,15,111,170,30,180,0,0           // movdqa        0xb41e(%edx),%xmm5
   .byte  102,15,219,238                      // pand          %xmm6,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,14,171,0,0                // mulps         0xab0e(%edx),%xmm5
-  .byte  102,15,219,178,30,171,0,0           // pand          0xab1e(%edx),%xmm6
+  .byte  15,89,170,46,180,0,0                // mulps         0xb42e(%edx),%xmm5
+  .byte  102,15,219,178,62,180,0,0           // pand          0xb43e(%edx),%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,178,46,171,0,0                // mulps         0xab2e(%edx),%xmm6
+  .byte  15,89,178,78,180,0,0                // mulps         0xb44e(%edx),%xmm6
   .byte  15,40,253                           // movaps        %xmm5,%xmm7
   .byte  15,93,254                           // minps         %xmm6,%xmm7
   .byte  15,40,226                           // movaps        %xmm2,%xmm4
@@ -65390,9 +66828,9 @@
   .byte  102,15,110,36,62                    // movd          (%esi,%edi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,162,32,168,0,0           // pand          0xa820(%edx),%xmm4
+  .byte  102,15,219,162,64,177,0,0           // pand          0xb140(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,96,169,0,0                // mulps         0xa960(%edx),%xmm4
+  .byte  15,89,162,128,178,0,0               // mulps         0xb280(%edx),%xmm4
   .byte  15,40,105,16                        // movaps        0x10(%ecx),%xmm5
   .byte  15,40,113,32                        // movaps        0x20(%ecx),%xmm6
   .byte  15,92,197                           // subps         %xmm5,%xmm0
@@ -65470,17 +66908,17 @@
   .byte  15,133,211,0,0,0                    // jne           2ace <_sk_lerp_565_sse2+0x113>
   .byte  243,15,126,44,126                   // movq          (%esi,%edi,2),%xmm5
   .byte  102,15,97,232                       // punpcklwd     %xmm0,%xmm5
-  .byte  102,15,111,162,134,168,0,0          // movdqa        0xa886(%edx),%xmm4
+  .byte  102,15,111,162,166,177,0,0          // movdqa        0xb1a6(%edx),%xmm4
   .byte  102,15,219,229                      // pand          %xmm5,%xmm4
   .byte  15,91,204                           // cvtdq2ps      %xmm4,%xmm1
-  .byte  15,89,138,150,168,0,0               // mulps         0xa896(%edx),%xmm1
-  .byte  102,15,111,178,166,168,0,0          // movdqa        0xa8a6(%edx),%xmm6
+  .byte  15,89,138,182,177,0,0               // mulps         0xb1b6(%edx),%xmm1
+  .byte  102,15,111,178,198,177,0,0          // movdqa        0xb1c6(%edx),%xmm6
   .byte  102,15,219,245                      // pand          %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,89,178,182,168,0,0               // mulps         0xa8b6(%edx),%xmm6
-  .byte  102,15,219,170,198,168,0,0          // pand          0xa8c6(%edx),%xmm5
+  .byte  15,89,178,214,177,0,0               // mulps         0xb1d6(%edx),%xmm6
+  .byte  102,15,219,170,230,177,0,0          // pand          0xb1e6(%edx),%xmm5
   .byte  15,91,213                           // cvtdq2ps      %xmm5,%xmm2
-  .byte  15,89,146,214,168,0,0               // mulps         0xa8d6(%edx),%xmm2
+  .byte  15,89,146,246,177,0,0               // mulps         0xb1f6(%edx),%xmm2
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  15,41,93,200                        // movaps        %xmm3,-0x38(%ebp)
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
@@ -65568,7 +67006,7 @@
   .byte  139,30                              // mov           (%esi),%ebx
   .byte  15,133,22,1,0,0                     // jne           2c52 <_sk_load_tables_sse2+0x13c>
   .byte  243,15,111,28,187                   // movdqu        (%ebx,%edi,4),%xmm3
-  .byte  102,15,111,146,236,165,0,0          // movdqa        0xa5ec(%edx),%xmm2
+  .byte  102,15,111,146,12,175,0,0           // movdqa        0xaf0c(%edx),%xmm2
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
@@ -65624,7 +67062,7 @@
   .byte  102,15,20,215                       // unpcklpd      %xmm7,%xmm2
   .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,154,44,167,0,0                // mulps         0xa72c(%edx),%xmm3
+  .byte  15,89,154,76,176,0,0                // mulps         0xb04c(%edx),%xmm3
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  137,194                             // mov           %eax,%edx
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -65699,7 +67137,7 @@
   .byte  242,15,112,192,232                  // pshuflw       $0xe8,%xmm0,%xmm0
   .byte  243,15,112,192,232                  // pshufhw       $0xe8,%xmm0,%xmm0
   .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
-  .byte  102,15,111,146,43,166,0,0           // movdqa        0xa62b(%edx),%xmm2
+  .byte  102,15,111,146,75,175,0,0           // movdqa        0xaf4b(%edx),%xmm2
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,97,196                       // punpcklwd     %xmm4,%xmm0
@@ -65771,7 +67209,7 @@
   .byte  243,15,126,93,232                   // movq          -0x18(%ebp),%xmm3
   .byte  102,15,97,220                       // punpcklwd     %xmm4,%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,154,59,166,0,0                // mulps         0xa63b(%edx),%xmm3
+  .byte  15,89,154,91,175,0,0                // mulps         0xaf5b(%edx),%xmm3
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  137,194                             // mov           %eax,%edx
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -65844,7 +67282,7 @@
   .byte  242,15,112,194,232                  // pshuflw       $0xe8,%xmm2,%xmm0
   .byte  243,15,112,192,232                  // pshufhw       $0xe8,%xmm0,%xmm0
   .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
-  .byte  102,15,111,146,4,164,0,0            // movdqa        0xa404(%edx),%xmm2
+  .byte  102,15,111,146,36,173,0,0           // movdqa        0xad24(%edx),%xmm2
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,97,195                       // punpcklwd     %xmm3,%xmm0
@@ -65909,7 +67347,7 @@
   .byte  137,198                             // mov           %eax,%esi
   .byte  141,70,8                            // lea           0x8(%esi),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,154,148,161,0,0               // movaps        0xa194(%edx),%xmm3
+  .byte  15,40,154,180,170,0,0               // movaps        0xaab4(%edx),%xmm3
   .byte  80                                  // push          %eax
   .byte  81                                  // push          %ecx
   .byte  255,86,4                            // call          *0x4(%esi)
@@ -65952,7 +67390,7 @@
   .byte  90                                  // pop           %edx
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  139,8                               // mov           (%eax),%ecx
-  .byte  15,40,162,55,160,0,0                // movaps        0xa037(%edx),%xmm4
+  .byte  15,40,162,87,169,0,0                // movaps        0xa957(%edx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  102,15,112,232,78                   // pshufd        $0x4e,%xmm0,%xmm5
@@ -65974,7 +67412,7 @@
   .byte  102,15,110,192                      // movd          %eax,%xmm0
   .byte  139,65,4                            // mov           0x4(%ecx),%eax
   .byte  102,15,196,195,1                    // pinsrw        $0x1,%ebx,%xmm0
-  .byte  15,40,170,103,161,0,0               // movaps        0xa167(%edx),%xmm5
+  .byte  15,40,170,135,170,0,0               // movaps        0xaa87(%edx),%xmm5
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
   .byte  102,15,112,241,78                   // pshufd        $0x4e,%xmm1,%xmm6
@@ -66104,7 +67542,7 @@
   .byte  102,15,110,199                      // movd          %edi,%xmm0
   .byte  139,121,4                           // mov           0x4(%ecx),%edi
   .byte  102,15,196,198,1                    // pinsrw        $0x1,%esi,%xmm0
-  .byte  15,40,170,107,159,0,0               // movaps        0x9f6b(%edx),%xmm5
+  .byte  15,40,170,139,168,0,0               // movaps        0xa88b(%edx),%xmm5
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
   .byte  102,15,91,201                       // cvtps2dq      %xmm1,%xmm1
   .byte  102,15,112,241,78                   // pshufd        $0x4e,%xmm1,%xmm6
@@ -66360,15 +67798,15 @@
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,88,245                           // addps         %xmm5,%xmm6
   .byte  15,91,238                           // cvtdq2ps      %xmm6,%xmm5
-  .byte  15,89,170,43,157,0,0                // mulps         0x9d2b(%edx),%xmm5
-  .byte  15,84,178,59,157,0,0                // andps         0x9d3b(%edx),%xmm6
-  .byte  15,86,178,139,154,0,0               // orps          0x9a8b(%edx),%xmm6
-  .byte  15,88,170,75,157,0,0                // addps         0x9d4b(%edx),%xmm5
-  .byte  15,40,186,91,157,0,0                // movaps        0x9d5b(%edx),%xmm7
+  .byte  15,89,170,75,166,0,0                // mulps         0xa64b(%edx),%xmm5
+  .byte  15,84,178,91,166,0,0                // andps         0xa65b(%edx),%xmm6
+  .byte  15,86,178,171,163,0,0               // orps          0xa3ab(%edx),%xmm6
+  .byte  15,88,170,107,166,0,0               // addps         0xa66b(%edx),%xmm5
+  .byte  15,40,186,123,166,0,0               // movaps        0xa67b(%edx),%xmm7
   .byte  15,89,254                           // mulps         %xmm6,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,88,178,107,157,0,0               // addps         0x9d6b(%edx),%xmm6
-  .byte  15,40,186,123,157,0,0               // movaps        0x9d7b(%edx),%xmm7
+  .byte  15,88,178,139,166,0,0               // addps         0xa68b(%edx),%xmm6
+  .byte  15,40,186,155,166,0,0               // movaps        0xa69b(%edx),%xmm7
   .byte  15,94,254                           // divps         %xmm6,%xmm7
   .byte  243,15,16,49                        // movss         (%ecx),%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
@@ -66378,21 +67816,21 @@
   .byte  15,91,254                           // cvtdq2ps      %xmm6,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
-  .byte  15,40,178,155,154,0,0               // movaps        0x9a9b(%edx),%xmm6
+  .byte  15,40,178,187,163,0,0               // movaps        0xa3bb(%edx),%xmm6
   .byte  15,84,198                           // andps         %xmm6,%xmm0
   .byte  15,92,248                           // subps         %xmm0,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,199                           // subps         %xmm7,%xmm0
-  .byte  15,88,170,139,157,0,0               // addps         0x9d8b(%edx),%xmm5
-  .byte  15,40,186,155,157,0,0               // movaps        0x9d9b(%edx),%xmm7
+  .byte  15,88,170,171,166,0,0               // addps         0xa6ab(%edx),%xmm5
+  .byte  15,40,186,187,166,0,0               // movaps        0xa6bb(%edx),%xmm7
   .byte  15,89,248                           // mulps         %xmm0,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,40,186,171,157,0,0               // movaps        0x9dab(%edx),%xmm7
+  .byte  15,40,186,203,166,0,0               // movaps        0xa6cb(%edx),%xmm7
   .byte  15,92,248                           // subps         %xmm0,%xmm7
-  .byte  15,40,130,187,157,0,0               // movaps        0x9dbb(%edx),%xmm0
+  .byte  15,40,130,219,166,0,0               // movaps        0xa6db(%edx),%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  15,88,197                           // addps         %xmm5,%xmm0
-  .byte  15,89,130,203,157,0,0               // mulps         0x9dcb(%edx),%xmm0
+  .byte  15,89,130,235,166,0,0               // mulps         0xa6eb(%edx),%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  243,15,16,105,20                    // movss         0x14(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
@@ -66446,15 +67884,15 @@
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,88,245                           // addps         %xmm5,%xmm6
   .byte  15,91,238                           // cvtdq2ps      %xmm6,%xmm5
-  .byte  15,89,170,244,155,0,0               // mulps         0x9bf4(%edx),%xmm5
-  .byte  15,84,178,4,156,0,0                 // andps         0x9c04(%edx),%xmm6
-  .byte  15,86,178,84,153,0,0                // orps          0x9954(%edx),%xmm6
-  .byte  15,88,170,20,156,0,0                // addps         0x9c14(%edx),%xmm5
-  .byte  15,40,186,36,156,0,0                // movaps        0x9c24(%edx),%xmm7
+  .byte  15,89,170,20,165,0,0                // mulps         0xa514(%edx),%xmm5
+  .byte  15,84,178,36,165,0,0                // andps         0xa524(%edx),%xmm6
+  .byte  15,86,178,116,162,0,0               // orps          0xa274(%edx),%xmm6
+  .byte  15,88,170,52,165,0,0                // addps         0xa534(%edx),%xmm5
+  .byte  15,40,186,68,165,0,0                // movaps        0xa544(%edx),%xmm7
   .byte  15,89,254                           // mulps         %xmm6,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,88,178,52,156,0,0                // addps         0x9c34(%edx),%xmm6
-  .byte  15,40,186,68,156,0,0                // movaps        0x9c44(%edx),%xmm7
+  .byte  15,88,178,84,165,0,0                // addps         0xa554(%edx),%xmm6
+  .byte  15,40,186,100,165,0,0               // movaps        0xa564(%edx),%xmm7
   .byte  15,94,254                           // divps         %xmm6,%xmm7
   .byte  243,15,16,49                        // movss         (%ecx),%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
@@ -66464,21 +67902,21 @@
   .byte  15,91,254                           // cvtdq2ps      %xmm6,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
-  .byte  15,40,178,100,153,0,0               // movaps        0x9964(%edx),%xmm6
+  .byte  15,40,178,132,162,0,0               // movaps        0xa284(%edx),%xmm6
   .byte  15,84,198                           // andps         %xmm6,%xmm0
   .byte  15,92,248                           // subps         %xmm0,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,199                           // subps         %xmm7,%xmm0
-  .byte  15,88,170,84,156,0,0                // addps         0x9c54(%edx),%xmm5
-  .byte  15,40,186,100,156,0,0               // movaps        0x9c64(%edx),%xmm7
+  .byte  15,88,170,116,165,0,0               // addps         0xa574(%edx),%xmm5
+  .byte  15,40,186,132,165,0,0               // movaps        0xa584(%edx),%xmm7
   .byte  15,89,248                           // mulps         %xmm0,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,40,186,116,156,0,0               // movaps        0x9c74(%edx),%xmm7
+  .byte  15,40,186,148,165,0,0               // movaps        0xa594(%edx),%xmm7
   .byte  15,92,248                           // subps         %xmm0,%xmm7
-  .byte  15,40,130,132,156,0,0               // movaps        0x9c84(%edx),%xmm0
+  .byte  15,40,130,164,165,0,0               // movaps        0xa5a4(%edx),%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  15,88,197                           // addps         %xmm5,%xmm0
-  .byte  15,89,130,148,156,0,0               // mulps         0x9c94(%edx),%xmm0
+  .byte  15,89,130,180,165,0,0               // mulps         0xa5b4(%edx),%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  243,15,16,105,20                    // movss         0x14(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
@@ -66531,15 +67969,15 @@
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,88,245                           // addps         %xmm5,%xmm6
   .byte  15,91,238                           // cvtdq2ps      %xmm6,%xmm5
-  .byte  15,89,170,192,154,0,0               // mulps         0x9ac0(%edx),%xmm5
-  .byte  15,84,178,208,154,0,0               // andps         0x9ad0(%edx),%xmm6
-  .byte  15,86,178,32,152,0,0                // orps          0x9820(%edx),%xmm6
-  .byte  15,88,170,224,154,0,0               // addps         0x9ae0(%edx),%xmm5
-  .byte  15,40,186,240,154,0,0               // movaps        0x9af0(%edx),%xmm7
+  .byte  15,89,170,224,163,0,0               // mulps         0xa3e0(%edx),%xmm5
+  .byte  15,84,178,240,163,0,0               // andps         0xa3f0(%edx),%xmm6
+  .byte  15,86,178,64,161,0,0                // orps          0xa140(%edx),%xmm6
+  .byte  15,88,170,0,164,0,0                 // addps         0xa400(%edx),%xmm5
+  .byte  15,40,186,16,164,0,0                // movaps        0xa410(%edx),%xmm7
   .byte  15,89,254                           // mulps         %xmm6,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,88,178,0,155,0,0                 // addps         0x9b00(%edx),%xmm6
-  .byte  15,40,186,16,155,0,0                // movaps        0x9b10(%edx),%xmm7
+  .byte  15,88,178,32,164,0,0                // addps         0xa420(%edx),%xmm6
+  .byte  15,40,186,48,164,0,0                // movaps        0xa430(%edx),%xmm7
   .byte  15,94,254                           // divps         %xmm6,%xmm7
   .byte  243,15,16,49                        // movss         (%ecx),%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
@@ -66549,21 +67987,21 @@
   .byte  15,91,254                           // cvtdq2ps      %xmm6,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
-  .byte  15,40,178,48,152,0,0                // movaps        0x9830(%edx),%xmm6
+  .byte  15,40,178,80,161,0,0                // movaps        0xa150(%edx),%xmm6
   .byte  15,84,198                           // andps         %xmm6,%xmm0
   .byte  15,92,248                           // subps         %xmm0,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,199                           // subps         %xmm7,%xmm0
-  .byte  15,88,170,32,155,0,0                // addps         0x9b20(%edx),%xmm5
-  .byte  15,40,186,48,155,0,0                // movaps        0x9b30(%edx),%xmm7
+  .byte  15,88,170,64,164,0,0                // addps         0xa440(%edx),%xmm5
+  .byte  15,40,186,80,164,0,0                // movaps        0xa450(%edx),%xmm7
   .byte  15,89,248                           // mulps         %xmm0,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,40,186,64,155,0,0                // movaps        0x9b40(%edx),%xmm7
+  .byte  15,40,186,96,164,0,0                // movaps        0xa460(%edx),%xmm7
   .byte  15,92,248                           // subps         %xmm0,%xmm7
-  .byte  15,40,130,80,155,0,0                // movaps        0x9b50(%edx),%xmm0
+  .byte  15,40,130,112,164,0,0               // movaps        0xa470(%edx),%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  15,88,197                           // addps         %xmm5,%xmm0
-  .byte  15,89,130,96,155,0,0                // mulps         0x9b60(%edx),%xmm0
+  .byte  15,89,130,128,164,0,0               // mulps         0xa480(%edx),%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  243,15,16,105,20                    // movss         0x14(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
@@ -66616,15 +68054,15 @@
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,88,245                           // addps         %xmm5,%xmm6
   .byte  15,91,238                           // cvtdq2ps      %xmm6,%xmm5
-  .byte  15,89,170,140,153,0,0               // mulps         0x998c(%edx),%xmm5
-  .byte  15,84,178,156,153,0,0               // andps         0x999c(%edx),%xmm6
-  .byte  15,86,178,236,150,0,0               // orps          0x96ec(%edx),%xmm6
-  .byte  15,88,170,172,153,0,0               // addps         0x99ac(%edx),%xmm5
-  .byte  15,40,186,188,153,0,0               // movaps        0x99bc(%edx),%xmm7
+  .byte  15,89,170,172,162,0,0               // mulps         0xa2ac(%edx),%xmm5
+  .byte  15,84,178,188,162,0,0               // andps         0xa2bc(%edx),%xmm6
+  .byte  15,86,178,12,160,0,0                // orps          0xa00c(%edx),%xmm6
+  .byte  15,88,170,204,162,0,0               // addps         0xa2cc(%edx),%xmm5
+  .byte  15,40,186,220,162,0,0               // movaps        0xa2dc(%edx),%xmm7
   .byte  15,89,254                           // mulps         %xmm6,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,88,178,204,153,0,0               // addps         0x99cc(%edx),%xmm6
-  .byte  15,40,186,220,153,0,0               // movaps        0x99dc(%edx),%xmm7
+  .byte  15,88,178,236,162,0,0               // addps         0xa2ec(%edx),%xmm6
+  .byte  15,40,186,252,162,0,0               // movaps        0xa2fc(%edx),%xmm7
   .byte  15,94,254                           // divps         %xmm6,%xmm7
   .byte  243,15,16,49                        // movss         (%ecx),%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
@@ -66634,21 +68072,21 @@
   .byte  15,91,254                           // cvtdq2ps      %xmm6,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,194,199,1                        // cmpltps       %xmm7,%xmm0
-  .byte  15,40,178,252,150,0,0               // movaps        0x96fc(%edx),%xmm6
+  .byte  15,40,178,28,160,0,0                // movaps        0xa01c(%edx),%xmm6
   .byte  15,84,198                           // andps         %xmm6,%xmm0
   .byte  15,92,248                           // subps         %xmm0,%xmm7
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,199                           // subps         %xmm7,%xmm0
-  .byte  15,88,170,236,153,0,0               // addps         0x99ec(%edx),%xmm5
-  .byte  15,40,186,252,153,0,0               // movaps        0x99fc(%edx),%xmm7
+  .byte  15,88,170,12,163,0,0                // addps         0xa30c(%edx),%xmm5
+  .byte  15,40,186,28,163,0,0                // movaps        0xa31c(%edx),%xmm7
   .byte  15,89,248                           // mulps         %xmm0,%xmm7
   .byte  15,92,239                           // subps         %xmm7,%xmm5
-  .byte  15,40,186,12,154,0,0                // movaps        0x9a0c(%edx),%xmm7
+  .byte  15,40,186,44,163,0,0                // movaps        0xa32c(%edx),%xmm7
   .byte  15,92,248                           // subps         %xmm0,%xmm7
-  .byte  15,40,130,28,154,0,0                // movaps        0x9a1c(%edx),%xmm0
+  .byte  15,40,130,60,163,0,0                // movaps        0xa33c(%edx),%xmm0
   .byte  15,94,199                           // divps         %xmm7,%xmm0
   .byte  15,88,197                           // addps         %xmm5,%xmm0
-  .byte  15,89,130,44,154,0,0                // mulps         0x9a2c(%edx),%xmm0
+  .byte  15,89,130,76,163,0,0                // mulps         0xa34c(%edx),%xmm0
   .byte  102,15,91,192                       // cvtps2dq      %xmm0,%xmm0
   .byte  243,15,16,105,20                    // movss         0x14(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
@@ -66683,15 +68121,15 @@
   .byte  232,0,0,0,0                         // call          3a85 <_sk_gamma_sse2+0x12>
   .byte  88                                  // pop           %eax
   .byte  15,91,195                           // cvtdq2ps      %xmm3,%xmm0
-  .byte  15,89,128,91,152,0,0                // mulps         0x985b(%eax),%xmm0
-  .byte  15,84,152,107,152,0,0               // andps         0x986b(%eax),%xmm3
-  .byte  15,86,152,187,149,0,0               // orps          0x95bb(%eax),%xmm3
-  .byte  15,88,128,123,152,0,0               // addps         0x987b(%eax),%xmm0
+  .byte  15,89,128,123,161,0,0               // mulps         0xa17b(%eax),%xmm0
+  .byte  15,84,152,139,161,0,0               // andps         0xa18b(%eax),%xmm3
+  .byte  15,86,152,219,158,0,0               // orps          0x9edb(%eax),%xmm3
+  .byte  15,88,128,155,161,0,0               // addps         0xa19b(%eax),%xmm0
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
-  .byte  15,89,160,139,152,0,0               // mulps         0x988b(%eax),%xmm4
+  .byte  15,89,160,171,161,0,0               // mulps         0xa1ab(%eax),%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  15,88,152,155,152,0,0               // addps         0x989b(%eax),%xmm3
-  .byte  15,40,160,171,152,0,0               // movaps        0x98ab(%eax),%xmm4
+  .byte  15,88,152,187,161,0,0               // addps         0xa1bb(%eax),%xmm3
+  .byte  15,40,160,203,161,0,0               // movaps        0xa1cb(%eax),%xmm4
   .byte  15,94,227                           // divps         %xmm3,%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
@@ -66704,30 +68142,30 @@
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
   .byte  15,194,227,1                        // cmpltps       %xmm3,%xmm4
-  .byte  15,84,160,203,149,0,0               // andps         0x95cb(%eax),%xmm4
+  .byte  15,84,160,235,158,0,0               // andps         0x9eeb(%eax),%xmm4
   .byte  15,92,220                           // subps         %xmm4,%xmm3
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
   .byte  15,92,227                           // subps         %xmm3,%xmm4
-  .byte  15,88,128,187,152,0,0               // addps         0x98bb(%eax),%xmm0
-  .byte  15,40,176,219,152,0,0               // movaps        0x98db(%eax),%xmm6
+  .byte  15,88,128,219,161,0,0               // addps         0xa1db(%eax),%xmm0
+  .byte  15,40,176,251,161,0,0               // movaps        0xa1fb(%eax),%xmm6
   .byte  15,92,244                           // subps         %xmm4,%xmm6
-  .byte  15,40,168,203,152,0,0               // movaps        0x98cb(%eax),%xmm5
+  .byte  15,40,168,235,161,0,0               // movaps        0xa1eb(%eax),%xmm5
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  15,40,152,235,152,0,0               // movaps        0x98eb(%eax),%xmm3
+  .byte  15,40,152,11,162,0,0                // movaps        0xa20b(%eax),%xmm3
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  15,94,230                           // divps         %xmm6,%xmm4
   .byte  15,88,224                           // addps         %xmm0,%xmm4
   .byte  15,91,193                           // cvtdq2ps      %xmm1,%xmm0
-  .byte  15,89,128,91,152,0,0                // mulps         0x985b(%eax),%xmm0
-  .byte  15,84,136,107,152,0,0               // andps         0x986b(%eax),%xmm1
-  .byte  15,86,136,187,149,0,0               // orps          0x95bb(%eax),%xmm1
-  .byte  15,88,128,123,152,0,0               // addps         0x987b(%eax),%xmm0
+  .byte  15,89,128,123,161,0,0               // mulps         0xa17b(%eax),%xmm0
+  .byte  15,84,136,139,161,0,0               // andps         0xa18b(%eax),%xmm1
+  .byte  15,86,136,219,158,0,0               // orps          0x9edb(%eax),%xmm1
+  .byte  15,88,128,155,161,0,0               // addps         0xa19b(%eax),%xmm0
   .byte  15,40,241                           // movaps        %xmm1,%xmm6
-  .byte  15,89,176,139,152,0,0               // mulps         0x988b(%eax),%xmm6
+  .byte  15,89,176,171,161,0,0               // mulps         0xa1ab(%eax),%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
-  .byte  15,88,136,155,152,0,0               // addps         0x989b(%eax),%xmm1
-  .byte  15,40,176,171,152,0,0               // movaps        0x98ab(%eax),%xmm6
+  .byte  15,88,136,187,161,0,0               // addps         0xa1bb(%eax),%xmm1
+  .byte  15,40,176,203,161,0,0               // movaps        0xa1cb(%eax),%xmm6
   .byte  15,94,241                           // divps         %xmm1,%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
@@ -66735,12 +68173,12 @@
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,194,241,1                        // cmpltps       %xmm1,%xmm6
-  .byte  15,84,176,203,149,0,0               // andps         0x95cb(%eax),%xmm6
+  .byte  15,84,176,235,158,0,0               // andps         0x9eeb(%eax),%xmm6
   .byte  15,92,206                           // subps         %xmm6,%xmm1
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,92,241                           // subps         %xmm1,%xmm6
-  .byte  15,88,128,187,152,0,0               // addps         0x98bb(%eax),%xmm0
-  .byte  15,40,184,219,152,0,0               // movaps        0x98db(%eax),%xmm7
+  .byte  15,88,128,219,161,0,0               // addps         0xa1db(%eax),%xmm0
+  .byte  15,40,184,251,161,0,0               // movaps        0xa1fb(%eax),%xmm7
   .byte  15,92,254                           // subps         %xmm6,%xmm7
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
@@ -66748,15 +68186,15 @@
   .byte  15,94,207                           // divps         %xmm7,%xmm1
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  15,91,194                           // cvtdq2ps      %xmm2,%xmm0
-  .byte  15,89,128,91,152,0,0                // mulps         0x985b(%eax),%xmm0
-  .byte  15,84,144,107,152,0,0               // andps         0x986b(%eax),%xmm2
-  .byte  15,86,144,187,149,0,0               // orps          0x95bb(%eax),%xmm2
-  .byte  15,88,128,123,152,0,0               // addps         0x987b(%eax),%xmm0
-  .byte  15,40,176,139,152,0,0               // movaps        0x988b(%eax),%xmm6
+  .byte  15,89,128,123,161,0,0               // mulps         0xa17b(%eax),%xmm0
+  .byte  15,84,144,139,161,0,0               // andps         0xa18b(%eax),%xmm2
+  .byte  15,86,144,219,158,0,0               // orps          0x9edb(%eax),%xmm2
+  .byte  15,88,128,155,161,0,0               // addps         0xa19b(%eax),%xmm0
+  .byte  15,40,176,171,161,0,0               // movaps        0xa1ab(%eax),%xmm6
   .byte  15,89,242                           // mulps         %xmm2,%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
-  .byte  15,88,144,155,152,0,0               // addps         0x989b(%eax),%xmm2
-  .byte  15,40,176,171,152,0,0               // movaps        0x98ab(%eax),%xmm6
+  .byte  15,88,144,187,161,0,0               // addps         0xa1bb(%eax),%xmm2
+  .byte  15,40,176,203,161,0,0               // movaps        0xa1cb(%eax),%xmm6
   .byte  15,94,242                           // divps         %xmm2,%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
   .byte  15,89,69,232                        // mulps         -0x18(%ebp),%xmm0
@@ -66764,18 +68202,18 @@
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,194,242,1                        // cmpltps       %xmm2,%xmm6
-  .byte  15,84,176,203,149,0,0               // andps         0x95cb(%eax),%xmm6
+  .byte  15,84,176,235,158,0,0               // andps         0x9eeb(%eax),%xmm6
   .byte  15,92,214                           // subps         %xmm6,%xmm2
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,92,242                           // subps         %xmm2,%xmm6
-  .byte  15,88,128,187,152,0,0               // addps         0x98bb(%eax),%xmm0
+  .byte  15,88,128,219,161,0,0               // addps         0xa1db(%eax),%xmm0
   .byte  15,89,238                           // mulps         %xmm6,%xmm5
   .byte  15,92,197                           // subps         %xmm5,%xmm0
-  .byte  15,40,144,219,152,0,0               // movaps        0x98db(%eax),%xmm2
+  .byte  15,40,144,251,161,0,0               // movaps        0xa1fb(%eax),%xmm2
   .byte  15,92,214                           // subps         %xmm6,%xmm2
   .byte  15,94,218                           // divps         %xmm2,%xmm3
   .byte  15,88,216                           // addps         %xmm0,%xmm3
-  .byte  15,40,128,251,152,0,0               // movaps        0x98fb(%eax),%xmm0
+  .byte  15,40,128,27,162,0,0                // movaps        0xa21b(%eax),%xmm0
   .byte  15,89,224                           // mulps         %xmm0,%xmm4
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  15,89,216                           // mulps         %xmm0,%xmm3
@@ -66804,29 +68242,29 @@
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  232,0,0,0,0                         // call          3c6d <_sk_lab_to_xyz_sse2+0x15>
   .byte  88                                  // pop           %eax
-  .byte  15,89,144,35,151,0,0                // mulps         0x9723(%eax),%xmm2
-  .byte  15,40,160,179,148,0,0               // movaps        0x94b3(%eax),%xmm4
+  .byte  15,89,144,67,160,0,0                // mulps         0xa043(%eax),%xmm2
+  .byte  15,40,160,211,157,0,0               // movaps        0x9dd3(%eax),%xmm4
   .byte  15,89,204                           // mulps         %xmm4,%xmm1
-  .byte  15,40,168,51,151,0,0                // movaps        0x9733(%eax),%xmm5
+  .byte  15,40,168,83,160,0,0                // movaps        0xa053(%eax),%xmm5
   .byte  15,88,205                           // addps         %xmm5,%xmm1
   .byte  15,89,220                           // mulps         %xmm4,%xmm3
   .byte  15,88,221                           // addps         %xmm5,%xmm3
-  .byte  15,88,144,67,151,0,0                // addps         0x9743(%eax),%xmm2
-  .byte  15,89,144,83,151,0,0                // mulps         0x9753(%eax),%xmm2
-  .byte  15,89,136,99,151,0,0                // mulps         0x9763(%eax),%xmm1
+  .byte  15,88,144,99,160,0,0                // addps         0xa063(%eax),%xmm2
+  .byte  15,89,144,115,160,0,0               // mulps         0xa073(%eax),%xmm2
+  .byte  15,89,136,131,160,0,0               // mulps         0xa083(%eax),%xmm1
   .byte  15,88,202                           // addps         %xmm2,%xmm1
-  .byte  15,89,152,115,151,0,0               // mulps         0x9773(%eax),%xmm3
+  .byte  15,89,152,147,160,0,0               // mulps         0xa093(%eax),%xmm3
   .byte  15,40,234                           // movaps        %xmm2,%xmm5
   .byte  15,92,235                           // subps         %xmm3,%xmm5
   .byte  15,40,193                           // movaps        %xmm1,%xmm0
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  15,89,193                           // mulps         %xmm1,%xmm0
-  .byte  15,40,152,131,151,0,0               // movaps        0x9783(%eax),%xmm3
+  .byte  15,40,152,163,160,0,0               // movaps        0xa0a3(%eax),%xmm3
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  15,194,224,1                        // cmpltps       %xmm0,%xmm4
-  .byte  15,40,176,147,151,0,0               // movaps        0x9793(%eax),%xmm6
+  .byte  15,40,176,179,160,0,0               // movaps        0xa0b3(%eax),%xmm6
   .byte  15,88,206                           // addps         %xmm6,%xmm1
-  .byte  15,40,184,163,151,0,0               // movaps        0x97a3(%eax),%xmm7
+  .byte  15,40,184,195,160,0,0               // movaps        0xa0c3(%eax),%xmm7
   .byte  15,89,207                           // mulps         %xmm7,%xmm1
   .byte  15,84,196                           // andps         %xmm4,%xmm0
   .byte  15,85,225                           // andnps        %xmm1,%xmm4
@@ -66851,8 +68289,8 @@
   .byte  15,85,221                           // andnps        %xmm5,%xmm3
   .byte  15,86,216                           // orps          %xmm0,%xmm3
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,89,160,179,151,0,0               // mulps         0x97b3(%eax),%xmm4
-  .byte  15,89,152,195,151,0,0               // mulps         0x97c3(%eax),%xmm3
+  .byte  15,89,160,211,160,0,0               // mulps         0xa0d3(%eax),%xmm4
+  .byte  15,89,152,227,160,0,0               // mulps         0xa0e3(%eax),%xmm3
   .byte  141,65,4                            // lea           0x4(%ecx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  15,40,196                           // movaps        %xmm4,%xmm0
@@ -66890,9 +68328,9 @@
   .byte  102,15,110,4,62                     // movd          (%esi,%edi,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  102,15,97,192                       // punpcklwd     %xmm0,%xmm0
-  .byte  102,15,219,130,179,147,0,0          // pand          0x93b3(%edx),%xmm0
+  .byte  102,15,219,130,211,156,0,0          // pand          0x9cd3(%edx),%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,154,243,148,0,0               // mulps         0x94f3(%edx),%xmm3
+  .byte  15,89,154,19,158,0,0                // mulps         0x9e13(%edx),%xmm3
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -66957,9 +68395,9 @@
   .byte  102,15,110,36,55                    // movd          (%edi,%esi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,162,223,146,0,0          // pand          0x92df(%edx),%xmm4
+  .byte  102,15,219,162,255,155,0,0          // pand          0x9bff(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,31,148,0,0                // mulps         0x941f(%edx),%xmm4
+  .byte  15,89,162,63,157,0,0                // mulps         0x9d3f(%edx),%xmm4
   .byte  15,41,97,64                         // movaps        %xmm4,0x40(%ecx)
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -67050,7 +68488,7 @@
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,216                           // cvtdq2ps      %xmm0,%xmm3
-  .byte  15,89,153,242,146,0,0               // mulps         0x92f2(%ecx),%xmm3
+  .byte  15,89,153,18,156,0,0                // mulps         0x9c12(%ecx),%xmm3
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
@@ -67084,7 +68522,7 @@
   .byte  15,175,81,4                         // imul          0x4(%ecx),%edx
   .byte  3,23                                // add           (%edi),%edx
   .byte  139,89,8                            // mov           0x8(%ecx),%ebx
-  .byte  15,40,166,126,145,0,0               // movaps        0x917e(%esi),%xmm4
+  .byte  15,40,166,158,154,0,0               // movaps        0x9a9e(%esi),%xmm4
   .byte  15,89,227                           // mulps         %xmm3,%xmm4
   .byte  102,15,91,228                       // cvtps2dq      %xmm4,%xmm4
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
@@ -67119,7 +68557,7 @@
   .byte  102,15,197,220,4                    // pextrw        $0x4,%xmm4,%ebx
   .byte  136,92,58,2                         // mov           %bl,0x2(%edx,%edi,1)
   .byte  102,15,112,228,212                  // pshufd        $0xd4,%xmm4,%xmm4
-  .byte  102,15,219,166,158,148,0,0          // pand          0x949e(%esi),%xmm4
+  .byte  102,15,219,166,190,157,0,0          // pand          0x9dbe(%esi),%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
@@ -67155,12 +68593,12 @@
   .byte  102,15,110,4,62                     // movd          (%esi,%edi,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
   .byte  102,15,97,192                       // punpcklwd     %xmm0,%xmm0
-  .byte  102,15,219,130,188,144,0,0          // pand          0x90bc(%edx),%xmm0
+  .byte  102,15,219,130,220,153,0,0          // pand          0x99dc(%edx),%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,130,252,145,0,0               // mulps         0x91fc(%edx),%xmm0
+  .byte  15,89,130,28,155,0,0                // mulps         0x9b1c(%edx),%xmm0
   .byte  141,112,8                           // lea           0x8(%eax),%esi
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,154,252,143,0,0               // movaps        0x8ffc(%edx),%xmm3
+  .byte  15,40,154,28,153,0,0                // movaps        0x991c(%edx),%xmm3
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  86                                  // push          %esi
@@ -67218,13 +68656,13 @@
   .byte  102,15,110,36,62                    // movd          (%esi,%edi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,219,162,16,144,0,0           // pand          0x9010(%edx),%xmm4
+  .byte  102,15,219,162,48,153,0,0           // pand          0x9930(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,80,145,0,0                // mulps         0x9150(%edx),%xmm4
+  .byte  15,89,162,112,154,0,0               // mulps         0x9a70(%edx),%xmm4
   .byte  15,41,97,48                         // movaps        %xmm4,0x30(%ecx)
   .byte  15,41,97,32                         // movaps        %xmm4,0x20(%ecx)
   .byte  15,41,97,16                         // movaps        %xmm4,0x10(%ecx)
-  .byte  15,40,162,80,143,0,0                // movaps        0x8f50(%edx),%xmm4
+  .byte  15,40,162,112,152,0,0               // movaps        0x9870(%edx),%xmm4
   .byte  15,41,97,64                         // movaps        %xmm4,0x40(%ecx)
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -67316,10 +68754,10 @@
   .byte  102,15,96,193                       // punpcklbw     %xmm1,%xmm0
   .byte  102,15,97,193                       // punpcklwd     %xmm1,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,129,249,143,0,0               // mulps         0x8ff9(%ecx),%xmm0
+  .byte  15,89,129,25,153,0,0                // mulps         0x9919(%ecx),%xmm0
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,153,249,141,0,0               // movaps        0x8df9(%ecx),%xmm3
+  .byte  15,40,153,25,151,0,0                // movaps        0x9719(%ecx),%xmm3
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  82                                  // push          %edx
@@ -67356,20 +68794,20 @@
   .byte  117,97                              // jne           4320 <_sk_load_565_sse2+0x8c>
   .byte  243,15,126,20,126                   // movq          (%esi,%edi,2),%xmm2
   .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
-  .byte  102,15,111,130,190,143,0,0          // movdqa        0x8fbe(%edx),%xmm0
+  .byte  102,15,111,130,222,152,0,0          // movdqa        0x98de(%edx),%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,130,206,143,0,0               // mulps         0x8fce(%edx),%xmm0
-  .byte  102,15,111,138,222,143,0,0          // movdqa        0x8fde(%edx),%xmm1
+  .byte  15,89,130,238,152,0,0               // mulps         0x98ee(%edx),%xmm0
+  .byte  102,15,111,138,254,152,0,0          // movdqa        0x98fe(%edx),%xmm1
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,138,238,143,0,0               // mulps         0x8fee(%edx),%xmm1
-  .byte  102,15,219,146,254,143,0,0          // pand          0x8ffe(%edx),%xmm2
+  .byte  15,89,138,14,153,0,0                // mulps         0x990e(%edx),%xmm1
+  .byte  102,15,219,146,30,153,0,0           // pand          0x991e(%edx),%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,146,14,144,0,0                // mulps         0x900e(%edx),%xmm2
+  .byte  15,89,146,46,153,0,0                // mulps         0x992e(%edx),%xmm2
   .byte  141,112,8                           // lea           0x8(%eax),%esi
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,154,174,141,0,0               // movaps        0x8dae(%edx),%xmm3
+  .byte  15,40,154,206,150,0,0               // movaps        0x96ce(%edx),%xmm3
   .byte  86                                  // push          %esi
   .byte  81                                  // push          %ecx
   .byte  255,80,4                            // call          *0x4(%eax)
@@ -67423,21 +68861,21 @@
   .byte  117,113                             // jne           4400 <_sk_load_565_dst_sse2+0x9c>
   .byte  243,15,126,36,126                   // movq          (%esi,%edi,2),%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,111,170,238,142,0,0          // movdqa        0x8eee(%edx),%xmm5
+  .byte  102,15,111,170,14,152,0,0           // movdqa        0x980e(%edx),%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,254,142,0,0               // mulps         0x8efe(%edx),%xmm5
+  .byte  15,89,170,30,152,0,0                // mulps         0x981e(%edx),%xmm5
   .byte  15,41,105,16                        // movaps        %xmm5,0x10(%ecx)
-  .byte  102,15,111,170,14,143,0,0           // movdqa        0x8f0e(%edx),%xmm5
+  .byte  102,15,111,170,46,152,0,0           // movdqa        0x982e(%edx),%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,30,143,0,0                // mulps         0x8f1e(%edx),%xmm5
+  .byte  15,89,170,62,152,0,0                // mulps         0x983e(%edx),%xmm5
   .byte  15,41,105,32                        // movaps        %xmm5,0x20(%ecx)
-  .byte  102,15,219,162,46,143,0,0           // pand          0x8f2e(%edx),%xmm4
+  .byte  102,15,219,162,78,152,0,0           // pand          0x984e(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,62,143,0,0                // mulps         0x8f3e(%edx),%xmm4
+  .byte  15,89,162,94,152,0,0                // mulps         0x985e(%edx),%xmm4
   .byte  15,41,97,48                         // movaps        %xmm4,0x30(%ecx)
-  .byte  15,40,162,222,140,0,0               // movaps        0x8cde(%edx),%xmm4
+  .byte  15,40,162,254,149,0,0               // movaps        0x95fe(%edx),%xmm4
   .byte  15,41,97,64                         // movaps        %xmm4,0x40(%ecx)
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -67522,20 +68960,20 @@
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,209,3                    // pinsrw        $0x3,%ecx,%xmm2
   .byte  102,15,97,208                       // punpcklwd     %xmm0,%xmm2
-  .byte  102,15,111,130,108,141,0,0          // movdqa        0x8d6c(%edx),%xmm0
+  .byte  102,15,111,130,140,150,0,0          // movdqa        0x968c(%edx),%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,130,124,141,0,0               // mulps         0x8d7c(%edx),%xmm0
-  .byte  102,15,111,138,140,141,0,0          // movdqa        0x8d8c(%edx),%xmm1
+  .byte  15,89,130,156,150,0,0               // mulps         0x969c(%edx),%xmm0
+  .byte  102,15,111,138,172,150,0,0          // movdqa        0x96ac(%edx),%xmm1
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,138,156,141,0,0               // mulps         0x8d9c(%edx),%xmm1
-  .byte  102,15,219,146,172,141,0,0          // pand          0x8dac(%edx),%xmm2
+  .byte  15,89,138,188,150,0,0               // mulps         0x96bc(%edx),%xmm1
+  .byte  102,15,219,146,204,150,0,0          // pand          0x96cc(%edx),%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,146,188,141,0,0               // mulps         0x8dbc(%edx),%xmm2
+  .byte  15,89,146,220,150,0,0               // mulps         0x96dc(%edx),%xmm2
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,154,92,139,0,0                // movaps        0x8b5c(%edx),%xmm3
+  .byte  15,40,154,124,148,0,0               // movaps        0x947c(%edx),%xmm3
   .byte  81                                  // push          %ecx
   .byte  255,117,8                           // pushl         0x8(%ebp)
   .byte  255,80,4                            // call          *0x4(%eax)
@@ -67563,8 +69001,8 @@
   .byte  1,210                               // add           %edx,%edx
   .byte  3,22                                // add           (%esi),%edx
   .byte  139,49                              // mov           (%ecx),%esi
-  .byte  15,40,167,233,142,0,0               // movaps        0x8ee9(%edi),%xmm4
-  .byte  15,40,175,249,142,0,0               // movaps        0x8ef9(%edi),%xmm5
+  .byte  15,40,167,9,152,0,0                 // movaps        0x9809(%edi),%xmm4
+  .byte  15,40,175,25,152,0,0                // movaps        0x9819(%edi),%xmm5
   .byte  139,89,8                            // mov           0x8(%ecx),%ebx
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,89,244                           // mulps         %xmm4,%xmm6
@@ -67644,21 +69082,21 @@
   .byte  117,112                             // jne           46d6 <_sk_load_4444_sse2+0x9b>
   .byte  243,15,126,28,126                   // movq          (%esi,%edi,2),%xmm3
   .byte  102,15,97,216                       // punpcklwd     %xmm0,%xmm3
-  .byte  102,15,111,130,39,142,0,0           // movdqa        0x8e27(%edx),%xmm0
+  .byte  102,15,111,130,71,151,0,0           // movdqa        0x9747(%edx),%xmm0
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,130,55,142,0,0                // mulps         0x8e37(%edx),%xmm0
-  .byte  102,15,111,138,71,142,0,0           // movdqa        0x8e47(%edx),%xmm1
+  .byte  15,89,130,87,151,0,0                // mulps         0x9757(%edx),%xmm0
+  .byte  102,15,111,138,103,151,0,0          // movdqa        0x9767(%edx),%xmm1
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,138,87,142,0,0                // mulps         0x8e57(%edx),%xmm1
-  .byte  102,15,111,146,103,142,0,0          // movdqa        0x8e67(%edx),%xmm2
+  .byte  15,89,138,119,151,0,0               // mulps         0x9777(%edx),%xmm1
+  .byte  102,15,111,146,135,151,0,0          // movdqa        0x9787(%edx),%xmm2
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,146,119,142,0,0               // mulps         0x8e77(%edx),%xmm2
-  .byte  102,15,219,154,135,142,0,0          // pand          0x8e87(%edx),%xmm3
+  .byte  15,89,146,151,151,0,0               // mulps         0x9797(%edx),%xmm2
+  .byte  102,15,219,154,167,151,0,0          // pand          0x97a7(%edx),%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,154,151,142,0,0               // mulps         0x8e97(%edx),%xmm3
+  .byte  15,89,154,183,151,0,0               // mulps         0x97b7(%edx),%xmm3
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  82                                  // push          %edx
@@ -67714,24 +69152,24 @@
   .byte  15,133,128,0,0,0                    // jne           47c9 <_sk_load_4444_dst_sse2+0xaf>
   .byte  243,15,126,36,126                   // movq          (%esi,%edi,2),%xmm4
   .byte  102,15,97,224                       // punpcklwd     %xmm0,%xmm4
-  .byte  102,15,111,170,72,141,0,0           // movdqa        0x8d48(%edx),%xmm5
+  .byte  102,15,111,170,104,150,0,0          // movdqa        0x9668(%edx),%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,88,141,0,0                // mulps         0x8d58(%edx),%xmm5
+  .byte  15,89,170,120,150,0,0               // mulps         0x9678(%edx),%xmm5
   .byte  15,41,105,16                        // movaps        %xmm5,0x10(%ecx)
-  .byte  102,15,111,170,104,141,0,0          // movdqa        0x8d68(%edx),%xmm5
+  .byte  102,15,111,170,136,150,0,0          // movdqa        0x9688(%edx),%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,120,141,0,0               // mulps         0x8d78(%edx),%xmm5
+  .byte  15,89,170,152,150,0,0               // mulps         0x9698(%edx),%xmm5
   .byte  15,41,105,32                        // movaps        %xmm5,0x20(%ecx)
-  .byte  102,15,111,170,136,141,0,0          // movdqa        0x8d88(%edx),%xmm5
+  .byte  102,15,111,170,168,150,0,0          // movdqa        0x96a8(%edx),%xmm5
   .byte  102,15,219,236                      // pand          %xmm4,%xmm5
   .byte  15,91,237                           // cvtdq2ps      %xmm5,%xmm5
-  .byte  15,89,170,152,141,0,0               // mulps         0x8d98(%edx),%xmm5
+  .byte  15,89,170,184,150,0,0               // mulps         0x96b8(%edx),%xmm5
   .byte  15,41,105,48                        // movaps        %xmm5,0x30(%ecx)
-  .byte  102,15,219,162,168,141,0,0          // pand          0x8da8(%edx),%xmm4
+  .byte  102,15,219,162,200,150,0,0          // pand          0x96c8(%edx),%xmm4
   .byte  15,91,228                           // cvtdq2ps      %xmm4,%xmm4
-  .byte  15,89,162,184,141,0,0               // mulps         0x8db8(%edx),%xmm4
+  .byte  15,89,162,216,150,0,0               // mulps         0x96d8(%edx),%xmm4
   .byte  15,41,97,64                         // movaps        %xmm4,0x40(%ecx)
   .byte  141,80,8                            // lea           0x8(%eax),%edx
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -67816,21 +69254,21 @@
   .byte  89                                  // pop           %ecx
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,216                       // punpcklwd     %xmm0,%xmm3
-  .byte  102,15,111,129,170,139,0,0          // movdqa        0x8baa(%ecx),%xmm0
+  .byte  102,15,111,129,202,148,0,0          // movdqa        0x94ca(%ecx),%xmm0
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,89,129,186,139,0,0               // mulps         0x8bba(%ecx),%xmm0
-  .byte  102,15,111,137,202,139,0,0          // movdqa        0x8bca(%ecx),%xmm1
+  .byte  15,89,129,218,148,0,0               // mulps         0x94da(%ecx),%xmm0
+  .byte  102,15,111,137,234,148,0,0          // movdqa        0x94ea(%ecx),%xmm1
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
   .byte  15,91,201                           // cvtdq2ps      %xmm1,%xmm1
-  .byte  15,89,137,218,139,0,0               // mulps         0x8bda(%ecx),%xmm1
-  .byte  102,15,111,145,234,139,0,0          // movdqa        0x8bea(%ecx),%xmm2
+  .byte  15,89,137,250,148,0,0               // mulps         0x94fa(%ecx),%xmm1
+  .byte  102,15,111,145,10,149,0,0           // movdqa        0x950a(%ecx),%xmm2
   .byte  102,15,219,211                      // pand          %xmm3,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,89,145,250,139,0,0               // mulps         0x8bfa(%ecx),%xmm2
-  .byte  102,15,219,153,10,140,0,0           // pand          0x8c0a(%ecx),%xmm3
+  .byte  15,89,145,26,149,0,0                // mulps         0x951a(%ecx),%xmm2
+  .byte  102,15,219,153,42,149,0,0           // pand          0x952a(%ecx),%xmm3
   .byte  15,91,219                           // cvtdq2ps      %xmm3,%xmm3
-  .byte  15,89,153,26,140,0,0                // mulps         0x8c1a(%ecx),%xmm3
+  .byte  15,89,153,58,149,0,0                // mulps         0x953a(%ecx),%xmm3
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  81                                  // push          %ecx
@@ -67860,7 +69298,7 @@
   .byte  1,210                               // add           %edx,%edx
   .byte  3,22                                // add           (%esi),%edx
   .byte  139,49                              // mov           (%ecx),%esi
-  .byte  15,40,167,173,139,0,0               // movaps        0x8bad(%edi),%xmm4
+  .byte  15,40,167,205,148,0,0               // movaps        0x94cd(%edi),%xmm4
   .byte  139,89,8                            // mov           0x8(%ecx),%ebx
   .byte  15,40,232                           // movaps        %xmm0,%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
@@ -67945,11 +69383,11 @@
   .byte  139,57                              // mov           (%ecx),%edi
   .byte  117,102                             // jne           4ab8 <_sk_load_8888_sse2+0x92>
   .byte  243,15,111,28,190                   // movdqu        (%esi,%edi,4),%xmm3
-  .byte  102,15,111,146,220,134,0,0          // movdqa        0x86dc(%edx),%xmm2
+  .byte  102,15,111,146,252,143,0,0          // movdqa        0x8ffc(%edx),%xmm2
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,40,162,28,136,0,0                // movaps        0x881c(%edx),%xmm4
+  .byte  15,40,162,60,145,0,0                // movaps        0x913c(%edx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -68014,11 +69452,11 @@
   .byte  139,57                              // mov           (%ecx),%edi
   .byte  117,118                             // jne           4b8f <_sk_load_8888_dst_sse2+0xa2>
   .byte  243,15,111,36,190                   // movdqu        (%esi,%edi,4),%xmm4
-  .byte  102,15,111,170,21,134,0,0           // movdqa        0x8615(%edx),%xmm5
+  .byte  102,15,111,170,53,143,0,0           // movdqa        0x8f35(%edx),%xmm5
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  102,15,219,245                      // pand          %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,40,186,85,135,0,0                // movaps        0x8755(%edx),%xmm7
+  .byte  15,40,186,117,144,0,0               // movaps        0x9075(%edx),%xmm7
   .byte  15,89,247                           // mulps         %xmm7,%xmm6
   .byte  15,41,113,16                        // movaps        %xmm6,0x10(%ecx)
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
@@ -68113,11 +69551,11 @@
   .byte  102,15,98,194                       // punpckldq     %xmm2,%xmm0
   .byte  102,15,98,217                       // punpckldq     %xmm1,%xmm3
   .byte  102,15,108,216                      // punpcklqdq    %xmm0,%xmm3
-  .byte  102,15,111,145,162,132,0,0          // movdqa        0x84a2(%ecx),%xmm2
+  .byte  102,15,111,145,194,141,0,0          // movdqa        0x8dc2(%ecx),%xmm2
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,219,194                      // pand          %xmm2,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,40,161,226,133,0,0               // movaps        0x85e2(%ecx),%xmm4
+  .byte  15,40,161,2,143,0,0                 // movaps        0x8f02(%ecx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -68161,7 +69599,7 @@
   .byte  193,226,2                           // shl           $0x2,%edx
   .byte  3,22                                // add           (%esi),%edx
   .byte  139,49                              // mov           (%ecx),%esi
-  .byte  15,40,167,55,132,0,0                // movaps        0x8437(%edi),%xmm4
+  .byte  15,40,167,87,141,0,0                // movaps        0x8d57(%edi),%xmm4
   .byte  139,89,8                            // mov           0x8(%ecx),%ebx
   .byte  15,40,232                           // movaps        %xmm0,%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
@@ -68232,11 +69670,11 @@
   .byte  139,57                              // mov           (%ecx),%edi
   .byte  117,102                             // jne           4e23 <_sk_load_bgra_sse2+0x92>
   .byte  243,15,111,28,190                   // movdqu        (%esi,%edi,4),%xmm3
-  .byte  102,15,111,130,113,131,0,0          // movdqa        0x8371(%edx),%xmm0
+  .byte  102,15,111,130,145,140,0,0          // movdqa        0x8c91(%edx),%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
   .byte  15,91,209                           // cvtdq2ps      %xmm1,%xmm2
-  .byte  15,40,162,177,132,0,0               // movaps        0x84b1(%edx),%xmm4
+  .byte  15,40,162,209,141,0,0               // movaps        0x8dd1(%edx),%xmm4
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -68301,11 +69739,11 @@
   .byte  139,57                              // mov           (%ecx),%edi
   .byte  117,118                             // jne           4efa <_sk_load_bgra_dst_sse2+0xa2>
   .byte  243,15,111,36,190                   // movdqu        (%esi,%edi,4),%xmm4
-  .byte  102,15,111,170,170,130,0,0          // movdqa        0x82aa(%edx),%xmm5
+  .byte  102,15,111,170,202,139,0,0          // movdqa        0x8bca(%edx),%xmm5
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  102,15,219,245                      // pand          %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
-  .byte  15,40,186,234,131,0,0               // movaps        0x83ea(%edx),%xmm7
+  .byte  15,40,186,10,141,0,0                // movaps        0x8d0a(%edx),%xmm7
   .byte  15,89,247                           // mulps         %xmm7,%xmm6
   .byte  15,41,113,48                        // movaps        %xmm6,0x30(%ecx)
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
@@ -68400,11 +69838,11 @@
   .byte  102,15,98,194                       // punpckldq     %xmm2,%xmm0
   .byte  102,15,98,217                       // punpckldq     %xmm1,%xmm3
   .byte  102,15,108,216                      // punpcklqdq    %xmm0,%xmm3
-  .byte  102,15,111,129,55,129,0,0           // movdqa        0x8137(%ecx),%xmm0
+  .byte  102,15,111,129,87,138,0,0           // movdqa        0x8a57(%ecx),%xmm0
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
   .byte  15,91,209                           // cvtdq2ps      %xmm1,%xmm2
-  .byte  15,40,161,119,130,0,0               // movaps        0x8277(%ecx),%xmm4
+  .byte  15,40,161,151,139,0,0               // movaps        0x8b97(%ecx),%xmm4
   .byte  15,89,212                           // mulps         %xmm4,%xmm2
   .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
   .byte  102,15,114,209,8                    // psrld         $0x8,%xmm1
@@ -68448,7 +69886,7 @@
   .byte  193,226,2                           // shl           $0x2,%edx
   .byte  3,22                                // add           (%esi),%edx
   .byte  139,49                              // mov           (%ecx),%esi
-  .byte  15,40,167,204,128,0,0               // movaps        0x80cc(%edi),%xmm4
+  .byte  15,40,167,236,137,0,0               // movaps        0x89ec(%edi),%xmm4
   .byte  139,89,8                            // mov           0x8(%ecx),%ebx
   .byte  15,40,234                           // movaps        %xmm2,%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
@@ -68547,11 +69985,11 @@
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,97,200                       // punpcklwd     %xmm0,%xmm1
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
-  .byte  102,15,111,170,246,131,0,0          // movdqa        0x83f6(%edx),%xmm5
+  .byte  102,15,111,170,22,141,0,0           // movdqa        0x8d16(%edx),%xmm5
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
   .byte  102,15,219,245                      // pand          %xmm5,%xmm6
-  .byte  102,15,219,138,6,132,0,0            // pand          0x8406(%edx),%xmm1
-  .byte  102,15,111,130,22,132,0,0           // movdqa        0x8416(%edx),%xmm0
+  .byte  102,15,219,138,38,141,0,0           // pand          0x8d26(%edx),%xmm1
+  .byte  102,15,111,130,54,141,0,0           // movdqa        0x8d36(%edx),%xmm0
   .byte  102,15,127,69,184                   // movdqa        %xmm0,-0x48(%ebp)
   .byte  102,15,114,246,16                   // pslld         $0x10,%xmm6
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
@@ -68563,7 +70001,7 @@
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  102,15,111,212                      // movdqa        %xmm4,%xmm2
   .byte  102,15,219,213                      // pand          %xmm5,%xmm2
-  .byte  102,15,111,154,6,132,0,0            // movdqa        0x8406(%edx),%xmm3
+  .byte  102,15,111,154,38,141,0,0           // movdqa        0x8d26(%edx),%xmm3
   .byte  102,15,219,227                      // pand          %xmm3,%xmm4
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,102,196                      // pcmpgtd       %xmm4,%xmm0
@@ -68576,7 +70014,7 @@
   .byte  102,15,219,253                      // pand          %xmm5,%xmm7
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  102,15,114,247,16                   // pslld         $0x10,%xmm7
-  .byte  102,15,111,146,22,132,0,0           // movdqa        0x8416(%edx),%xmm2
+  .byte  102,15,111,146,54,141,0,0           // movdqa        0x8d36(%edx),%xmm2
   .byte  102,15,102,208                      // pcmpgtd       %xmm0,%xmm2
   .byte  102,15,114,240,13                   // pslld         $0xd,%xmm0
   .byte  102,15,235,199                      // por           %xmm7,%xmm0
@@ -68591,7 +70029,7 @@
   .byte  102,15,127,93,184                   // movdqa        %xmm3,-0x48(%ebp)
   .byte  102,15,114,247,13                   // pslld         $0xd,%xmm7
   .byte  102,15,235,253                      // por           %xmm5,%xmm7
-  .byte  102,15,111,154,38,132,0,0           // movdqa        0x8426(%edx),%xmm3
+  .byte  102,15,111,154,70,141,0,0           // movdqa        0x8d46(%edx),%xmm3
   .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
   .byte  102,15,254,227                      // paddd         %xmm3,%xmm4
   .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
@@ -68680,18 +70118,18 @@
   .byte  102,15,126,85,232                   // movd          %xmm2,-0x18(%ebp)
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
   .byte  102,15,97,202                       // punpcklwd     %xmm2,%xmm1
-  .byte  102,15,111,186,217,129,0,0          // movdqa        0x81d9(%edx),%xmm7
+  .byte  102,15,111,186,249,138,0,0          // movdqa        0x8af9(%edx),%xmm7
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,219,215                      // pand          %xmm7,%xmm2
-  .byte  102,15,111,154,233,129,0,0          // movdqa        0x81e9(%edx),%xmm3
+  .byte  102,15,111,154,9,139,0,0            // movdqa        0x8b09(%edx),%xmm3
   .byte  102,15,219,203                      // pand          %xmm3,%xmm1
-  .byte  102,15,111,162,249,129,0,0          // movdqa        0x81f9(%edx),%xmm4
+  .byte  102,15,111,162,25,139,0,0           // movdqa        0x8b19(%edx),%xmm4
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,102,233                      // pcmpgtd       %xmm1,%xmm5
   .byte  102,15,114,241,13                   // pslld         $0xd,%xmm1
   .byte  102,15,235,202                      // por           %xmm2,%xmm1
-  .byte  102,15,111,146,9,130,0,0            // movdqa        0x8209(%edx),%xmm2
+  .byte  102,15,111,146,41,139,0,0           // movdqa        0x8b29(%edx),%xmm2
   .byte  102,15,127,85,184                   // movdqa        %xmm2,-0x48(%ebp)
   .byte  102,15,254,202                      // paddd         %xmm2,%xmm1
   .byte  102,15,223,233                      // pandn         %xmm1,%xmm5
@@ -68837,11 +70275,11 @@
   .byte  102,15,126,69,248                   // movd          %xmm0,-0x8(%ebp)
   .byte  102,15,97,206                       // punpcklwd     %xmm6,%xmm1
   .byte  102,15,239,210                      // pxor          %xmm2,%xmm2
-  .byte  102,15,111,177,61,127,0,0           // movdqa        0x7f3d(%ecx),%xmm6
+  .byte  102,15,111,177,93,136,0,0           // movdqa        0x885d(%ecx),%xmm6
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,219,230                      // pand          %xmm6,%xmm4
-  .byte  102,15,219,137,77,127,0,0           // pand          0x7f4d(%ecx),%xmm1
-  .byte  102,15,111,129,93,127,0,0           // movdqa        0x7f5d(%ecx),%xmm0
+  .byte  102,15,219,137,109,136,0,0          // pand          0x886d(%ecx),%xmm1
+  .byte  102,15,111,129,125,136,0,0          // movdqa        0x887d(%ecx),%xmm0
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
   .byte  102,15,102,217                      // pcmpgtd       %xmm1,%xmm3
@@ -68852,7 +70290,7 @@
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  102,15,111,213                      // movdqa        %xmm5,%xmm2
   .byte  102,15,219,214                      // pand          %xmm6,%xmm2
-  .byte  102,15,111,153,77,127,0,0           // movdqa        0x7f4d(%ecx),%xmm3
+  .byte  102,15,111,153,109,136,0,0          // movdqa        0x886d(%ecx),%xmm3
   .byte  102,15,219,235                      // pand          %xmm3,%xmm5
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,111,224                      // movdqa        %xmm0,%xmm4
@@ -68866,7 +70304,7 @@
   .byte  102,15,219,254                      // pand          %xmm6,%xmm7
   .byte  102,15,219,195                      // pand          %xmm3,%xmm0
   .byte  102,15,114,247,16                   // pslld         $0x10,%xmm7
-  .byte  102,15,111,145,93,127,0,0           // movdqa        0x7f5d(%ecx),%xmm2
+  .byte  102,15,111,145,125,136,0,0          // movdqa        0x887d(%ecx),%xmm2
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
   .byte  102,15,102,216                      // pcmpgtd       %xmm0,%xmm3
   .byte  102,15,114,240,13                   // pslld         $0xd,%xmm0
@@ -68875,12 +70313,12 @@
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,97,252                       // punpcklwd     %xmm4,%xmm7
   .byte  102,15,219,247                      // pand          %xmm7,%xmm6
-  .byte  102,15,219,185,77,127,0,0           // pand          0x7f4d(%ecx),%xmm7
+  .byte  102,15,219,185,109,136,0,0          // pand          0x886d(%ecx),%xmm7
   .byte  102,15,114,246,16                   // pslld         $0x10,%xmm6
   .byte  102,15,102,215                      // pcmpgtd       %xmm7,%xmm2
   .byte  102,15,114,247,13                   // pslld         $0xd,%xmm7
   .byte  102,15,235,254                      // por           %xmm6,%xmm7
-  .byte  102,15,111,161,109,127,0,0          // movdqa        0x7f6d(%ecx),%xmm4
+  .byte  102,15,111,161,141,136,0,0          // movdqa        0x888d(%ecx),%xmm4
   .byte  102,15,254,204                      // paddd         %xmm4,%xmm1
   .byte  102,15,254,236                      // paddd         %xmm4,%xmm5
   .byte  102,15,254,196                      // paddd         %xmm4,%xmm0
@@ -68919,22 +70357,22 @@
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  232,0,0,0,0                         // call          5787 <_sk_store_f16_sse2+0x19>
   .byte  88                                  // pop           %eax
-  .byte  102,15,111,152,185,125,0,0          // movdqa        0x7db9(%eax),%xmm3
+  .byte  102,15,111,152,217,134,0,0          // movdqa        0x86d9(%eax),%xmm3
   .byte  102,15,111,224                      // movdqa        %xmm0,%xmm4
   .byte  102,15,219,227                      // pand          %xmm3,%xmm4
-  .byte  102,15,111,136,233,125,0,0          // movdqa        0x7de9(%eax),%xmm1
+  .byte  102,15,111,136,9,135,0,0            // movdqa        0x8709(%eax),%xmm1
   .byte  102,15,127,77,232                   // movdqa        %xmm1,-0x18(%ebp)
   .byte  102,15,111,240                      // movdqa        %xmm0,%xmm6
   .byte  102,15,114,246,3                    // pslld         $0x3,%xmm6
   .byte  102,15,219,241                      // pand          %xmm1,%xmm6
   .byte  102,15,254,244                      // paddd         %xmm4,%xmm6
-  .byte  102,15,111,184,201,125,0,0          // movdqa        0x7dc9(%eax),%xmm7
+  .byte  102,15,111,184,233,134,0,0          // movdqa        0x86e9(%eax),%xmm7
   .byte  102,15,127,69,168                   // movdqa        %xmm0,-0x58(%ebp)
   .byte  102,15,219,199                      // pand          %xmm7,%xmm0
-  .byte  102,15,111,168,217,125,0,0          // movdqa        0x7dd9(%eax),%xmm5
+  .byte  102,15,111,168,249,134,0,0          // movdqa        0x86f9(%eax),%xmm5
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
   .byte  102,15,102,200                      // pcmpgtd       %xmm0,%xmm1
-  .byte  102,15,111,128,105,122,0,0          // movdqa        0x7a69(%eax),%xmm0
+  .byte  102,15,111,128,137,131,0,0          // movdqa        0x8389(%eax),%xmm0
   .byte  102,15,127,69,216                   // movdqa        %xmm0,-0x28(%ebp)
   .byte  102,15,254,240                      // paddd         %xmm0,%xmm6
   .byte  102,15,114,230,16                   // psrad         $0x10,%xmm6
@@ -69068,7 +70506,7 @@
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,97,195                       // punpcklwd     %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,40,162,124,121,0,0               // movaps        0x797c(%edx),%xmm4
+  .byte  15,40,162,156,130,0,0               // movaps        0x829c(%edx),%xmm4
   .byte  15,89,196                           // mulps         %xmm4,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -69172,7 +70610,7 @@
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,97,195                       // punpcklwd     %xmm3,%xmm0
   .byte  15,91,192                           // cvtdq2ps      %xmm0,%xmm0
-  .byte  15,40,170,21,120,0,0                // movaps        0x7815(%edx),%xmm5
+  .byte  15,40,170,53,129,0,0                // movaps        0x8135(%edx),%xmm5
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
   .byte  102,15,111,204                      // movdqa        %xmm4,%xmm1
   .byte  102,15,113,241,8                    // psllw         $0x8,%xmm1
@@ -69198,7 +70636,7 @@
   .byte  15,89,213                           // mulps         %xmm5,%xmm2
   .byte  141,112,8                           // lea           0x8(%eax),%esi
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,154,149,117,0,0               // movaps        0x7595(%edx),%xmm3
+  .byte  15,40,154,181,126,0,0               // movaps        0x7eb5(%edx),%xmm3
   .byte  86                                  // push          %esi
   .byte  81                                  // push          %ecx
   .byte  255,80,4                            // call          *0x4(%eax)
@@ -69245,7 +70683,7 @@
   .byte  1,210                               // add           %edx,%edx
   .byte  3,22                                // add           (%esi),%edx
   .byte  139,49                              // mov           (%ecx),%esi
-  .byte  15,40,183,113,121,0,0               // movaps        0x7971(%edi),%xmm6
+  .byte  15,40,183,145,130,0,0               // movaps        0x8291(%edi),%xmm6
   .byte  139,121,8                           // mov           0x8(%ecx),%edi
   .byte  193,230,2                           // shl           $0x2,%esi
   .byte  15,40,224                           // movaps        %xmm0,%xmm4
@@ -69525,7 +70963,7 @@
   .byte  243,15,91,245                       // cvttps2dq     %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  15,194,238,1                        // cmpltps       %xmm6,%xmm5
-  .byte  15,84,168,29,113,0,0                // andps         0x711d(%eax),%xmm5
+  .byte  15,84,168,61,122,0,0                // andps         0x7a3d(%eax),%xmm5
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
   .byte  15,89,230                           // mulps         %xmm6,%xmm4
@@ -69557,7 +70995,7 @@
   .byte  243,15,91,245                       // cvttps2dq     %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  15,194,238,1                        // cmpltps       %xmm6,%xmm5
-  .byte  15,84,168,203,112,0,0               // andps         0x70cb(%eax),%xmm5
+  .byte  15,84,168,235,121,0,0               // andps         0x79eb(%eax),%xmm5
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
   .byte  15,89,230                           // mulps         %xmm6,%xmm4
@@ -69587,13 +71025,13 @@
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
   .byte  15,92,198                           // subps         %xmm6,%xmm0
-  .byte  243,15,89,169,25,120,0,0            // mulss         0x7819(%ecx),%xmm5
+  .byte  243,15,89,169,57,129,0,0            // mulss         0x8139(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,89,232                           // mulps         %xmm0,%xmm5
   .byte  243,15,91,253                       // cvttps2dq     %xmm5,%xmm7
   .byte  15,91,255                           // cvtdq2ps      %xmm7,%xmm7
   .byte  15,194,239,1                        // cmpltps       %xmm7,%xmm5
-  .byte  15,84,169,121,112,0,0               // andps         0x7079(%ecx),%xmm5
+  .byte  15,84,169,153,121,0,0               // andps         0x7999(%ecx),%xmm5
   .byte  15,92,253                           // subps         %xmm5,%xmm7
   .byte  243,15,88,228                       // addss         %xmm4,%xmm4
   .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
@@ -69628,13 +71066,13 @@
   .byte  15,40,244                           // movaps        %xmm4,%xmm6
   .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
   .byte  15,92,206                           // subps         %xmm6,%xmm1
-  .byte  243,15,89,169,165,119,0,0           // mulss         0x77a5(%ecx),%xmm5
+  .byte  243,15,89,169,197,128,0,0           // mulss         0x80c5(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,89,233                           // mulps         %xmm1,%xmm5
   .byte  243,15,91,253                       // cvttps2dq     %xmm5,%xmm7
   .byte  15,91,255                           // cvtdq2ps      %xmm7,%xmm7
   .byte  15,194,239,1                        // cmpltps       %xmm7,%xmm5
-  .byte  15,84,169,5,112,0,0                 // andps         0x7005(%ecx),%xmm5
+  .byte  15,84,169,37,121,0,0                // andps         0x7925(%ecx),%xmm5
   .byte  15,92,253                           // subps         %xmm5,%xmm7
   .byte  243,15,88,228                       // addss         %xmm4,%xmm4
   .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
@@ -69665,7 +71103,7 @@
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,95,224                           // maxps         %xmm0,%xmm4
-  .byte  15,93,160,145,111,0,0               // minps         0x6f91(%eax),%xmm4
+  .byte  15,93,160,177,120,0,0               // minps         0x78b1(%eax),%xmm4
   .byte  141,65,4                            // lea           0x4(%ecx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  15,40,196                           // movaps        %xmm4,%xmm0
@@ -69690,7 +71128,7 @@
   .byte  15,91,236                           // cvtdq2ps      %xmm4,%xmm5
   .byte  15,40,240                           // movaps        %xmm0,%xmm6
   .byte  15,194,245,1                        // cmpltps       %xmm5,%xmm6
-  .byte  15,40,184,97,111,0,0                // movaps        0x6f61(%eax),%xmm7
+  .byte  15,40,184,129,120,0,0               // movaps        0x7881(%eax),%xmm7
   .byte  15,84,247                           // andps         %xmm7,%xmm6
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,92,238                           // subps         %xmm6,%xmm5
@@ -69716,14 +71154,14 @@
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  232,0,0,0,0                         // call          6139 <_sk_mirror_x_1_sse2+0xb>
   .byte  88                                  // pop           %eax
-  .byte  15,40,160,135,111,0,0               // movaps        0x6f87(%eax),%xmm4
+  .byte  15,40,160,167,120,0,0               // movaps        0x78a7(%eax),%xmm4
   .byte  15,88,196                           // addps         %xmm4,%xmm0
-  .byte  15,40,168,7,111,0,0                 // movaps        0x6f07(%eax),%xmm5
+  .byte  15,40,168,39,120,0,0                // movaps        0x7827(%eax),%xmm5
   .byte  15,89,232                           // mulps         %xmm0,%xmm5
   .byte  243,15,91,245                       // cvttps2dq     %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  15,194,238,1                        // cmpltps       %xmm6,%xmm5
-  .byte  15,40,184,23,111,0,0                // movaps        0x6f17(%eax),%xmm7
+  .byte  15,40,184,55,120,0,0                // movaps        0x7837(%eax),%xmm7
   .byte  15,84,239                           // andps         %xmm7,%xmm5
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  15,88,246                           // addps         %xmm6,%xmm6
@@ -69756,10 +71194,10 @@
   .byte  232,0,0,0,0                         // call          61a3 <_sk_luminance_to_alpha_sse2+0xe>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,89,128,237,115,0,0               // mulps         0x73ed(%eax),%xmm0
-  .byte  15,89,136,253,115,0,0               // mulps         0x73fd(%eax),%xmm1
+  .byte  15,89,128,13,125,0,0                // mulps         0x7d0d(%eax),%xmm0
+  .byte  15,89,136,29,125,0,0                // mulps         0x7d1d(%eax),%xmm1
   .byte  15,88,200                           // addps         %xmm0,%xmm1
-  .byte  15,89,152,13,116,0,0                // mulps         0x740d(%eax),%xmm3
+  .byte  15,89,152,45,125,0,0                // mulps         0x7d2d(%eax),%xmm3
   .byte  15,88,217                           // addps         %xmm1,%xmm3
   .byte  141,65,4                            // lea           0x4(%ecx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -70155,12 +71593,12 @@
   .byte  78                                  // dec           %esi
   .byte  102,15,110,206                      // movd          %esi,%xmm1
   .byte  102,15,112,201,0                    // pshufd        $0x0,%xmm1,%xmm1
-  .byte  102,15,111,146,135,111,0,0          // movdqa        0x6f87(%edx),%xmm2
+  .byte  102,15,111,146,167,120,0,0          // movdqa        0x78a7(%edx),%xmm2
   .byte  102,15,219,209                      // pand          %xmm1,%xmm2
-  .byte  102,15,235,146,71,109,0,0           // por           0x6d47(%edx),%xmm2
+  .byte  102,15,235,146,103,118,0,0          // por           0x7667(%edx),%xmm2
   .byte  102,15,114,209,16                   // psrld         $0x10,%xmm1
-  .byte  102,15,235,138,151,111,0,0          // por           0x6f97(%edx),%xmm1
-  .byte  15,88,138,167,111,0,0               // addps         0x6fa7(%edx),%xmm1
+  .byte  102,15,235,138,183,120,0,0          // por           0x78b7(%edx),%xmm1
+  .byte  15,88,138,199,120,0,0               // addps         0x78c7(%edx),%xmm1
   .byte  15,88,202                           // addps         %xmm2,%xmm1
   .byte  15,89,200                           // mulps         %xmm0,%xmm1
   .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
@@ -70269,15 +71707,15 @@
   .byte  232,0,0,0,0                         // call          67f2 <_sk_gauss_a_to_rgba_sse2+0xb>
   .byte  88                                  // pop           %eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
-  .byte  15,40,128,254,109,0,0               // movaps        0x6dfe(%eax),%xmm0
+  .byte  15,40,128,30,119,0,0                // movaps        0x771e(%eax),%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,128,14,110,0,0                // addps         0x6e0e(%eax),%xmm0
+  .byte  15,88,128,46,119,0,0                // addps         0x772e(%eax),%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,128,30,110,0,0                // addps         0x6e1e(%eax),%xmm0
+  .byte  15,88,128,62,119,0,0                // addps         0x773e(%eax),%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,128,46,110,0,0                // addps         0x6e2e(%eax),%xmm0
+  .byte  15,88,128,78,119,0,0                // addps         0x774e(%eax),%xmm0
   .byte  15,89,195                           // mulps         %xmm3,%xmm0
-  .byte  15,88,128,62,110,0,0                // addps         0x6e3e(%eax),%xmm0
+  .byte  15,88,128,94,119,0,0                // addps         0x775e(%eax),%xmm0
   .byte  141,65,4                            // lea           0x4(%ecx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
@@ -70484,30 +71922,30 @@
   .byte  15,94,247                           // divps         %xmm7,%xmm6
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
   .byte  15,89,255                           // mulps         %xmm7,%xmm7
-  .byte  15,40,128,191,107,0,0               // movaps        0x6bbf(%eax),%xmm0
+  .byte  15,40,128,223,116,0,0               // movaps        0x74df(%eax),%xmm0
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  15,88,128,207,107,0,0               // addps         0x6bcf(%eax),%xmm0
+  .byte  15,88,128,239,116,0,0               // addps         0x74ef(%eax),%xmm0
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  15,88,128,223,107,0,0               // addps         0x6bdf(%eax),%xmm0
+  .byte  15,88,128,255,116,0,0               // addps         0x74ff(%eax),%xmm0
   .byte  15,89,199                           // mulps         %xmm7,%xmm0
-  .byte  15,88,128,239,107,0,0               // addps         0x6bef(%eax),%xmm0
+  .byte  15,88,128,15,117,0,0                // addps         0x750f(%eax),%xmm0
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
   .byte  15,194,236,1                        // cmpltps       %xmm4,%xmm5
-  .byte  15,40,176,255,107,0,0               // movaps        0x6bff(%eax),%xmm6
+  .byte  15,40,176,31,117,0,0                // movaps        0x751f(%eax),%xmm6
   .byte  15,92,240                           // subps         %xmm0,%xmm6
   .byte  15,84,245                           // andps         %xmm5,%xmm6
   .byte  15,85,232                           // andnps        %xmm0,%xmm5
   .byte  15,87,228                           // xorps         %xmm4,%xmm4
   .byte  15,86,238                           // orps          %xmm6,%xmm5
   .byte  15,194,204,1                        // cmpltps       %xmm4,%xmm1
-  .byte  15,40,128,191,101,0,0               // movaps        0x65bf(%eax),%xmm0
+  .byte  15,40,128,223,110,0,0               // movaps        0x6edf(%eax),%xmm0
   .byte  15,92,197                           // subps         %xmm5,%xmm0
   .byte  15,84,193                           // andps         %xmm1,%xmm0
   .byte  15,85,205                           // andnps        %xmm5,%xmm1
   .byte  15,86,200                           // orps          %xmm0,%xmm1
   .byte  15,40,194                           // movaps        %xmm2,%xmm0
   .byte  15,194,196,1                        // cmpltps       %xmm4,%xmm0
-  .byte  15,40,168,207,101,0,0               // movaps        0x65cf(%eax),%xmm5
+  .byte  15,40,168,239,110,0,0               // movaps        0x6eef(%eax),%xmm5
   .byte  15,92,233                           // subps         %xmm1,%xmm5
   .byte  15,84,232                           // andps         %xmm0,%xmm5
   .byte  15,85,193                           // andnps        %xmm1,%xmm0
@@ -70575,8 +72013,8 @@
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  243,15,16,65,64                     // movss         0x40(%ecx),%xmm0
   .byte  15,198,192,0                        // shufps        $0x0,%xmm0,%xmm0
-  .byte  15,89,162,242,106,0,0               // mulps         0x6af2(%edx),%xmm4
-  .byte  15,89,130,2,107,0,0                 // mulps         0x6b02(%edx),%xmm0
+  .byte  15,89,162,18,116,0,0                // mulps         0x7412(%edx),%xmm4
+  .byte  15,89,130,34,116,0,0                // mulps         0x7422(%edx),%xmm0
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,237                           // mulps         %xmm5,%xmm5
@@ -70584,11 +72022,11 @@
   .byte  15,81,237                           // sqrtps        %xmm5,%xmm5
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  15,87,162,162,105,0,0               // xorps         0x69a2(%edx),%xmm4
+  .byte  15,87,162,194,114,0,0               // xorps         0x72c2(%edx),%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  243,15,16,105,68                    // movss         0x44(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
-  .byte  15,89,170,162,100,0,0               // mulps         0x64a2(%edx),%xmm5
+  .byte  15,89,170,194,109,0,0               // mulps         0x6dc2(%edx),%xmm5
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,95,196                           // maxps         %xmm4,%xmm0
@@ -70626,8 +72064,8 @@
   .byte  15,92,245                           // subps         %xmm5,%xmm6
   .byte  243,15,16,65,64                     // movss         0x40(%ecx),%xmm0
   .byte  15,198,192,0                        // shufps        $0x0,%xmm0,%xmm0
-  .byte  15,89,162,84,106,0,0                // mulps         0x6a54(%edx),%xmm4
-  .byte  15,89,130,100,106,0,0               // mulps         0x6a64(%edx),%xmm0
+  .byte  15,89,162,116,115,0,0               // mulps         0x7374(%edx),%xmm4
+  .byte  15,89,130,132,115,0,0               // mulps         0x7384(%edx),%xmm0
   .byte  15,89,198                           // mulps         %xmm6,%xmm0
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,237                           // mulps         %xmm5,%xmm5
@@ -70635,11 +72073,11 @@
   .byte  15,81,237                           // sqrtps        %xmm5,%xmm5
   .byte  15,40,197                           // movaps        %xmm5,%xmm0
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  15,87,162,4,105,0,0                 // xorps         0x6904(%edx),%xmm4
+  .byte  15,87,162,36,114,0,0                // xorps         0x7224(%edx),%xmm4
   .byte  15,92,229                           // subps         %xmm5,%xmm4
   .byte  243,15,16,105,68                    // movss         0x44(%ecx),%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
-  .byte  15,89,170,4,100,0,0                 // mulps         0x6404(%edx),%xmm5
+  .byte  15,89,170,36,109,0,0                // mulps         0x6d24(%edx),%xmm5
   .byte  15,89,197                           // mulps         %xmm5,%xmm0
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,93,196                           // minps         %xmm4,%xmm0
@@ -70668,7 +72106,7 @@
   .byte  243,15,89,236                       // mulss         %xmm4,%xmm5
   .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
   .byte  15,88,232                           // addps         %xmm0,%xmm5
-  .byte  15,89,169,220,105,0,0               // mulps         0x69dc(%ecx),%xmm5
+  .byte  15,89,169,252,114,0,0               // mulps         0x72fc(%ecx),%xmm5
   .byte  15,89,192                           // mulps         %xmm0,%xmm0
   .byte  15,40,241                           // movaps        %xmm1,%xmm6
   .byte  15,89,246                           // mulps         %xmm6,%xmm6
@@ -70676,7 +72114,7 @@
   .byte  243,15,89,228                       // mulss         %xmm4,%xmm4
   .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
   .byte  15,92,196                           // subps         %xmm4,%xmm0
-  .byte  15,87,129,140,104,0,0               // xorps         0x688c(%ecx),%xmm0
+  .byte  15,87,129,172,113,0,0               // xorps         0x71ac(%ecx),%xmm0
   .byte  15,94,197                           // divps         %xmm5,%xmm0
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -70755,14 +72193,14 @@
   .byte  15,40,218                           // movaps        %xmm2,%xmm3
   .byte  232,0,0,0,0                         // call          6d9d <_sk_save_xy_sse2+0x12>
   .byte  88                                  // pop           %eax
-  .byte  15,40,160,163,98,0,0                // movaps        0x62a3(%eax),%xmm4
+  .byte  15,40,160,195,107,0,0               // movaps        0x6bc3(%eax),%xmm4
   .byte  15,40,232                           // movaps        %xmm0,%xmm5
   .byte  15,88,236                           // addps         %xmm4,%xmm5
   .byte  243,15,91,245                       // cvttps2dq     %xmm5,%xmm6
   .byte  15,91,246                           // cvtdq2ps      %xmm6,%xmm6
   .byte  15,40,253                           // movaps        %xmm5,%xmm7
   .byte  15,194,254,1                        // cmpltps       %xmm6,%xmm7
-  .byte  15,40,144,179,98,0,0                // movaps        0x62b3(%eax),%xmm2
+  .byte  15,40,144,211,107,0,0               // movaps        0x6bd3(%eax),%xmm2
   .byte  15,84,250                           // andps         %xmm2,%xmm7
   .byte  15,92,247                           // subps         %xmm7,%xmm6
   .byte  15,92,238                           // subps         %xmm6,%xmm5
@@ -70844,8 +72282,8 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,166,128,0,0,0                 // movups        0x80(%esi),%xmm4
-  .byte  15,88,128,43,104,0,0                // addps         0x682b(%eax),%xmm0
-  .byte  15,40,168,203,97,0,0                // movaps        0x61cb(%eax),%xmm5
+  .byte  15,88,128,75,113,0,0                // addps         0x714b(%eax),%xmm0
+  .byte  15,40,168,235,106,0,0               // movaps        0x6aeb(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,17,174,0,1,0,0                   // movups        %xmm5,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -70873,7 +72311,7 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,166,128,0,0,0                 // movups        0x80(%esi),%xmm4
-  .byte  15,88,128,117,97,0,0                // addps         0x6175(%eax),%xmm0
+  .byte  15,88,128,149,106,0,0               // addps         0x6a95(%eax),%xmm0
   .byte  15,17,166,0,1,0,0                   // movups        %xmm4,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -70900,8 +72338,8 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,166,192,0,0,0                 // movups        0xc0(%esi),%xmm4
-  .byte  15,88,136,169,103,0,0               // addps         0x67a9(%eax),%xmm1
-  .byte  15,40,168,73,97,0,0                 // movaps        0x6149(%eax),%xmm5
+  .byte  15,88,136,201,112,0,0               // addps         0x70c9(%eax),%xmm1
+  .byte  15,40,168,105,106,0,0               // movaps        0x6a69(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,17,174,64,1,0,0                  // movups        %xmm5,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -70929,7 +72367,7 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,166,192,0,0,0                 // movups        0xc0(%esi),%xmm4
-  .byte  15,88,136,242,96,0,0                // addps         0x60f2(%eax),%xmm1
+  .byte  15,88,136,18,106,0,0                // addps         0x6a12(%eax),%xmm1
   .byte  15,17,166,64,1,0,0                  // movups        %xmm4,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -70956,13 +72394,13 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,166,128,0,0,0                 // movups        0x80(%esi),%xmm4
-  .byte  15,88,128,53,103,0,0                // addps         0x6735(%eax),%xmm0
-  .byte  15,40,168,197,96,0,0                // movaps        0x60c5(%eax),%xmm5
+  .byte  15,88,128,85,112,0,0                // addps         0x7055(%eax),%xmm0
+  .byte  15,40,168,229,105,0,0               // movaps        0x69e5(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,40,229                           // movaps        %xmm5,%xmm4
   .byte  15,89,228                           // mulps         %xmm4,%xmm4
-  .byte  15,89,168,69,103,0,0                // mulps         0x6745(%eax),%xmm5
-  .byte  15,88,168,181,98,0,0                // addps         0x62b5(%eax),%xmm5
+  .byte  15,89,168,101,112,0,0               // mulps         0x7065(%eax),%xmm5
+  .byte  15,88,168,213,107,0,0               // addps         0x6bd5(%eax),%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
   .byte  15,17,174,0,1,0,0                   // movups        %xmm5,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -70990,16 +72428,16 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,166,128,0,0,0                 // movups        0x80(%esi),%xmm4
-  .byte  15,88,128,200,102,0,0               // addps         0x66c8(%eax),%xmm0
-  .byte  15,40,168,104,96,0,0                // movaps        0x6068(%eax),%xmm5
+  .byte  15,88,128,232,111,0,0               // addps         0x6fe8(%eax),%xmm0
+  .byte  15,40,168,136,105,0,0               // movaps        0x6988(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
-  .byte  15,40,160,248,102,0,0               // movaps        0x66f8(%eax),%xmm4
+  .byte  15,40,160,24,112,0,0                // movaps        0x7018(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,8,103,0,0                 // addps         0x6708(%eax),%xmm4
+  .byte  15,88,160,40,112,0,0                // addps         0x7028(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,88,96,0,0                 // addps         0x6058(%eax),%xmm4
+  .byte  15,88,160,120,105,0,0               // addps         0x6978(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,24,103,0,0                // addps         0x6718(%eax),%xmm4
+  .byte  15,88,160,56,112,0,0                // addps         0x7038(%eax),%xmm4
   .byte  15,17,166,0,1,0,0                   // movups        %xmm4,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -71024,17 +72462,17 @@
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,12                           // mov           0xc(%ebp),%edx
   .byte  139,50                              // mov           (%edx),%esi
-  .byte  15,40,160,237,95,0,0                // movaps        0x5fed(%eax),%xmm4
+  .byte  15,40,160,13,105,0,0                // movaps        0x690d(%eax),%xmm4
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,174,128,0,0,0                 // movups        0x80(%esi),%xmm5
   .byte  15,88,196                           // addps         %xmm4,%xmm0
-  .byte  15,40,176,141,102,0,0               // movaps        0x668d(%eax),%xmm6
+  .byte  15,40,176,173,111,0,0               // movaps        0x6fad(%eax),%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
-  .byte  15,88,176,157,102,0,0               // addps         0x669d(%eax),%xmm6
+  .byte  15,88,176,189,111,0,0               // addps         0x6fbd(%eax),%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
   .byte  15,88,244                           // addps         %xmm4,%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
-  .byte  15,88,176,173,102,0,0               // addps         0x66ad(%eax),%xmm6
+  .byte  15,88,176,205,111,0,0               // addps         0x6fcd(%eax),%xmm6
   .byte  15,17,182,0,1,0,0                   // movups        %xmm6,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -71061,11 +72499,11 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,6                             // movups        (%esi),%xmm0
   .byte  15,16,166,128,0,0,0                 // movups        0x80(%esi),%xmm4
-  .byte  15,88,128,61,102,0,0                // addps         0x663d(%eax),%xmm0
+  .byte  15,88,128,93,111,0,0                // addps         0x6f5d(%eax),%xmm0
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,237                           // mulps         %xmm5,%xmm5
-  .byte  15,89,160,29,102,0,0                // mulps         0x661d(%eax),%xmm4
-  .byte  15,88,160,141,97,0,0                // addps         0x618d(%eax),%xmm4
+  .byte  15,89,160,61,111,0,0                // mulps         0x6f3d(%eax),%xmm4
+  .byte  15,88,160,173,106,0,0               // addps         0x6aad(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,17,166,0,1,0,0                   // movups        %xmm4,0x100(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -71093,13 +72531,13 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,166,192,0,0,0                 // movups        0xc0(%esi),%xmm4
-  .byte  15,88,136,186,101,0,0               // addps         0x65ba(%eax),%xmm1
-  .byte  15,40,168,74,95,0,0                 // movaps        0x5f4a(%eax),%xmm5
+  .byte  15,88,136,218,110,0,0               // addps         0x6eda(%eax),%xmm1
+  .byte  15,40,168,106,104,0,0               // movaps        0x686a(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
   .byte  15,40,229                           // movaps        %xmm5,%xmm4
   .byte  15,89,228                           // mulps         %xmm4,%xmm4
-  .byte  15,89,168,202,101,0,0               // mulps         0x65ca(%eax),%xmm5
-  .byte  15,88,168,58,97,0,0                 // addps         0x613a(%eax),%xmm5
+  .byte  15,89,168,234,110,0,0               // mulps         0x6eea(%eax),%xmm5
+  .byte  15,88,168,90,106,0,0                // addps         0x6a5a(%eax),%xmm5
   .byte  15,89,236                           // mulps         %xmm4,%xmm5
   .byte  15,17,174,64,1,0,0                  // movups        %xmm5,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -71127,16 +72565,16 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,166,192,0,0,0                 // movups        0xc0(%esi),%xmm4
-  .byte  15,88,136,76,101,0,0                // addps         0x654c(%eax),%xmm1
-  .byte  15,40,168,236,94,0,0                // movaps        0x5eec(%eax),%xmm5
+  .byte  15,88,136,108,110,0,0               // addps         0x6e6c(%eax),%xmm1
+  .byte  15,40,168,12,104,0,0                // movaps        0x680c(%eax),%xmm5
   .byte  15,92,236                           // subps         %xmm4,%xmm5
-  .byte  15,40,160,124,101,0,0               // movaps        0x657c(%eax),%xmm4
+  .byte  15,40,160,156,110,0,0               // movaps        0x6e9c(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,140,101,0,0               // addps         0x658c(%eax),%xmm4
+  .byte  15,88,160,172,110,0,0               // addps         0x6eac(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,220,94,0,0                // addps         0x5edc(%eax),%xmm4
+  .byte  15,88,160,252,103,0,0               // addps         0x67fc(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
-  .byte  15,88,160,156,101,0,0               // addps         0x659c(%eax),%xmm4
+  .byte  15,88,160,188,110,0,0               // addps         0x6ebc(%eax),%xmm4
   .byte  15,17,166,64,1,0,0                  // movups        %xmm4,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -71161,17 +72599,17 @@
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,12                           // mov           0xc(%ebp),%edx
   .byte  139,50                              // mov           (%edx),%esi
-  .byte  15,40,160,112,94,0,0                // movaps        0x5e70(%eax),%xmm4
+  .byte  15,40,160,144,103,0,0               // movaps        0x6790(%eax),%xmm4
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,174,192,0,0,0                 // movups        0xc0(%esi),%xmm5
   .byte  15,88,204                           // addps         %xmm4,%xmm1
-  .byte  15,40,176,16,101,0,0                // movaps        0x6510(%eax),%xmm6
+  .byte  15,40,176,48,110,0,0                // movaps        0x6e30(%eax),%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
-  .byte  15,88,176,32,101,0,0                // addps         0x6520(%eax),%xmm6
+  .byte  15,88,176,64,110,0,0                // addps         0x6e40(%eax),%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
   .byte  15,88,244                           // addps         %xmm4,%xmm6
   .byte  15,89,245                           // mulps         %xmm5,%xmm6
-  .byte  15,88,176,48,101,0,0                // addps         0x6530(%eax),%xmm6
+  .byte  15,88,176,80,110,0,0                // addps         0x6e50(%eax),%xmm6
   .byte  15,17,182,64,1,0,0                  // movups        %xmm6,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
@@ -71198,11 +72636,11 @@
   .byte  139,50                              // mov           (%edx),%esi
   .byte  15,16,78,64                         // movups        0x40(%esi),%xmm1
   .byte  15,16,166,192,0,0,0                 // movups        0xc0(%esi),%xmm4
-  .byte  15,88,136,191,100,0,0               // addps         0x64bf(%eax),%xmm1
+  .byte  15,88,136,223,109,0,0               // addps         0x6ddf(%eax),%xmm1
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
   .byte  15,89,237                           // mulps         %xmm5,%xmm5
-  .byte  15,89,160,159,100,0,0               // mulps         0x649f(%eax),%xmm4
-  .byte  15,88,160,15,96,0,0                 // addps         0x600f(%eax),%xmm4
+  .byte  15,89,160,191,109,0,0               // mulps         0x6dbf(%eax),%xmm4
+  .byte  15,88,160,47,105,0,0                // addps         0x692f(%eax),%xmm4
   .byte  15,89,229                           // mulps         %xmm5,%xmm4
   .byte  15,17,166,64,1,0,0                  // movups        %xmm4,0x140(%esi)
   .byte  141,66,8                            // lea           0x8(%edx),%eax
@@ -71353,7 +72791,7 @@
   .byte  102,15,254,216                      // paddd         %xmm0,%xmm3
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,254,195                      // paddd         %xmm3,%xmm0
-  .byte  102,15,111,162,9,99,0,0             // movdqa        0x6309(%edx),%xmm4
+  .byte  102,15,111,162,41,108,0,0           // movdqa        0x6c29(%edx),%xmm4
   .byte  102,15,127,165,120,255,255,255      // movdqa        %xmm4,-0x88(%ebp)
   .byte  102,15,112,200,245                  // pshufd        $0xf5,%xmm0,%xmm1
   .byte  102,15,244,196                      // pmuludq       %xmm4,%xmm0
@@ -71393,7 +72831,7 @@
   .byte  243,15,16,4,129                     // movss         (%ecx,%eax,4),%xmm0
   .byte  15,20,248                           // unpcklps      %xmm0,%xmm7
   .byte  102,15,20,252                       // unpcklpd      %xmm4,%xmm7
-  .byte  102,15,111,130,121,92,0,0           // movdqa        0x5c79(%edx),%xmm0
+  .byte  102,15,111,130,153,101,0,0          // movdqa        0x6599(%edx),%xmm0
   .byte  102,15,127,133,104,255,255,255      // movdqa        %xmm0,-0x98(%ebp)
   .byte  102,15,254,200                      // paddd         %xmm0,%xmm1
   .byte  102,15,112,193,229                  // pshufd        $0xe5,%xmm1,%xmm0
@@ -71410,7 +72848,7 @@
   .byte  243,15,16,4,129                     // movss         (%ecx,%eax,4),%xmm0
   .byte  15,20,240                           // unpcklps      %xmm0,%xmm6
   .byte  102,15,20,241                       // unpcklpd      %xmm1,%xmm6
-  .byte  15,40,130,249,98,0,0                // movaps        0x62f9(%edx),%xmm0
+  .byte  15,40,130,25,108,0,0                // movaps        0x6c19(%edx),%xmm0
   .byte  15,40,141,248,254,255,255           // movaps        -0x108(%ebp),%xmm1
   .byte  15,88,200                           // addps         %xmm0,%xmm1
   .byte  15,41,77,200                        // movaps        %xmm1,-0x38(%ebp)
@@ -72041,7 +73479,7 @@
   .byte  102,15,111,212                      // movdqa        %xmm4,%xmm2
   .byte  102,15,254,214                      // paddd         %xmm6,%xmm2
   .byte  102,15,254,218                      // paddd         %xmm2,%xmm3
-  .byte  102,15,111,161,125,87,0,0           // movdqa        0x577d(%ecx),%xmm4
+  .byte  102,15,111,161,157,96,0,0           // movdqa        0x609d(%ecx),%xmm4
   .byte  102,15,112,203,245                  // pshufd        $0xf5,%xmm3,%xmm1
   .byte  102,15,244,220                      // pmuludq       %xmm4,%xmm3
   .byte  102,15,244,204                      // pmuludq       %xmm4,%xmm1
@@ -72083,7 +73521,7 @@
   .byte  15,20,223                           // unpcklps      %xmm7,%xmm3
   .byte  102,15,20,222                       // unpcklpd      %xmm6,%xmm3
   .byte  102,15,41,157,232,254,255,255       // movapd        %xmm3,-0x118(%ebp)
-  .byte  102,15,254,137,237,80,0,0           // paddd         0x50ed(%ecx),%xmm1
+  .byte  102,15,254,137,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm1
   .byte  102,15,112,241,229                  // pshufd        $0xe5,%xmm1,%xmm6
   .byte  102,15,126,240                      // movd          %xmm6,%eax
   .byte  102,15,112,241,78                   // pshufd        $0x4e,%xmm1,%xmm6
@@ -72098,7 +73536,7 @@
   .byte  243,15,16,60,130                    // movss         (%edx,%eax,4),%xmm7
   .byte  15,20,223                           // unpcklps      %xmm7,%xmm3
   .byte  102,15,20,222                       // unpcklpd      %xmm6,%xmm3
-  .byte  15,40,185,109,87,0,0                // movaps        0x576d(%ecx),%xmm7
+  .byte  15,40,185,141,96,0,0                // movaps        0x608d(%ecx),%xmm7
   .byte  15,88,199                           // addps         %xmm7,%xmm0
   .byte  15,41,133,120,255,255,255           // movaps        %xmm0,-0x88(%ebp)
   .byte  15,40,133,248,254,255,255           // movaps        -0x108(%ebp),%xmm0
@@ -72119,7 +73557,7 @@
   .byte  102,15,127,101,168                  // movdqa        %xmm4,-0x58(%ebp)
   .byte  102,15,254,212                      // paddd         %xmm4,%xmm2
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,161,125,87,0,0           // movdqa        0x577d(%ecx),%xmm4
+  .byte  102,15,111,161,157,96,0,0           // movdqa        0x609d(%ecx),%xmm4
   .byte  102,15,244,212                      // pmuludq       %xmm4,%xmm2
   .byte  102,15,244,196                      // pmuludq       %xmm4,%xmm0
   .byte  102,15,112,240,232                  // pshufd        $0xe8,%xmm0,%xmm6
@@ -72155,7 +73593,7 @@
   .byte  243,15,16,44,130                    // movss         (%edx,%eax,4),%xmm5
   .byte  15,20,213                           // unpcklps      %xmm5,%xmm2
   .byte  102,15,20,214                       // unpcklpd      %xmm6,%xmm2
-  .byte  102,15,254,129,237,80,0,0           // paddd         0x50ed(%ecx),%xmm0
+  .byte  102,15,254,129,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm0
   .byte  102,15,112,232,229                  // pshufd        $0xe5,%xmm0,%xmm5
   .byte  102,15,126,232                      // movd          %xmm5,%eax
   .byte  102,15,112,232,78                   // pshufd        $0x4e,%xmm0,%xmm5
@@ -72239,7 +73677,7 @@
   .byte  243,15,16,44,130                    // movss         (%edx,%eax,4),%xmm5
   .byte  15,20,245                           // unpcklps      %xmm5,%xmm6
   .byte  102,15,20,240                       // unpcklpd      %xmm0,%xmm6
-  .byte  102,15,254,137,237,80,0,0           // paddd         0x50ed(%ecx),%xmm1
+  .byte  102,15,254,137,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm1
   .byte  102,15,112,193,229                  // pshufd        $0xe5,%xmm1,%xmm0
   .byte  102,15,126,192                      // movd          %xmm0,%eax
   .byte  102,15,112,193,78                   // pshufd        $0x4e,%xmm1,%xmm0
@@ -72257,7 +73695,7 @@
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,15,254,77,168                   // paddd         -0x58(%ebp),%xmm1
   .byte  102,15,112,193,245                  // pshufd        $0xf5,%xmm1,%xmm0
-  .byte  102,15,111,145,125,87,0,0           // movdqa        0x577d(%ecx),%xmm2
+  .byte  102,15,111,145,157,96,0,0           // movdqa        0x609d(%ecx),%xmm2
   .byte  102,15,244,202                      // pmuludq       %xmm2,%xmm1
   .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
   .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
@@ -72293,7 +73731,7 @@
   .byte  243,15,16,4,130                     // movss         (%edx,%eax,4),%xmm0
   .byte  15,20,216                           // unpcklps      %xmm0,%xmm3
   .byte  102,15,20,217                       // unpcklpd      %xmm1,%xmm3
-  .byte  102,15,254,169,237,80,0,0           // paddd         0x50ed(%ecx),%xmm5
+  .byte  102,15,254,169,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm5
   .byte  102,15,112,197,229                  // pshufd        $0xe5,%xmm5,%xmm0
   .byte  102,15,126,192                      // movd          %xmm0,%eax
   .byte  102,15,112,197,78                   // pshufd        $0x4e,%xmm5,%xmm0
@@ -72353,7 +73791,7 @@
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,254,85,152                   // paddd         -0x68(%ebp),%xmm2
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,153,125,87,0,0           // movdqa        0x577d(%ecx),%xmm3
+  .byte  102,15,111,153,157,96,0,0           // movdqa        0x609d(%ecx),%xmm3
   .byte  102,15,244,211                      // pmuludq       %xmm3,%xmm2
   .byte  102,15,244,195                      // pmuludq       %xmm3,%xmm0
   .byte  102,15,111,251                      // movdqa        %xmm3,%xmm7
@@ -72392,7 +73830,7 @@
   .byte  243,15,16,44,130                    // movss         (%edx,%eax,4),%xmm5
   .byte  15,20,229                           // unpcklps      %xmm5,%xmm4
   .byte  102,15,20,227                       // unpcklpd      %xmm3,%xmm4
-  .byte  102,15,254,129,237,80,0,0           // paddd         0x50ed(%ecx),%xmm0
+  .byte  102,15,254,129,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm0
   .byte  102,15,112,216,229                  // pshufd        $0xe5,%xmm0,%xmm3
   .byte  102,15,126,216                      // movd          %xmm3,%eax
   .byte  102,15,112,216,78                   // pshufd        $0x4e,%xmm0,%xmm3
@@ -72444,7 +73882,7 @@
   .byte  243,15,16,52,130                    // movss         (%edx,%eax,4),%xmm6
   .byte  15,20,238                           // unpcklps      %xmm6,%xmm5
   .byte  102,15,20,235                       // unpcklpd      %xmm3,%xmm5
-  .byte  102,15,111,137,237,80,0,0           // movdqa        0x50ed(%ecx),%xmm1
+  .byte  102,15,111,137,13,90,0,0            // movdqa        0x5a0d(%ecx),%xmm1
   .byte  102,15,254,249                      // paddd         %xmm1,%xmm7
   .byte  102,15,112,223,229                  // pshufd        $0xe5,%xmm7,%xmm3
   .byte  102,15,126,216                      // movd          %xmm3,%eax
@@ -72478,7 +73916,7 @@
   .byte  102,15,111,226                      // movdqa        %xmm2,%xmm4
   .byte  102,15,254,69,152                   // paddd         -0x68(%ebp),%xmm0
   .byte  102,15,112,208,245                  // pshufd        $0xf5,%xmm0,%xmm2
-  .byte  102,15,111,153,125,87,0,0           // movdqa        0x577d(%ecx),%xmm3
+  .byte  102,15,111,153,157,96,0,0           // movdqa        0x609d(%ecx),%xmm3
   .byte  102,15,244,195                      // pmuludq       %xmm3,%xmm0
   .byte  102,15,244,211                      // pmuludq       %xmm3,%xmm2
   .byte  102,15,112,218,232                  // pshufd        $0xe8,%xmm2,%xmm3
@@ -72535,7 +73973,7 @@
   .byte  102,15,111,212                      // movdqa        %xmm4,%xmm2
   .byte  102,15,254,85,168                   // paddd         -0x58(%ebp),%xmm2
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,137,125,87,0,0           // movdqa        0x577d(%ecx),%xmm1
+  .byte  102,15,111,137,157,96,0,0           // movdqa        0x609d(%ecx),%xmm1
   .byte  102,15,244,209                      // pmuludq       %xmm1,%xmm2
   .byte  102,15,244,193                      // pmuludq       %xmm1,%xmm0
   .byte  102,15,112,240,232                  // pshufd        $0xe8,%xmm0,%xmm6
@@ -72572,7 +74010,7 @@
   .byte  243,15,16,36,130                    // movss         (%edx,%eax,4),%xmm4
   .byte  15,20,212                           // unpcklps      %xmm4,%xmm2
   .byte  102,15,20,214                       // unpcklpd      %xmm6,%xmm2
-  .byte  102,15,254,129,237,80,0,0           // paddd         0x50ed(%ecx),%xmm0
+  .byte  102,15,254,129,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm0
   .byte  102,15,112,224,229                  // pshufd        $0xe5,%xmm0,%xmm4
   .byte  102,15,126,224                      // movd          %xmm4,%eax
   .byte  102,15,112,224,78                   // pshufd        $0x4e,%xmm0,%xmm4
@@ -72640,7 +74078,7 @@
   .byte  102,15,111,93,152                   // movdqa        -0x68(%ebp),%xmm3
   .byte  102,15,254,217                      // paddd         %xmm1,%xmm3
   .byte  102,15,112,211,245                  // pshufd        $0xf5,%xmm3,%xmm2
-  .byte  102,15,111,129,125,87,0,0           // movdqa        0x577d(%ecx),%xmm0
+  .byte  102,15,111,129,157,96,0,0           // movdqa        0x609d(%ecx),%xmm0
   .byte  102,15,244,216                      // pmuludq       %xmm0,%xmm3
   .byte  102,15,244,208                      // pmuludq       %xmm0,%xmm2
   .byte  102,15,111,248                      // movdqa        %xmm0,%xmm7
@@ -72679,7 +74117,7 @@
   .byte  243,15,16,28,130                    // movss         (%edx,%eax,4),%xmm3
   .byte  15,20,235                           // unpcklps      %xmm3,%xmm5
   .byte  102,15,20,238                       // unpcklpd      %xmm6,%xmm5
-  .byte  102,15,254,145,237,80,0,0           // paddd         0x50ed(%ecx),%xmm2
+  .byte  102,15,254,145,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm2
   .byte  102,15,112,218,229                  // pshufd        $0xe5,%xmm2,%xmm3
   .byte  102,15,126,216                      // movd          %xmm3,%eax
   .byte  102,15,112,218,78                   // pshufd        $0x4e,%xmm2,%xmm3
@@ -72731,7 +74169,7 @@
   .byte  243,15,16,4,130                     // movss         (%edx,%eax,4),%xmm0
   .byte  15,20,208                           // unpcklps      %xmm0,%xmm2
   .byte  102,15,20,215                       // unpcklpd      %xmm7,%xmm2
-  .byte  102,15,111,137,237,80,0,0           // movdqa        0x50ed(%ecx),%xmm1
+  .byte  102,15,111,137,13,90,0,0            // movdqa        0x5a0d(%ecx),%xmm1
   .byte  102,15,254,217                      // paddd         %xmm1,%xmm3
   .byte  102,15,112,195,229                  // pshufd        $0xe5,%xmm3,%xmm0
   .byte  102,15,126,192                      // movd          %xmm0,%eax
@@ -72766,7 +74204,7 @@
   .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
   .byte  102,15,254,69,152                   // paddd         -0x68(%ebp),%xmm0
   .byte  102,15,112,224,245                  // pshufd        $0xf5,%xmm0,%xmm4
-  .byte  102,15,111,169,125,87,0,0           // movdqa        0x577d(%ecx),%xmm5
+  .byte  102,15,111,169,157,96,0,0           // movdqa        0x609d(%ecx),%xmm5
   .byte  102,15,111,221                      // movdqa        %xmm5,%xmm3
   .byte  102,15,244,195                      // pmuludq       %xmm3,%xmm0
   .byte  102,15,244,227                      // pmuludq       %xmm3,%xmm4
@@ -72822,7 +74260,7 @@
   .byte  102,15,111,202                      // movdqa        %xmm2,%xmm1
   .byte  102,15,254,77,168                   // paddd         -0x58(%ebp),%xmm1
   .byte  102,15,112,193,245                  // pshufd        $0xf5,%xmm1,%xmm0
-  .byte  102,15,111,145,125,87,0,0           // movdqa        0x577d(%ecx),%xmm2
+  .byte  102,15,111,145,157,96,0,0           // movdqa        0x609d(%ecx),%xmm2
   .byte  102,15,244,202                      // pmuludq       %xmm2,%xmm1
   .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
   .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
@@ -72858,7 +74296,7 @@
   .byte  243,15,16,4,130                     // movss         (%edx,%eax,4),%xmm0
   .byte  15,20,216                           // unpcklps      %xmm0,%xmm3
   .byte  102,15,20,217                       // unpcklpd      %xmm1,%xmm3
-  .byte  102,15,254,185,237,80,0,0           // paddd         0x50ed(%ecx),%xmm7
+  .byte  102,15,254,185,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm7
   .byte  102,15,112,199,229                  // pshufd        $0xe5,%xmm7,%xmm0
   .byte  102,15,126,192                      // movd          %xmm0,%eax
   .byte  102,15,112,199,78                   // pshufd        $0x4e,%xmm7,%xmm0
@@ -72908,7 +74346,7 @@
   .byte  102,15,111,233                      // movdqa        %xmm1,%xmm5
   .byte  102,15,254,69,152                   // paddd         -0x68(%ebp),%xmm0
   .byte  102,15,112,200,245                  // pshufd        $0xf5,%xmm0,%xmm1
-  .byte  102,15,111,145,125,87,0,0           // movdqa        0x577d(%ecx),%xmm2
+  .byte  102,15,111,145,157,96,0,0           // movdqa        0x609d(%ecx),%xmm2
   .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
   .byte  102,15,244,202                      // pmuludq       %xmm2,%xmm1
   .byte  102,15,111,250                      // movdqa        %xmm2,%xmm7
@@ -72947,7 +74385,7 @@
   .byte  243,15,16,36,130                    // movss         (%edx,%eax,4),%xmm4
   .byte  15,20,204                           // unpcklps      %xmm4,%xmm1
   .byte  102,15,20,203                       // unpcklpd      %xmm3,%xmm1
-  .byte  102,15,254,145,237,80,0,0           // paddd         0x50ed(%ecx),%xmm2
+  .byte  102,15,254,145,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm2
   .byte  102,15,112,218,229                  // pshufd        $0xe5,%xmm2,%xmm3
   .byte  102,15,126,216                      // movd          %xmm3,%eax
   .byte  102,15,112,218,78                   // pshufd        $0x4e,%xmm2,%xmm3
@@ -73001,7 +74439,7 @@
   .byte  243,15,16,36,130                    // movss         (%edx,%eax,4),%xmm4
   .byte  15,20,252                           // unpcklps      %xmm4,%xmm7
   .byte  102,15,20,254                       // unpcklpd      %xmm6,%xmm7
-  .byte  102,15,254,153,237,80,0,0           // paddd         0x50ed(%ecx),%xmm3
+  .byte  102,15,254,153,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm3
   .byte  102,15,112,227,229                  // pshufd        $0xe5,%xmm3,%xmm4
   .byte  102,15,126,224                      // movd          %xmm4,%eax
   .byte  102,15,112,227,78                   // pshufd        $0x4e,%xmm3,%xmm4
@@ -73034,7 +74472,7 @@
   .byte  102,15,254,208                      // paddd         %xmm0,%xmm2
   .byte  102,15,111,216                      // movdqa        %xmm0,%xmm3
   .byte  102,15,112,194,245                  // pshufd        $0xf5,%xmm2,%xmm0
-  .byte  102,15,111,137,125,87,0,0           // movdqa        0x577d(%ecx),%xmm1
+  .byte  102,15,111,137,157,96,0,0           // movdqa        0x609d(%ecx),%xmm1
   .byte  102,15,244,209                      // pmuludq       %xmm1,%xmm2
   .byte  102,15,244,193                      // pmuludq       %xmm1,%xmm0
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
@@ -73074,7 +74512,7 @@
   .byte  15,20,225                           // unpcklps      %xmm1,%xmm4
   .byte  102,15,20,226                       // unpcklpd      %xmm2,%xmm4
   .byte  102,15,41,165,88,255,255,255        // movapd        %xmm4,-0xa8(%ebp)
-  .byte  102,15,254,129,237,80,0,0           // paddd         0x50ed(%ecx),%xmm0
+  .byte  102,15,254,129,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm0
   .byte  102,15,112,200,229                  // pshufd        $0xe5,%xmm0,%xmm1
   .byte  102,15,126,200                      // movd          %xmm1,%eax
   .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
@@ -73128,7 +74566,7 @@
   .byte  243,15,16,52,130                    // movss         (%edx,%eax,4),%xmm6
   .byte  15,20,206                           // unpcklps      %xmm6,%xmm1
   .byte  102,15,20,203                       // unpcklpd      %xmm3,%xmm1
-  .byte  102,15,254,145,237,80,0,0           // paddd         0x50ed(%ecx),%xmm2
+  .byte  102,15,254,145,13,90,0,0            // paddd         0x5a0d(%ecx),%xmm2
   .byte  102,15,112,218,229                  // pshufd        $0xe5,%xmm2,%xmm3
   .byte  102,15,126,216                      // movd          %xmm3,%eax
   .byte  102,15,112,218,78                   // pshufd        $0x4e,%xmm2,%xmm3
@@ -73197,7 +74635,7 @@
   .byte  139,85,12                           // mov           0xc(%ebp),%edx
   .byte  141,66,8                            // lea           0x8(%edx),%eax
   .byte  131,236,8                           // sub           $0x8,%esp
-  .byte  15,40,153,173,80,0,0                // movaps        0x50ad(%ecx),%xmm3
+  .byte  15,40,153,205,89,0,0                // movaps        0x59cd(%ecx),%xmm3
   .byte  80                                  // push          %eax
   .byte  255,117,8                           // pushl         0x8(%ebp)
   .byte  255,82,4                            // call          *0x4(%edx)
@@ -73327,7 +74765,7 @@
   .byte  102,15,110,210                      // movd          %edx,%xmm2
   .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
   .byte  15,91,210                           // cvtdq2ps      %xmm2,%xmm2
-  .byte  15,88,145,100,59,0,0                // addps         0x3b64(%ecx),%xmm2
+  .byte  15,88,145,132,68,0,0                // addps         0x4484(%ecx),%xmm2
   .byte  141,78,8                            // lea           0x8(%esi),%ecx
   .byte  15,41,116,36,64                     // movaps        %xmm6,0x40(%esp)
   .byte  15,41,108,36,48                     // movaps        %xmm5,0x30(%esp)
@@ -73345,6 +74783,98 @@
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
 
+HIDDEN _sk_matrix_translate_sse2_lowp
+.globl _sk_matrix_translate_sse2_lowp
+FUNCTION(_sk_matrix_translate_sse2_lowp)
+_sk_matrix_translate_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,80                          // sub           $0x50,%esp
+  .byte  139,69,8                            // mov           0x8(%ebp),%eax
+  .byte  139,77,16                           // mov           0x10(%ebp),%ecx
+  .byte  139,85,20                           // mov           0x14(%ebp),%edx
+  .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
+  .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
+  .byte  15,40,117,56                        // movaps        0x38(%ebp),%xmm6
+  .byte  139,117,12                          // mov           0xc(%ebp),%esi
+  .byte  139,62                              // mov           (%esi),%edi
+  .byte  243,15,16,63                        // movss         (%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,88,199                           // addps         %xmm7,%xmm0
+  .byte  15,88,207                           // addps         %xmm7,%xmm1
+  .byte  243,15,16,127,4                     // movss         0x4(%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,88,215                           // addps         %xmm7,%xmm2
+  .byte  15,88,223                           // addps         %xmm7,%xmm3
+  .byte  15,40,125,72                        // movaps        0x48(%ebp),%xmm7
+  .byte  141,126,8                           // lea           0x8(%esi),%edi
+  .byte  15,41,124,36,64                     // movaps        %xmm7,0x40(%esp)
+  .byte  15,41,116,36,48                     // movaps        %xmm6,0x30(%esp)
+  .byte  15,41,108,36,32                     // movaps        %xmm5,0x20(%esp)
+  .byte  15,41,100,36,16                     // movaps        %xmm4,0x10(%esp)
+  .byte  137,84,36,12                        // mov           %edx,0xc(%esp)
+  .byte  137,76,36,8                         // mov           %ecx,0x8(%esp)
+  .byte  137,124,36,4                        // mov           %edi,0x4(%esp)
+  .byte  137,4,36                            // mov           %eax,(%esp)
+  .byte  255,86,4                            // call          *0x4(%esi)
+  .byte  131,196,80                          // add           $0x50,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
+HIDDEN _sk_matrix_scale_translate_sse2_lowp
+.globl _sk_matrix_scale_translate_sse2_lowp
+FUNCTION(_sk_matrix_scale_translate_sse2_lowp)
+_sk_matrix_scale_translate_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,80                          // sub           $0x50,%esp
+  .byte  139,69,8                            // mov           0x8(%ebp),%eax
+  .byte  139,85,16                           // mov           0x10(%ebp),%edx
+  .byte  139,117,20                          // mov           0x14(%ebp),%esi
+  .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
+  .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
+  .byte  15,40,117,56                        // movaps        0x38(%ebp),%xmm6
+  .byte  139,77,12                           // mov           0xc(%ebp),%ecx
+  .byte  139,57                              // mov           (%ecx),%edi
+  .byte  243,15,16,63                        // movss         (%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,89,207                           // mulps         %xmm7,%xmm1
+  .byte  15,89,199                           // mulps         %xmm7,%xmm0
+  .byte  243,15,16,127,8                     // movss         0x8(%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,88,199                           // addps         %xmm7,%xmm0
+  .byte  15,88,207                           // addps         %xmm7,%xmm1
+  .byte  243,15,16,127,4                     // movss         0x4(%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,89,223                           // mulps         %xmm7,%xmm3
+  .byte  15,89,215                           // mulps         %xmm7,%xmm2
+  .byte  243,15,16,127,12                    // movss         0xc(%edi),%xmm7
+  .byte  15,198,255,0                        // shufps        $0x0,%xmm7,%xmm7
+  .byte  15,88,215                           // addps         %xmm7,%xmm2
+  .byte  15,88,223                           // addps         %xmm7,%xmm3
+  .byte  15,40,125,72                        // movaps        0x48(%ebp),%xmm7
+  .byte  141,121,8                           // lea           0x8(%ecx),%edi
+  .byte  15,41,124,36,64                     // movaps        %xmm7,0x40(%esp)
+  .byte  15,41,116,36,48                     // movaps        %xmm6,0x30(%esp)
+  .byte  15,41,108,36,32                     // movaps        %xmm5,0x20(%esp)
+  .byte  15,41,100,36,16                     // movaps        %xmm4,0x10(%esp)
+  .byte  137,116,36,12                       // mov           %esi,0xc(%esp)
+  .byte  137,84,36,8                         // mov           %edx,0x8(%esp)
+  .byte  137,124,36,4                        // mov           %edi,0x4(%esp)
+  .byte  137,4,36                            // mov           %eax,(%esp)
+  .byte  255,81,4                            // call          *0x4(%ecx)
+  .byte  131,196,80                          // add           $0x50,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_matrix_2x3_sse2_lowp
 .globl _sk_matrix_2x3_sse2_lowp
 FUNCTION(_sk_matrix_2x3_sse2_lowp)
@@ -73414,6 +74944,99 @@
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
 
+HIDDEN _sk_matrix_perspective_sse2_lowp
+.globl _sk_matrix_perspective_sse2_lowp
+FUNCTION(_sk_matrix_perspective_sse2_lowp)
+_sk_matrix_perspective_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,112                         // sub           $0x70,%esp
+  .byte  15,40,235                           // movaps        %xmm3,%xmm5
+  .byte  15,41,77,232                        // movaps        %xmm1,-0x18(%ebp)
+  .byte  15,40,216                           // movaps        %xmm0,%xmm3
+  .byte  139,69,12                           // mov           0xc(%ebp),%eax
+  .byte  139,8                               // mov           (%eax),%ecx
+  .byte  243,15,16,9                         // movss         (%ecx),%xmm1
+  .byte  243,15,16,113,4                     // movss         0x4(%ecx),%xmm6
+  .byte  15,198,201,0                        // shufps        $0x0,%xmm1,%xmm1
+  .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
+  .byte  243,15,16,97,8                      // movss         0x8(%ecx),%xmm4
+  .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
+  .byte  15,40,253                           // movaps        %xmm5,%xmm7
+  .byte  15,89,254                           // mulps         %xmm6,%xmm7
+  .byte  15,89,242                           // mulps         %xmm2,%xmm6
+  .byte  15,41,85,216                        // movaps        %xmm2,-0x28(%ebp)
+  .byte  15,88,244                           // addps         %xmm4,%xmm6
+  .byte  15,88,252                           // addps         %xmm4,%xmm7
+  .byte  15,40,227                           // movaps        %xmm3,%xmm4
+  .byte  15,89,225                           // mulps         %xmm1,%xmm4
+  .byte  15,88,230                           // addps         %xmm6,%xmm4
+  .byte  15,89,77,232                        // mulps         -0x18(%ebp),%xmm1
+  .byte  15,88,207                           // addps         %xmm7,%xmm1
+  .byte  243,15,16,113,16                    // movss         0x10(%ecx),%xmm6
+  .byte  15,198,246,0                        // shufps        $0x0,%xmm6,%xmm6
+  .byte  243,15,16,65,20                     // movss         0x14(%ecx),%xmm0
+  .byte  15,198,192,0                        // shufps        $0x0,%xmm0,%xmm0
+  .byte  15,40,253                           // movaps        %xmm5,%xmm7
+  .byte  15,89,254                           // mulps         %xmm6,%xmm7
+  .byte  15,89,242                           // mulps         %xmm2,%xmm6
+  .byte  15,88,240                           // addps         %xmm0,%xmm6
+  .byte  15,88,248                           // addps         %xmm0,%xmm7
+  .byte  243,15,16,65,12                     // movss         0xc(%ecx),%xmm0
+  .byte  15,198,192,0                        // shufps        $0x0,%xmm0,%xmm0
+  .byte  15,40,211                           // movaps        %xmm3,%xmm2
+  .byte  15,89,208                           // mulps         %xmm0,%xmm2
+  .byte  15,88,242                           // addps         %xmm2,%xmm6
+  .byte  15,89,69,232                        // mulps         -0x18(%ebp),%xmm0
+  .byte  15,88,248                           // addps         %xmm0,%xmm7
+  .byte  243,15,16,65,28                     // movss         0x1c(%ecx),%xmm0
+  .byte  15,198,192,0                        // shufps        $0x0,%xmm0,%xmm0
+  .byte  15,89,232                           // mulps         %xmm0,%xmm5
+  .byte  15,89,69,216                        // mulps         -0x28(%ebp),%xmm0
+  .byte  243,15,16,81,32                     // movss         0x20(%ecx),%xmm2
+  .byte  15,198,210,0                        // shufps        $0x0,%xmm2,%xmm2
+  .byte  15,88,194                           // addps         %xmm2,%xmm0
+  .byte  15,88,234                           // addps         %xmm2,%xmm5
+  .byte  243,15,16,81,24                     // movss         0x18(%ecx),%xmm2
+  .byte  15,198,210,0                        // shufps        $0x0,%xmm2,%xmm2
+  .byte  15,89,218                           // mulps         %xmm2,%xmm3
+  .byte  15,88,216                           // addps         %xmm0,%xmm3
+  .byte  139,77,8                            // mov           0x8(%ebp),%ecx
+  .byte  139,85,16                           // mov           0x10(%ebp),%edx
+  .byte  139,117,20                          // mov           0x14(%ebp),%esi
+  .byte  15,89,85,232                        // mulps         -0x18(%ebp),%xmm2
+  .byte  15,40,69,24                         // movaps        0x18(%ebp),%xmm0
+  .byte  15,88,213                           // addps         %xmm5,%xmm2
+  .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
+  .byte  15,83,219                           // rcpps         %xmm3,%xmm3
+  .byte  15,89,227                           // mulps         %xmm3,%xmm4
+  .byte  15,89,243                           // mulps         %xmm3,%xmm6
+  .byte  15,40,93,56                         // movaps        0x38(%ebp),%xmm3
+  .byte  15,83,210                           // rcpps         %xmm2,%xmm2
+  .byte  15,89,202                           // mulps         %xmm2,%xmm1
+  .byte  15,89,250                           // mulps         %xmm2,%xmm7
+  .byte  15,40,85,72                         // movaps        0x48(%ebp),%xmm2
+  .byte  141,120,8                           // lea           0x8(%eax),%edi
+  .byte  15,41,84,36,64                      // movaps        %xmm2,0x40(%esp)
+  .byte  15,41,92,36,48                      // movaps        %xmm3,0x30(%esp)
+  .byte  15,41,108,36,32                     // movaps        %xmm5,0x20(%esp)
+  .byte  15,41,68,36,16                      // movaps        %xmm0,0x10(%esp)
+  .byte  137,116,36,12                       // mov           %esi,0xc(%esp)
+  .byte  137,84,36,8                         // mov           %edx,0x8(%esp)
+  .byte  137,124,36,4                        // mov           %edi,0x4(%esp)
+  .byte  137,12,36                           // mov           %ecx,(%esp)
+  .byte  15,40,196                           // movaps        %xmm4,%xmm0
+  .byte  15,40,214                           // movaps        %xmm6,%xmm2
+  .byte  15,40,223                           // movaps        %xmm7,%xmm3
+  .byte  255,80,4                            // call          *0x4(%eax)
+  .byte  131,196,112                         // add           $0x70,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_uniform_color_sse2_lowp
 .globl _sk_uniform_color_sse2_lowp
 FUNCTION(_sk_uniform_color_sse2_lowp)
@@ -73472,7 +75095,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,92                          // sub           $0x5c,%esp
-  .byte  232,0,0,0,0                         // call          96d0 <_sk_black_color_sse2_lowp+0xe>
+  .byte  232,0,0,0,0                         // call          98fb <_sk_black_color_sse2_lowp+0xe>
   .byte  88                                  // pop           %eax
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
@@ -73491,7 +75114,7 @@
   .byte  137,84,36,8                         // mov           %edx,0x8(%esp)
   .byte  137,92,36,4                         // mov           %ebx,0x4(%esp)
   .byte  137,12,36                           // mov           %ecx,(%esp)
-  .byte  15,40,152,96,64,0,0                 // movaps        0x4060(%eax),%xmm3
+  .byte  15,40,152,85,71,0,0                 // movaps        0x4755(%eax),%xmm3
   .byte  15,87,192                           // xorps         %xmm0,%xmm0
   .byte  15,87,201                           // xorps         %xmm1,%xmm1
   .byte  15,87,210                           // xorps         %xmm2,%xmm2
@@ -73513,7 +75136,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,92                          // sub           $0x5c,%esp
-  .byte  232,0,0,0,0                         // call          973b <_sk_white_color_sse2_lowp+0xe>
+  .byte  232,0,0,0,0                         // call          9966 <_sk_white_color_sse2_lowp+0xe>
   .byte  88                                  // pop           %eax
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
@@ -73532,7 +75155,7 @@
   .byte  137,84,36,8                         // mov           %edx,0x8(%esp)
   .byte  137,92,36,4                         // mov           %ebx,0x4(%esp)
   .byte  137,12,36                           // mov           %ecx,(%esp)
-  .byte  15,40,128,245,63,0,0                // movaps        0x3ff5(%eax),%xmm0
+  .byte  15,40,128,234,70,0,0                // movaps        0x46ea(%eax),%xmm0
   .byte  15,40,200                           // movaps        %xmm0,%xmm1
   .byte  15,40,208                           // movaps        %xmm0,%xmm2
   .byte  15,40,216                           // movaps        %xmm0,%xmm3
@@ -73553,16 +75176,16 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          97a5 <_sk_set_rgb_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          99d0 <_sk_set_rgb_sse2_lowp+0xd>
   .byte  94                                  // pop           %esi
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,57                              // mov           (%ecx),%edi
-  .byte  243,15,16,150,79,64,0,0             // movss         0x404f(%esi),%xmm2
+  .byte  243,15,16,150,68,71,0,0             // movss         0x4744(%esi),%xmm2
   .byte  243,15,16,7                         // movss         (%edi),%xmm0
   .byte  243,15,89,194                       // mulss         %xmm2,%xmm0
-  .byte  243,15,16,142,75,64,0,0             // movss         0x404b(%esi),%xmm1
+  .byte  243,15,16,142,64,71,0,0             // movss         0x4740(%esi),%xmm1
   .byte  243,15,88,193                       // addss         %xmm1,%xmm0
   .byte  243,15,44,240                       // cvttss2si     %xmm0,%esi
   .byte  102,15,110,198                      // movd          %esi,%xmm0
@@ -73611,7 +75234,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9869 <_sk_premul_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9a94 <_sk_premul_sse2_lowp+0xd>
   .byte  90                                  // pop           %edx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,77,16                           // mov           0x10(%ebp),%ecx
@@ -73620,7 +75243,7 @@
   .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
   .byte  15,40,117,56                        // movaps        0x38(%ebp),%xmm6
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,15,111,186,199,62,0,0           // movdqa        0x3ec7(%edx),%xmm7
+  .byte  102,15,111,186,188,69,0,0           // movdqa        0x45bc(%edx),%xmm7
   .byte  102,15,253,199                      // paddw         %xmm7,%xmm0
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
   .byte  102,15,253,207                      // paddw         %xmm7,%xmm1
@@ -73754,12 +75377,12 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          99e3 <_sk_invert_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9c0e <_sk_invert_sse2_lowp+0xd>
   .byte  88                                  // pop           %eax
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
   .byte  139,117,20                          // mov           0x14(%ebp),%esi
-  .byte  102,15,111,160,77,61,0,0            // movdqa        0x3d4d(%eax),%xmm4
+  .byte  102,15,111,160,66,68,0,0            // movdqa        0x4442(%eax),%xmm4
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,249,232                      // psubw         %xmm0,%xmm5
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
@@ -73841,11 +75464,11 @@
   .byte  102,15,111,234                      // movdqa        %xmm2,%xmm5
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          9adb <_sk_srcatop_sse2_lowp+0x19>
+  .byte  232,0,0,0,0                         // call          9d06 <_sk_srcatop_sse2_lowp+0x19>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,117,72                   // movdqa        0x48(%ebp),%xmm6
   .byte  102,15,213,206                      // pmullw        %xmm6,%xmm1
-  .byte  102,15,111,128,85,60,0,0            // movdqa        0x3c55(%eax),%xmm0
+  .byte  102,15,111,128,74,67,0,0            // movdqa        0x434a(%eax),%xmm0
   .byte  102,15,253,200                      // paddw         %xmm0,%xmm1
   .byte  102,15,213,214                      // pmullw        %xmm6,%xmm2
   .byte  102,15,253,208                      // paddw         %xmm0,%xmm2
@@ -73903,12 +75526,12 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9baf <_sk_dstatop_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9dda <_sk_dstatop_sse2_lowp+0xd>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,101,40                   // movdqa        0x28(%ebp),%xmm4
   .byte  102,15,111,117,24                   // movdqa        0x18(%ebp),%xmm6
   .byte  102,15,213,243                      // pmullw        %xmm3,%xmm6
-  .byte  102,15,111,184,129,59,0,0           // movdqa        0x3b81(%eax),%xmm7
+  .byte  102,15,111,184,118,66,0,0           // movdqa        0x4276(%eax),%xmm7
   .byte  102,15,253,247                      // paddw         %xmm7,%xmm6
   .byte  102,15,213,227                      // pmullw        %xmm3,%xmm4
   .byte  102,15,253,231                      // paddw         %xmm7,%xmm4
@@ -73961,7 +75584,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9c70 <_sk_srcin_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9e9b <_sk_srcin_sse2_lowp+0xd>
   .byte  94                                  // pop           %esi
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,77,16                           // mov           0x10(%ebp),%ecx
@@ -73970,7 +75593,7 @@
   .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
   .byte  102,15,111,117,72                   // movdqa        0x48(%ebp),%xmm6
   .byte  102,15,213,198                      // pmullw        %xmm6,%xmm0
-  .byte  102,15,111,190,192,58,0,0           // movdqa        0x3ac0(%esi),%xmm7
+  .byte  102,15,111,190,181,65,0,0           // movdqa        0x41b5(%esi),%xmm7
   .byte  102,15,253,199                      // paddw         %xmm7,%xmm0
   .byte  102,15,213,206                      // pmullw        %xmm6,%xmm1
   .byte  102,15,253,207                      // paddw         %xmm7,%xmm1
@@ -74009,14 +75632,14 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9d07 <_sk_dstin_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9f32 <_sk_dstin_sse2_lowp+0xd>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,85,56                    // movdqa        0x38(%ebp),%xmm2
   .byte  102,15,111,109,40                   // movdqa        0x28(%ebp),%xmm5
   .byte  102,15,111,117,24                   // movdqa        0x18(%ebp),%xmm6
   .byte  102,15,111,198                      // movdqa        %xmm6,%xmm0
   .byte  102,15,213,195                      // pmullw        %xmm3,%xmm0
-  .byte  102,15,111,184,41,58,0,0            // movdqa        0x3a29(%eax),%xmm7
+  .byte  102,15,111,184,30,65,0,0            // movdqa        0x411e(%eax),%xmm7
   .byte  102,15,253,199                      // paddw         %xmm7,%xmm0
   .byte  102,15,111,205                      // movdqa        %xmm5,%xmm1
   .byte  102,15,213,203                      // pmullw        %xmm3,%xmm1
@@ -74060,14 +75683,14 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9daf <_sk_srcout_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          9fda <_sk_srcout_sse2_lowp+0xd>
   .byte  94                                  // pop           %esi
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,77,16                           // mov           0x10(%ebp),%ecx
   .byte  139,85,20                           // mov           0x14(%ebp),%edx
   .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
   .byte  102,15,111,109,72                   // movdqa        0x48(%ebp),%xmm5
-  .byte  102,15,111,182,129,57,0,0           // movdqa        0x3981(%esi),%xmm6
+  .byte  102,15,111,182,118,64,0,0           // movdqa        0x4076(%esi),%xmm6
   .byte  102,15,111,254                      // movdqa        %xmm6,%xmm7
   .byte  102,15,249,253                      // psubw         %xmm5,%xmm7
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
@@ -74111,11 +75734,11 @@
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
-  .byte  232,0,0,0,0                         // call          9e52 <_sk_dstout_sse2_lowp+0x11>
+  .byte  232,0,0,0,0                         // call          a07d <_sk_dstout_sse2_lowp+0x11>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,56                   // movdqa        0x38(%ebp),%xmm5
   .byte  102,15,111,117,24                   // movdqa        0x18(%ebp),%xmm6
-  .byte  102,15,111,184,222,56,0,0           // movdqa        0x38de(%eax),%xmm7
+  .byte  102,15,111,184,211,63,0,0           // movdqa        0x3fd3(%eax),%xmm7
   .byte  102,15,111,223                      // movdqa        %xmm7,%xmm3
   .byte  102,15,249,216                      // psubw         %xmm0,%xmm3
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
@@ -74167,9 +75790,9 @@
   .byte  15,41,85,232                        // movaps        %xmm2,-0x18(%ebp)
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,111,224                      // movdqa        %xmm0,%xmm4
-  .byte  232,0,0,0,0                         // call          9f0e <_sk_srcover_sse2_lowp+0x19>
+  .byte  232,0,0,0,0                         // call          a139 <_sk_srcover_sse2_lowp+0x19>
   .byte  88                                  // pop           %eax
-  .byte  102,15,111,136,34,56,0,0            // movdqa        0x3822(%eax),%xmm1
+  .byte  102,15,111,136,23,63,0,0            // movdqa        0x3f17(%eax),%xmm1
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,249,195                      // psubw         %xmm3,%xmm0
   .byte  102,15,111,248                      // movdqa        %xmm0,%xmm7
@@ -74228,7 +75851,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          9fde <_sk_dstover_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          a209 <_sk_dstover_sse2_lowp+0xd>
   .byte  95                                  // pop           %edi
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
@@ -74236,7 +75859,7 @@
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  102,15,111,101,56                   // movdqa        0x38(%ebp),%xmm4
   .byte  102,15,111,109,72                   // movdqa        0x48(%ebp),%xmm5
-  .byte  102,15,111,183,82,55,0,0            // movdqa        0x3752(%edi),%xmm6
+  .byte  102,15,111,183,71,62,0,0            // movdqa        0x3e47(%edi),%xmm6
   .byte  102,15,111,254                      // movdqa        %xmm6,%xmm7
   .byte  102,15,249,253                      // psubw         %xmm5,%xmm7
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
@@ -74282,12 +75905,12 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          a093 <_sk_modulate_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          a2be <_sk_modulate_sse2_lowp+0xd>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,40                   // movdqa        0x28(%ebp),%xmm5
   .byte  102,15,111,117,24                   // movdqa        0x18(%ebp),%xmm6
   .byte  102,15,213,198                      // pmullw        %xmm6,%xmm0
-  .byte  102,15,111,184,157,54,0,0           // movdqa        0x369d(%eax),%xmm7
+  .byte  102,15,111,184,146,61,0,0           // movdqa        0x3d92(%eax),%xmm7
   .byte  102,15,253,199                      // paddw         %xmm7,%xmm0
   .byte  102,15,213,205                      // pmullw        %xmm5,%xmm1
   .byte  102,15,253,207                      // paddw         %xmm7,%xmm1
@@ -74330,10 +75953,10 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          a12f <_sk_multiply_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          a35a <_sk_multiply_sse2_lowp+0xd>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
-  .byte  102,15,111,160,1,54,0,0             // movdqa        0x3601(%eax),%xmm4
+  .byte  102,15,111,160,246,60,0,0           // movdqa        0x3cf6(%eax),%xmm4
   .byte  102,15,111,245                      // movdqa        %xmm5,%xmm6
   .byte  102,15,253,244                      // paddw         %xmm4,%xmm6
   .byte  102,15,249,117,72                   // psubw         0x48(%ebp),%xmm6
@@ -74403,17 +76026,17 @@
   .byte  131,236,96                          // sub           $0x60,%esp
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a230 <_sk_plus__sse2_lowp+0x15>
+  .byte  232,0,0,0,0                         // call          a45b <_sk_plus__sse2_lowp+0x15>
   .byte  88                                  // pop           %eax
   .byte  102,15,253,77,24                    // paddw         0x18(%ebp),%xmm1
-  .byte  102,15,111,168,16,53,0,0            // movdqa        0x3510(%eax),%xmm5
+  .byte  102,15,111,168,5,60,0,0             // movdqa        0x3c05(%eax),%xmm5
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,217,197                      // psubusw       %xmm5,%xmm0
   .byte  102,15,127,109,232                  // movdqa        %xmm5,-0x18(%ebp)
   .byte  102,15,239,255                      // pxor          %xmm7,%xmm7
   .byte  102,15,117,199                      // pcmpeqw       %xmm7,%xmm0
   .byte  102,15,219,200                      // pand          %xmm0,%xmm1
-  .byte  102,15,111,176,0,53,0,0             // movdqa        0x3500(%eax),%xmm6
+  .byte  102,15,111,176,245,59,0,0           // movdqa        0x3bf5(%eax),%xmm6
   .byte  102,15,223,198                      // pandn         %xmm6,%xmm0
   .byte  102,15,235,193                      // por           %xmm1,%xmm0
   .byte  102,15,253,101,40                   // paddw         0x28(%ebp),%xmm4
@@ -74474,14 +76097,14 @@
   .byte  131,236,80                          // sub           $0x50,%esp
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a328 <_sk_screen_sse2_lowp+0x15>
+  .byte  232,0,0,0,0                         // call          a553 <_sk_screen_sse2_lowp+0x15>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,40                   // movdqa        0x28(%ebp),%xmm5
   .byte  102,15,111,117,24                   // movdqa        0x18(%ebp),%xmm6
   .byte  102,15,111,198                      // movdqa        %xmm6,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,206                      // pmullw        %xmm6,%xmm1
-  .byte  102,15,111,176,8,52,0,0             // movdqa        0x3408(%eax),%xmm6
+  .byte  102,15,111,176,253,58,0,0           // movdqa        0x3afd(%eax),%xmm6
   .byte  102,15,253,206                      // paddw         %xmm6,%xmm1
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
@@ -74539,9 +76162,9 @@
   .byte  86                                  // push          %esi
   .byte  131,236,96                          // sub           $0x60,%esp
   .byte  102,15,127,93,232                   // movdqa        %xmm3,-0x18(%ebp)
-  .byte  232,0,0,0,0                         // call          a408 <_sk_xor__sse2_lowp+0x12>
+  .byte  232,0,0,0,0                         // call          a633 <_sk_xor__sse2_lowp+0x12>
   .byte  88                                  // pop           %eax
-  .byte  102,15,111,176,40,51,0,0            // movdqa        0x3328(%eax),%xmm6
+  .byte  102,15,111,176,29,58,0,0            // movdqa        0x3a1d(%eax),%xmm6
   .byte  102,15,111,254                      // movdqa        %xmm6,%xmm7
   .byte  102,15,249,125,72                   // psubw         0x48(%ebp),%xmm7
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
@@ -74608,7 +76231,7 @@
   .byte  15,41,85,200                        // movaps        %xmm2,-0x38(%ebp)
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a4fd <_sk_darken_sse2_lowp+0x25>
+  .byte  232,0,0,0,0                         // call          a728 <_sk_darken_sse2_lowp+0x25>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,117,72                   // movdqa        0x48(%ebp),%xmm6
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
@@ -74619,7 +76242,7 @@
   .byte  102,15,111,222                      // movdqa        %xmm6,%xmm3
   .byte  102,15,111,197                      // movdqa        %xmm5,%xmm0
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,15,111,184,83,50,0,0            // movdqa        0x3253(%eax),%xmm7
+  .byte  102,15,111,184,72,57,0,0            // movdqa        0x3948(%eax),%xmm7
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
   .byte  102,15,239,247                      // pxor          %xmm7,%xmm6
   .byte  102,15,111,232                      // movdqa        %xmm0,%xmm5
@@ -74628,7 +76251,7 @@
   .byte  102,15,219,197                      // pand          %xmm5,%xmm0
   .byte  102,15,223,233                      // pandn         %xmm1,%xmm5
   .byte  102,15,235,232                      // por           %xmm0,%xmm5
-  .byte  102,15,111,128,51,50,0,0            // movdqa        0x3233(%eax),%xmm0
+  .byte  102,15,111,128,40,57,0,0            // movdqa        0x3928(%eax),%xmm0
   .byte  102,15,127,69,232                   // movdqa        %xmm0,-0x18(%ebp)
   .byte  102,15,253,232                      // paddw         %xmm0,%xmm5
   .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
@@ -74719,7 +76342,7 @@
   .byte  15,41,85,200                        // movaps        %xmm2,-0x38(%ebp)
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a6a6 <_sk_lighten_sse2_lowp+0x25>
+  .byte  232,0,0,0,0                         // call          a8d1 <_sk_lighten_sse2_lowp+0x25>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
   .byte  102,15,111,197                      // movdqa        %xmm5,%xmm0
@@ -74728,7 +76351,7 @@
   .byte  102,15,213,77,72                    // pmullw        0x48(%ebp),%xmm1
   .byte  102,15,111,197                      // movdqa        %xmm5,%xmm0
   .byte  102,15,213,198                      // pmullw        %xmm6,%xmm0
-  .byte  102,15,111,184,170,48,0,0           // movdqa        0x30aa(%eax),%xmm7
+  .byte  102,15,111,184,159,55,0,0           // movdqa        0x379f(%eax),%xmm7
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
   .byte  102,15,239,247                      // pxor          %xmm7,%xmm6
   .byte  102,15,111,232                      // movdqa        %xmm0,%xmm5
@@ -74737,7 +76360,7 @@
   .byte  102,15,219,205                      // pand          %xmm5,%xmm1
   .byte  102,15,223,232                      // pandn         %xmm0,%xmm5
   .byte  102,15,235,233                      // por           %xmm1,%xmm5
-  .byte  102,15,111,128,138,48,0,0           // movdqa        0x308a(%eax),%xmm0
+  .byte  102,15,111,128,127,55,0,0           // movdqa        0x377f(%eax),%xmm0
   .byte  102,15,253,232                      // paddw         %xmm0,%xmm5
   .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
   .byte  102,15,127,85,232                   // movdqa        %xmm2,-0x18(%ebp)
@@ -74827,7 +76450,7 @@
   .byte  15,41,85,184                        // movaps        %xmm2,-0x48(%ebp)
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a844 <_sk_difference_sse2_lowp+0x25>
+  .byte  232,0,0,0,0                         // call          aa6f <_sk_difference_sse2_lowp+0x25>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,117,72                   // movdqa        0x48(%ebp),%xmm6
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
@@ -74838,7 +76461,7 @@
   .byte  102,15,111,222                      // movdqa        %xmm6,%xmm3
   .byte  102,15,111,197                      // movdqa        %xmm5,%xmm0
   .byte  102,15,213,199                      // pmullw        %xmm7,%xmm0
-  .byte  102,15,111,184,12,47,0,0            // movdqa        0x2f0c(%eax),%xmm7
+  .byte  102,15,111,184,1,54,0,0             // movdqa        0x3601(%eax),%xmm7
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
   .byte  102,15,239,247                      // pxor          %xmm7,%xmm6
   .byte  102,15,111,232                      // movdqa        %xmm0,%xmm5
@@ -74847,11 +76470,11 @@
   .byte  102,15,219,205                      // pand          %xmm5,%xmm1
   .byte  102,15,223,232                      // pandn         %xmm0,%xmm5
   .byte  102,15,235,233                      // por           %xmm1,%xmm5
-  .byte  102,15,111,128,236,46,0,0           // movdqa        0x2eec(%eax),%xmm0
+  .byte  102,15,111,128,225,53,0,0           // movdqa        0x35e1(%eax),%xmm0
   .byte  102,15,127,69,216                   // movdqa        %xmm0,-0x28(%ebp)
   .byte  102,15,253,232                      // paddw         %xmm0,%xmm5
   .byte  102,15,113,213,7                    // psrlw         $0x7,%xmm5
-  .byte  102,15,111,128,28,47,0,0            // movdqa        0x2f1c(%eax),%xmm0
+  .byte  102,15,111,128,17,54,0,0            // movdqa        0x3611(%eax),%xmm0
   .byte  102,15,127,69,232                   // movdqa        %xmm0,-0x18(%ebp)
   .byte  102,15,219,232                      // pand          %xmm0,%xmm5
   .byte  102,15,249,213                      // psubw         %xmm5,%xmm2
@@ -74940,16 +76563,16 @@
   .byte  131,236,80                          // sub           $0x50,%esp
   .byte  102,15,111,225                      // movdqa        %xmm1,%xmm4
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          a9f8 <_sk_exclusion_sse2_lowp+0x15>
+  .byte  232,0,0,0,0                         // call          ac23 <_sk_exclusion_sse2_lowp+0x15>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
   .byte  102,15,111,197                      // movdqa        %xmm5,%xmm0
   .byte  102,15,253,193                      // paddw         %xmm1,%xmm0
   .byte  102,15,213,205                      // pmullw        %xmm5,%xmm1
-  .byte  102,15,111,184,56,45,0,0            // movdqa        0x2d38(%eax),%xmm7
+  .byte  102,15,111,184,45,52,0,0            // movdqa        0x342d(%eax),%xmm7
   .byte  102,15,253,207                      // paddw         %xmm7,%xmm1
   .byte  102,15,113,209,7                    // psrlw         $0x7,%xmm1
-  .byte  102,15,111,176,104,45,0,0           // movdqa        0x2d68(%eax),%xmm6
+  .byte  102,15,111,176,93,52,0,0            // movdqa        0x345d(%eax),%xmm6
   .byte  102,15,219,206                      // pand          %xmm6,%xmm1
   .byte  102,15,249,193                      // psubw         %xmm1,%xmm0
   .byte  102,15,111,109,40                   // movdqa        0x28(%ebp),%xmm5
@@ -75010,9 +76633,9 @@
   .byte  15,41,85,152                        // movaps        %xmm2,-0x68(%ebp)
   .byte  102,15,111,233                      // movdqa        %xmm1,%xmm5
   .byte  102,15,111,224                      // movdqa        %xmm0,%xmm4
-  .byte  232,0,0,0,0                         // call          aaf2 <_sk_hardlight_sse2_lowp+0x1c>
+  .byte  232,0,0,0,0                         // call          ad1d <_sk_hardlight_sse2_lowp+0x1c>
   .byte  88                                  // pop           %eax
-  .byte  102,15,111,184,62,44,0,0            // movdqa        0x2c3e(%eax),%xmm7
+  .byte  102,15,111,184,51,51,0,0            // movdqa        0x3333(%eax),%xmm7
   .byte  102,15,127,125,200                  // movdqa        %xmm7,-0x38(%ebp)
   .byte  102,15,111,215                      // movdqa        %xmm7,%xmm2
   .byte  102,15,249,85,72                    // psubw         0x48(%ebp),%xmm2
@@ -75039,7 +76662,7 @@
   .byte  102,15,111,214                      // movdqa        %xmm6,%xmm2
   .byte  102,15,249,209                      // psubw         %xmm1,%xmm2
   .byte  102,15,253,228                      // paddw         %xmm4,%xmm4
-  .byte  102,15,111,128,94,44,0,0            // movdqa        0x2c5e(%eax),%xmm0
+  .byte  102,15,111,128,83,51,0,0            // movdqa        0x3353(%eax),%xmm0
   .byte  102,15,239,216                      // pxor          %xmm0,%xmm3
   .byte  102,15,127,93,184                   // movdqa        %xmm3,-0x48(%ebp)
   .byte  102,15,111,204                      // movdqa        %xmm4,%xmm1
@@ -75146,10 +76769,10 @@
   .byte  15,41,85,152                        // movaps        %xmm2,-0x68(%ebp)
   .byte  15,41,141,120,255,255,255           // movaps        %xmm1,-0x88(%ebp)
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          ad14 <_sk_overlay_sse2_lowp+0x1f>
+  .byte  232,0,0,0,0                         // call          af3f <_sk_overlay_sse2_lowp+0x1f>
   .byte  88                                  // pop           %eax
   .byte  102,15,111,69,72                    // movdqa        0x48(%ebp),%xmm0
-  .byte  102,15,111,184,28,42,0,0            // movdqa        0x2a1c(%eax),%xmm7
+  .byte  102,15,111,184,17,49,0,0            // movdqa        0x3111(%eax),%xmm7
   .byte  102,15,127,125,232                  // movdqa        %xmm7,-0x18(%ebp)
   .byte  102,15,111,215                      // movdqa        %xmm7,%xmm2
   .byte  102,15,249,208                      // psubw         %xmm0,%xmm2
@@ -75176,7 +76799,7 @@
   .byte  102,15,249,234                      // psubw         %xmm2,%xmm5
   .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
   .byte  102,15,253,192                      // paddw         %xmm0,%xmm0
-  .byte  102,15,111,144,60,42,0,0            // movdqa        0x2a3c(%eax),%xmm2
+  .byte  102,15,111,144,49,49,0,0            // movdqa        0x3131(%eax),%xmm2
   .byte  102,15,127,85,184                   // movdqa        %xmm2,-0x48(%ebp)
   .byte  102,15,239,226                      // pxor          %xmm2,%xmm4
   .byte  102,15,239,194                      // pxor          %xmm2,%xmm0
@@ -75306,26 +76929,26 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          af83 <_sk_load_8888_sse2_lowp+0x30>
+  .byte  232,0,0,0,0                         // call          b1ae <_sk_load_8888_sse2_lowp+0x30>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,32                              // ja            afa9 <_sk_load_8888_sse2_lowp+0x56>
+  .byte  119,32                              // ja            b1d4 <_sk_load_8888_sse2_lowp+0x56>
   .byte  102,15,239,246                      // pxor          %xmm6,%xmm6
   .byte  15,182,201                          // movzbl        %cl,%ecx
-  .byte  139,140,143,53,1,0,0                // mov           0x135(%edi,%ecx,4),%ecx
+  .byte  139,140,143,50,1,0,0                // mov           0x132(%edi,%ecx,4),%ecx
   .byte  1,249                               // add           %edi,%ecx
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  255,225                             // jmp           *%ecx
   .byte  243,15,16,28,179                    // movss         (%ebx,%esi,4),%xmm3
   .byte  15,87,246                           // xorps         %xmm6,%xmm6
-  .byte  235,73                              // jmp           aff2 <_sk_load_8888_sse2_lowp+0x9f>
+  .byte  235,73                              // jmp           b21d <_sk_load_8888_sse2_lowp+0x9f>
   .byte  102,15,16,28,179                    // movupd        (%ebx,%esi,4),%xmm3
   .byte  15,16,116,179,16                    // movups        0x10(%ebx,%esi,4),%xmm6
-  .byte  235,61                              // jmp           aff2 <_sk_load_8888_sse2_lowp+0x9f>
+  .byte  235,61                              // jmp           b21d <_sk_load_8888_sse2_lowp+0x9f>
   .byte  102,15,110,68,179,8                 // movd          0x8(%ebx,%esi,4),%xmm0
   .byte  102,15,112,216,69                   // pshufd        $0x45,%xmm0,%xmm3
   .byte  102,15,18,28,179                    // movlpd        (%ebx,%esi,4),%xmm3
-  .byte  235,43                              // jmp           aff2 <_sk_load_8888_sse2_lowp+0x9f>
+  .byte  235,43                              // jmp           b21d <_sk_load_8888_sse2_lowp+0x9f>
   .byte  102,15,110,68,179,24                // movd          0x18(%ebx,%esi,4),%xmm0
   .byte  102,15,112,240,69                   // pshufd        $0x45,%xmm0,%xmm6
   .byte  243,15,16,68,179,20                 // movss         0x14(%ebx,%esi,4),%xmm0
@@ -75342,7 +76965,7 @@
   .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
   .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
   .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
-  .byte  102,15,111,175,173,39,0,0           // movdqa        0x27ad(%edi),%xmm5
+  .byte  102,15,111,175,162,46,0,0           // movdqa        0x2ea2(%edi),%xmm5
   .byte  102,15,111,193                      // movdqa        %xmm1,%xmm0
   .byte  102,15,219,197                      // pand          %xmm5,%xmm0
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
@@ -75383,7 +77006,6 @@
   .byte  91                                  // pop           %ebx
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
-  .byte  15,31,0                             // nopl          (%eax)
   .byte  28,0                                // sbb           $0x0,%al
   .byte  0,0                                 // add           %al,(%eax)
   .byte  61,0,0,0,50                         // cmp           $0x32000000,%eax
@@ -75422,10 +77044,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          b111 <_sk_load_8888_dst_sse2_lowp+0x3d>
+  .byte  232,0,0,0,0                         // call          b339 <_sk_load_8888_dst_sse2_lowp+0x3d>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,32                              // ja            b137 <_sk_load_8888_dst_sse2_lowp+0x63>
+  .byte  119,32                              // ja            b35f <_sk_load_8888_dst_sse2_lowp+0x63>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,59,1,0,0                // mov           0x13b(%edi,%ecx,4),%ecx
@@ -75434,14 +77056,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  243,15,16,60,179                    // movss         (%ebx,%esi,4),%xmm7
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
-  .byte  235,73                              // jmp           b180 <_sk_load_8888_dst_sse2_lowp+0xac>
+  .byte  235,73                              // jmp           b3a8 <_sk_load_8888_dst_sse2_lowp+0xac>
   .byte  102,15,16,60,179                    // movupd        (%ebx,%esi,4),%xmm7
   .byte  15,16,92,179,16                     // movups        0x10(%ebx,%esi,4),%xmm3
-  .byte  235,61                              // jmp           b180 <_sk_load_8888_dst_sse2_lowp+0xac>
+  .byte  235,61                              // jmp           b3a8 <_sk_load_8888_dst_sse2_lowp+0xac>
   .byte  102,15,110,84,179,8                 // movd          0x8(%ebx,%esi,4),%xmm2
   .byte  102,15,112,250,69                   // pshufd        $0x45,%xmm2,%xmm7
   .byte  102,15,18,60,179                    // movlpd        (%ebx,%esi,4),%xmm7
-  .byte  235,43                              // jmp           b180 <_sk_load_8888_dst_sse2_lowp+0xac>
+  .byte  235,43                              // jmp           b3a8 <_sk_load_8888_dst_sse2_lowp+0xac>
   .byte  102,15,110,84,179,24                // movd          0x18(%ebx,%esi,4),%xmm2
   .byte  102,15,112,218,69                   // pshufd        $0x45,%xmm2,%xmm3
   .byte  243,15,16,84,179,20                 // movss         0x14(%ebx,%esi,4),%xmm2
@@ -75458,7 +77080,7 @@
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,114,226,16                   // psrad         $0x10,%xmm2
   .byte  102,15,107,242                      // packssdw      %xmm2,%xmm6
-  .byte  102,15,111,151,31,38,0,0            // movdqa        0x261f(%edi),%xmm2
+  .byte  102,15,111,151,23,45,0,0            // movdqa        0x2d17(%edi),%xmm2
   .byte  102,15,111,206                      // movdqa        %xmm6,%xmm1
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  102,15,113,214,8                    // psrlw         $0x8,%xmm6
@@ -75552,19 +77174,19 @@
   .byte  128,227,7                           // and           $0x7,%bl
   .byte  254,203                             // dec           %bl
   .byte  128,251,6                           // cmp           $0x6,%bl
-  .byte  232,0,0,0,0                         // call          b2dc <_sk_store_8888_sse2_lowp+0x74>
+  .byte  232,0,0,0,0                         // call          b504 <_sk_store_8888_sse2_lowp+0x74>
   .byte  89                                  // pop           %ecx
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
   .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
-  .byte  119,19                              // ja            b2f9 <_sk_store_8888_sse2_lowp+0x91>
+  .byte  119,19                              // ja            b521 <_sk_store_8888_sse2_lowp+0x91>
   .byte  15,182,219                          // movzbl        %bl,%ebx
   .byte  3,140,153,184,0,0,0                 // add           0xb8(%ecx,%ebx,4),%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  102,15,126,60,183                   // movd          %xmm7,(%edi,%esi,4)
-  .byte  235,93                              // jmp           b356 <_sk_store_8888_sse2_lowp+0xee>
+  .byte  235,93                              // jmp           b57e <_sk_store_8888_sse2_lowp+0xee>
   .byte  243,15,127,60,183                   // movdqu        %xmm7,(%edi,%esi,4)
   .byte  243,15,127,116,183,16               // movdqu        %xmm6,0x10(%edi,%esi,4)
-  .byte  235,80                              // jmp           b356 <_sk_store_8888_sse2_lowp+0xee>
+  .byte  235,80                              // jmp           b57e <_sk_store_8888_sse2_lowp+0xee>
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  102,15,112,223,78                   // pshufd        $0x4e,%xmm7,%xmm3
@@ -75572,7 +77194,7 @@
   .byte  15,40,220                           // movaps        %xmm4,%xmm3
   .byte  102,15,111,230                      // movdqa        %xmm6,%xmm4
   .byte  102,15,214,60,183                   // movq          %xmm7,(%edi,%esi,4)
-  .byte  235,48                              // jmp           b356 <_sk_store_8888_sse2_lowp+0xee>
+  .byte  235,48                              // jmp           b57e <_sk_store_8888_sse2_lowp+0xee>
   .byte  102,15,112,222,78                   // pshufd        $0x4e,%xmm6,%xmm3
   .byte  102,15,126,92,183,24                // movd          %xmm3,0x18(%edi,%esi,4)
   .byte  15,40,93,72                         // movaps        0x48(%ebp),%xmm3
@@ -75609,7 +77231,7 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  42,0                                // sub           (%eax),%al
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  117,0                               // jne           b3a2 <_sk_store_8888_sse2_lowp+0x13a>
+  .byte  117,0                               // jne           b5ca <_sk_store_8888_sse2_lowp+0x13a>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  111                                 // outsl         %ds:(%esi),(%dx)
   .byte  0,0                                 // add           %al,(%eax)
@@ -75641,10 +77263,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          b3e0 <_sk_load_bgra_sse2_lowp+0x30>
+  .byte  232,0,0,0,0                         // call          b608 <_sk_load_bgra_sse2_lowp+0x30>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,32                              // ja            b406 <_sk_load_bgra_sse2_lowp+0x56>
+  .byte  119,32                              // ja            b62e <_sk_load_bgra_sse2_lowp+0x56>
   .byte  102,15,239,246                      // pxor          %xmm6,%xmm6
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,52,1,0,0                // mov           0x134(%edi,%ecx,4),%ecx
@@ -75653,14 +77275,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  243,15,16,28,179                    // movss         (%ebx,%esi,4),%xmm3
   .byte  15,87,246                           // xorps         %xmm6,%xmm6
-  .byte  235,73                              // jmp           b44f <_sk_load_bgra_sse2_lowp+0x9f>
+  .byte  235,73                              // jmp           b677 <_sk_load_bgra_sse2_lowp+0x9f>
   .byte  102,15,16,28,179                    // movupd        (%ebx,%esi,4),%xmm3
   .byte  15,16,116,179,16                    // movups        0x10(%ebx,%esi,4),%xmm6
-  .byte  235,61                              // jmp           b44f <_sk_load_bgra_sse2_lowp+0x9f>
+  .byte  235,61                              // jmp           b677 <_sk_load_bgra_sse2_lowp+0x9f>
   .byte  102,15,110,68,179,8                 // movd          0x8(%ebx,%esi,4),%xmm0
   .byte  102,15,112,216,69                   // pshufd        $0x45,%xmm0,%xmm3
   .byte  102,15,18,28,179                    // movlpd        (%ebx,%esi,4),%xmm3
-  .byte  235,43                              // jmp           b44f <_sk_load_bgra_sse2_lowp+0x9f>
+  .byte  235,43                              // jmp           b677 <_sk_load_bgra_sse2_lowp+0x9f>
   .byte  102,15,110,68,179,24                // movd          0x18(%ebx,%esi,4),%xmm0
   .byte  102,15,112,240,69                   // pshufd        $0x45,%xmm0,%xmm6
   .byte  243,15,16,68,179,20                 // movss         0x14(%ebx,%esi,4),%xmm0
@@ -75677,7 +77299,7 @@
   .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
   .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
   .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
-  .byte  102,15,111,175,80,35,0,0            // movdqa        0x2350(%edi),%xmm5
+  .byte  102,15,111,175,72,42,0,0            // movdqa        0x2a48(%edi),%xmm5
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,219,213                      // pand          %xmm5,%xmm2
   .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
@@ -75757,10 +77379,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          b56d <_sk_load_bgra_dst_sse2_lowp+0x3d>
+  .byte  232,0,0,0,0                         // call          b795 <_sk_load_bgra_dst_sse2_lowp+0x3d>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,32                              // ja            b593 <_sk_load_bgra_dst_sse2_lowp+0x63>
+  .byte  119,32                              // ja            b7bb <_sk_load_bgra_dst_sse2_lowp+0x63>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,59,1,0,0                // mov           0x13b(%edi,%ecx,4),%ecx
@@ -75769,14 +77391,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  243,15,16,60,179                    // movss         (%ebx,%esi,4),%xmm7
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
-  .byte  235,73                              // jmp           b5dc <_sk_load_bgra_dst_sse2_lowp+0xac>
+  .byte  235,73                              // jmp           b804 <_sk_load_bgra_dst_sse2_lowp+0xac>
   .byte  102,15,16,60,179                    // movupd        (%ebx,%esi,4),%xmm7
   .byte  15,16,92,179,16                     // movups        0x10(%ebx,%esi,4),%xmm3
-  .byte  235,61                              // jmp           b5dc <_sk_load_bgra_dst_sse2_lowp+0xac>
+  .byte  235,61                              // jmp           b804 <_sk_load_bgra_dst_sse2_lowp+0xac>
   .byte  102,15,110,84,179,8                 // movd          0x8(%ebx,%esi,4),%xmm2
   .byte  102,15,112,250,69                   // pshufd        $0x45,%xmm2,%xmm7
   .byte  102,15,18,60,179                    // movlpd        (%ebx,%esi,4),%xmm7
-  .byte  235,43                              // jmp           b5dc <_sk_load_bgra_dst_sse2_lowp+0xac>
+  .byte  235,43                              // jmp           b804 <_sk_load_bgra_dst_sse2_lowp+0xac>
   .byte  102,15,110,84,179,24                // movd          0x18(%ebx,%esi,4),%xmm2
   .byte  102,15,112,218,69                   // pshufd        $0x45,%xmm2,%xmm3
   .byte  243,15,16,84,179,20                 // movss         0x14(%ebx,%esi,4),%xmm2
@@ -75793,7 +77415,7 @@
   .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
   .byte  102,15,114,226,16                   // psrad         $0x10,%xmm2
   .byte  102,15,107,242                      // packssdw      %xmm2,%xmm6
-  .byte  102,15,111,151,195,33,0,0           // movdqa        0x21c3(%edi),%xmm2
+  .byte  102,15,111,151,187,40,0,0           // movdqa        0x28bb(%edi),%xmm2
   .byte  102,15,111,206                      // movdqa        %xmm6,%xmm1
   .byte  102,15,219,202                      // pand          %xmm2,%xmm1
   .byte  102,15,113,214,8                    // psrlw         $0x8,%xmm6
@@ -75887,19 +77509,19 @@
   .byte  128,227,7                           // and           $0x7,%bl
   .byte  254,203                             // dec           %bl
   .byte  128,251,6                           // cmp           $0x6,%bl
-  .byte  232,0,0,0,0                         // call          b738 <_sk_store_bgra_sse2_lowp+0x74>
+  .byte  232,0,0,0,0                         // call          b960 <_sk_store_bgra_sse2_lowp+0x74>
   .byte  89                                  // pop           %ecx
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
   .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
-  .byte  119,19                              // ja            b755 <_sk_store_bgra_sse2_lowp+0x91>
+  .byte  119,19                              // ja            b97d <_sk_store_bgra_sse2_lowp+0x91>
   .byte  15,182,219                          // movzbl        %bl,%ebx
   .byte  3,140,153,184,0,0,0                 // add           0xb8(%ecx,%ebx,4),%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  102,15,126,60,183                   // movd          %xmm7,(%edi,%esi,4)
-  .byte  235,93                              // jmp           b7b2 <_sk_store_bgra_sse2_lowp+0xee>
+  .byte  235,93                              // jmp           b9da <_sk_store_bgra_sse2_lowp+0xee>
   .byte  243,15,127,60,183                   // movdqu        %xmm7,(%edi,%esi,4)
   .byte  243,15,127,116,183,16               // movdqu        %xmm6,0x10(%edi,%esi,4)
-  .byte  235,80                              // jmp           b7b2 <_sk_store_bgra_sse2_lowp+0xee>
+  .byte  235,80                              // jmp           b9da <_sk_store_bgra_sse2_lowp+0xee>
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  15,40,227                           // movaps        %xmm3,%xmm4
   .byte  102,15,112,223,78                   // pshufd        $0x4e,%xmm7,%xmm3
@@ -75907,7 +77529,7 @@
   .byte  15,40,220                           // movaps        %xmm4,%xmm3
   .byte  102,15,111,230                      // movdqa        %xmm6,%xmm4
   .byte  102,15,214,60,183                   // movq          %xmm7,(%edi,%esi,4)
-  .byte  235,48                              // jmp           b7b2 <_sk_store_bgra_sse2_lowp+0xee>
+  .byte  235,48                              // jmp           b9da <_sk_store_bgra_sse2_lowp+0xee>
   .byte  102,15,112,222,78                   // pshufd        $0x4e,%xmm6,%xmm3
   .byte  102,15,126,92,183,24                // movd          %xmm3,0x18(%edi,%esi,4)
   .byte  15,40,93,72                         // movaps        0x48(%ebp),%xmm3
@@ -75944,7 +77566,7 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  42,0                                // sub           (%eax),%al
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  117,0                               // jne           b7fe <_sk_store_bgra_sse2_lowp+0x13a>
+  .byte  117,0                               // jne           ba26 <_sk_store_bgra_sse2_lowp+0x13a>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  111                                 // outsl         %ds:(%esi),(%dx)
   .byte  0,0                                 // add           %al,(%eax)
@@ -75963,7 +77585,7 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,96                          // sub           $0x60,%esp
-  .byte  232,0,0,0,0                         // call          b819 <_sk_gather_8888_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          ba41 <_sk_gather_8888_sse2_lowp+0xd>
   .byte  90                                  // pop           %edx
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
@@ -76055,7 +77677,7 @@
   .byte  102,15,114,243,16                   // pslld         $0x10,%xmm3
   .byte  102,15,114,227,16                   // psrad         $0x10,%xmm3
   .byte  102,15,107,216                      // packssdw      %xmm0,%xmm3
-  .byte  102,15,111,186,23,31,0,0            // movdqa        0x1f17(%edx),%xmm7
+  .byte  102,15,111,186,15,38,0,0            // movdqa        0x260f(%edx),%xmm7
   .byte  102,15,114,210,16                   // psrld         $0x10,%xmm2
   .byte  102,15,114,212,16                   // psrld         $0x10,%xmm4
   .byte  102,15,114,244,16                   // pslld         $0x10,%xmm4
@@ -76085,6 +77707,136 @@
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
 
+HIDDEN _sk_gather_bgra_sse2_lowp
+.globl _sk_gather_bgra_sse2_lowp
+FUNCTION(_sk_gather_bgra_sse2_lowp)
+_sk_gather_bgra_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,80                          // sub           $0x50,%esp
+  .byte  232,0,0,0,0                         // call          bc3c <_sk_gather_bgra_sse2_lowp+0xd>
+  .byte  90                                  // pop           %edx
+  .byte  139,69,8                            // mov           0x8(%ebp),%eax
+  .byte  139,77,12                           // mov           0xc(%ebp),%ecx
+  .byte  139,57                              // mov           (%ecx),%edi
+  .byte  243,15,16,103,8                     // movss         0x8(%edi),%xmm4
+  .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
+  .byte  102,15,118,237                      // pcmpeqd       %xmm5,%xmm5
+  .byte  102,15,254,229                      // paddd         %xmm5,%xmm4
+  .byte  15,87,246                           // xorps         %xmm6,%xmm6
+  .byte  15,95,198                           // maxps         %xmm6,%xmm0
+  .byte  15,95,206                           // maxps         %xmm6,%xmm1
+  .byte  15,93,204                           // minps         %xmm4,%xmm1
+  .byte  15,93,196                           // minps         %xmm4,%xmm0
+  .byte  243,15,16,103,12                    // movss         0xc(%edi),%xmm4
+  .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
+  .byte  102,15,254,229                      // paddd         %xmm5,%xmm4
+  .byte  15,95,222                           // maxps         %xmm6,%xmm3
+  .byte  15,95,214                           // maxps         %xmm6,%xmm2
+  .byte  15,93,212                           // minps         %xmm4,%xmm2
+  .byte  15,93,220                           // minps         %xmm4,%xmm3
+  .byte  139,55                              // mov           (%edi),%esi
+  .byte  243,15,91,219                       // cvttps2dq     %xmm3,%xmm3
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,103,4                    // movd          0x4(%edi),%xmm4
+  .byte  102,15,112,228,0                    // pshufd        $0x0,%xmm4,%xmm4
+  .byte  102,15,112,234,245                  // pshufd        $0xf5,%xmm2,%xmm5
+  .byte  102,15,244,236                      // pmuludq       %xmm4,%xmm5
+  .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
+  .byte  102,15,112,251,245                  // pshufd        $0xf5,%xmm3,%xmm7
+  .byte  102,15,244,252                      // pmuludq       %xmm4,%xmm7
+  .byte  102,15,244,226                      // pmuludq       %xmm2,%xmm4
+  .byte  102,15,112,212,232                  // pshufd        $0xe8,%xmm4,%xmm2
+  .byte  102,15,112,229,232                  // pshufd        $0xe8,%xmm5,%xmm4
+  .byte  102,15,98,212                       // punpckldq     %xmm4,%xmm2
+  .byte  102,15,244,243                      // pmuludq       %xmm3,%xmm6
+  .byte  102,15,112,222,232                  // pshufd        $0xe8,%xmm6,%xmm3
+  .byte  102,15,112,231,232                  // pshufd        $0xe8,%xmm7,%xmm4
+  .byte  102,15,98,220                       // punpckldq     %xmm4,%xmm3
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
+  .byte  102,15,112,208,78                   // pshufd        $0x4e,%xmm0,%xmm2
+  .byte  102,15,112,216,231                  // pshufd        $0xe7,%xmm0,%xmm3
+  .byte  102,15,126,223                      // movd          %xmm3,%edi
+  .byte  102,15,110,36,190                   // movd          (%esi,%edi,4),%xmm4
+  .byte  102,15,126,215                      // movd          %xmm2,%edi
+  .byte  102,15,110,52,190                   // movd          (%esi,%edi,4),%xmm6
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  102,15,112,192,229                  // pshufd        $0xe5,%xmm0,%xmm0
+  .byte  102,15,110,28,190                   // movd          (%esi,%edi,4),%xmm3
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  102,15,112,193,78                   // pshufd        $0x4e,%xmm1,%xmm0
+  .byte  102,15,112,209,231                  // pshufd        $0xe7,%xmm1,%xmm2
+  .byte  102,15,110,44,190                   // movd          (%esi,%edi,4),%xmm5
+  .byte  102,15,126,215                      // movd          %xmm2,%edi
+  .byte  102,15,110,60,190                   // movd          (%esi,%edi,4),%xmm7
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  102,15,110,4,190                    // movd          (%esi,%edi,4),%xmm0
+  .byte  102,15,126,207                      // movd          %xmm1,%edi
+  .byte  102,15,112,201,229                  // pshufd        $0xe5,%xmm1,%xmm1
+  .byte  102,15,110,20,190                   // movd          (%esi,%edi,4),%xmm2
+  .byte  102,15,126,207                      // movd          %xmm1,%edi
+  .byte  102,15,110,12,190                   // movd          (%esi,%edi,4),%xmm1
+  .byte  139,117,16                          // mov           0x10(%ebp),%esi
+  .byte  139,125,20                          // mov           0x14(%ebp),%edi
+  .byte  102,15,98,244                       // punpckldq     %xmm4,%xmm6
+  .byte  102,15,98,221                       // punpckldq     %xmm5,%xmm3
+  .byte  15,40,109,40                        // movaps        0x28(%ebp),%xmm5
+  .byte  102,15,108,222                      // punpcklqdq    %xmm6,%xmm3
+  .byte  15,40,117,56                        // movaps        0x38(%ebp),%xmm6
+  .byte  102,15,98,199                       // punpckldq     %xmm7,%xmm0
+  .byte  102,15,98,209                       // punpckldq     %xmm1,%xmm2
+  .byte  102,15,108,208                      // punpcklqdq    %xmm0,%xmm2
+  .byte  102,15,111,194                      // movdqa        %xmm2,%xmm0
+  .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
+  .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
+  .byte  102,15,111,203                      // movdqa        %xmm3,%xmm1
+  .byte  102,15,114,241,16                   // pslld         $0x10,%xmm1
+  .byte  102,15,114,225,16                   // psrad         $0x10,%xmm1
+  .byte  102,15,107,200                      // packssdw      %xmm0,%xmm1
+  .byte  102,15,111,195                      // movdqa        %xmm3,%xmm0
+  .byte  102,15,111,250                      // movdqa        %xmm2,%xmm7
+  .byte  102,15,114,210,24                   // psrld         $0x18,%xmm2
+  .byte  102,15,114,242,16                   // pslld         $0x10,%xmm2
+  .byte  102,15,114,226,16                   // psrad         $0x10,%xmm2
+  .byte  102,15,114,211,24                   // psrld         $0x18,%xmm3
+  .byte  102,15,114,243,16                   // pslld         $0x10,%xmm3
+  .byte  102,15,114,227,16                   // psrad         $0x10,%xmm3
+  .byte  102,15,107,218                      // packssdw      %xmm2,%xmm3
+  .byte  102,15,111,162,20,36,0,0            // movdqa        0x2414(%edx),%xmm4
+  .byte  102,15,114,208,16                   // psrld         $0x10,%xmm0
+  .byte  102,15,114,215,16                   // psrld         $0x10,%xmm7
+  .byte  102,15,114,247,16                   // pslld         $0x10,%xmm7
+  .byte  102,15,114,231,16                   // psrad         $0x10,%xmm7
+  .byte  102,15,114,240,16                   // pslld         $0x10,%xmm0
+  .byte  102,15,114,224,16                   // psrad         $0x10,%xmm0
+  .byte  102,15,107,199                      // packssdw      %xmm7,%xmm0
+  .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
+  .byte  102,15,219,212                      // pand          %xmm4,%xmm2
+  .byte  102,15,219,196                      // pand          %xmm4,%xmm0
+  .byte  15,40,101,72                        // movaps        0x48(%ebp),%xmm4
+  .byte  102,15,113,209,8                    // psrlw         $0x8,%xmm1
+  .byte  141,81,8                            // lea           0x8(%ecx),%edx
+  .byte  15,41,100,36,64                     // movaps        %xmm4,0x40(%esp)
+  .byte  15,41,116,36,48                     // movaps        %xmm6,0x30(%esp)
+  .byte  15,41,108,36,32                     // movaps        %xmm5,0x20(%esp)
+  .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
+  .byte  15,41,100,36,16                     // movaps        %xmm4,0x10(%esp)
+  .byte  137,124,36,12                       // mov           %edi,0xc(%esp)
+  .byte  137,116,36,8                        // mov           %esi,0x8(%esp)
+  .byte  137,84,36,4                         // mov           %edx,0x4(%esp)
+  .byte  137,4,36                            // mov           %eax,(%esp)
+  .byte  255,81,4                            // call          *0x4(%ecx)
+  .byte  131,196,80                          // add           $0x50,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_load_565_sse2_lowp
 .globl _sk_load_565_sse2_lowp
 FUNCTION(_sk_load_565_sse2_lowp)
@@ -76110,25 +77862,25 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          ba42 <_sk_load_565_sse2_lowp+0x3b>
+  .byte  232,0,0,0,0                         // call          be60 <_sk_load_565_sse2_lowp+0x3b>
   .byte  94                                  // pop           %esi
   .byte  139,125,16                          // mov           0x10(%ebp),%edi
-  .byte  119,28                              // ja            ba64 <_sk_load_565_sse2_lowp+0x5d>
+  .byte  119,28                              // ja            be82 <_sk_load_565_sse2_lowp+0x5d>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,201                          // movzbl        %cl,%ecx
-  .byte  139,140,142,246,0,0,0               // mov           0xf6(%esi,%ecx,4),%ecx
+  .byte  139,140,142,244,0,0,0               // mov           0xf4(%esi,%ecx,4),%ecx
   .byte  1,241                               // add           %esi,%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  15,183,12,123                       // movzwl        (%ebx,%edi,2),%ecx
   .byte  102,15,110,193                      // movd          %ecx,%xmm0
-  .byte  235,59                              // jmp           ba9f <_sk_load_565_sse2_lowp+0x98>
+  .byte  235,59                              // jmp           bebd <_sk_load_565_sse2_lowp+0x98>
   .byte  243,15,111,4,123                    // movdqu        (%ebx,%edi,2),%xmm0
-  .byte  235,52                              // jmp           ba9f <_sk_load_565_sse2_lowp+0x98>
+  .byte  235,52                              // jmp           bebd <_sk_load_565_sse2_lowp+0x98>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,68,123,4,2               // pinsrw        $0x2,0x4(%ebx,%edi,2),%xmm0
   .byte  243,15,16,12,123                    // movss         (%ebx,%edi,2),%xmm1
   .byte  243,15,16,193                       // movss         %xmm1,%xmm0
-  .byte  235,30                              // jmp           ba9f <_sk_load_565_sse2_lowp+0x98>
+  .byte  235,30                              // jmp           bebd <_sk_load_565_sse2_lowp+0x98>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,68,123,12,6              // pinsrw        $0x6,0xc(%ebx,%edi,2),%xmm0
   .byte  102,15,196,68,123,10,5              // pinsrw        $0x5,0xa(%ebx,%edi,2),%xmm0
@@ -76136,11 +77888,11 @@
   .byte  102,15,18,4,123                     // movlpd        (%ebx,%edi,2),%xmm0
   .byte  102,15,111,248                      // movdqa        %xmm0,%xmm7
   .byte  102,15,113,215,8                    // psrlw         $0x8,%xmm7
-  .byte  102,15,219,190,46,29,0,0            // pand          0x1d2e(%esi),%xmm7
+  .byte  102,15,219,190,48,34,0,0            // pand          0x2230(%esi),%xmm7
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
-  .byte  102,15,219,142,62,29,0,0            // pand          0x1d3e(%esi),%xmm1
-  .byte  102,15,111,150,78,29,0,0            // movdqa        0x1d4e(%esi),%xmm2
+  .byte  102,15,219,142,64,34,0,0            // pand          0x2240(%esi),%xmm1
+  .byte  102,15,111,150,80,34,0,0            // movdqa        0x2250(%esi),%xmm2
   .byte  102,15,219,208                      // pand          %xmm0,%xmm2
   .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
   .byte  102,15,235,199                      // por           %xmm7,%xmm0
@@ -76162,7 +77914,7 @@
   .byte  137,124,36,8                        // mov           %edi,0x8(%esp)
   .byte  137,76,36,4                         // mov           %ecx,0x4(%esp)
   .byte  137,20,36                           // mov           %edx,(%esp)
-  .byte  15,40,158,238,28,0,0                // movaps        0x1cee(%esi),%xmm3
+  .byte  15,40,158,240,33,0,0                // movaps        0x21f0(%esi),%xmm3
   .byte  255,80,4                            // call          *0x4(%eax)
   .byte  131,196,92                          // add           $0x5c,%esp
   .byte  94                                  // pop           %esi
@@ -76170,7 +77922,7 @@
   .byte  91                                  // pop           %ebx
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
-  .byte  15,31,0                             // nopl          (%eax)
+  .byte  144                                 // nop
   .byte  24,0                                // sbb           %al,(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  52,0                                // xor           $0x0,%al
@@ -76208,10 +77960,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          bb7f <_sk_load_565_dst_sse2_lowp+0x2b>
+  .byte  232,0,0,0,0                         // call          bf9b <_sk_load_565_dst_sse2_lowp+0x2b>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            bba1 <_sk_load_565_dst_sse2_lowp+0x4d>
+  .byte  119,28                              // ja            bfbd <_sk_load_565_dst_sse2_lowp+0x4d>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,249,0,0,0               // mov           0xf9(%edi,%ecx,4),%ecx
@@ -76219,14 +77971,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,183,12,115                       // movzwl        (%ebx,%esi,2),%ecx
   .byte  102,15,110,225                      // movd          %ecx,%xmm4
-  .byte  235,59                              // jmp           bbdc <_sk_load_565_dst_sse2_lowp+0x88>
+  .byte  235,59                              // jmp           bff8 <_sk_load_565_dst_sse2_lowp+0x88>
   .byte  243,15,111,36,115                   // movdqu        (%ebx,%esi,2),%xmm4
-  .byte  235,52                              // jmp           bbdc <_sk_load_565_dst_sse2_lowp+0x88>
+  .byte  235,52                              // jmp           bff8 <_sk_load_565_dst_sse2_lowp+0x88>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,100,115,4,2              // pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm4
   .byte  243,15,16,44,115                    // movss         (%ebx,%esi,2),%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,30                              // jmp           bbdc <_sk_load_565_dst_sse2_lowp+0x88>
+  .byte  235,30                              // jmp           bff8 <_sk_load_565_dst_sse2_lowp+0x88>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,100,115,12,6             // pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm4
   .byte  102,15,196,100,115,10,5             // pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm4
@@ -76234,11 +77986,11 @@
   .byte  102,15,18,36,115                    // movlpd        (%ebx,%esi,2),%xmm4
   .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
   .byte  102,15,113,214,8                    // psrlw         $0x8,%xmm6
-  .byte  102,15,219,183,241,27,0,0           // pand          0x1bf1(%edi),%xmm6
+  .byte  102,15,219,183,245,32,0,0           // pand          0x20f5(%edi),%xmm6
   .byte  102,15,111,236                      // movdqa        %xmm4,%xmm5
   .byte  102,15,113,213,5                    // psrlw         $0x5,%xmm5
-  .byte  102,15,219,175,1,28,0,0             // pand          0x1c01(%edi),%xmm5
-  .byte  102,15,111,191,17,28,0,0            // movdqa        0x1c11(%edi),%xmm7
+  .byte  102,15,219,175,5,33,0,0             // pand          0x2105(%edi),%xmm5
+  .byte  102,15,111,191,21,33,0,0            // movdqa        0x2115(%edi),%xmm7
   .byte  102,15,219,252                      // pand          %xmm4,%xmm7
   .byte  102,15,113,212,13                   // psrlw         $0xd,%xmm4
   .byte  102,15,235,230                      // por           %xmm6,%xmm4
@@ -76251,7 +78003,7 @@
   .byte  102,15,113,215,2                    // psrlw         $0x2,%xmm7
   .byte  102,15,235,254                      // por           %xmm6,%xmm7
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
-  .byte  15,40,183,177,27,0,0                // movaps        0x1bb1(%edi),%xmm6
+  .byte  15,40,183,181,32,0,0                // movaps        0x20b5(%edi),%xmm6
   .byte  15,41,116,36,64                     // movaps        %xmm6,0x40(%esp)
   .byte  139,125,20                          // mov           0x14(%ebp),%edi
   .byte  137,124,36,12                       // mov           %edi,0xc(%esp)
@@ -76296,15 +78048,15 @@
   .byte  86                                  // push          %esi
   .byte  131,236,92                          // sub           $0x5c,%esp
   .byte  102,15,111,226                      // movdqa        %xmm2,%xmm4
-  .byte  232,0,0,0,0                         // call          bca6 <_sk_store_565_sse2_lowp+0x12>
+  .byte  232,0,0,0,0                         // call          c0c2 <_sk_store_565_sse2_lowp+0x12>
   .byte  95                                  // pop           %edi
   .byte  15,40,109,72                        // movaps        0x48(%ebp),%xmm5
   .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
   .byte  102,15,113,242,8                    // psllw         $0x8,%xmm2
-  .byte  102,15,219,151,250,26,0,0           // pand          0x1afa(%edi),%xmm2
+  .byte  102,15,219,151,254,31,0,0           // pand          0x1ffe(%edi),%xmm2
   .byte  102,15,111,241                      // movdqa        %xmm1,%xmm6
   .byte  102,15,113,246,3                    // psllw         $0x3,%xmm6
-  .byte  102,15,219,183,10,27,0,0            // pand          0x1b0a(%edi),%xmm6
+  .byte  102,15,219,183,14,32,0,0            // pand          0x200e(%edi),%xmm6
   .byte  102,15,235,242                      // por           %xmm2,%xmm6
   .byte  102,15,111,212                      // movdqa        %xmm4,%xmm2
   .byte  102,15,113,210,3                    // psrlw         $0x3,%xmm2
@@ -76323,19 +78075,19 @@
   .byte  128,249,6                           // cmp           $0x6,%cl
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
   .byte  15,40,125,24                        // movaps        0x18(%ebp),%xmm7
-  .byte  119,22                              // ja            bd1e <_sk_store_565_sse2_lowp+0x8a>
+  .byte  119,22                              // ja            c13a <_sk_store_565_sse2_lowp+0x8a>
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  3,188,143,242,0,0,0                 // add           0xf2(%edi,%ecx,4),%edi
   .byte  255,231                             // jmp           *%edi
   .byte  102,15,126,209                      // movd          %xmm2,%ecx
   .byte  102,137,12,115                      // mov           %cx,(%ebx,%esi,2)
-  .byte  235,59                              // jmp           bd59 <_sk_store_565_sse2_lowp+0xc5>
+  .byte  235,59                              // jmp           c175 <_sk_store_565_sse2_lowp+0xc5>
   .byte  243,15,127,20,115                   // movdqu        %xmm2,(%ebx,%esi,2)
-  .byte  235,52                              // jmp           bd59 <_sk_store_565_sse2_lowp+0xc5>
+  .byte  235,52                              // jmp           c175 <_sk_store_565_sse2_lowp+0xc5>
   .byte  102,15,197,202,2                    // pextrw        $0x2,%xmm2,%ecx
   .byte  102,137,76,115,4                    // mov           %cx,0x4(%ebx,%esi,2)
   .byte  102,15,126,20,115                   // movd          %xmm2,(%ebx,%esi,2)
-  .byte  235,35                              // jmp           bd59 <_sk_store_565_sse2_lowp+0xc5>
+  .byte  235,35                              // jmp           c175 <_sk_store_565_sse2_lowp+0xc5>
   .byte  102,15,197,202,6                    // pextrw        $0x6,%xmm2,%ecx
   .byte  102,137,76,115,12                   // mov           %cx,0xc(%ebx,%esi,2)
   .byte  102,15,197,202,5                    // pextrw        $0x5,%xmm2,%ecx
@@ -76373,6 +78125,140 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0                                   // .byte         0x0
 
+HIDDEN _sk_gather_565_sse2_lowp
+.globl _sk_gather_565_sse2_lowp
+FUNCTION(_sk_gather_565_sse2_lowp)
+_sk_gather_565_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  83                                  // push          %ebx
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,108                         // sub           $0x6c,%esp
+  .byte  139,69,12                           // mov           0xc(%ebp),%eax
+  .byte  139,16                              // mov           (%eax),%edx
+  .byte  243,15,16,98,8                      // movss         0x8(%edx),%xmm4
+  .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
+  .byte  102,15,118,237                      // pcmpeqd       %xmm5,%xmm5
+  .byte  102,15,254,229                      // paddd         %xmm5,%xmm4
+  .byte  15,87,246                           // xorps         %xmm6,%xmm6
+  .byte  15,95,198                           // maxps         %xmm6,%xmm0
+  .byte  15,95,206                           // maxps         %xmm6,%xmm1
+  .byte  15,93,204                           // minps         %xmm4,%xmm1
+  .byte  15,93,196                           // minps         %xmm4,%xmm0
+  .byte  243,15,16,98,12                     // movss         0xc(%edx),%xmm4
+  .byte  15,198,228,0                        // shufps        $0x0,%xmm4,%xmm4
+  .byte  102,15,254,229                      // paddd         %xmm5,%xmm4
+  .byte  15,95,222                           // maxps         %xmm6,%xmm3
+  .byte  15,95,214                           // maxps         %xmm6,%xmm2
+  .byte  15,93,212                           // minps         %xmm4,%xmm2
+  .byte  15,93,220                           // minps         %xmm4,%xmm3
+  .byte  243,15,91,219                       // cvttps2dq     %xmm3,%xmm3
+  .byte  243,15,91,210                       // cvttps2dq     %xmm2,%xmm2
+  .byte  102,15,110,98,4                     // movd          0x4(%edx),%xmm4
+  .byte  102,15,112,228,0                    // pshufd        $0x0,%xmm4,%xmm4
+  .byte  102,15,112,234,245                  // pshufd        $0xf5,%xmm2,%xmm5
+  .byte  102,15,244,236                      // pmuludq       %xmm4,%xmm5
+  .byte  102,15,111,244                      // movdqa        %xmm4,%xmm6
+  .byte  102,15,244,242                      // pmuludq       %xmm2,%xmm6
+  .byte  102,15,112,211,245                  // pshufd        $0xf5,%xmm3,%xmm2
+  .byte  102,15,244,212                      // pmuludq       %xmm4,%xmm2
+  .byte  232,0,0,0,0                         // call          c248 <_sk_gather_565_sse2_lowp+0x78>
+  .byte  89                                  // pop           %ecx
+  .byte  139,50                              // mov           (%edx),%esi
+  .byte  102,15,112,246,232                  // pshufd        $0xe8,%xmm6,%xmm6
+  .byte  102,15,112,237,232                  // pshufd        $0xe8,%xmm5,%xmm5
+  .byte  102,15,98,245                       // punpckldq     %xmm5,%xmm6
+  .byte  102,15,244,227                      // pmuludq       %xmm3,%xmm4
+  .byte  102,15,112,220,232                  // pshufd        $0xe8,%xmm4,%xmm3
+  .byte  102,15,112,210,232                  // pshufd        $0xe8,%xmm2,%xmm2
+  .byte  102,15,98,218                       // punpckldq     %xmm2,%xmm3
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,198                      // paddd         %xmm6,%xmm0
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,15,254,203                      // paddd         %xmm3,%xmm1
+  .byte  102,15,126,194                      // movd          %xmm0,%edx
+  .byte  102,15,112,208,229                  // pshufd        $0xe5,%xmm0,%xmm2
+  .byte  102,15,127,85,216                   // movdqa        %xmm2,-0x28(%ebp)
+  .byte  102,15,112,216,78                   // pshufd        $0x4e,%xmm0,%xmm3
+  .byte  102,15,112,192,231                  // pshufd        $0xe7,%xmm0,%xmm0
+  .byte  102,15,112,225,78                   // pshufd        $0x4e,%xmm1,%xmm4
+  .byte  102,15,112,233,231                  // pshufd        $0xe7,%xmm1,%xmm5
+  .byte  102,15,126,239                      // movd          %xmm5,%edi
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,239                      // movd          %edi,%xmm5
+  .byte  102,15,126,231                      // movd          %xmm4,%edi
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,231                      // movd          %edi,%xmm4
+  .byte  102,15,126,207                      // movd          %xmm1,%edi
+  .byte  102,15,112,201,229                  // pshufd        $0xe5,%xmm1,%xmm1
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,247                      // movd          %edi,%xmm6
+  .byte  102,15,126,207                      // movd          %xmm1,%edi
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,215                      // movd          %edi,%xmm2
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,255                      // movd          %edi,%xmm7
+  .byte  102,15,126,223                      // movd          %xmm3,%edi
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  102,15,110,207                      // movd          %edi,%xmm1
+  .byte  102,15,111,69,216                   // movdqa        -0x28(%ebp),%xmm0
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  15,183,20,86                        // movzwl        (%esi,%edx,2),%edx
+  .byte  102,15,110,194                      // movd          %edx,%xmm0
+  .byte  139,85,8                            // mov           0x8(%ebp),%edx
+  .byte  15,183,60,126                       // movzwl        (%esi,%edi,2),%edi
+  .byte  139,117,16                          // mov           0x10(%ebp),%esi
+  .byte  102,15,97,229                       // punpcklwd     %xmm5,%xmm4
+  .byte  102,15,110,239                      // movd          %edi,%xmm5
+  .byte  139,125,20                          // mov           0x14(%ebp),%edi
+  .byte  102,15,97,242                       // punpcklwd     %xmm2,%xmm6
+  .byte  15,40,93,24                         // movaps        0x18(%ebp),%xmm3
+  .byte  102,15,98,244                       // punpckldq     %xmm4,%xmm6
+  .byte  15,40,101,40                        // movaps        0x28(%ebp),%xmm4
+  .byte  102,15,97,207                       // punpcklwd     %xmm7,%xmm1
+  .byte  15,40,125,56                        // movaps        0x38(%ebp),%xmm7
+  .byte  102,15,97,197                       // punpcklwd     %xmm5,%xmm0
+  .byte  102,15,98,193                       // punpckldq     %xmm1,%xmm0
+  .byte  102,15,108,198                      // punpcklqdq    %xmm6,%xmm0
+  .byte  102,15,111,232                      // movdqa        %xmm0,%xmm5
+  .byte  102,15,113,213,8                    // psrlw         $0x8,%xmm5
+  .byte  102,15,219,169,72,30,0,0            // pand          0x1e48(%ecx),%xmm5
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,113,209,5                    // psrlw         $0x5,%xmm1
+  .byte  102,15,219,137,88,30,0,0            // pand          0x1e58(%ecx),%xmm1
+  .byte  102,15,111,145,104,30,0,0           // movdqa        0x1e68(%ecx),%xmm2
+  .byte  102,15,219,208                      // pand          %xmm0,%xmm2
+  .byte  102,15,113,208,13                   // psrlw         $0xd,%xmm0
+  .byte  102,15,235,197                      // por           %xmm5,%xmm0
+  .byte  102,15,111,233                      // movdqa        %xmm1,%xmm5
+  .byte  102,15,113,245,2                    // psllw         $0x2,%xmm5
+  .byte  102,15,113,209,4                    // psrlw         $0x4,%xmm1
+  .byte  102,15,235,205                      // por           %xmm5,%xmm1
+  .byte  102,15,111,234                      // movdqa        %xmm2,%xmm5
+  .byte  102,15,113,245,3                    // psllw         $0x3,%xmm5
+  .byte  102,15,113,210,2                    // psrlw         $0x2,%xmm2
+  .byte  102,15,235,213                      // por           %xmm5,%xmm2
+  .byte  15,40,109,72                        // movaps        0x48(%ebp),%xmm5
+  .byte  141,88,8                            // lea           0x8(%eax),%ebx
+  .byte  15,41,108,36,64                     // movaps        %xmm5,0x40(%esp)
+  .byte  15,41,124,36,48                     // movaps        %xmm7,0x30(%esp)
+  .byte  15,41,100,36,32                     // movaps        %xmm4,0x20(%esp)
+  .byte  15,41,92,36,16                      // movaps        %xmm3,0x10(%esp)
+  .byte  137,124,36,12                       // mov           %edi,0xc(%esp)
+  .byte  137,116,36,8                        // mov           %esi,0x8(%esp)
+  .byte  137,92,36,4                         // mov           %ebx,0x4(%esp)
+  .byte  137,20,36                           // mov           %edx,(%esp)
+  .byte  15,40,153,8,30,0,0                  // movaps        0x1e08(%ecx),%xmm3
+  .byte  255,80,4                            // call          *0x4(%eax)
+  .byte  131,196,108                         // add           $0x6c,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  91                                  // pop           %ebx
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_load_a8_sse2_lowp
 .globl _sk_load_a8_sse2_lowp
 FUNCTION(_sk_load_a8_sse2_lowp)
@@ -76397,21 +78283,21 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          bdee <_sk_load_a8_sse2_lowp+0x3a>
+  .byte  232,0,0,0,0                         // call          c405 <_sk_load_a8_sse2_lowp+0x3a>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            be10 <_sk_load_a8_sse2_lowp+0x5c>
+  .byte  119,28                              // ja            c427 <_sk_load_a8_sse2_lowp+0x5c>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
-  .byte  139,140,143,198,0,0,0               // mov           0xc6(%edi,%ecx,4),%ecx
+  .byte  139,140,143,195,0,0,0               // mov           0xc3(%edi,%ecx,4),%ecx
   .byte  1,249                               // add           %edi,%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,51                        // movzbl        (%ebx,%esi,1),%ecx
   .byte  102,15,110,217                      // movd          %ecx,%xmm3
-  .byte  235,90                              // jmp           be6a <_sk_load_a8_sse2_lowp+0xb6>
+  .byte  235,90                              // jmp           c481 <_sk_load_a8_sse2_lowp+0xb6>
   .byte  243,15,126,28,51                    // movq          (%ebx,%esi,1),%xmm3
   .byte  102,15,96,216                       // punpcklbw     %xmm0,%xmm3
-  .byte  235,79                              // jmp           be6a <_sk_load_a8_sse2_lowp+0xb6>
+  .byte  235,79                              // jmp           c481 <_sk_load_a8_sse2_lowp+0xb6>
   .byte  15,182,76,51,2                      // movzbl        0x2(%ebx,%esi,1),%ecx
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,217,2                    // pinsrw        $0x2,%ecx,%xmm3
@@ -76419,7 +78305,7 @@
   .byte  102,15,110,233                      // movd          %ecx,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,221                       // movss         %xmm5,%xmm3
-  .byte  235,47                              // jmp           be6a <_sk_load_a8_sse2_lowp+0xb6>
+  .byte  235,47                              // jmp           c481 <_sk_load_a8_sse2_lowp+0xb6>
   .byte  15,182,76,51,6                      // movzbl        0x6(%ebx,%esi,1),%ecx
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,217,6                    // pinsrw        $0x6,%ecx,%xmm3
@@ -76430,7 +78316,7 @@
   .byte  102,15,110,44,51                    // movd          (%ebx,%esi,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,221                       // movsd         %xmm5,%xmm3
-  .byte  102,15,219,159,66,25,0,0            // pand          0x1942(%edi),%xmm3
+  .byte  102,15,219,159,75,28,0,0            // pand          0x1c4b(%edi),%xmm3
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  15,41,100,36,64                     // movaps        %xmm4,0x40(%esp)
   .byte  15,41,84,36,48                      // movaps        %xmm2,0x30(%esp)
@@ -76451,7 +78337,6 @@
   .byte  91                                  // pop           %ebx
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
-  .byte  15,31,0                             // nopl          (%eax)
   .byte  24,0                                // sbb           %al,(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  59,0                                // cmp           (%eax),%eax
@@ -76485,10 +78370,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          bef9 <_sk_load_a8_dst_sse2_lowp+0x29>
+  .byte  232,0,0,0,0                         // call          c50d <_sk_load_a8_dst_sse2_lowp+0x29>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            bf1b <_sk_load_a8_dst_sse2_lowp+0x4b>
+  .byte  119,28                              // ja            c52f <_sk_load_a8_dst_sse2_lowp+0x4b>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,195,0,0,0               // mov           0xc3(%edi,%ecx,4),%ecx
@@ -76496,10 +78381,10 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,51                        // movzbl        (%ebx,%esi,1),%ecx
   .byte  102,15,110,225                      // movd          %ecx,%xmm4
-  .byte  235,90                              // jmp           bf75 <_sk_load_a8_dst_sse2_lowp+0xa5>
+  .byte  235,90                              // jmp           c589 <_sk_load_a8_dst_sse2_lowp+0xa5>
   .byte  243,15,126,36,51                    // movq          (%ebx,%esi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
-  .byte  235,79                              // jmp           bf75 <_sk_load_a8_dst_sse2_lowp+0xa5>
+  .byte  235,79                              // jmp           c589 <_sk_load_a8_dst_sse2_lowp+0xa5>
   .byte  15,182,76,51,2                      // movzbl        0x2(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,2                    // pinsrw        $0x2,%ecx,%xmm4
@@ -76507,7 +78392,7 @@
   .byte  102,15,110,233                      // movd          %ecx,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,47                              // jmp           bf75 <_sk_load_a8_dst_sse2_lowp+0xa5>
+  .byte  235,47                              // jmp           c589 <_sk_load_a8_dst_sse2_lowp+0xa5>
   .byte  15,182,76,51,6                      // movzbl        0x6(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,6                    // pinsrw        $0x6,%ecx,%xmm4
@@ -76518,7 +78403,7 @@
   .byte  102,15,110,44,51                    // movd          (%ebx,%esi,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,229                       // movsd         %xmm5,%xmm4
-  .byte  102,15,219,167,55,24,0,0            // pand          0x1837(%edi),%xmm4
+  .byte  102,15,219,167,67,27,0,0            // pand          0x1b43(%edi),%xmm4
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  102,15,127,100,36,64                // movdqa        %xmm4,0x40(%esp)
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
@@ -76575,32 +78460,32 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c011 <_sk_store_a8_sse2_lowp+0x39>
+  .byte  232,0,0,0,0                         // call          c625 <_sk_store_a8_sse2_lowp+0x39>
   .byte  91                                  // pop           %ebx
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,26                              // ja            c031 <_sk_store_a8_sse2_lowp+0x59>
+  .byte  119,26                              // ja            c645 <_sk_store_a8_sse2_lowp+0x59>
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,139,223,0,0,0               // mov           0xdf(%ebx,%ecx,4),%ecx
   .byte  1,217                               // add           %ebx,%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  102,15,126,217                      // movd          %xmm3,%ecx
   .byte  136,12,55                           // mov           %cl,(%edi,%esi,1)
-  .byte  233,136,0,0,0                       // jmp           c0b9 <_sk_store_a8_sse2_lowp+0xe1>
-  .byte  102,15,111,163,31,23,0,0            // movdqa        0x171f(%ebx),%xmm4
+  .byte  233,136,0,0,0                       // jmp           c6cd <_sk_store_a8_sse2_lowp+0xe1>
+  .byte  102,15,111,163,43,26,0,0            // movdqa        0x1a2b(%ebx),%xmm4
   .byte  102,15,219,227                      // pand          %xmm3,%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
   .byte  102,15,214,36,55                    // movq          %xmm4,(%edi,%esi,1)
   .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
-  .byte  235,109                             // jmp           c0b9 <_sk_store_a8_sse2_lowp+0xe1>
+  .byte  235,109                             // jmp           c6cd <_sk_store_a8_sse2_lowp+0xe1>
   .byte  102,15,197,203,2                    // pextrw        $0x2,%xmm3,%ecx
   .byte  136,76,55,2                         // mov           %cl,0x2(%edi,%esi,1)
-  .byte  102,15,111,163,31,23,0,0            // movdqa        0x171f(%ebx),%xmm4
+  .byte  102,15,111,163,43,26,0,0            // movdqa        0x1a2b(%ebx),%xmm4
   .byte  102,15,219,227                      // pand          %xmm3,%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
   .byte  102,15,126,225                      // movd          %xmm4,%ecx
   .byte  15,40,101,24                        // movaps        0x18(%ebp),%xmm4
   .byte  102,137,12,55                       // mov           %cx,(%edi,%esi,1)
-  .byte  235,70                              // jmp           c0b9 <_sk_store_a8_sse2_lowp+0xe1>
+  .byte  235,70                              // jmp           c6cd <_sk_store_a8_sse2_lowp+0xe1>
   .byte  102,15,197,203,6                    // pextrw        $0x6,%xmm3,%ecx
   .byte  136,76,55,6                         // mov           %cl,0x6(%edi,%esi,1)
   .byte  102,15,197,203,5                    // pextrw        $0x5,%xmm3,%ecx
@@ -76610,7 +78495,7 @@
   .byte  15,40,254                           // movaps        %xmm6,%xmm7
   .byte  15,40,245                           // movaps        %xmm5,%xmm6
   .byte  15,40,236                           // movaps        %xmm4,%xmm5
-  .byte  102,15,111,163,31,23,0,0            // movdqa        0x171f(%ebx),%xmm4
+  .byte  102,15,111,163,43,26,0,0            // movdqa        0x1a2b(%ebx),%xmm4
   .byte  102,15,219,227                      // pand          %xmm3,%xmm4
   .byte  102,15,103,228                      // packuswb      %xmm4,%xmm4
   .byte  102,15,126,36,55                    // movd          %xmm4,(%edi,%esi,1)
@@ -76644,12 +78529,122 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0,125,0                             // add           %bh,0x0(%ebp)
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  116,0                               // je            c102 <_sk_store_a8_sse2_lowp+0x12a>
+  .byte  116,0                               // je            c716 <_sk_store_a8_sse2_lowp+0x12a>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  107,0,0                             // imul          $0x0,(%eax),%eax
   .byte  0,98,0                              // add           %ah,0x0(%edx)
   .byte  0,0                                 // add           %al,(%eax)
 
+HIDDEN _sk_gather_a8_sse2_lowp
+.globl _sk_gather_a8_sse2_lowp
+FUNCTION(_sk_gather_a8_sse2_lowp)
+_sk_gather_a8_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,80                          // sub           $0x50,%esp
+  .byte  139,69,12                           // mov           0xc(%ebp),%eax
+  .byte  139,8                               // mov           (%eax),%ecx
+  .byte  243,15,16,105,8                     // movss         0x8(%ecx),%xmm5
+  .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
+  .byte  102,15,118,246                      // pcmpeqd       %xmm6,%xmm6
+  .byte  102,15,254,238                      // paddd         %xmm6,%xmm5
+  .byte  15,87,228                           // xorps         %xmm4,%xmm4
+  .byte  15,95,196                           // maxps         %xmm4,%xmm0
+  .byte  15,95,204                           // maxps         %xmm4,%xmm1
+  .byte  15,93,205                           // minps         %xmm5,%xmm1
+  .byte  15,93,197                           // minps         %xmm5,%xmm0
+  .byte  243,15,16,105,12                    // movss         0xc(%ecx),%xmm5
+  .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
+  .byte  102,15,254,238                      // paddd         %xmm6,%xmm5
+  .byte  15,95,220                           // maxps         %xmm4,%xmm3
+  .byte  15,95,212                           // maxps         %xmm4,%xmm2
+  .byte  15,93,213                           // minps         %xmm5,%xmm2
+  .byte  15,93,221                           // minps         %xmm5,%xmm3
+  .byte  243,15,91,234                       // cvttps2dq     %xmm2,%xmm5
+  .byte  102,15,110,81,4                     // movd          0x4(%ecx),%xmm2
+  .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
+  .byte  102,15,111,242                      // movdqa        %xmm2,%xmm6
+  .byte  102,15,244,245                      // pmuludq       %xmm5,%xmm6
+  .byte  102,15,112,237,245                  // pshufd        $0xf5,%xmm5,%xmm5
+  .byte  102,15,244,234                      // pmuludq       %xmm2,%xmm5
+  .byte  102,15,112,246,232                  // pshufd        $0xe8,%xmm6,%xmm6
+  .byte  102,15,112,237,232                  // pshufd        $0xe8,%xmm5,%xmm5
+  .byte  102,15,98,245                       // punpckldq     %xmm5,%xmm6
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,198                      // paddd         %xmm6,%xmm0
+  .byte  102,15,112,232,78                   // pshufd        $0x4e,%xmm0,%xmm5
+  .byte  102,15,126,234                      // movd          %xmm5,%edx
+  .byte  102,15,112,232,231                  // pshufd        $0xe7,%xmm0,%xmm5
+  .byte  102,15,126,238                      // movd          %xmm5,%esi
+  .byte  139,9                               // mov           (%ecx),%ecx
+  .byte  243,15,91,219                       // cvttps2dq     %xmm3,%xmm3
+  .byte  102,15,126,199                      // movd          %xmm0,%edi
+  .byte  102,15,112,192,229                  // pshufd        $0xe5,%xmm0,%xmm0
+  .byte  15,182,20,17                        // movzbl        (%ecx,%edx,1),%edx
+  .byte  15,182,52,49                        // movzbl        (%ecx,%esi,1),%esi
+  .byte  193,230,8                           // shl           $0x8,%esi
+  .byte  9,214                               // or            %edx,%esi
+  .byte  102,15,126,194                      // movd          %xmm0,%edx
+  .byte  102,15,112,195,245                  // pshufd        $0xf5,%xmm3,%xmm0
+  .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
+  .byte  15,182,60,57                        // movzbl        (%ecx,%edi,1),%edi
+  .byte  15,182,20,17                        // movzbl        (%ecx,%edx,1),%edx
+  .byte  193,226,8                           // shl           $0x8,%edx
+  .byte  9,250                               // or            %edi,%edx
+  .byte  139,125,8                           // mov           0x8(%ebp),%edi
+  .byte  102,15,244,211                      // pmuludq       %xmm3,%xmm2
+  .byte  102,15,112,210,232                  // pshufd        $0xe8,%xmm2,%xmm2
+  .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
+  .byte  102,15,98,208                       // punpckldq     %xmm0,%xmm2
+  .byte  243,15,91,193                       // cvttps2dq     %xmm1,%xmm0
+  .byte  102,15,254,194                      // paddd         %xmm2,%xmm0
+  .byte  102,15,110,218                      // movd          %edx,%xmm3
+  .byte  102,15,126,194                      // movd          %xmm0,%edx
+  .byte  102,15,112,200,229                  // pshufd        $0xe5,%xmm0,%xmm1
+  .byte  102,15,196,222,1                    // pinsrw        $0x1,%esi,%xmm3
+  .byte  102,15,126,206                      // movd          %xmm1,%esi
+  .byte  102,15,112,200,78                   // pshufd        $0x4e,%xmm0,%xmm1
+  .byte  15,182,20,17                        // movzbl        (%ecx,%edx,1),%edx
+  .byte  15,182,52,49                        // movzbl        (%ecx,%esi,1),%esi
+  .byte  193,230,8                           // shl           $0x8,%esi
+  .byte  9,214                               // or            %edx,%esi
+  .byte  102,15,126,202                      // movd          %xmm1,%edx
+  .byte  102,15,112,192,231                  // pshufd        $0xe7,%xmm0,%xmm0
+  .byte  102,15,196,222,2                    // pinsrw        $0x2,%esi,%xmm3
+  .byte  102,15,126,198                      // movd          %xmm0,%esi
+  .byte  15,182,20,17                        // movzbl        (%ecx,%edx,1),%edx
+  .byte  15,182,12,49                        // movzbl        (%ecx,%esi,1),%ecx
+  .byte  139,117,16                          // mov           0x10(%ebp),%esi
+  .byte  193,225,8                           // shl           $0x8,%ecx
+  .byte  9,209                               // or            %edx,%ecx
+  .byte  139,85,20                           // mov           0x14(%ebp),%edx
+  .byte  15,40,69,24                         // movaps        0x18(%ebp),%xmm0
+  .byte  15,40,77,40                         // movaps        0x28(%ebp),%xmm1
+  .byte  15,40,85,56                         // movaps        0x38(%ebp),%xmm2
+  .byte  102,15,196,217,3                    // pinsrw        $0x3,%ecx,%xmm3
+  .byte  102,15,96,220                       // punpcklbw     %xmm4,%xmm3
+  .byte  15,40,101,72                        // movaps        0x48(%ebp),%xmm4
+  .byte  141,72,8                            // lea           0x8(%eax),%ecx
+  .byte  15,41,100,36,64                     // movaps        %xmm4,0x40(%esp)
+  .byte  15,41,84,36,48                      // movaps        %xmm2,0x30(%esp)
+  .byte  15,41,76,36,32                      // movaps        %xmm1,0x20(%esp)
+  .byte  15,41,68,36,16                      // movaps        %xmm0,0x10(%esp)
+  .byte  137,84,36,12                        // mov           %edx,0xc(%esp)
+  .byte  137,116,36,8                        // mov           %esi,0x8(%esp)
+  .byte  137,76,36,4                         // mov           %ecx,0x4(%esp)
+  .byte  137,60,36                           // mov           %edi,(%esp)
+  .byte  15,87,192                           // xorps         %xmm0,%xmm0
+  .byte  15,87,201                           // xorps         %xmm1,%xmm1
+  .byte  15,87,210                           // xorps         %xmm2,%xmm2
+  .byte  255,80,4                            // call          *0x4(%eax)
+  .byte  131,196,80                          // add           $0x50,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_load_g8_sse2_lowp
 .globl _sk_load_g8_sse2_lowp
 FUNCTION(_sk_load_g8_sse2_lowp)
@@ -76674,21 +78669,21 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c145 <_sk_load_g8_sse2_lowp+0x39>
+  .byte  232,0,0,0,0                         // call          c8d8 <_sk_load_g8_sse2_lowp+0x39>
   .byte  94                                  // pop           %esi
   .byte  139,125,16                          // mov           0x10(%ebp),%edi
-  .byte  119,28                              // ja            c167 <_sk_load_g8_sse2_lowp+0x5b>
+  .byte  119,28                              // ja            c8fa <_sk_load_g8_sse2_lowp+0x5b>
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  15,182,201                          // movzbl        %cl,%ecx
-  .byte  139,140,142,199,0,0,0               // mov           0xc7(%esi,%ecx,4),%ecx
+  .byte  139,140,142,200,0,0,0               // mov           0xc8(%esi,%ecx,4),%ecx
   .byte  1,241                               // add           %esi,%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,59                        // movzbl        (%ebx,%edi,1),%ecx
   .byte  102,15,110,193                      // movd          %ecx,%xmm0
-  .byte  235,90                              // jmp           c1c1 <_sk_load_g8_sse2_lowp+0xb5>
+  .byte  235,90                              // jmp           c954 <_sk_load_g8_sse2_lowp+0xb5>
   .byte  243,15,126,4,59                     // movq          (%ebx,%edi,1),%xmm0
   .byte  102,15,96,192                       // punpcklbw     %xmm0,%xmm0
-  .byte  235,79                              // jmp           c1c1 <_sk_load_g8_sse2_lowp+0xb5>
+  .byte  235,79                              // jmp           c954 <_sk_load_g8_sse2_lowp+0xb5>
   .byte  15,182,76,59,2                      // movzbl        0x2(%ebx,%edi,1),%ecx
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,193,2                    // pinsrw        $0x2,%ecx,%xmm0
@@ -76696,7 +78691,7 @@
   .byte  102,15,110,233                      // movd          %ecx,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,197                       // movss         %xmm5,%xmm0
-  .byte  235,47                              // jmp           c1c1 <_sk_load_g8_sse2_lowp+0xb5>
+  .byte  235,47                              // jmp           c954 <_sk_load_g8_sse2_lowp+0xb5>
   .byte  15,182,76,59,6                      // movzbl        0x6(%ebx,%edi,1),%ecx
   .byte  102,15,239,192                      // pxor          %xmm0,%xmm0
   .byte  102,15,196,193,6                    // pinsrw        $0x6,%ecx,%xmm0
@@ -76707,7 +78702,7 @@
   .byte  102,15,110,44,59                    // movd          (%ebx,%edi,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,197                       // movsd         %xmm5,%xmm0
-  .byte  102,15,219,134,235,21,0,0           // pand          0x15eb(%esi),%xmm0
+  .byte  102,15,219,134,120,23,0,0           // pand          0x1778(%esi),%xmm0
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
   .byte  15,41,100,36,64                     // movaps        %xmm4,0x40(%esp)
   .byte  15,41,92,36,48                      // movaps        %xmm3,0x30(%esp)
@@ -76718,7 +78713,7 @@
   .byte  137,124,36,8                        // mov           %edi,0x8(%esp)
   .byte  137,76,36,4                         // mov           %ecx,0x4(%esp)
   .byte  137,20,36                           // mov           %edx,(%esp)
-  .byte  15,40,158,235,21,0,0                // movaps        0x15eb(%esi),%xmm3
+  .byte  15,40,158,120,23,0,0                // movaps        0x1778(%esi),%xmm3
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
   .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
   .byte  255,80,4                            // call          *0x4(%eax)
@@ -76728,6 +78723,7 @@
   .byte  91                                  // pop           %ebx
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
+  .byte  144                                 // nop
   .byte  24,0                                // sbb           %al,(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  59,0                                // cmp           (%eax),%eax
@@ -76761,10 +78757,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c251 <_sk_load_g8_dst_sse2_lowp+0x29>
+  .byte  232,0,0,0,0                         // call          c9e5 <_sk_load_g8_dst_sse2_lowp+0x29>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            c273 <_sk_load_g8_dst_sse2_lowp+0x4b>
+  .byte  119,28                              // ja            ca07 <_sk_load_g8_dst_sse2_lowp+0x4b>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,195,0,0,0               // mov           0xc3(%edi,%ecx,4),%ecx
@@ -76772,10 +78768,10 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,51                        // movzbl        (%ebx,%esi,1),%ecx
   .byte  102,15,110,225                      // movd          %ecx,%xmm4
-  .byte  235,90                              // jmp           c2cd <_sk_load_g8_dst_sse2_lowp+0xa5>
+  .byte  235,90                              // jmp           ca61 <_sk_load_g8_dst_sse2_lowp+0xa5>
   .byte  243,15,126,36,51                    // movq          (%ebx,%esi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
-  .byte  235,79                              // jmp           c2cd <_sk_load_g8_dst_sse2_lowp+0xa5>
+  .byte  235,79                              // jmp           ca61 <_sk_load_g8_dst_sse2_lowp+0xa5>
   .byte  15,182,76,51,2                      // movzbl        0x2(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,2                    // pinsrw        $0x2,%ecx,%xmm4
@@ -76783,7 +78779,7 @@
   .byte  102,15,110,233                      // movd          %ecx,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,47                              // jmp           c2cd <_sk_load_g8_dst_sse2_lowp+0xa5>
+  .byte  235,47                              // jmp           ca61 <_sk_load_g8_dst_sse2_lowp+0xa5>
   .byte  15,182,76,51,6                      // movzbl        0x6(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,6                    // pinsrw        $0x6,%ecx,%xmm4
@@ -76794,9 +78790,9 @@
   .byte  102,15,110,44,51                    // movd          (%ebx,%esi,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,229                       // movsd         %xmm5,%xmm4
-  .byte  102,15,219,167,223,20,0,0           // pand          0x14df(%edi),%xmm4
+  .byte  102,15,219,167,107,22,0,0           // pand          0x166b(%edi),%xmm4
   .byte  141,72,8                            // lea           0x8(%eax),%ecx
-  .byte  15,40,175,223,20,0,0                // movaps        0x14df(%edi),%xmm5
+  .byte  15,40,175,107,22,0,0                // movaps        0x166b(%edi),%xmm5
   .byte  15,41,108,36,64                     // movaps        %xmm5,0x40(%esp)
   .byte  102,15,127,100,36,48                // movdqa        %xmm4,0x30(%esp)
   .byte  102,15,127,100,36,32                // movdqa        %xmm4,0x20(%esp)
@@ -76837,7 +78833,7 @@
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
-  .byte  232,0,0,0,0                         // call          c341 <_sk_luminance_to_alpha_sse2_lowp+0x11>
+  .byte  232,0,0,0,0                         // call          cad5 <_sk_luminance_to_alpha_sse2_lowp+0x11>
   .byte  88                                  // pop           %eax
   .byte  139,77,8                            // mov           0x8(%ebp),%ecx
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
@@ -76847,10 +78843,10 @@
   .byte  15,40,109,56                        // movaps        0x38(%ebp),%xmm5
   .byte  15,40,117,72                        // movaps        0x48(%ebp),%xmm6
   .byte  139,125,12                          // mov           0xc(%ebp),%edi
-  .byte  102,15,213,128,127,20,0,0           // pmullw        0x147f(%eax),%xmm0
-  .byte  102,15,213,136,143,20,0,0           // pmullw        0x148f(%eax),%xmm1
+  .byte  102,15,213,128,11,22,0,0            // pmullw        0x160b(%eax),%xmm0
+  .byte  102,15,213,136,27,22,0,0            // pmullw        0x161b(%eax),%xmm1
   .byte  102,15,253,200                      // paddw         %xmm0,%xmm1
-  .byte  102,15,213,152,159,20,0,0           // pmullw        0x149f(%eax),%xmm3
+  .byte  102,15,213,152,43,22,0,0            // pmullw        0x162b(%eax),%xmm3
   .byte  102,15,253,217                      // paddw         %xmm1,%xmm3
   .byte  102,15,113,211,8                    // psrlw         $0x8,%xmm3
   .byte  141,71,4                            // lea           0x4(%edi),%eax
@@ -76872,6 +78868,120 @@
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
 
+HIDDEN _sk_gather_g8_sse2_lowp
+.globl _sk_gather_g8_sse2_lowp
+FUNCTION(_sk_gather_g8_sse2_lowp)
+_sk_gather_g8_sse2_lowp:
+  .byte  85                                  // push          %ebp
+  .byte  137,229                             // mov           %esp,%ebp
+  .byte  83                                  // push          %ebx
+  .byte  87                                  // push          %edi
+  .byte  86                                  // push          %esi
+  .byte  131,236,92                          // sub           $0x5c,%esp
+  .byte  139,69,12                           // mov           0xc(%ebp),%eax
+  .byte  139,8                               // mov           (%eax),%ecx
+  .byte  243,15,16,105,8                     // movss         0x8(%ecx),%xmm5
+  .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
+  .byte  102,15,118,246                      // pcmpeqd       %xmm6,%xmm6
+  .byte  102,15,254,238                      // paddd         %xmm6,%xmm5
+  .byte  15,87,228                           // xorps         %xmm4,%xmm4
+  .byte  15,95,196                           // maxps         %xmm4,%xmm0
+  .byte  15,95,204                           // maxps         %xmm4,%xmm1
+  .byte  15,93,205                           // minps         %xmm5,%xmm1
+  .byte  15,93,197                           // minps         %xmm5,%xmm0
+  .byte  243,15,16,105,12                    // movss         0xc(%ecx),%xmm5
+  .byte  15,198,237,0                        // shufps        $0x0,%xmm5,%xmm5
+  .byte  102,15,254,238                      // paddd         %xmm6,%xmm5
+  .byte  15,95,220                           // maxps         %xmm4,%xmm3
+  .byte  15,95,212                           // maxps         %xmm4,%xmm2
+  .byte  15,93,213                           // minps         %xmm5,%xmm2
+  .byte  15,93,221                           // minps         %xmm5,%xmm3
+  .byte  243,15,91,234                       // cvttps2dq     %xmm2,%xmm5
+  .byte  102,15,110,81,4                     // movd          0x4(%ecx),%xmm2
+  .byte  102,15,112,210,0                    // pshufd        $0x0,%xmm2,%xmm2
+  .byte  102,15,111,242                      // movdqa        %xmm2,%xmm6
+  .byte  102,15,244,245                      // pmuludq       %xmm5,%xmm6
+  .byte  102,15,112,237,245                  // pshufd        $0xf5,%xmm5,%xmm5
+  .byte  102,15,244,234                      // pmuludq       %xmm2,%xmm5
+  .byte  102,15,112,246,232                  // pshufd        $0xe8,%xmm6,%xmm6
+  .byte  102,15,112,237,232                  // pshufd        $0xe8,%xmm5,%xmm5
+  .byte  102,15,98,245                       // punpckldq     %xmm5,%xmm6
+  .byte  243,15,91,192                       // cvttps2dq     %xmm0,%xmm0
+  .byte  102,15,254,198                      // paddd         %xmm6,%xmm0
+  .byte  102,15,112,232,78                   // pshufd        $0x4e,%xmm0,%xmm5
+  .byte  102,15,126,234                      // movd          %xmm5,%edx
+  .byte  102,15,112,232,231                  // pshufd        $0xe7,%xmm0,%xmm5
+  .byte  102,15,126,238                      // movd          %xmm5,%esi
+  .byte  139,57                              // mov           (%ecx),%edi
+  .byte  243,15,91,219                       // cvttps2dq     %xmm3,%xmm3
+  .byte  102,15,126,193                      // movd          %xmm0,%ecx
+  .byte  102,15,112,192,229                  // pshufd        $0xe5,%xmm0,%xmm0
+  .byte  15,182,20,23                        // movzbl        (%edi,%edx,1),%edx
+  .byte  15,182,52,55                        // movzbl        (%edi,%esi,1),%esi
+  .byte  193,230,8                           // shl           $0x8,%esi
+  .byte  9,214                               // or            %edx,%esi
+  .byte  102,15,126,194                      // movd          %xmm0,%edx
+  .byte  102,15,112,195,245                  // pshufd        $0xf5,%xmm3,%xmm0
+  .byte  102,15,244,194                      // pmuludq       %xmm2,%xmm0
+  .byte  15,182,12,15                        // movzbl        (%edi,%ecx,1),%ecx
+  .byte  15,182,20,23                        // movzbl        (%edi,%edx,1),%edx
+  .byte  193,226,8                           // shl           $0x8,%edx
+  .byte  9,202                               // or            %ecx,%edx
+  .byte  232,0,0,0,0                         // call          cc1a <_sk_gather_g8_sse2_lowp+0xc9>
+  .byte  89                                  // pop           %ecx
+  .byte  102,15,244,211                      // pmuludq       %xmm3,%xmm2
+  .byte  102,15,112,210,232                  // pshufd        $0xe8,%xmm2,%xmm2
+  .byte  102,15,112,192,232                  // pshufd        $0xe8,%xmm0,%xmm0
+  .byte  102,15,98,208                       // punpckldq     %xmm0,%xmm2
+  .byte  243,15,91,201                       // cvttps2dq     %xmm1,%xmm1
+  .byte  102,15,254,202                      // paddd         %xmm2,%xmm1
+  .byte  102,15,110,194                      // movd          %edx,%xmm0
+  .byte  102,15,126,202                      // movd          %xmm1,%edx
+  .byte  102,15,112,209,229                  // pshufd        $0xe5,%xmm1,%xmm2
+  .byte  102,15,196,198,1                    // pinsrw        $0x1,%esi,%xmm0
+  .byte  102,15,126,214                      // movd          %xmm2,%esi
+  .byte  102,15,112,209,78                   // pshufd        $0x4e,%xmm1,%xmm2
+  .byte  15,182,20,23                        // movzbl        (%edi,%edx,1),%edx
+  .byte  15,182,52,55                        // movzbl        (%edi,%esi,1),%esi
+  .byte  193,230,8                           // shl           $0x8,%esi
+  .byte  9,214                               // or            %edx,%esi
+  .byte  102,15,126,210                      // movd          %xmm2,%edx
+  .byte  102,15,112,201,231                  // pshufd        $0xe7,%xmm1,%xmm1
+  .byte  102,15,196,198,2                    // pinsrw        $0x2,%esi,%xmm0
+  .byte  102,15,126,206                      // movd          %xmm1,%esi
+  .byte  15,182,20,23                        // movzbl        (%edi,%edx,1),%edx
+  .byte  15,182,52,55                        // movzbl        (%edi,%esi,1),%esi
+  .byte  139,125,8                           // mov           0x8(%ebp),%edi
+  .byte  193,230,8                           // shl           $0x8,%esi
+  .byte  9,214                               // or            %edx,%esi
+  .byte  139,85,16                           // mov           0x10(%ebp),%edx
+  .byte  102,15,196,198,3                    // pinsrw        $0x3,%esi,%xmm0
+  .byte  139,117,20                          // mov           0x14(%ebp),%esi
+  .byte  15,40,77,24                         // movaps        0x18(%ebp),%xmm1
+  .byte  15,40,85,40                         // movaps        0x28(%ebp),%xmm2
+  .byte  15,40,93,56                         // movaps        0x38(%ebp),%xmm3
+  .byte  102,15,96,196                       // punpcklbw     %xmm4,%xmm0
+  .byte  15,40,101,72                        // movaps        0x48(%ebp),%xmm4
+  .byte  141,88,8                            // lea           0x8(%eax),%ebx
+  .byte  15,41,100,36,64                     // movaps        %xmm4,0x40(%esp)
+  .byte  15,41,92,36,48                      // movaps        %xmm3,0x30(%esp)
+  .byte  15,41,84,36,32                      // movaps        %xmm2,0x20(%esp)
+  .byte  15,41,76,36,16                      // movaps        %xmm1,0x10(%esp)
+  .byte  137,116,36,12                       // mov           %esi,0xc(%esp)
+  .byte  137,84,36,8                         // mov           %edx,0x8(%esp)
+  .byte  137,92,36,4                         // mov           %ebx,0x4(%esp)
+  .byte  137,60,36                           // mov           %edi,(%esp)
+  .byte  15,40,153,54,20,0,0                 // movaps        0x1436(%ecx),%xmm3
+  .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
+  .byte  102,15,111,208                      // movdqa        %xmm0,%xmm2
+  .byte  255,80,4                            // call          *0x4(%eax)
+  .byte  131,196,92                          // add           $0x5c,%esp
+  .byte  94                                  // pop           %esi
+  .byte  95                                  // pop           %edi
+  .byte  91                                  // pop           %ebx
+  .byte  93                                  // pop           %ebp
+  .byte  195                                 // ret
+
 HIDDEN _sk_scale_1_float_sse2_lowp
 .globl _sk_scale_1_float_sse2_lowp
 FUNCTION(_sk_scale_1_float_sse2_lowp)
@@ -76881,15 +78991,15 @@
   .byte  87                                  // push          %edi
   .byte  86                                  // push          %esi
   .byte  131,236,80                          // sub           $0x50,%esp
-  .byte  232,0,0,0,0                         // call          c3ca <_sk_scale_1_float_sse2_lowp+0xd>
+  .byte  232,0,0,0,0                         // call          cceb <_sk_scale_1_float_sse2_lowp+0xd>
   .byte  94                                  // pop           %esi
   .byte  139,69,8                            // mov           0x8(%ebp),%eax
   .byte  139,85,16                           // mov           0x10(%ebp),%edx
   .byte  139,77,12                           // mov           0xc(%ebp),%ecx
   .byte  139,57                              // mov           (%ecx),%edi
   .byte  243,15,16,39                        // movss         (%edi),%xmm4
-  .byte  243,15,89,166,42,20,0,0             // mulss         0x142a(%esi),%xmm4
-  .byte  243,15,88,166,38,20,0,0             // addss         0x1426(%esi),%xmm4
+  .byte  243,15,89,166,41,20,0,0             // mulss         0x1429(%esi),%xmm4
+  .byte  243,15,88,166,37,20,0,0             // addss         0x1425(%esi),%xmm4
   .byte  243,15,44,252                       // cvttss2si     %xmm4,%edi
   .byte  102,15,110,239                      // movd          %edi,%xmm5
   .byte  139,125,20                          // mov           0x14(%ebp),%edi
@@ -76902,7 +79012,7 @@
   .byte  102,15,213,205                      // pmullw        %xmm5,%xmm1
   .byte  102,15,213,213                      // pmullw        %xmm5,%xmm2
   .byte  102,15,213,221                      // pmullw        %xmm5,%xmm3
-  .byte  102,15,111,174,102,19,0,0           // movdqa        0x1366(%esi),%xmm5
+  .byte  102,15,111,174,101,19,0,0           // movdqa        0x1365(%esi),%xmm5
   .byte  102,15,253,197                      // paddw         %xmm5,%xmm0
   .byte  102,15,253,205                      // paddw         %xmm5,%xmm1
   .byte  102,15,253,213                      // paddw         %xmm5,%xmm2
@@ -76941,19 +79051,19 @@
   .byte  102,15,111,218                      // movdqa        %xmm2,%xmm3
   .byte  102,15,111,209                      // movdqa        %xmm1,%xmm2
   .byte  102,15,111,200                      // movdqa        %xmm0,%xmm1
-  .byte  232,0,0,0,0                         // call          c498 <_sk_lerp_1_float_sse2_lowp+0x1d>
+  .byte  232,0,0,0,0                         // call          cdb9 <_sk_lerp_1_float_sse2_lowp+0x1d>
   .byte  89                                  // pop           %ecx
   .byte  102,15,111,109,24                   // movdqa        0x18(%ebp),%xmm5
   .byte  139,69,12                           // mov           0xc(%ebp),%eax
   .byte  139,16                              // mov           (%eax),%edx
   .byte  243,15,16,50                        // movss         (%edx),%xmm6
-  .byte  243,15,89,177,92,19,0,0             // mulss         0x135c(%ecx),%xmm6
-  .byte  243,15,88,177,88,19,0,0             // addss         0x1358(%ecx),%xmm6
+  .byte  243,15,89,177,91,19,0,0             // mulss         0x135b(%ecx),%xmm6
+  .byte  243,15,88,177,87,19,0,0             // addss         0x1357(%ecx),%xmm6
   .byte  243,15,44,214                       // cvttss2si     %xmm6,%edx
   .byte  102,15,110,242                      // movd          %edx,%xmm6
   .byte  242,15,112,246,0                    // pshuflw       $0x0,%xmm6,%xmm6
   .byte  102,15,112,246,80                   // pshufd        $0x50,%xmm6,%xmm6
-  .byte  102,15,111,129,152,18,0,0           // movdqa        0x1298(%ecx),%xmm0
+  .byte  102,15,111,129,151,18,0,0           // movdqa        0x1297(%ecx),%xmm0
   .byte  102,15,213,206                      // pmullw        %xmm6,%xmm1
   .byte  102,15,253,200                      // paddw         %xmm0,%xmm1
   .byte  102,15,213,214                      // pmullw        %xmm6,%xmm2
@@ -77026,21 +79136,21 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c5c1 <_sk_scale_u8_sse2_lowp+0x31>
+  .byte  232,0,0,0,0                         // call          cee2 <_sk_scale_u8_sse2_lowp+0x31>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            c5e3 <_sk_scale_u8_sse2_lowp+0x53>
+  .byte  119,28                              // ja            cf04 <_sk_scale_u8_sse2_lowp+0x53>
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  15,182,201                          // movzbl        %cl,%ecx
-  .byte  139,140,143,3,1,0,0                 // mov           0x103(%edi,%ecx,4),%ecx
+  .byte  139,140,143,2,1,0,0                 // mov           0x102(%edi,%ecx,4),%ecx
   .byte  1,249                               // add           %edi,%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,51                        // movzbl        (%ebx,%esi,1),%ecx
   .byte  102,15,110,225                      // movd          %ecx,%xmm4
-  .byte  235,90                              // jmp           c63d <_sk_scale_u8_sse2_lowp+0xad>
+  .byte  235,90                              // jmp           cf5e <_sk_scale_u8_sse2_lowp+0xad>
   .byte  243,15,126,36,51                    // movq          (%ebx,%esi,1),%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
-  .byte  235,79                              // jmp           c63d <_sk_scale_u8_sse2_lowp+0xad>
+  .byte  235,79                              // jmp           cf5e <_sk_scale_u8_sse2_lowp+0xad>
   .byte  15,182,76,51,2                      // movzbl        0x2(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,2                    // pinsrw        $0x2,%ecx,%xmm4
@@ -77048,7 +79158,7 @@
   .byte  102,15,110,233                      // movd          %ecx,%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  243,15,16,229                       // movss         %xmm5,%xmm4
-  .byte  235,47                              // jmp           c63d <_sk_scale_u8_sse2_lowp+0xad>
+  .byte  235,47                              // jmp           cf5e <_sk_scale_u8_sse2_lowp+0xad>
   .byte  15,182,76,51,6                      // movzbl        0x6(%ebx,%esi,1),%ecx
   .byte  102,15,239,228                      // pxor          %xmm4,%xmm4
   .byte  102,15,196,225,6                    // pinsrw        $0x6,%ecx,%xmm4
@@ -77059,9 +79169,9 @@
   .byte  102,15,110,44,51                    // movd          (%ebx,%esi,1),%xmm5
   .byte  102,15,96,232                       // punpcklbw     %xmm0,%xmm5
   .byte  242,15,16,229                       // movsd         %xmm5,%xmm4
-  .byte  102,15,219,167,111,17,0,0           // pand          0x116f(%edi),%xmm4
+  .byte  102,15,219,167,110,17,0,0           // pand          0x116e(%edi),%xmm4
   .byte  102,15,213,196                      // pmullw        %xmm4,%xmm0
-  .byte  102,15,111,175,111,17,0,0           // movdqa        0x116f(%edi),%xmm5
+  .byte  102,15,111,175,110,17,0,0           // movdqa        0x116e(%edi),%xmm5
   .byte  102,15,253,197                      // paddw         %xmm5,%xmm0
   .byte  102,15,113,208,8                    // psrlw         $0x8,%xmm0
   .byte  102,15,213,204                      // pmullw        %xmm4,%xmm1
@@ -77093,7 +79203,7 @@
   .byte  91                                  // pop           %ebx
   .byte  93                                  // pop           %ebp
   .byte  195                                 // ret
-  .byte  15,31,0                             // nopl          (%eax)
+  .byte  102,144                             // xchg          %ax,%ax
   .byte  24,0                                // sbb           %al,(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  59,0                                // cmp           (%eax),%eax
@@ -77128,10 +79238,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c70e <_sk_lerp_u8_sse2_lowp+0x2e>
+  .byte  232,0,0,0,0                         // call          d02e <_sk_lerp_u8_sse2_lowp+0x2e>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            c730 <_sk_lerp_u8_sse2_lowp+0x50>
+  .byte  119,28                              // ja            d050 <_sk_lerp_u8_sse2_lowp+0x50>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,74,1,0,0                // mov           0x14a(%edi,%ecx,4),%ecx
@@ -77139,10 +79249,10 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,182,12,51                        // movzbl        (%ebx,%esi,1),%ecx
   .byte  102,15,110,217                      // movd          %ecx,%xmm3
-  .byte  235,90                              // jmp           c78a <_sk_lerp_u8_sse2_lowp+0xaa>
+  .byte  235,90                              // jmp           d0aa <_sk_lerp_u8_sse2_lowp+0xaa>
   .byte  243,15,126,28,51                    // movq          (%ebx,%esi,1),%xmm3
   .byte  102,15,96,216                       // punpcklbw     %xmm0,%xmm3
-  .byte  235,79                              // jmp           c78a <_sk_lerp_u8_sse2_lowp+0xaa>
+  .byte  235,79                              // jmp           d0aa <_sk_lerp_u8_sse2_lowp+0xaa>
   .byte  15,182,76,51,2                      // movzbl        0x2(%ebx,%esi,1),%ecx
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,217,2                    // pinsrw        $0x2,%ecx,%xmm3
@@ -77150,7 +79260,7 @@
   .byte  102,15,110,225                      // movd          %ecx,%xmm4
   .byte  102,15,96,224                       // punpcklbw     %xmm0,%xmm4
   .byte  243,15,16,220                       // movss         %xmm4,%xmm3
-  .byte  235,47                              // jmp           c78a <_sk_lerp_u8_sse2_lowp+0xaa>
+  .byte  235,47                              // jmp           d0aa <_sk_lerp_u8_sse2_lowp+0xaa>
   .byte  15,182,76,51,6                      // movzbl        0x6(%ebx,%esi,1),%ecx
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,217,6                    // pinsrw        $0x6,%ecx,%xmm3
@@ -77251,10 +79361,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          c8b7 <_sk_scale_565_sse2_lowp+0x43>
+  .byte  232,0,0,0,0                         // call          d1d7 <_sk_scale_565_sse2_lowp+0x43>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            c8d9 <_sk_scale_565_sse2_lowp+0x65>
+  .byte  119,28                              // ja            d1f9 <_sk_scale_565_sse2_lowp+0x65>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,249,1,0,0               // mov           0x1f9(%edi,%ecx,4),%ecx
@@ -77262,14 +79372,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,183,12,115                       // movzwl        (%ebx,%esi,2),%ecx
   .byte  102,15,110,217                      // movd          %ecx,%xmm3
-  .byte  235,59                              // jmp           c914 <_sk_scale_565_sse2_lowp+0xa0>
+  .byte  235,59                              // jmp           d234 <_sk_scale_565_sse2_lowp+0xa0>
   .byte  243,15,111,28,115                   // movdqu        (%ebx,%esi,2),%xmm3
-  .byte  235,52                              // jmp           c914 <_sk_scale_565_sse2_lowp+0xa0>
+  .byte  235,52                              // jmp           d234 <_sk_scale_565_sse2_lowp+0xa0>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,92,115,4,2               // pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm3
   .byte  243,15,16,12,115                    // movss         (%ebx,%esi,2),%xmm1
   .byte  243,15,16,217                       // movss         %xmm1,%xmm3
-  .byte  235,30                              // jmp           c914 <_sk_scale_565_sse2_lowp+0xa0>
+  .byte  235,30                              // jmp           d234 <_sk_scale_565_sse2_lowp+0xa0>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,92,115,12,6              // pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm3
   .byte  102,15,196,92,115,10,5              // pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm3
@@ -77414,10 +79524,10 @@
   .byte  128,225,7                           // and           $0x7,%cl
   .byte  254,201                             // dec           %cl
   .byte  128,249,6                           // cmp           $0x6,%cl
-  .byte  232,0,0,0,0                         // call          cb0f <_sk_lerp_565_sse2_lowp+0x43>
+  .byte  232,0,0,0,0                         // call          d42f <_sk_lerp_565_sse2_lowp+0x43>
   .byte  95                                  // pop           %edi
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
-  .byte  119,28                              // ja            cb31 <_sk_lerp_565_sse2_lowp+0x65>
+  .byte  119,28                              // ja            d451 <_sk_lerp_565_sse2_lowp+0x65>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  15,182,201                          // movzbl        %cl,%ecx
   .byte  139,140,143,73,2,0,0                // mov           0x249(%edi,%ecx,4),%ecx
@@ -77425,14 +79535,14 @@
   .byte  255,225                             // jmp           *%ecx
   .byte  15,183,12,115                       // movzwl        (%ebx,%esi,2),%ecx
   .byte  102,15,110,217                      // movd          %ecx,%xmm3
-  .byte  235,59                              // jmp           cb6c <_sk_lerp_565_sse2_lowp+0xa0>
+  .byte  235,59                              // jmp           d48c <_sk_lerp_565_sse2_lowp+0xa0>
   .byte  243,15,111,28,115                   // movdqu        (%ebx,%esi,2),%xmm3
-  .byte  235,52                              // jmp           cb6c <_sk_lerp_565_sse2_lowp+0xa0>
+  .byte  235,52                              // jmp           d48c <_sk_lerp_565_sse2_lowp+0xa0>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,92,115,4,2               // pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm3
   .byte  243,15,16,12,115                    // movss         (%ebx,%esi,2),%xmm1
   .byte  243,15,16,217                       // movss         %xmm1,%xmm3
-  .byte  235,30                              // jmp           cb6c <_sk_lerp_565_sse2_lowp+0xa0>
+  .byte  235,30                              // jmp           d48c <_sk_lerp_565_sse2_lowp+0xa0>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  102,15,196,92,115,12,6              // pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm3
   .byte  102,15,196,92,115,10,5              // pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm3
@@ -77594,11 +79704,11 @@
   .byte  128,227,7                           // and           $0x7,%bl
   .byte  254,203                             // dec           %bl
   .byte  128,251,6                           // cmp           $0x6,%bl
-  .byte  232,0,0,0,0                         // call          cdb3 <_sk_srcover_rgba_8888_sse2_lowp+0x3f>
+  .byte  232,0,0,0,0                         // call          d6d3 <_sk_srcover_rgba_8888_sse2_lowp+0x3f>
   .byte  89                                  // pop           %ecx
   .byte  139,117,16                          // mov           0x10(%ebp),%esi
   .byte  15,182,211                          // movzbl        %bl,%edx
-  .byte  119,29                              // ja            cdd9 <_sk_srcover_rgba_8888_sse2_lowp+0x65>
+  .byte  119,29                              // ja            d6f9 <_sk_srcover_rgba_8888_sse2_lowp+0x65>
   .byte  102,15,239,219                      // pxor          %xmm3,%xmm3
   .byte  139,132,145,77,2,0,0                // mov           0x24d(%ecx,%edx,4),%eax
   .byte  1,200                               // add           %ecx,%eax
@@ -77606,14 +79716,14 @@
   .byte  255,224                             // jmp           *%eax
   .byte  243,15,16,52,183                    // movss         (%edi,%esi,4),%xmm6
   .byte  15,87,219                           // xorps         %xmm3,%xmm3
-  .byte  235,73                              // jmp           ce22 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  .byte  235,73                              // jmp           d742 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   .byte  102,15,16,52,183                    // movupd        (%edi,%esi,4),%xmm6
   .byte  15,16,92,183,16                     // movups        0x10(%edi,%esi,4),%xmm3
-  .byte  235,61                              // jmp           ce22 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  .byte  235,61                              // jmp           d742 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   .byte  102,15,110,68,183,8                 // movd          0x8(%edi,%esi,4),%xmm0
   .byte  102,15,112,240,69                   // pshufd        $0x45,%xmm0,%xmm6
   .byte  102,15,18,52,183                    // movlpd        (%edi,%esi,4),%xmm6
-  .byte  235,43                              // jmp           ce22 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  .byte  235,43                              // jmp           d742 <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   .byte  102,15,110,68,183,24                // movd          0x18(%edi,%esi,4),%xmm0
   .byte  102,15,112,216,69                   // pshufd        $0x45,%xmm0,%xmm3
   .byte  243,15,16,68,183,20                 // movss         0x14(%edi,%esi,4),%xmm0
@@ -77691,18 +79801,18 @@
   .byte  102,15,105,220                      // punpckhwd     %xmm4,%xmm3
   .byte  102,15,235,222                      // por           %xmm6,%xmm3
   .byte  128,251,6                           // cmp           $0x6,%bl
-  .byte  119,16                              // ja            cf63 <_sk_srcover_rgba_8888_sse2_lowp+0x1ef>
+  .byte  119,16                              // ja            d883 <_sk_srcover_rgba_8888_sse2_lowp+0x1ef>
   .byte  3,140,145,105,2,0,0                 // add           0x269(%ecx,%edx,4),%ecx
   .byte  255,225                             // jmp           *%ecx
   .byte  102,15,126,4,183                    // movd          %xmm0,(%edi,%esi,4)
-  .byte  235,64                              // jmp           cfa3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  .byte  235,64                              // jmp           d8c3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   .byte  243,15,127,4,183                    // movdqu        %xmm0,(%edi,%esi,4)
   .byte  243,15,127,92,183,16                // movdqu        %xmm3,0x10(%edi,%esi,4)
-  .byte  235,51                              // jmp           cfa3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  .byte  235,51                              // jmp           d8c3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   .byte  102,15,112,208,78                   // pshufd        $0x4e,%xmm0,%xmm2
   .byte  102,15,126,84,183,8                 // movd          %xmm2,0x8(%edi,%esi,4)
   .byte  102,15,214,4,183                    // movq          %xmm0,(%edi,%esi,4)
-  .byte  235,33                              // jmp           cfa3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  .byte  235,33                              // jmp           d8c3 <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   .byte  102,15,112,211,78                   // pshufd        $0x4e,%xmm3,%xmm2
   .byte  102,15,126,84,183,24                // movd          %xmm2,0x18(%edi,%esi,4)
   .byte  102,15,112,211,229                  // pshufd        $0xe5,%xmm3,%xmm2
@@ -77822,11 +79932,11 @@
   .byte  0,128,191,0,0,128                   // add           %al,-0x7fffff41(%eax)
   .byte  191,0,0,224,64                      // mov           $0x40e00000,%edi
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,64                              // loopne        d118 <.literal16+0xd8>
+  .byte  224,64                              // loopne        da38 <.literal16+0xd8>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,64                              // loopne        d11c <.literal16+0xdc>
+  .byte  224,64                              // loopne        da3c <.literal16+0xdc>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,64                              // loopne        d120 <.literal16+0xe0>
+  .byte  224,64                              // loopne        da40 <.literal16+0xe0>
   .byte  154,153,153,62,154,153,153          // lcall         $0x9999,$0x9a3e9999
   .byte  62,154,153,153,62,154,153,153       // ds            lcall $0x9999,$0x9a3e9999
   .byte  62,61,10,23,63,61                   // ds            cmp $0x3d3f170a,%eax
@@ -77837,16 +79947,16 @@
   .byte  63                                  // aas
   .byte  174                                 // scas          %es:(%edi),%al
   .byte  71                                  // inc           %edi
-  .byte  225,61                              // loope         d141 <.literal16+0x101>
+  .byte  225,61                              // loope         da61 <.literal16+0x101>
   .byte  174                                 // scas          %es:(%edi),%al
   .byte  71                                  // inc           %edi
-  .byte  225,61                              // loope         d145 <.literal16+0x105>
+  .byte  225,61                              // loope         da65 <.literal16+0x105>
   .byte  174                                 // scas          %es:(%edi),%al
   .byte  71                                  // inc           %edi
-  .byte  225,61                              // loope         d149 <.literal16+0x109>
+  .byte  225,61                              // loope         da69 <.literal16+0x109>
   .byte  174                                 // scas          %es:(%edi),%al
   .byte  71                                  // inc           %edi
-  .byte  225,61                              // loope         d14d <.literal16+0x10d>
+  .byte  225,61                              // loope         da6d <.literal16+0x10d>
   .byte  255,0                               // incl          (%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  255,0                               // incl          (%eax)
@@ -77856,13 +79966,13 @@
   .byte  255,0                               // incl          (%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  127,67                              // jg            d167 <.literal16+0x127>
+  .byte  127,67                              // jg            da87 <.literal16+0x127>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  127,67                              // jg            d16b <.literal16+0x12b>
+  .byte  127,67                              // jg            da8b <.literal16+0x12b>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  127,67                              // jg            d16f <.literal16+0x12f>
+  .byte  127,67                              // jg            da8f <.literal16+0x12f>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  127,67                              // jg            d173 <.literal16+0x133>
+  .byte  127,67                              // jg            da93 <.literal16+0x133>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  128,127,0,0                         // cmpb          $0x0,0x0(%edi)
   .byte  128,127,0,0                         // cmpb          $0x0,0x0(%edi)
@@ -78018,13 +80128,13 @@
   .byte  132,55                              // test          %dh,(%edi)
   .byte  8,33                                // or            %ah,(%ecx)
   .byte  132,55                              // test          %dh,(%edi)
-  .byte  224,7                               // loopne        d289 <.literal16+0x249>
+  .byte  224,7                               // loopne        dba9 <.literal16+0x249>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,7                               // loopne        d28d <.literal16+0x24d>
+  .byte  224,7                               // loopne        dbad <.literal16+0x24d>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,7                               // loopne        d291 <.literal16+0x251>
+  .byte  224,7                               // loopne        dbb1 <.literal16+0x251>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  224,7                               // loopne        d295 <.literal16+0x255>
+  .byte  224,7                               // loopne        dbb5 <.literal16+0x255>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  33,8                                // and           %ecx,(%eax)
   .byte  2,58                                // add           (%edx),%bh
@@ -78073,17 +80183,17 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  52,255                              // xor           $0xff,%al
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d2f4 <.literal16+0x2b4>
+  .byte  127,0                               // jg            dc14 <.literal16+0x2b4>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d2f8 <.literal16+0x2b8>
+  .byte  127,0                               // jg            dc18 <.literal16+0x2b8>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d2fc <.literal16+0x2bc>
+  .byte  127,0                               // jg            dc1c <.literal16+0x2bc>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d300 <.literal16+0x2c0>
-  .byte  119,115                             // ja            d375 <.literal16+0x335>
+  .byte  127,0                               // jg            dc20 <.literal16+0x2c0>
+  .byte  119,115                             // ja            dc95 <.literal16+0x335>
   .byte  248                                 // clc
   .byte  194,119,115                         // ret           $0x7377
   .byte  248                                 // clc
@@ -78094,7 +80204,7 @@
   .byte  194,117,191                         // ret           $0xbf75
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // aas
-  .byte  117,191                             // jne           d2d9 <.literal16+0x299>
+  .byte  117,191                             // jne           dbf9 <.literal16+0x299>
   .byte  191,63,117,191,191                  // mov           $0xbfbf753f,%edi
   .byte  63                                  // aas
   .byte  249                                 // stc
@@ -78110,7 +80220,7 @@
   .byte  68                                  // inc           %esp
   .byte  180,62                              // mov           $0x3e,%ah
   .byte  163,233,220,63,163                  // mov           %eax,0xa33fdce9
-  .byte  233,220,63,163,233                  // jmp           e9a41316 <_sk_srcover_rgba_8888_sse2_lowp+0xe9a345a2>
+  .byte  233,220,63,163,233                  // jmp           e9a41c36 <_sk_srcover_rgba_8888_sse2_lowp+0xe9a345a2>
   .byte  220,63                              // fdivrl        (%edi)
   .byte  163,233,220,63,81                   // mov           %eax,0x513fdce9
   .byte  140,242                             // mov           %?,%edx
@@ -78206,16 +80316,16 @@
   .byte  128,3,62                            // addb          $0x3e,(%ebx)
   .byte  31                                  // pop           %ds
   .byte  215                                 // xlat          %ds:(%ebx)
-  .byte  118,63                              // jbe           d463 <.literal16+0x423>
+  .byte  118,63                              // jbe           dd83 <.literal16+0x423>
   .byte  31                                  // pop           %ds
   .byte  215                                 // xlat          %ds:(%ebx)
-  .byte  118,63                              // jbe           d467 <.literal16+0x427>
+  .byte  118,63                              // jbe           dd87 <.literal16+0x427>
   .byte  31                                  // pop           %ds
   .byte  215                                 // xlat          %ds:(%ebx)
-  .byte  118,63                              // jbe           d46b <.literal16+0x42b>
+  .byte  118,63                              // jbe           dd8b <.literal16+0x42b>
   .byte  31                                  // pop           %ds
   .byte  215                                 // xlat          %ds:(%ebx)
-  .byte  118,63                              // jbe           d46f <.literal16+0x42f>
+  .byte  118,63                              // jbe           dd8f <.literal16+0x42f>
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%eax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%eax)
   .byte  246,64,83,63                        // testb         $0x3f,0x53(%eax)
@@ -78241,13 +80351,13 @@
   .byte  248                                 // clc
   .byte  65                                  // inc           %ecx
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  124,66                              // jl            d4a6 <.literal16+0x466>
+  .byte  124,66                              // jl            ddc6 <.literal16+0x466>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  124,66                              // jl            d4aa <.literal16+0x46a>
+  .byte  124,66                              // jl            ddca <.literal16+0x46a>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  124,66                              // jl            d4ae <.literal16+0x46e>
+  .byte  124,66                              // jl            ddce <.literal16+0x46e>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  124,66                              // jl            d4b2 <.literal16+0x472>
+  .byte  124,66                              // jl            ddd2 <.literal16+0x472>
   .byte  0,240                               // add           %dh,%al
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0,240                               // add           %dh,%al
@@ -78293,25 +80403,25 @@
   .byte  136,136,61,137,136,136              // mov           %cl,-0x777776c3(%eax)
   .byte  61,137,136,136,61                   // cmp           $0x3d888889,%eax
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  112,65                              // jo            d535 <.literal16+0x4f5>
+  .byte  112,65                              // jo            de55 <.literal16+0x4f5>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  112,65                              // jo            d539 <.literal16+0x4f9>
+  .byte  112,65                              // jo            de59 <.literal16+0x4f9>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  112,65                              // jo            d53d <.literal16+0x4fd>
+  .byte  112,65                              // jo            de5d <.literal16+0x4fd>
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  112,65                              // jo            d541 <.literal16+0x501>
+  .byte  112,65                              // jo            de61 <.literal16+0x501>
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0,128,0,0,0,128                     // add           %al,-0x80000000(%eax)
   .byte  0,0                                 // add           %al,(%eax)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d513 <.literal16+0x4d3>
+  .byte  127,0                               // jg            de33 <.literal16+0x4d3>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            d517 <.literal16+0x4d7>
+  .byte  127,0                               // jg            de37 <.literal16+0x4d7>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            d51b <.literal16+0x4db>
+  .byte  127,0                               // jg            de3b <.literal16+0x4db>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,0                               // jg            d51f <.literal16+0x4df>
+  .byte  127,0                               // jg            de3f <.literal16+0x4df>
   .byte  0,0                                 // add           %al,(%eax)
   .byte  4,0                                 // add           $0x0,%al
   .byte  0,0                                 // add           %al,(%eax)
@@ -78337,13 +80447,13 @@
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,255                             // jg            d558 <.literal16+0x518>
+  .byte  127,255                             // jg            de78 <.literal16+0x518>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,255                             // jg            d55c <.literal16+0x51c>
+  .byte  127,255                             // jg            de7c <.literal16+0x51c>
   .byte  255                                 // (bad)
   .byte  255                                 // (bad)
-  .byte  127,0                               // jg            d561 <.literal16+0x521>
+  .byte  127,0                               // jg            de81 <.literal16+0x521>
   .byte  0,128,56,0,0,128                    // add           %al,-0x7fffffc8(%eax)
   .byte  56,0                                // cmp           %al,(%eax)
   .byte  0,128,56,0,0,128                    // add           %al,-0x7fffffc8(%eax)
@@ -78357,13 +80467,13 @@
   .byte  0,255                               // add           %bh,%bh
   .byte  255,0                               // incl          (%eax)
   .byte  255                                 // (bad)
-  .byte  127,71                              // jg            d5cb <.literal16+0x58b>
+  .byte  127,71                              // jg            deeb <.literal16+0x58b>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            d5cf <.literal16+0x58f>
+  .byte  127,71                              // jg            deef <.literal16+0x58f>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            d5d3 <.literal16+0x593>
+  .byte  127,71                              // jg            def3 <.literal16+0x593>
   .byte  0,255                               // add           %bh,%bh
-  .byte  127,71                              // jg            d5d7 <.literal16+0x597>
+  .byte  127,71                              // jg            def7 <.literal16+0x597>
   .byte  208                                 // (bad)
   .byte  179,89                              // mov           $0x59,%bl
   .byte  62,208                              // ds            (bad)
@@ -78416,7 +80526,7 @@
   .byte  192,45,16,17,192,45,16              // shrb          $0x10,0x2dc01110
   .byte  17,192                              // adc           %eax,%eax
   .byte  45,16,17,192,18                     // sub           $0x12c01110,%eax
-  .byte  120,57                              // js            d63c <.literal16+0x5fc>
+  .byte  120,57                              // js            df5c <.literal16+0x5fc>
   .byte  64                                  // inc           %eax
   .byte  18,120,57                           // adc           0x39(%eax),%bh
   .byte  64                                  // inc           %eax
@@ -78487,13 +80597,13 @@
   .byte  192,191,0,0,192,191,0               // sarb          $0x0,-0x40400000(%edi)
   .byte  0,192                               // add           %al,%al
   .byte  191,0,0,192,191                     // mov           $0xbfc00000,%edi
-  .byte  114,28                              // jb            d6ee <.literal16+0x6ae>
+  .byte  114,28                              // jb            e00e <.literal16+0x6ae>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         d6f2 <.literal16+0x6b2>
+  .byte  62,114,28                           // jb,pt         e012 <.literal16+0x6b2>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         d6f6 <.literal16+0x6b6>
+  .byte  62,114,28                           // jb,pt         e016 <.literal16+0x6b6>
   .byte  199                                 // (bad)
-  .byte  62,114,28                           // jb,pt         d6fa <.literal16+0x6ba>
+  .byte  62,114,28                           // jb,pt         e01a <.literal16+0x6ba>
   .byte  199                                 // (bad)
   .byte  62,85                               // ds            push %ebp
   .byte  85                                  // push          %ebp
@@ -78514,13 +80624,13 @@
   .byte  57,142,99,61,57,142                 // cmp           %ecx,-0x71c6c29d(%esi)
   .byte  99,61,57,142,99,61                  // arpl          %di,0x3d638e39
   .byte  57,142,99,61,114,249                // cmp           %ecx,-0x68dc29d(%esi)
-  .byte  127,63                              // jg            d753 <.literal16+0x713>
-  .byte  114,249                             // jb            d70f <.literal16+0x6cf>
-  .byte  127,63                              // jg            d757 <.literal16+0x717>
-  .byte  114,249                             // jb            d713 <.literal16+0x6d3>
-  .byte  127,63                              // jg            d75b <.literal16+0x71b>
-  .byte  114,249                             // jb            d717 <.literal16+0x6d7>
-  .byte  127,63                              // jg            d75f <.literal16+0x71f>
+  .byte  127,63                              // jg            e073 <.literal16+0x713>
+  .byte  114,249                             // jb            e02f <.literal16+0x6cf>
+  .byte  127,63                              // jg            e077 <.literal16+0x717>
+  .byte  114,249                             // jb            e033 <.literal16+0x6d3>
+  .byte  127,63                              // jg            e07b <.literal16+0x71b>
+  .byte  114,249                             // jb            e037 <.literal16+0x6d7>
+  .byte  127,63                              // jg            e07f <.literal16+0x71f>
   .byte  3,0                                 // add           (%eax),%eax
   .byte  0,0                                 // add           %al,(%eax)
   .byte  3,0                                 // add           (%eax),%eax
@@ -78588,14 +80698,14 @@
   .byte  0,248                               // add           %bh,%al
   .byte  0,248                               // add           %bh,%al
   .byte  0,248                               // add           %bh,%al
-  .byte  224,255                             // loopne        d7b1 <.literal16+0x771>
-  .byte  224,255                             // loopne        d7b3 <.literal16+0x773>
-  .byte  224,255                             // loopne        d7b5 <.literal16+0x775>
-  .byte  224,255                             // loopne        d7b7 <.literal16+0x777>
-  .byte  224,255                             // loopne        d7b9 <.literal16+0x779>
-  .byte  224,255                             // loopne        d7bb <.literal16+0x77b>
-  .byte  224,255                             // loopne        d7bd <.literal16+0x77d>
-  .byte  224,255                             // loopne        d7bf <.literal16+0x77f>
+  .byte  224,255                             // loopne        e0d1 <.literal16+0x771>
+  .byte  224,255                             // loopne        e0d3 <.literal16+0x773>
+  .byte  224,255                             // loopne        e0d5 <.literal16+0x775>
+  .byte  224,255                             // loopne        e0d7 <.literal16+0x777>
+  .byte  224,255                             // loopne        e0d9 <.literal16+0x779>
+  .byte  224,255                             // loopne        e0db <.literal16+0x77b>
+  .byte  224,255                             // loopne        e0dd <.literal16+0x77d>
+  .byte  224,255                             // loopne        e0df <.literal16+0x77f>
   .byte  54,0,54                             // add           %dh,%ss:(%esi)
   .byte  0,54                                // add           %dh,(%esi)
   .byte  0,54                                // add           %dh,(%esi)
@@ -78619,5 +80729,5 @@
   .byte  0,0                                 // add           %al,(%eax)
   .byte  0,63                                // add           %bh,(%edi)
   .byte  0,0                                 // add           %al,(%eax)
-  .byte  127,67                              // jg            d83b <_sk_srcover_rgba_8888_sse2_lowp+0xac7>
+  .byte  127,67                              // jg            e15b <_sk_srcover_rgba_8888_sse2_lowp+0xac7>
 #endif
diff --git a/src/jumper/SkJumper_generated_win.S b/src/jumper/SkJumper_generated_win.S
index f9edf53..baf2d95 100644
--- a/src/jumper/SkJumper_generated_win.S
+++ b/src/jumper/SkJumper_generated_win.S
@@ -120,10 +120,10 @@
   DB  197,249,110,201                     ; vmovd         %ecx,%xmm1
   DB  196,226,125,88,201                  ; vpbroadcastd  %xmm1,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,3,187,2,0         ; vbroadcastss  0x2bb03(%rip),%ymm2        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,21,175,209,2,0       ; vbroadcastss  0x2d1af(%rip),%ymm2        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,21,248,186,2,0       ; vbroadcastss  0x2baf8(%rip),%ymm2        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,21,164,209,2,0       ; vbroadcastss  0x2d1a4(%rip),%ymm2        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
   DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
@@ -136,17 +136,17 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,121,110,194                     ; vmovd         %edx,%xmm8
   DB  196,66,125,88,192                   ; vpbroadcastd  %xmm8,%ymm8
-  DB  197,61,254,5,119,188,2,0            ; vpaddd        0x2bc77(%rip),%ymm8,%ymm8        # 2be60 <_sk_srcover_rgba_8888_sse2_lowp+0x458>
+  DB  197,61,254,5,23,211,2,0             ; vpaddd        0x2d317(%rip),%ymm8,%ymm8        # 2d500 <_sk_srcover_rgba_8888_sse2_lowp+0x44c>
   DB  197,121,110,201                     ; vmovd         %ecx,%xmm9
   DB  196,66,125,88,201                   ; vpbroadcastd  %xmm9,%ymm9
   DB  196,65,53,239,200                   ; vpxor         %ymm8,%ymm9,%ymm9
-  DB  196,98,125,88,21,188,186,2,0        ; vpbroadcastd  0x2babc(%rip),%ymm10        # 2bcbc <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  DB  196,98,125,88,21,104,209,2,0        ; vpbroadcastd  0x2d168(%rip),%ymm10        # 2d368 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   DB  196,65,53,219,218                   ; vpand         %ymm10,%ymm9,%ymm11
   DB  196,193,37,114,243,5                ; vpslld        $0x5,%ymm11,%ymm11
   DB  196,65,61,219,210                   ; vpand         %ymm10,%ymm8,%ymm10
   DB  196,193,45,114,242,4                ; vpslld        $0x4,%ymm10,%ymm10
-  DB  196,98,125,88,37,161,186,2,0        ; vpbroadcastd  0x2baa1(%rip),%ymm12        # 2bcc0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
-  DB  196,98,125,88,45,156,186,2,0        ; vpbroadcastd  0x2ba9c(%rip),%ymm13        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,98,125,88,37,77,209,2,0         ; vpbroadcastd  0x2d14d(%rip),%ymm12        # 2d36c <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
+  DB  196,98,125,88,45,72,209,2,0         ; vpbroadcastd  0x2d148(%rip),%ymm13        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  196,65,53,219,245                   ; vpand         %ymm13,%ymm9,%ymm14
   DB  196,193,13,114,246,2                ; vpslld        $0x2,%ymm14,%ymm14
   DB  196,65,37,235,222                   ; vpor          %ymm14,%ymm11,%ymm11
@@ -161,8 +161,8 @@
   DB  196,65,61,235,195                   ; vpor          %ymm11,%ymm8,%ymm8
   DB  196,65,61,235,193                   ; vpor          %ymm9,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,78,186,2,0         ; vbroadcastss  0x2ba4e(%rip),%ymm9        # 2bcc8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
-  DB  196,98,125,24,21,73,186,2,0         ; vbroadcastss  0x2ba49(%rip),%ymm10        # 2bccc <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
+  DB  196,98,125,24,13,250,208,2,0        ; vbroadcastss  0x2d0fa(%rip),%ymm9        # 2d374 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
+  DB  196,98,125,24,21,245,208,2,0        ; vbroadcastss  0x2d0f5(%rip),%ymm10        # 2d378 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
   DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
   DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
   DB  196,65,44,89,192                    ; vmulps        %ymm8,%ymm10,%ymm8
@@ -192,7 +192,7 @@
 PUBLIC _sk_black_color_hsw
 _sk_black_color_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,209,185,2,0       ; vbroadcastss  0x2b9d1(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,125,208,2,0       ; vbroadcastss  0x2d07d(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
@@ -201,7 +201,7 @@
 PUBLIC _sk_white_color_hsw
 _sk_white_color_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,184,185,2,0        ; vbroadcastss  0x2b9b8(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,100,208,2,0        ; vbroadcastss  0x2d064(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  197,252,40,216                      ; vmovaps       %ymm0,%ymm3
@@ -239,7 +239,7 @@
 PUBLIC _sk_srcatop_hsw
 _sk_srcatop_hsw LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,87,185,2,0          ; vbroadcastss  0x2b957(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,3,208,2,0           ; vbroadcastss  0x2d003(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,226,61,184,196                  ; vfmadd231ps   %ymm4,%ymm8,%ymm0
   DB  197,244,89,207                      ; vmulps        %ymm7,%ymm1,%ymm1
@@ -253,7 +253,7 @@
 
 PUBLIC _sk_dstatop_hsw
 _sk_dstatop_hsw LABEL PROC
-  DB  196,98,125,24,5,38,185,2,0          ; vbroadcastss  0x2b926(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,210,207,2,0         ; vbroadcastss  0x2cfd2(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  196,226,101,184,196                 ; vfmadd231ps   %ymm4,%ymm3,%ymm0
@@ -286,7 +286,7 @@
 
 PUBLIC _sk_srcout_hsw
 _sk_srcout_hsw LABEL PROC
-  DB  196,98,125,24,5,201,184,2,0         ; vbroadcastss  0x2b8c9(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,117,207,2,0         ; vbroadcastss  0x2cf75(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
@@ -297,7 +297,7 @@
 
 PUBLIC _sk_dstout_hsw
 _sk_dstout_hsw LABEL PROC
-  DB  196,226,125,24,5,168,184,2,0        ; vbroadcastss  0x2b8a8(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,84,207,2,0         ; vbroadcastss  0x2cf54(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,92,219                      ; vsubps        %ymm3,%ymm0,%ymm3
   DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
   DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
@@ -308,7 +308,7 @@
 
 PUBLIC _sk_srcover_hsw
 _sk_srcover_hsw LABEL PROC
-  DB  196,98,125,24,5,135,184,2,0         ; vbroadcastss  0x2b887(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,51,207,2,0          ; vbroadcastss  0x2cf33(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,93,184,192                  ; vfmadd231ps   %ymm8,%ymm4,%ymm0
   DB  196,194,85,184,200                  ; vfmadd231ps   %ymm8,%ymm5,%ymm1
@@ -319,7 +319,7 @@
 
 PUBLIC _sk_dstover_hsw
 _sk_dstover_hsw LABEL PROC
-  DB  196,98,125,24,5,98,184,2,0          ; vbroadcastss  0x2b862(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,14,207,2,0          ; vbroadcastss  0x2cf0e(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
   DB  196,226,61,168,196                  ; vfmadd213ps   %ymm4,%ymm8,%ymm0
   DB  196,226,61,168,205                  ; vfmadd213ps   %ymm5,%ymm8,%ymm1
@@ -339,7 +339,7 @@
 
 PUBLIC _sk_multiply_hsw
 _sk_multiply_hsw LABEL PROC
-  DB  196,98,125,24,5,41,184,2,0          ; vbroadcastss  0x2b829(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,213,206,2,0         ; vbroadcastss  0x2ced5(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,52,89,208                       ; vmulps        %ymm0,%ymm9,%ymm10
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -360,7 +360,7 @@
 PUBLIC _sk_plus__hsw
 _sk_plus__hsw LABEL PROC
   DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
-  DB  196,98,125,24,5,216,183,2,0         ; vbroadcastss  0x2b7d8(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,132,206,2,0         ; vbroadcastss  0x2ce84(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
@@ -386,7 +386,7 @@
 
 PUBLIC _sk_xor__hsw
 _sk_xor__hsw LABEL PROC
-  DB  196,98,125,24,5,131,183,2,0         ; vbroadcastss  0x2b783(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,47,206,2,0          ; vbroadcastss  0x2ce2f(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -418,7 +418,7 @@
   DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
   DB  196,193,108,95,209                  ; vmaxps        %ymm9,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,7,183,2,0           ; vbroadcastss  0x2b707(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,179,205,2,0         ; vbroadcastss  0x2cdb3(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -441,7 +441,7 @@
   DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
   DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,178,182,2,0         ; vbroadcastss  0x2b6b2(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,94,205,2,0          ; vbroadcastss  0x2cd5e(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -467,7 +467,7 @@
   DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
   DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,81,182,2,0          ; vbroadcastss  0x2b651(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,253,204,2,0         ; vbroadcastss  0x2ccfd(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -487,7 +487,7 @@
   DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
   DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,11,182,2,0          ; vbroadcastss  0x2b60b(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,183,204,2,0         ; vbroadcastss  0x2ccb7(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  196,194,69,184,216                  ; vfmadd231ps   %ymm8,%ymm7,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -495,7 +495,7 @@
 
 PUBLIC _sk_colorburn_hsw
 _sk_colorburn_hsw LABEL PROC
-  DB  196,98,125,24,5,245,181,2,0         ; vbroadcastss  0x2b5f5(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,161,204,2,0         ; vbroadcastss  0x2cca1(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,52,89,216                       ; vmulps        %ymm0,%ymm9,%ymm11
   DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
@@ -551,7 +551,7 @@
 PUBLIC _sk_colordodge_hsw
 _sk_colordodge_hsw LABEL PROC
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
-  DB  196,98,125,24,13,5,181,2,0          ; vbroadcastss  0x2b505(%rip),%ymm9        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,13,177,203,2,0        ; vbroadcastss  0x2cbb1(%rip),%ymm9        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,52,92,215                       ; vsubps        %ymm7,%ymm9,%ymm10
   DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
   DB  197,52,92,203                       ; vsubps        %ymm3,%ymm9,%ymm9
@@ -602,7 +602,7 @@
 
 PUBLIC _sk_hardlight_hsw
 _sk_hardlight_hsw LABEL PROC
-  DB  196,98,125,24,5,31,180,2,0          ; vbroadcastss  0x2b41f(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,203,202,2,0         ; vbroadcastss  0x2cacb(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
   DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -651,7 +651,7 @@
 
 PUBLIC _sk_overlay_hsw
 _sk_overlay_hsw LABEL PROC
-  DB  196,98,125,24,5,83,179,2,0          ; vbroadcastss  0x2b353(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,255,201,2,0         ; vbroadcastss  0x2c9ff(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
   DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -711,10 +711,10 @@
   DB  196,65,20,88,197                    ; vaddps        %ymm13,%ymm13,%ymm8
   DB  196,65,60,88,192                    ; vaddps        %ymm8,%ymm8,%ymm8
   DB  196,66,61,168,192                   ; vfmadd213ps   %ymm8,%ymm8,%ymm8
-  DB  196,98,125,24,29,106,178,2,0        ; vbroadcastss  0x2b26a(%rip),%ymm11        # 2bcd0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  DB  196,98,125,24,29,22,201,2,0         ; vbroadcastss  0x2c916(%rip),%ymm11        # 2d37c <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   DB  196,65,20,88,227                    ; vaddps        %ymm11,%ymm13,%ymm12
   DB  196,65,28,89,192                    ; vmulps        %ymm8,%ymm12,%ymm8
-  DB  196,98,125,24,37,91,178,2,0         ; vbroadcastss  0x2b25b(%rip),%ymm12        # 2bcd4 <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
+  DB  196,98,125,24,37,7,201,2,0          ; vbroadcastss  0x2c907(%rip),%ymm12        # 2d380 <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
   DB  196,66,21,184,196                   ; vfmadd231ps   %ymm12,%ymm13,%ymm8
   DB  196,65,124,82,245                   ; vrsqrtps      %ymm13,%ymm14
   DB  196,65,124,83,246                   ; vrcpps        %ymm14,%ymm14
@@ -724,7 +724,7 @@
   DB  197,4,194,255,2                     ; vcmpleps      %ymm7,%ymm15,%ymm15
   DB  196,67,13,74,240,240                ; vblendvps     %ymm15,%ymm8,%ymm14,%ymm14
   DB  197,116,88,249                      ; vaddps        %ymm1,%ymm1,%ymm15
-  DB  196,98,125,24,5,10,178,2,0          ; vbroadcastss  0x2b20a(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,182,200,2,0         ; vbroadcastss  0x2c8b6(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,65,60,92,237                    ; vsubps        %ymm13,%ymm8,%ymm13
   DB  197,132,92,195                      ; vsubps        %ymm3,%ymm15,%ymm0
   DB  196,98,125,168,235                  ; vfmadd213ps   %ymm3,%ymm0,%ymm13
@@ -837,11 +837,11 @@
   DB  196,65,28,89,210                    ; vmulps        %ymm10,%ymm12,%ymm10
   DB  196,65,44,94,214                    ; vdivps        %ymm14,%ymm10,%ymm10
   DB  196,67,45,74,224,240                ; vblendvps     %ymm15,%ymm8,%ymm10,%ymm12
-  DB  196,98,125,24,53,29,176,2,0         ; vbroadcastss  0x2b01d(%rip),%ymm14        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  DB  196,98,125,24,61,24,176,2,0         ; vbroadcastss  0x2b018(%rip),%ymm15        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,53,201,198,2,0        ; vbroadcastss  0x2c6c9(%rip),%ymm14        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,61,196,198,2,0        ; vbroadcastss  0x2c6c4(%rip),%ymm15        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,66,93,184,238                   ; vfmadd231ps   %ymm14,%ymm4,%ymm13
-  DB  196,226,125,24,5,9,176,2,0          ; vbroadcastss  0x2b009(%rip),%ymm0        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,226,125,24,5,181,198,2,0        ; vbroadcastss  0x2c6b5(%rip),%ymm0        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,98,77,184,232                   ; vfmadd231ps   %ymm0,%ymm6,%ymm13
   DB  196,65,116,89,215                   ; vmulps        %ymm15,%ymm1,%ymm10
   DB  196,66,53,184,214                   ; vfmadd231ps   %ymm14,%ymm9,%ymm10
@@ -896,7 +896,7 @@
   DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
   DB  196,65,36,95,200                    ; vmaxps        %ymm8,%ymm11,%ymm9
   DB  196,65,116,95,192                   ; vmaxps        %ymm8,%ymm1,%ymm8
-  DB  196,226,125,24,13,202,174,2,0       ; vbroadcastss  0x2aeca(%rip),%ymm1        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,13,118,197,2,0       ; vbroadcastss  0x2c576(%rip),%ymm1        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,116,92,215                      ; vsubps        %ymm7,%ymm1,%ymm10
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  197,116,92,219                      ; vsubps        %ymm3,%ymm1,%ymm11
@@ -950,11 +950,11 @@
   DB  196,65,28,89,210                    ; vmulps        %ymm10,%ymm12,%ymm10
   DB  196,65,44,94,214                    ; vdivps        %ymm14,%ymm10,%ymm10
   DB  196,67,45,74,224,240                ; vblendvps     %ymm15,%ymm8,%ymm10,%ymm12
-  DB  196,98,125,24,53,247,173,2,0        ; vbroadcastss  0x2adf7(%rip),%ymm14        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  DB  196,98,125,24,61,242,173,2,0        ; vbroadcastss  0x2adf2(%rip),%ymm15        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,53,163,196,2,0        ; vbroadcastss  0x2c4a3(%rip),%ymm14        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,61,158,196,2,0        ; vbroadcastss  0x2c49e(%rip),%ymm15        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,66,93,184,238                   ; vfmadd231ps   %ymm14,%ymm4,%ymm13
-  DB  196,226,125,24,5,227,173,2,0        ; vbroadcastss  0x2ade3(%rip),%ymm0        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,226,125,24,5,143,196,2,0        ; vbroadcastss  0x2c48f(%rip),%ymm0        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,98,77,184,232                   ; vfmadd231ps   %ymm0,%ymm6,%ymm13
   DB  196,65,116,89,215                   ; vmulps        %ymm15,%ymm1,%ymm10
   DB  196,66,53,184,214                   ; vfmadd231ps   %ymm14,%ymm9,%ymm10
@@ -1009,7 +1009,7 @@
   DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
   DB  196,65,36,95,200                    ; vmaxps        %ymm8,%ymm11,%ymm9
   DB  196,65,116,95,192                   ; vmaxps        %ymm8,%ymm1,%ymm8
-  DB  196,226,125,24,13,164,172,2,0       ; vbroadcastss  0x2aca4(%rip),%ymm1        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,13,80,195,2,0        ; vbroadcastss  0x2c350(%rip),%ymm1        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,116,92,215                      ; vsubps        %ymm7,%ymm1,%ymm10
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  197,116,92,219                      ; vsubps        %ymm3,%ymm1,%ymm11
@@ -1037,11 +1037,11 @@
   DB  197,108,89,199                      ; vmulps        %ymm7,%ymm2,%ymm8
   DB  197,116,89,215                      ; vmulps        %ymm7,%ymm1,%ymm10
   DB  197,52,89,223                       ; vmulps        %ymm7,%ymm9,%ymm11
-  DB  196,98,125,24,45,83,172,2,0         ; vbroadcastss  0x2ac53(%rip),%ymm13        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  DB  196,98,125,24,53,78,172,2,0         ; vbroadcastss  0x2ac4e(%rip),%ymm14        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,45,255,194,2,0        ; vbroadcastss  0x2c2ff(%rip),%ymm13        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,53,250,194,2,0        ; vbroadcastss  0x2c2fa(%rip),%ymm14        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,230                    ; vmulps        %ymm14,%ymm5,%ymm12
   DB  196,66,93,184,229                   ; vfmadd231ps   %ymm13,%ymm4,%ymm12
-  DB  196,98,125,24,61,63,172,2,0         ; vbroadcastss  0x2ac3f(%rip),%ymm15        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,98,125,24,61,235,194,2,0        ; vbroadcastss  0x2c2eb(%rip),%ymm15        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,66,77,184,231                   ; vfmadd231ps   %ymm15,%ymm6,%ymm12
   DB  196,65,44,89,206                    ; vmulps        %ymm14,%ymm10,%ymm9
   DB  196,66,61,184,205                   ; vfmadd231ps   %ymm13,%ymm8,%ymm9
@@ -1097,7 +1097,7 @@
   DB  196,193,116,95,206                  ; vmaxps        %ymm14,%ymm1,%ymm1
   DB  196,65,44,95,198                    ; vmaxps        %ymm14,%ymm10,%ymm8
   DB  196,65,124,95,206                   ; vmaxps        %ymm14,%ymm0,%ymm9
-  DB  196,226,125,24,5,245,170,2,0        ; vbroadcastss  0x2aaf5(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,161,193,2,0        ; vbroadcastss  0x2c1a1(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,124,92,215                      ; vsubps        %ymm7,%ymm0,%ymm10
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  197,124,92,219                      ; vsubps        %ymm3,%ymm0,%ymm11
@@ -1125,11 +1125,11 @@
   DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
   DB  197,100,89,213                      ; vmulps        %ymm5,%ymm3,%ymm10
   DB  197,100,89,222                      ; vmulps        %ymm6,%ymm3,%ymm11
-  DB  196,98,125,24,45,164,170,2,0        ; vbroadcastss  0x2aaa4(%rip),%ymm13        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
-  DB  196,98,125,24,53,159,170,2,0        ; vbroadcastss  0x2aa9f(%rip),%ymm14        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,45,80,193,2,0         ; vbroadcastss  0x2c150(%rip),%ymm13        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,53,75,193,2,0         ; vbroadcastss  0x2c14b(%rip),%ymm14        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,116,89,230                   ; vmulps        %ymm14,%ymm1,%ymm12
   DB  196,66,109,184,229                  ; vfmadd231ps   %ymm13,%ymm2,%ymm12
-  DB  196,98,125,24,61,144,170,2,0        ; vbroadcastss  0x2aa90(%rip),%ymm15        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,98,125,24,61,60,193,2,0         ; vbroadcastss  0x2c13c(%rip),%ymm15        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,66,53,184,231                   ; vfmadd231ps   %ymm15,%ymm9,%ymm12
   DB  196,65,44,89,206                    ; vmulps        %ymm14,%ymm10,%ymm9
   DB  196,66,61,184,205                   ; vfmadd231ps   %ymm13,%ymm8,%ymm9
@@ -1185,7 +1185,7 @@
   DB  196,193,116,95,206                  ; vmaxps        %ymm14,%ymm1,%ymm1
   DB  196,65,44,95,198                    ; vmaxps        %ymm14,%ymm10,%ymm8
   DB  196,65,124,95,206                   ; vmaxps        %ymm14,%ymm0,%ymm9
-  DB  196,226,125,24,5,70,169,2,0         ; vbroadcastss  0x2a946(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,242,191,2,0        ; vbroadcastss  0x2bff2(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,124,92,215                      ; vsubps        %ymm7,%ymm0,%ymm10
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  197,124,92,219                      ; vsubps        %ymm3,%ymm0,%ymm11
@@ -1215,17 +1215,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,179,0,0,0                    ; jne           148a <_sk_srcover_rgba_8888_hsw+0xd3>
   DB  196,129,126,111,60,152              ; vmovdqu       (%r8,%r11,4),%ymm7
-  DB  197,197,219,37,155,170,2,0          ; vpand         0x2aa9b(%rip),%ymm7,%ymm4        # 2be80 <_sk_srcover_rgba_8888_sse2_lowp+0x478>
+  DB  197,197,219,37,59,193,2,0           ; vpand         0x2c13b(%rip),%ymm7,%ymm4        # 2d520 <_sk_srcover_rgba_8888_sse2_lowp+0x46c>
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,69,0,45,174,170,2,0         ; vpshufb       0x2aaae(%rip),%ymm7,%ymm5        # 2bea0 <_sk_srcover_rgba_8888_sse2_lowp+0x498>
+  DB  196,226,69,0,45,78,193,2,0          ; vpshufb       0x2c14e(%rip),%ymm7,%ymm5        # 2d540 <_sk_srcover_rgba_8888_sse2_lowp+0x48c>
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,226,69,0,53,193,170,2,0         ; vpshufb       0x2aac1(%rip),%ymm7,%ymm6        # 2bec0 <_sk_srcover_rgba_8888_sse2_lowp+0x4b8>
+  DB  196,226,69,0,53,97,193,2,0          ; vpshufb       0x2c161(%rip),%ymm7,%ymm6        # 2d560 <_sk_srcover_rgba_8888_sse2_lowp+0x4ac>
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
   DB  197,197,114,215,24                  ; vpsrld        $0x18,%ymm7,%ymm7
   DB  197,252,91,255                      ; vcvtdq2ps     %ymm7,%ymm7
-  DB  196,98,125,24,5,163,168,2,0         ; vbroadcastss  0x2a8a3(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,79,191,2,0          ; vbroadcastss  0x2bf4f(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
-  DB  196,98,125,24,13,194,168,2,0        ; vbroadcastss  0x2a8c2(%rip),%ymm9        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,13,110,191,2,0        ; vbroadcastss  0x2bf6e(%rip),%ymm9        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  196,194,93,184,192                  ; vfmadd231ps   %ymm8,%ymm4,%ymm0
   DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
@@ -1320,7 +1320,7 @@
   DB  255                                 ; (bad)
   DB  255,104,255                         ; ljmp          *-0x1(%rax)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,163              ; lcall         *-0x5c000001(%rip)        # ffffffffa40015cc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffa3fd5bc4>
+  DB  255,29,255,255,255,163              ; lcall         *-0x5c000001(%rip)        # ffffffffa40015cc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffa3fd4518>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,155,255,255,255,215             ; lcall         *-0x28000001(%rbx)
@@ -1347,7 +1347,7 @@
 
 PUBLIC _sk_clamp_1_hsw
 _sk_clamp_1_hsw LABEL PROC
-  DB  196,98,125,24,5,174,166,2,0         ; vbroadcastss  0x2a6ae(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,90,189,2,0          ; vbroadcastss  0x2bd5a(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
@@ -1357,7 +1357,7 @@
 
 PUBLIC _sk_clamp_a_hsw
 _sk_clamp_a_hsw LABEL PROC
-  DB  196,98,125,24,5,141,166,2,0         ; vbroadcastss  0x2a68d(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,57,189,2,0          ; vbroadcastss  0x2bd39(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
   DB  197,252,93,195                      ; vminps        %ymm3,%ymm0,%ymm0
   DB  197,244,93,203                      ; vminps        %ymm3,%ymm1,%ymm1
@@ -1367,7 +1367,7 @@
 
 PUBLIC _sk_clamp_a_dst_hsw
 _sk_clamp_a_dst_hsw LABEL PROC
-  DB  196,98,125,24,5,111,166,2,0         ; vbroadcastss  0x2a66f(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,27,189,2,0          ; vbroadcastss  0x2bd1b(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,68,93,248                   ; vminps        %ymm8,%ymm7,%ymm7
   DB  197,220,93,231                      ; vminps        %ymm7,%ymm4,%ymm4
   DB  197,212,93,239                      ; vminps        %ymm7,%ymm5,%ymm5
@@ -1394,7 +1394,7 @@
 
 PUBLIC _sk_invert_hsw
 _sk_invert_hsw LABEL PROC
-  DB  196,98,125,24,5,42,166,2,0          ; vbroadcastss  0x2a62a(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,214,188,2,0         ; vbroadcastss  0x2bcd6(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
   DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
@@ -1438,9 +1438,9 @@
 
 PUBLIC _sk_unpremul_hsw
 _sk_unpremul_hsw LABEL PROC
-  DB  196,98,125,24,5,197,165,2,0         ; vbroadcastss  0x2a5c5(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,113,188,2,0         ; vbroadcastss  0x2bc71(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,94,195                       ; vdivps        %ymm3,%ymm8,%ymm8
-  DB  196,98,125,24,13,232,165,2,0        ; vbroadcastss  0x2a5e8(%rip),%ymm9        # 2bce8 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
+  DB  196,98,125,24,13,148,188,2,0        ; vbroadcastss  0x2bc94(%rip),%ymm9        # 2d394 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
   DB  196,65,60,194,201,1                 ; vcmpltps      %ymm9,%ymm8,%ymm9
   DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
   DB  196,67,45,74,192,144                ; vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
@@ -1452,16 +1452,16 @@
 
 PUBLIC _sk_from_srgb_hsw
 _sk_from_srgb_hsw LABEL PROC
-  DB  196,98,125,24,5,194,165,2,0         ; vbroadcastss  0x2a5c2(%rip),%ymm8        # 2bcec <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  DB  196,98,125,24,5,110,188,2,0         ; vbroadcastss  0x2bc6e(%rip),%ymm8        # 2d398 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,180,165,2,0        ; vbroadcastss  0x2a5b4(%rip),%ymm11        # 2bcf0 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  DB  196,98,125,24,37,147,165,2,0        ; vbroadcastss  0x2a593(%rip),%ymm12        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,29,96,188,2,0         ; vbroadcastss  0x2bc60(%rip),%ymm11        # 2d39c <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  DB  196,98,125,24,37,63,188,2,0         ; vbroadcastss  0x2bc3f(%rip),%ymm12        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,125,168,235                  ; vfmadd213ps   %ymm11,%ymm0,%ymm13
-  DB  196,98,125,24,53,156,165,2,0        ; vbroadcastss  0x2a59c(%rip),%ymm14        # 2bcf4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  DB  196,98,125,24,53,72,188,2,0         ; vbroadcastss  0x2bc48(%rip),%ymm14        # 2d3a0 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   DB  196,66,45,168,238                   ; vfmadd213ps   %ymm14,%ymm10,%ymm13
-  DB  196,98,125,24,21,146,165,2,0        ; vbroadcastss  0x2a592(%rip),%ymm10        # 2bcf8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  DB  196,98,125,24,21,62,188,2,0         ; vbroadcastss  0x2bc3e(%rip),%ymm10        # 2d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   DB  196,193,124,194,194,1               ; vcmpltps      %ymm10,%ymm0,%ymm0
   DB  196,195,21,74,193,0                 ; vblendvps     %ymm0,%ymm9,%ymm13,%ymm0
   DB  196,65,116,89,200                   ; vmulps        %ymm8,%ymm1,%ymm9
@@ -1482,16 +1482,16 @@
 
 PUBLIC _sk_from_srgb_dst_hsw
 _sk_from_srgb_dst_hsw LABEL PROC
-  DB  196,98,125,24,5,42,165,2,0          ; vbroadcastss  0x2a52a(%rip),%ymm8        # 2bcec <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  DB  196,98,125,24,5,214,187,2,0         ; vbroadcastss  0x2bbd6(%rip),%ymm8        # 2d398 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   DB  196,65,92,89,200                    ; vmulps        %ymm8,%ymm4,%ymm9
   DB  197,92,89,212                       ; vmulps        %ymm4,%ymm4,%ymm10
-  DB  196,98,125,24,29,28,165,2,0         ; vbroadcastss  0x2a51c(%rip),%ymm11        # 2bcf0 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
-  DB  196,98,125,24,37,251,164,2,0        ; vbroadcastss  0x2a4fb(%rip),%ymm12        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,29,200,187,2,0        ; vbroadcastss  0x2bbc8(%rip),%ymm11        # 2d39c <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  DB  196,98,125,24,37,167,187,2,0        ; vbroadcastss  0x2bba7(%rip),%ymm12        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,93,168,235                   ; vfmadd213ps   %ymm11,%ymm4,%ymm13
-  DB  196,98,125,24,53,4,165,2,0          ; vbroadcastss  0x2a504(%rip),%ymm14        # 2bcf4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  DB  196,98,125,24,53,176,187,2,0        ; vbroadcastss  0x2bbb0(%rip),%ymm14        # 2d3a0 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   DB  196,66,45,168,238                   ; vfmadd213ps   %ymm14,%ymm10,%ymm13
-  DB  196,98,125,24,21,250,164,2,0        ; vbroadcastss  0x2a4fa(%rip),%ymm10        # 2bcf8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  DB  196,98,125,24,21,166,187,2,0        ; vbroadcastss  0x2bba6(%rip),%ymm10        # 2d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   DB  196,193,92,194,226,1                ; vcmpltps      %ymm10,%ymm4,%ymm4
   DB  196,195,21,74,225,64                ; vblendvps     %ymm4,%ymm9,%ymm13,%ymm4
   DB  196,65,84,89,200                    ; vmulps        %ymm8,%ymm5,%ymm9
@@ -1513,19 +1513,19 @@
 PUBLIC _sk_to_srgb_hsw
 _sk_to_srgb_hsw LABEL PROC
   DB  197,124,82,200                      ; vrsqrtps      %ymm0,%ymm9
-  DB  196,98,125,24,5,158,164,2,0         ; vbroadcastss  0x2a49e(%rip),%ymm8        # 2bcfc <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
+  DB  196,98,125,24,5,74,187,2,0          ; vbroadcastss  0x2bb4a(%rip),%ymm8        # 2d3a8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
   DB  196,65,124,89,208                   ; vmulps        %ymm8,%ymm0,%ymm10
-  DB  196,98,125,24,29,148,164,2,0        ; vbroadcastss  0x2a494(%rip),%ymm11        # 2bd00 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
-  DB  196,98,125,24,37,143,164,2,0        ; vbroadcastss  0x2a48f(%rip),%ymm12        # 2bd04 <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
+  DB  196,98,125,24,29,64,187,2,0         ; vbroadcastss  0x2bb40(%rip),%ymm11        # 2d3ac <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
+  DB  196,98,125,24,37,59,187,2,0         ; vbroadcastss  0x2bb3b(%rip),%ymm12        # 2d3b0 <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
   DB  196,65,124,40,236                   ; vmovaps       %ymm12,%ymm13
   DB  196,66,53,168,235                   ; vfmadd213ps   %ymm11,%ymm9,%ymm13
-  DB  196,98,125,24,53,128,164,2,0        ; vbroadcastss  0x2a480(%rip),%ymm14        # 2bd08 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
+  DB  196,98,125,24,53,44,187,2,0         ; vbroadcastss  0x2bb2c(%rip),%ymm14        # 2d3b4 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
   DB  196,66,53,168,238                   ; vfmadd213ps   %ymm14,%ymm9,%ymm13
-  DB  196,98,125,24,61,118,164,2,0        ; vbroadcastss  0x2a476(%rip),%ymm15        # 2bd0c <_sk_srcover_rgba_8888_sse2_lowp+0x304>
+  DB  196,98,125,24,61,34,187,2,0         ; vbroadcastss  0x2bb22(%rip),%ymm15        # 2d3b8 <_sk_srcover_rgba_8888_sse2_lowp+0x304>
   DB  196,65,52,88,207                    ; vaddps        %ymm15,%ymm9,%ymm9
   DB  196,65,124,83,201                   ; vrcpps        %ymm9,%ymm9
   DB  196,65,20,89,201                    ; vmulps        %ymm9,%ymm13,%ymm9
-  DB  196,98,125,24,45,98,164,2,0         ; vbroadcastss  0x2a462(%rip),%ymm13        # 2bd10 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
+  DB  196,98,125,24,45,14,187,2,0         ; vbroadcastss  0x2bb0e(%rip),%ymm13        # 2d3bc <_sk_srcover_rgba_8888_sse2_lowp+0x308>
   DB  196,193,124,194,197,1               ; vcmpltps      %ymm13,%ymm0,%ymm0
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -1557,26 +1557,26 @@
   DB  197,116,93,202                      ; vminps        %ymm2,%ymm1,%ymm9
   DB  196,65,124,93,201                   ; vminps        %ymm9,%ymm0,%ymm9
   DB  196,65,60,92,209                    ; vsubps        %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,29,121,163,2,0        ; vbroadcastss  0x2a379(%rip),%ymm11        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,29,37,186,2,0         ; vbroadcastss  0x2ba25(%rip),%ymm11        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,65,36,94,218                    ; vdivps        %ymm10,%ymm11,%ymm11
   DB  197,116,92,226                      ; vsubps        %ymm2,%ymm1,%ymm12
   DB  197,116,194,234,1                   ; vcmpltps      %ymm2,%ymm1,%ymm13
-  DB  196,98,125,24,53,190,163,2,0        ; vbroadcastss  0x2a3be(%rip),%ymm14        # 2bd14 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  DB  196,98,125,24,53,106,186,2,0        ; vbroadcastss  0x2ba6a(%rip),%ymm14        # 2d3c0 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   DB  196,65,4,87,255                     ; vxorps        %ymm15,%ymm15,%ymm15
   DB  196,67,5,74,238,208                 ; vblendvps     %ymm13,%ymm14,%ymm15,%ymm13
   DB  196,66,37,168,229                   ; vfmadd213ps   %ymm13,%ymm11,%ymm12
   DB  197,236,92,208                      ; vsubps        %ymm0,%ymm2,%ymm2
   DB  197,124,92,233                      ; vsubps        %ymm1,%ymm0,%ymm13
-  DB  196,98,125,24,53,165,163,2,0        ; vbroadcastss  0x2a3a5(%rip),%ymm14        # 2bd1c <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  DB  196,98,125,24,53,81,186,2,0         ; vbroadcastss  0x2ba51(%rip),%ymm14        # 2d3c8 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   DB  196,66,37,168,238                   ; vfmadd213ps   %ymm14,%ymm11,%ymm13
-  DB  196,98,125,24,53,147,163,2,0        ; vbroadcastss  0x2a393(%rip),%ymm14        # 2bd18 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  DB  196,98,125,24,53,63,186,2,0         ; vbroadcastss  0x2ba3f(%rip),%ymm14        # 2d3c4 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   DB  196,194,37,168,214                  ; vfmadd213ps   %ymm14,%ymm11,%ymm2
   DB  197,188,194,201,0                   ; vcmpeqps      %ymm1,%ymm8,%ymm1
   DB  196,227,21,74,202,16                ; vblendvps     %ymm1,%ymm2,%ymm13,%ymm1
   DB  197,188,194,192,0                   ; vcmpeqps      %ymm0,%ymm8,%ymm0
   DB  196,195,117,74,196,0                ; vblendvps     %ymm0,%ymm12,%ymm1,%ymm0
   DB  196,193,60,88,201                   ; vaddps        %ymm9,%ymm8,%ymm1
-  DB  196,98,125,24,29,6,163,2,0          ; vbroadcastss  0x2a306(%rip),%ymm11        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,29,178,185,2,0        ; vbroadcastss  0x2b9b2(%rip),%ymm11        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,116,89,211                  ; vmulps        %ymm11,%ymm1,%ymm2
   DB  197,36,194,218,1                    ; vcmpltps      %ymm2,%ymm11,%ymm11
   DB  196,65,12,92,224                    ; vsubps        %ymm8,%ymm14,%ymm12
@@ -1586,7 +1586,7 @@
   DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
   DB  196,195,125,74,199,128              ; vblendvps     %ymm8,%ymm15,%ymm0,%ymm0
   DB  196,195,117,74,207,128              ; vblendvps     %ymm8,%ymm15,%ymm1,%ymm1
-  DB  196,98,125,24,5,57,163,2,0          ; vbroadcastss  0x2a339(%rip),%ymm8        # 2bd20 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  DB  196,98,125,24,5,229,185,2,0         ; vbroadcastss  0x2b9e5(%rip),%ymm8        # 2d3cc <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -1601,30 +1601,30 @@
   DB  197,252,17,28,36                    ; vmovups       %ymm3,(%rsp)
   DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
   DB  197,252,40,224                      ; vmovaps       %ymm0,%ymm4
-  DB  196,98,125,24,5,140,162,2,0         ; vbroadcastss  0x2a28c(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,56,185,2,0          ; vbroadcastss  0x2b938(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,60,194,202,2                    ; vcmpleps      %ymm2,%ymm8,%ymm9
   DB  197,84,89,210                       ; vmulps        %ymm2,%ymm5,%ymm10
   DB  196,65,84,92,218                    ; vsubps        %ymm10,%ymm5,%ymm11
   DB  196,67,45,74,203,144                ; vblendvps     %ymm9,%ymm11,%ymm10,%ymm9
   DB  197,52,88,210                       ; vaddps        %ymm2,%ymm9,%ymm10
-  DB  196,98,125,24,13,207,162,2,0        ; vbroadcastss  0x2a2cf(%rip),%ymm9        # 2bd18 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  DB  196,98,125,24,13,123,185,2,0        ; vbroadcastss  0x2b97b(%rip),%ymm9        # 2d3c4 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   DB  196,66,109,170,202                  ; vfmsub213ps   %ymm10,%ymm2,%ymm9
-  DB  196,98,125,24,29,205,162,2,0        ; vbroadcastss  0x2a2cd(%rip),%ymm11        # 2bd24 <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
+  DB  196,98,125,24,29,121,185,2,0        ; vbroadcastss  0x2b979(%rip),%ymm11        # 2d3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
   DB  196,65,92,88,219                    ; vaddps        %ymm11,%ymm4,%ymm11
   DB  196,67,125,8,227,1                  ; vroundps      $0x1,%ymm11,%ymm12
   DB  196,65,36,92,252                    ; vsubps        %ymm12,%ymm11,%ymm15
   DB  196,65,44,92,217                    ; vsubps        %ymm9,%ymm10,%ymm11
-  DB  196,98,125,24,45,159,162,2,0        ; vbroadcastss  0x2a29f(%rip),%ymm13        # 2bd14 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  DB  196,98,125,24,45,75,185,2,0         ; vbroadcastss  0x2b94b(%rip),%ymm13        # 2d3c0 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   DB  196,193,4,89,197                    ; vmulps        %ymm13,%ymm15,%ymm0
-  DB  196,98,125,24,53,153,162,2,0        ; vbroadcastss  0x2a299(%rip),%ymm14        # 2bd1c <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  DB  196,98,125,24,53,69,185,2,0         ; vbroadcastss  0x2b945(%rip),%ymm14        # 2d3c8 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   DB  197,12,92,224                       ; vsubps        %ymm0,%ymm14,%ymm12
   DB  196,66,37,168,225                   ; vfmadd213ps   %ymm9,%ymm11,%ymm12
-  DB  196,226,125,24,29,147,162,2,0       ; vbroadcastss  0x2a293(%rip),%ymm3        # 2bd28 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
+  DB  196,226,125,24,29,63,185,2,0        ; vbroadcastss  0x2b93f(%rip),%ymm3        # 2d3d4 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
   DB  196,193,100,194,255,2               ; vcmpleps      %ymm15,%ymm3,%ymm7
   DB  196,195,29,74,249,112               ; vblendvps     %ymm7,%ymm9,%ymm12,%ymm7
   DB  196,65,60,194,231,2                 ; vcmpleps      %ymm15,%ymm8,%ymm12
   DB  196,227,45,74,255,192               ; vblendvps     %ymm12,%ymm7,%ymm10,%ymm7
-  DB  196,98,125,24,37,106,162,2,0        ; vbroadcastss  0x2a26a(%rip),%ymm12        # 2bd20 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  DB  196,98,125,24,37,22,185,2,0         ; vbroadcastss  0x2b916(%rip),%ymm12        # 2d3cc <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   DB  196,65,28,194,255,2                 ; vcmpleps      %ymm15,%ymm12,%ymm15
   DB  196,194,37,168,193                  ; vfmadd213ps   %ymm9,%ymm11,%ymm0
   DB  196,99,125,74,255,240               ; vblendvps     %ymm15,%ymm7,%ymm0,%ymm15
@@ -1640,7 +1640,7 @@
   DB  197,156,194,192,2                   ; vcmpleps      %ymm0,%ymm12,%ymm0
   DB  196,194,37,168,249                  ; vfmadd213ps   %ymm9,%ymm11,%ymm7
   DB  196,227,69,74,201,0                 ; vblendvps     %ymm0,%ymm1,%ymm7,%ymm1
-  DB  196,226,125,24,5,30,162,2,0         ; vbroadcastss  0x2a21e(%rip),%ymm0        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,226,125,24,5,202,184,2,0        ; vbroadcastss  0x2b8ca(%rip),%ymm0        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  197,220,88,192                      ; vaddps        %ymm0,%ymm4,%ymm0
   DB  196,227,125,8,224,1                 ; vroundps      $0x1,%ymm0,%ymm4
   DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
@@ -1690,10 +1690,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,58                              ; jne           1c03 <_sk_scale_u8_hsw+0x52>
   DB  196,2,121,48,4,24                   ; vpmovzxbw     (%r8,%r11,1),%xmm8
-  DB  197,57,219,5,249,167,2,0            ; vpand         0x2a7f9(%rip),%xmm8,%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,57,219,5,217,190,2,0            ; vpand         0x2bed9(%rip),%xmm8,%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,66,125,51,192                   ; vpmovzxwd     %xmm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,70,161,2,0         ; vbroadcastss  0x2a146(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,242,183,2,0        ; vbroadcastss  0x2b7f2(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
@@ -1768,20 +1768,20 @@
   DB  15,133,161,0,0,0                    ; jne           1d78 <_sk_scale_565_hsw+0xc0>
   DB  196,1,122,111,4,88                  ; vmovdqu       (%r8,%r11,2),%xmm8
   DB  196,66,125,51,192                   ; vpmovzxwd     %xmm8,%ymm8
-  DB  196,98,125,88,13,73,160,2,0         ; vpbroadcastd  0x2a049(%rip),%ymm9        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,98,125,88,13,245,182,2,0        ; vpbroadcastd  0x2b6f5(%rip),%ymm9        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  196,65,61,219,201                   ; vpand         %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,58,160,2,0         ; vbroadcastss  0x2a03a(%rip),%ymm10        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,98,125,24,21,230,182,2,0        ; vbroadcastss  0x2b6e6(%rip),%ymm10        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,88,21,48,160,2,0         ; vpbroadcastd  0x2a030(%rip),%ymm10        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,98,125,88,21,220,182,2,0        ; vpbroadcastd  0x2b6dc(%rip),%ymm10        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  196,65,61,219,210                   ; vpand         %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,33,160,2,0         ; vbroadcastss  0x2a021(%rip),%ymm11        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,98,125,24,29,205,182,2,0        ; vbroadcastss  0x2b6cd(%rip),%ymm11        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,88,29,23,160,2,0         ; vpbroadcastd  0x2a017(%rip),%ymm11        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,98,125,88,29,195,182,2,0        ; vpbroadcastd  0x2b6c3(%rip),%ymm11        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  196,65,61,219,195                   ; vpand         %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,8,160,2,0          ; vbroadcastss  0x2a008(%rip),%ymm11        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,98,125,24,29,180,182,2,0        ; vbroadcastss  0x2b6b4(%rip),%ymm11        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  197,100,194,223,1                   ; vcmpltps      %ymm7,%ymm3,%ymm11
   DB  196,65,44,93,224                    ; vminps        %ymm8,%ymm10,%ymm12
@@ -1868,10 +1868,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,78                              ; jne           1eb1 <_sk_lerp_u8_hsw+0x66>
   DB  196,2,121,48,4,24                   ; vpmovzxbw     (%r8,%r11,1),%xmm8
-  DB  197,57,219,5,95,165,2,0             ; vpand         0x2a55f(%rip),%xmm8,%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,57,219,5,63,188,2,0             ; vpand         0x2bc3f(%rip),%xmm8,%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,66,125,51,192                   ; vpmovzxwd     %xmm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,172,158,2,0        ; vbroadcastss  0x29eac(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,88,181,2,0         ; vbroadcastss  0x2b558(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
   DB  196,226,61,168,196                  ; vfmadd213ps   %ymm4,%ymm8,%ymm0
@@ -1923,7 +1923,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,221                 ; callq         ffffffffde001f5c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd6554>
+  DB  232,255,255,255,221                 ; callq         ffffffffde001f5c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd4ea8>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,210                             ; callq         *%rdx
@@ -1947,20 +1947,20 @@
   DB  15,133,181,0,0,0                    ; jne           203c <_sk_lerp_565_hsw+0xd4>
   DB  196,1,122,111,4,88                  ; vmovdqu       (%r8,%r11,2),%xmm8
   DB  196,66,125,51,192                   ; vpmovzxwd     %xmm8,%ymm8
-  DB  196,98,125,88,13,153,157,2,0        ; vpbroadcastd  0x29d99(%rip),%ymm9        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,98,125,88,13,69,180,2,0         ; vpbroadcastd  0x2b445(%rip),%ymm9        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  196,65,61,219,201                   ; vpand         %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,138,157,2,0        ; vbroadcastss  0x29d8a(%rip),%ymm10        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,98,125,24,21,54,180,2,0         ; vbroadcastss  0x2b436(%rip),%ymm10        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,88,21,128,157,2,0        ; vpbroadcastd  0x29d80(%rip),%ymm10        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,98,125,88,21,44,180,2,0         ; vpbroadcastd  0x2b42c(%rip),%ymm10        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  196,65,61,219,210                   ; vpand         %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,113,157,2,0        ; vbroadcastss  0x29d71(%rip),%ymm11        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,98,125,24,29,29,180,2,0         ; vbroadcastss  0x2b41d(%rip),%ymm11        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,88,29,103,157,2,0        ; vpbroadcastd  0x29d67(%rip),%ymm11        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,98,125,88,29,19,180,2,0         ; vpbroadcastd  0x2b413(%rip),%ymm11        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  196,65,61,219,195                   ; vpand         %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,88,157,2,0         ; vbroadcastss  0x29d58(%rip),%ymm11        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,98,125,24,29,4,180,2,0          ; vbroadcastss  0x2b404(%rip),%ymm11        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  197,100,194,223,1                   ; vcmpltps      %ymm7,%ymm3,%ymm11
   DB  196,65,44,93,224                    ; vminps        %ymm8,%ymm10,%ymm12
@@ -2032,7 +2032,7 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,116                             ; jne           215e <_sk_load_tables_hsw+0x7e>
   DB  196,193,126,111,28,144              ; vmovdqu       (%r8,%rdx,4),%ymm3
-  DB  197,229,219,13,232,157,2,0          ; vpand         0x29de8(%rip),%ymm3,%ymm1        # 2bee0 <_sk_srcover_rgba_8888_sse2_lowp+0x4d8>
+  DB  197,229,219,13,136,180,2,0          ; vpand         0x2b488(%rip),%ymm3,%ymm1        # 2d580 <_sk_srcover_rgba_8888_sse2_lowp+0x4cc>
   DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
   DB  76,139,64,8                         ; mov           0x8(%rax),%r8
   DB  76,139,72,16                        ; mov           0x10(%rax),%r9
@@ -2040,16 +2040,16 @@
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
   DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
   DB  196,194,53,146,4,136                ; vgatherdps    %ymm9,(%r8,%ymm1,4),%ymm0
-  DB  196,98,101,0,13,223,157,2,0         ; vpshufb       0x29ddf(%rip),%ymm3,%ymm9        # 2bf00 <_sk_srcover_rgba_8888_sse2_lowp+0x4f8>
+  DB  196,98,101,0,13,127,180,2,0         ; vpshufb       0x2b47f(%rip),%ymm3,%ymm9        # 2d5a0 <_sk_srcover_rgba_8888_sse2_lowp+0x4ec>
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  196,65,45,118,210                   ; vpcmpeqd      %ymm10,%ymm10,%ymm10
   DB  196,130,45,146,12,137               ; vgatherdps    %ymm10,(%r9,%ymm9,4),%ymm1
   DB  72,139,64,24                        ; mov           0x18(%rax),%rax
-  DB  196,98,101,0,13,227,157,2,0         ; vpshufb       0x29de3(%rip),%ymm3,%ymm9        # 2bf20 <_sk_srcover_rgba_8888_sse2_lowp+0x518>
+  DB  196,98,101,0,13,131,180,2,0         ; vpshufb       0x2b483(%rip),%ymm3,%ymm9        # 2d5c0 <_sk_srcover_rgba_8888_sse2_lowp+0x50c>
   DB  196,162,61,146,20,136               ; vgatherdps    %ymm8,(%rax,%ymm9,4),%ymm2
   DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,219,155,2,0         ; vbroadcastss  0x29bdb(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,135,178,2,0         ; vbroadcastss  0x2b287(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2128,7 +2128,7 @@
   DB  197,185,108,200                     ; vpunpcklqdq   %xmm0,%xmm8,%xmm1
   DB  197,57,109,192                      ; vpunpckhqdq   %xmm0,%xmm8,%xmm8
   DB  196,65,49,108,214                   ; vpunpcklqdq   %xmm14,%xmm9,%xmm10
-  DB  197,121,111,29,64,161,2,0           ; vmovdqa       0x2a140(%rip),%xmm11        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,121,111,29,32,184,2,0           ; vmovdqa       0x2b820(%rip),%xmm11        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,193,113,219,195                 ; vpand         %xmm11,%xmm1,%xmm0
   DB  196,226,125,51,200                  ; vpmovzxwd     %xmm0,%ymm1
   DB  196,65,29,118,228                   ; vpcmpeqd      %ymm12,%ymm12,%ymm12
@@ -2153,7 +2153,7 @@
   DB  197,185,235,219                     ; vpor          %xmm3,%xmm8,%xmm3
   DB  196,226,125,51,219                  ; vpmovzxwd     %xmm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,64,154,2,0          ; vbroadcastss  0x29a40(%rip),%ymm8        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,98,125,24,5,236,176,2,0         ; vbroadcastss  0x2b0ec(%rip),%ymm8        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2211,7 +2211,7 @@
   DB  197,105,108,192                     ; vpunpcklqdq   %xmm0,%xmm2,%xmm8
   DB  197,241,108,195                     ; vpunpcklqdq   %xmm3,%xmm1,%xmm0
   DB  197,241,109,203                     ; vpunpckhqdq   %xmm3,%xmm1,%xmm1
-  DB  197,249,111,29,192,159,2,0          ; vmovdqa       0x29fc0(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,249,111,29,160,182,2,0          ; vmovdqa       0x2b6a0(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  197,249,219,195                     ; vpand         %xmm3,%xmm0,%xmm0
   DB  196,98,125,51,200                   ; vpmovzxwd     %xmm0,%ymm9
   DB  196,65,45,118,210                   ; vpcmpeqd      %ymm10,%ymm10,%ymm10
@@ -2231,7 +2231,7 @@
   DB  196,226,125,51,219                  ; vpmovzxwd     %xmm3,%ymm3
   DB  196,226,45,146,20,152               ; vgatherdps    %ymm10,(%rax,%ymm3,4),%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,73,152,2,0        ; vbroadcastss  0x29849(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,245,174,2,0       ; vbroadcastss  0x2aef5(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  196,129,121,110,4,72                ; vmovd         (%r8,%r9,2),%xmm0
   DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
@@ -2298,7 +2298,7 @@
 PUBLIC _sk_byte_tables_hsw
 _sk_byte_tables_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,62,151,2,0          ; vbroadcastss  0x2973e(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,234,173,2,0         ; vbroadcastss  0x2adea(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  197,125,91,200                      ; vcvtps2dq     %ymm0,%ymm9
   DB  196,65,249,126,200                  ; vmovq         %xmm9,%r8
@@ -2428,7 +2428,7 @@
   DB  67,15,182,4,17                      ; movzbl        (%r9,%r10,1),%eax
   DB  196,194,125,49,193                  ; vpmovzxbd     %xmm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,16,149,2,0          ; vbroadcastss  0x29510(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,188,171,2,0         ; vbroadcastss  0x2abbc(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  196,226,125,49,201                  ; vpmovzxbd     %xmm1,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
@@ -2548,7 +2548,7 @@
   DB  67,15,182,4,17                      ; movzbl        (%r9,%r10,1),%eax
   DB  196,194,125,49,193                  ; vpmovzxbd     %xmm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,228,146,2,0         ; vbroadcastss  0x292e4(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,144,169,2,0         ; vbroadcastss  0x2a990(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  196,226,125,49,201                  ; vpmovzxbd     %xmm1,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
@@ -2641,33 +2641,33 @@
   DB  196,66,125,168,211                  ; vfmadd213ps   %ymm11,%ymm0,%ymm10
   DB  196,226,125,24,0                    ; vbroadcastss  (%rax),%ymm0
   DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
-  DB  196,98,125,24,37,202,145,2,0        ; vbroadcastss  0x291ca(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  DB  196,98,125,24,45,197,145,2,0        ; vbroadcastss  0x291c5(%rip),%ymm13        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,118,168,2,0        ; vbroadcastss  0x2a876(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,45,113,168,2,0        ; vbroadcastss  0x2a871(%rip),%ymm13        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,23,145,2,0         ; vbroadcastss  0x29117(%rip),%ymm13        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,45,195,167,2,0        ; vbroadcastss  0x2a7c3(%rip),%ymm13        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,173,145,2,0        ; vbroadcastss  0x291ad(%rip),%ymm13        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,45,89,168,2,0         ; vbroadcastss  0x2a859(%rip),%ymm13        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,163,145,2,0        ; vbroadcastss  0x291a3(%rip),%ymm11        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,29,79,168,2,0         ; vbroadcastss  0x2a84f(%rip),%ymm11        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,153,145,2,0        ; vbroadcastss  0x29199(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,69,168,2,0         ; vbroadcastss  0x2a845(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,143,145,2,0        ; vbroadcastss  0x2918f(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,59,168,2,0         ; vbroadcastss  0x2a83b(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  196,65,28,94,210                    ; vdivps        %ymm10,%ymm12,%ymm10
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
   DB  196,193,124,89,194                  ; vmulps        %ymm10,%ymm0,%ymm0
   DB  196,99,125,8,208,1                  ; vroundps      $0x1,%ymm0,%ymm10
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,112,145,2,0        ; vbroadcastss  0x29170(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,28,168,2,0         ; vbroadcastss  0x2a81c(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,102,145,2,0        ; vbroadcastss  0x29166(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,18,168,2,0         ; vbroadcastss  0x2a812(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,98,45,172,216                   ; vfnmadd213ps  %ymm0,%ymm10,%ymm11
-  DB  196,226,125,24,5,92,145,2,0         ; vbroadcastss  0x2915c(%rip),%ymm0        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,5,8,168,2,0          ; vbroadcastss  0x2a808(%rip),%ymm0        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,193,124,92,194                  ; vsubps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,82,145,2,0         ; vbroadcastss  0x29152(%rip),%ymm10        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,21,254,167,2,0        ; vbroadcastss  0x2a7fe(%rip),%ymm10        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  197,172,94,192                      ; vdivps        %ymm0,%ymm10,%ymm0
   DB  197,164,88,192                      ; vaddps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,21,69,145,2,0         ; vbroadcastss  0x29145(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,241,167,2,0        ; vbroadcastss  0x2a7f1(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,124,89,194                  ; vmulps        %ymm10,%ymm0,%ymm0
   DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -2675,7 +2675,7 @@
   DB  196,195,125,74,193,128              ; vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,89,144,2,0          ; vbroadcastss  0x29059(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,5,167,2,0           ; vbroadcastss  0x2a705(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2693,33 +2693,33 @@
   DB  196,66,117,168,211                  ; vfmadd213ps   %ymm11,%ymm1,%ymm10
   DB  196,226,125,24,8                    ; vbroadcastss  (%rax),%ymm1
   DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
-  DB  196,98,125,24,37,165,144,2,0        ; vbroadcastss  0x290a5(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  DB  196,98,125,24,45,160,144,2,0        ; vbroadcastss  0x290a0(%rip),%ymm13        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,81,167,2,0         ; vbroadcastss  0x2a751(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,45,76,167,2,0         ; vbroadcastss  0x2a74c(%rip),%ymm13        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,242,143,2,0        ; vbroadcastss  0x28ff2(%rip),%ymm13        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,45,158,166,2,0        ; vbroadcastss  0x2a69e(%rip),%ymm13        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,136,144,2,0        ; vbroadcastss  0x29088(%rip),%ymm13        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,45,52,167,2,0         ; vbroadcastss  0x2a734(%rip),%ymm13        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,126,144,2,0        ; vbroadcastss  0x2907e(%rip),%ymm11        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,29,42,167,2,0         ; vbroadcastss  0x2a72a(%rip),%ymm11        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,116,144,2,0        ; vbroadcastss  0x29074(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,32,167,2,0         ; vbroadcastss  0x2a720(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,106,144,2,0        ; vbroadcastss  0x2906a(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,22,167,2,0         ; vbroadcastss  0x2a716(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  196,65,28,94,210                    ; vdivps        %ymm10,%ymm12,%ymm10
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
   DB  196,193,116,89,202                  ; vmulps        %ymm10,%ymm1,%ymm1
   DB  196,99,125,8,209,1                  ; vroundps      $0x1,%ymm1,%ymm10
   DB  196,65,116,92,210                   ; vsubps        %ymm10,%ymm1,%ymm10
-  DB  196,98,125,24,29,75,144,2,0         ; vbroadcastss  0x2904b(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,247,166,2,0        ; vbroadcastss  0x2a6f7(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,65,144,2,0         ; vbroadcastss  0x29041(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,237,166,2,0        ; vbroadcastss  0x2a6ed(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,98,45,172,217                   ; vfnmadd213ps  %ymm1,%ymm10,%ymm11
-  DB  196,226,125,24,13,55,144,2,0        ; vbroadcastss  0x29037(%rip),%ymm1        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,13,227,166,2,0       ; vbroadcastss  0x2a6e3(%rip),%ymm1        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,193,116,92,202                  ; vsubps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,45,144,2,0         ; vbroadcastss  0x2902d(%rip),%ymm10        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,21,217,166,2,0        ; vbroadcastss  0x2a6d9(%rip),%ymm10        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
   DB  197,164,88,201                      ; vaddps        %ymm1,%ymm11,%ymm1
-  DB  196,98,125,24,21,32,144,2,0         ; vbroadcastss  0x29020(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,204,166,2,0        ; vbroadcastss  0x2a6cc(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,116,89,202                  ; vmulps        %ymm10,%ymm1,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -2727,7 +2727,7 @@
   DB  196,195,117,74,201,128              ; vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,5,52,143,2,0          ; vbroadcastss  0x28f34(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,224,165,2,0         ; vbroadcastss  0x2a5e0(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2745,33 +2745,33 @@
   DB  196,66,109,168,211                  ; vfmadd213ps   %ymm11,%ymm2,%ymm10
   DB  196,226,125,24,16                   ; vbroadcastss  (%rax),%ymm2
   DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
-  DB  196,98,125,24,37,128,143,2,0        ; vbroadcastss  0x28f80(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  DB  196,98,125,24,45,123,143,2,0        ; vbroadcastss  0x28f7b(%rip),%ymm13        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,44,166,2,0         ; vbroadcastss  0x2a62c(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,45,39,166,2,0         ; vbroadcastss  0x2a627(%rip),%ymm13        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,205,142,2,0        ; vbroadcastss  0x28ecd(%rip),%ymm13        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,45,121,165,2,0        ; vbroadcastss  0x2a579(%rip),%ymm13        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,99,143,2,0         ; vbroadcastss  0x28f63(%rip),%ymm13        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,45,15,166,2,0         ; vbroadcastss  0x2a60f(%rip),%ymm13        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,89,143,2,0         ; vbroadcastss  0x28f59(%rip),%ymm11        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,29,5,166,2,0          ; vbroadcastss  0x2a605(%rip),%ymm11        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,79,143,2,0         ; vbroadcastss  0x28f4f(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,251,165,2,0        ; vbroadcastss  0x2a5fb(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,69,143,2,0         ; vbroadcastss  0x28f45(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,241,165,2,0        ; vbroadcastss  0x2a5f1(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  196,65,28,94,210                    ; vdivps        %ymm10,%ymm12,%ymm10
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
   DB  196,193,108,89,210                  ; vmulps        %ymm10,%ymm2,%ymm2
   DB  196,99,125,8,210,1                  ; vroundps      $0x1,%ymm2,%ymm10
   DB  196,65,108,92,210                   ; vsubps        %ymm10,%ymm2,%ymm10
-  DB  196,98,125,24,29,38,143,2,0         ; vbroadcastss  0x28f26(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,210,165,2,0        ; vbroadcastss  0x2a5d2(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,28,143,2,0         ; vbroadcastss  0x28f1c(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,200,165,2,0        ; vbroadcastss  0x2a5c8(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,98,45,172,218                   ; vfnmadd213ps  %ymm2,%ymm10,%ymm11
-  DB  196,226,125,24,21,18,143,2,0        ; vbroadcastss  0x28f12(%rip),%ymm2        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,21,190,165,2,0       ; vbroadcastss  0x2a5be(%rip),%ymm2        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,193,108,92,210                  ; vsubps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,8,143,2,0          ; vbroadcastss  0x28f08(%rip),%ymm10        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,21,180,165,2,0        ; vbroadcastss  0x2a5b4(%rip),%ymm10        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  197,172,94,210                      ; vdivps        %ymm2,%ymm10,%ymm2
   DB  197,164,88,210                      ; vaddps        %ymm2,%ymm11,%ymm2
-  DB  196,98,125,24,21,251,142,2,0        ; vbroadcastss  0x28efb(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,167,165,2,0        ; vbroadcastss  0x2a5a7(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,108,89,210                  ; vmulps        %ymm10,%ymm2,%ymm2
   DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -2779,7 +2779,7 @@
   DB  196,195,109,74,209,128              ; vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,15,142,2,0          ; vbroadcastss  0x28e0f(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,187,164,2,0         ; vbroadcastss  0x2a4bb(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2797,33 +2797,33 @@
   DB  196,66,101,168,211                  ; vfmadd213ps   %ymm11,%ymm3,%ymm10
   DB  196,226,125,24,24                   ; vbroadcastss  (%rax),%ymm3
   DB  196,65,124,91,218                   ; vcvtdq2ps     %ymm10,%ymm11
-  DB  196,98,125,24,37,91,142,2,0         ; vbroadcastss  0x28e5b(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  DB  196,98,125,24,45,86,142,2,0         ; vbroadcastss  0x28e56(%rip),%ymm13        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,7,165,2,0          ; vbroadcastss  0x2a507(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,45,2,165,2,0          ; vbroadcastss  0x2a502(%rip),%ymm13        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,65,44,84,213                    ; vandps        %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,168,141,2,0        ; vbroadcastss  0x28da8(%rip),%ymm13        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,45,84,164,2,0         ; vbroadcastss  0x2a454(%rip),%ymm13        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,86,213                    ; vorps         %ymm13,%ymm10,%ymm10
-  DB  196,98,125,24,45,62,142,2,0         ; vbroadcastss  0x28e3e(%rip),%ymm13        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,45,234,164,2,0        ; vbroadcastss  0x2a4ea(%rip),%ymm13        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,66,37,184,236                   ; vfmadd231ps   %ymm12,%ymm11,%ymm13
-  DB  196,98,125,24,29,52,142,2,0         ; vbroadcastss  0x28e34(%rip),%ymm11        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,29,224,164,2,0        ; vbroadcastss  0x2a4e0(%rip),%ymm11        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,66,45,172,221                   ; vfnmadd213ps  %ymm13,%ymm10,%ymm11
-  DB  196,98,125,24,37,42,142,2,0         ; vbroadcastss  0x28e2a(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,214,164,2,0        ; vbroadcastss  0x2a4d6(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,32,142,2,0         ; vbroadcastss  0x28e20(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,204,164,2,0        ; vbroadcastss  0x2a4cc(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  196,65,28,94,210                    ; vdivps        %ymm10,%ymm12,%ymm10
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
   DB  196,193,100,89,218                  ; vmulps        %ymm10,%ymm3,%ymm3
   DB  196,99,125,8,211,1                  ; vroundps      $0x1,%ymm3,%ymm10
   DB  196,65,100,92,210                   ; vsubps        %ymm10,%ymm3,%ymm10
-  DB  196,98,125,24,29,1,142,2,0          ; vbroadcastss  0x28e01(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,173,164,2,0        ; vbroadcastss  0x2a4ad(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,247,141,2,0        ; vbroadcastss  0x28df7(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,163,164,2,0        ; vbroadcastss  0x2a4a3(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,98,45,172,219                   ; vfnmadd213ps  %ymm3,%ymm10,%ymm11
-  DB  196,226,125,24,29,237,141,2,0       ; vbroadcastss  0x28ded(%rip),%ymm3        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,29,153,164,2,0       ; vbroadcastss  0x2a499(%rip),%ymm3        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,193,100,92,218                  ; vsubps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,227,141,2,0        ; vbroadcastss  0x28de3(%rip),%ymm10        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,21,143,164,2,0        ; vbroadcastss  0x2a48f(%rip),%ymm10        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  197,172,94,219                      ; vdivps        %ymm3,%ymm10,%ymm3
   DB  197,164,88,219                      ; vaddps        %ymm3,%ymm11,%ymm3
-  DB  196,98,125,24,21,214,141,2,0        ; vbroadcastss  0x28dd6(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,130,164,2,0        ; vbroadcastss  0x2a482(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,100,89,218                  ; vmulps        %ymm10,%ymm3,%ymm3
   DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -2831,7 +2831,7 @@
   DB  196,195,101,74,217,128              ; vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,100,95,216                  ; vmaxps        %ymm8,%ymm3,%ymm3
-  DB  196,98,125,24,5,234,140,2,0         ; vbroadcastss  0x28cea(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,150,163,2,0         ; vbroadcastss  0x2a396(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2848,35 +2848,35 @@
   DB  197,252,40,241                      ; vmovaps       %ymm1,%ymm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,124,91,208                      ; vcvtdq2ps     %ymm0,%ymm10
-  DB  196,98,125,24,29,59,141,2,0         ; vbroadcastss  0x28d3b(%rip),%ymm11        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
-  DB  196,226,125,24,45,54,141,2,0        ; vbroadcastss  0x28d36(%rip),%ymm5        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,29,231,163,2,0        ; vbroadcastss  0x2a3e7(%rip),%ymm11        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,226,125,24,45,226,163,2,0       ; vbroadcastss  0x2a3e2(%rip),%ymm5        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  197,124,84,205                      ; vandps        %ymm5,%ymm0,%ymm9
-  DB  196,226,125,24,37,137,140,2,0       ; vbroadcastss  0x28c89(%rip),%ymm4        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,37,53,163,2,0        ; vbroadcastss  0x2a335(%rip),%ymm4        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,52,86,228                       ; vorps         %ymm4,%ymm9,%ymm12
-  DB  196,98,125,24,45,32,141,2,0         ; vbroadcastss  0x28d20(%rip),%ymm13        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,45,204,163,2,0        ; vbroadcastss  0x2a3cc(%rip),%ymm13        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,66,37,168,213                   ; vfmadd213ps   %ymm13,%ymm11,%ymm10
-  DB  196,98,125,24,13,22,141,2,0         ; vbroadcastss  0x28d16(%rip),%ymm9        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,13,194,163,2,0        ; vbroadcastss  0x2a3c2(%rip),%ymm9        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,66,29,188,209                   ; vfnmadd231ps  %ymm9,%ymm12,%ymm10
   DB  197,124,91,246                      ; vcvtdq2ps     %ymm6,%ymm14
   DB  196,66,37,168,245                   ; vfmadd213ps   %ymm13,%ymm11,%ymm14
   DB  197,124,91,255                      ; vcvtdq2ps     %ymm7,%ymm15
   DB  196,66,37,168,253                   ; vfmadd213ps   %ymm13,%ymm11,%ymm15
-  DB  196,98,125,24,29,250,140,2,0        ; vbroadcastss  0x28cfa(%rip),%ymm11        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,29,166,163,2,0        ; vbroadcastss  0x2a3a6(%rip),%ymm11        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,65,28,88,227                    ; vaddps        %ymm11,%ymm12,%ymm12
-  DB  196,98,125,24,45,240,140,2,0        ; vbroadcastss  0x28cf0(%rip),%ymm13        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,45,156,163,2,0        ; vbroadcastss  0x2a39c(%rip),%ymm13        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  196,65,20,94,228                    ; vdivps        %ymm12,%ymm13,%ymm12
   DB  196,65,44,92,212                    ; vsubps        %ymm12,%ymm10,%ymm10
   DB  196,98,125,24,32                    ; vbroadcastss  (%rax),%ymm12
   DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
   DB  196,67,125,8,194,1                  ; vroundps      $0x1,%ymm10,%ymm8
   DB  196,65,44,92,192                    ; vsubps        %ymm8,%ymm10,%ymm8
-  DB  196,226,125,24,21,204,140,2,0       ; vbroadcastss  0x28ccc(%rip),%ymm2        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,226,125,24,21,120,163,2,0       ; vbroadcastss  0x2a378(%rip),%ymm2        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  197,44,88,210                       ; vaddps        %ymm2,%ymm10,%ymm10
-  DB  196,226,125,24,29,195,140,2,0       ; vbroadcastss  0x28cc3(%rip),%ymm3        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,226,125,24,29,111,163,2,0       ; vbroadcastss  0x2a36f(%rip),%ymm3        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,98,61,188,211                   ; vfnmadd231ps  %ymm3,%ymm8,%ymm10
-  DB  196,226,125,24,13,185,140,2,0       ; vbroadcastss  0x28cb9(%rip),%ymm1        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,13,101,163,2,0       ; vbroadcastss  0x2a365(%rip),%ymm1        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,65,116,92,192                   ; vsubps        %ymm8,%ymm1,%ymm8
-  DB  196,226,125,24,5,175,140,2,0        ; vbroadcastss  0x28caf(%rip),%ymm0        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,226,125,24,5,91,163,2,0         ; vbroadcastss  0x2a35b(%rip),%ymm0        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  196,65,124,94,192                   ; vdivps        %ymm8,%ymm0,%ymm8
   DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
   DB  197,204,84,245                      ; vandps        %ymm5,%ymm6,%ymm6
@@ -2907,7 +2907,7 @@
   DB  197,244,92,205                      ; vsubps        %ymm5,%ymm1,%ymm1
   DB  197,252,94,193                      ; vdivps        %ymm1,%ymm0,%ymm0
   DB  197,236,88,192                      ; vaddps        %ymm0,%ymm2,%ymm0
-  DB  196,226,125,24,13,32,140,2,0        ; vbroadcastss  0x28c20(%rip),%ymm1        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,226,125,24,13,204,162,2,0       ; vbroadcastss  0x2a2cc(%rip),%ymm1        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  197,188,89,209                      ; vmulps        %ymm1,%ymm8,%ymm2
   DB  197,204,89,217                      ; vmulps        %ymm1,%ymm6,%ymm3
   DB  197,252,89,225                      ; vmulps        %ymm1,%ymm0,%ymm4
@@ -2925,26 +2925,26 @@
 
 PUBLIC _sk_lab_to_xyz_hsw
 _sk_lab_to_xyz_hsw LABEL PROC
-  DB  196,98,125,24,5,216,139,2,0         ; vbroadcastss  0x28bd8(%rip),%ymm8        # 2bd7c <_sk_srcover_rgba_8888_sse2_lowp+0x374>
-  DB  196,98,125,24,13,55,139,2,0         ; vbroadcastss  0x28b37(%rip),%ymm9        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  196,98,125,24,21,202,139,2,0        ; vbroadcastss  0x28bca(%rip),%ymm10        # 2bd80 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
+  DB  196,98,125,24,5,132,162,2,0         ; vbroadcastss  0x2a284(%rip),%ymm8        # 2d428 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
+  DB  196,98,125,24,13,227,161,2,0        ; vbroadcastss  0x2a1e3(%rip),%ymm9        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,21,118,162,2,0        ; vbroadcastss  0x2a276(%rip),%ymm10        # 2d42c <_sk_srcover_rgba_8888_sse2_lowp+0x378>
   DB  196,194,53,168,202                  ; vfmadd213ps   %ymm10,%ymm9,%ymm1
   DB  196,194,53,168,210                  ; vfmadd213ps   %ymm10,%ymm9,%ymm2
-  DB  196,98,125,24,13,187,139,2,0        ; vbroadcastss  0x28bbb(%rip),%ymm9        # 2bd84 <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
+  DB  196,98,125,24,13,103,162,2,0        ; vbroadcastss  0x2a267(%rip),%ymm9        # 2d430 <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
   DB  196,66,125,184,200                  ; vfmadd231ps   %ymm8,%ymm0,%ymm9
-  DB  196,226,125,24,5,177,139,2,0        ; vbroadcastss  0x28bb1(%rip),%ymm0        # 2bd88 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
+  DB  196,226,125,24,5,93,162,2,0         ; vbroadcastss  0x2a25d(%rip),%ymm0        # 2d434 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
-  DB  196,98,125,24,5,168,139,2,0         ; vbroadcastss  0x28ba8(%rip),%ymm8        # 2bd8c <_sk_srcover_rgba_8888_sse2_lowp+0x384>
+  DB  196,98,125,24,5,84,162,2,0          ; vbroadcastss  0x2a254(%rip),%ymm8        # 2d438 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
   DB  196,98,117,168,192                  ; vfmadd213ps   %ymm0,%ymm1,%ymm8
-  DB  196,98,125,24,13,158,139,2,0        ; vbroadcastss  0x28b9e(%rip),%ymm9        # 2bd90 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
+  DB  196,98,125,24,13,74,162,2,0         ; vbroadcastss  0x2a24a(%rip),%ymm9        # 2d43c <_sk_srcover_rgba_8888_sse2_lowp+0x388>
   DB  196,98,109,172,200                  ; vfnmadd213ps  %ymm0,%ymm2,%ymm9
   DB  196,193,60,89,200                   ; vmulps        %ymm8,%ymm8,%ymm1
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
-  DB  196,226,125,24,21,139,139,2,0       ; vbroadcastss  0x28b8b(%rip),%ymm2        # 2bd94 <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
+  DB  196,226,125,24,21,55,162,2,0        ; vbroadcastss  0x2a237(%rip),%ymm2        # 2d440 <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
   DB  197,108,194,209,1                   ; vcmpltps      %ymm1,%ymm2,%ymm10
-  DB  196,98,125,24,29,129,139,2,0        ; vbroadcastss  0x28b81(%rip),%ymm11        # 2bd98 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
+  DB  196,98,125,24,29,45,162,2,0         ; vbroadcastss  0x2a22d(%rip),%ymm11        # 2d444 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
   DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
-  DB  196,98,125,24,37,119,139,2,0        ; vbroadcastss  0x28b77(%rip),%ymm12        # 2bd9c <_sk_srcover_rgba_8888_sse2_lowp+0x394>
+  DB  196,98,125,24,37,35,162,2,0         ; vbroadcastss  0x2a223(%rip),%ymm12        # 2d448 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
   DB  196,65,60,89,196                    ; vmulps        %ymm12,%ymm8,%ymm8
   DB  196,99,61,74,193,160                ; vblendvps     %ymm10,%ymm1,%ymm8,%ymm8
   DB  197,252,89,200                      ; vmulps        %ymm0,%ymm0,%ymm1
@@ -2959,9 +2959,9 @@
   DB  196,65,52,88,203                    ; vaddps        %ymm11,%ymm9,%ymm9
   DB  196,65,52,89,204                    ; vmulps        %ymm12,%ymm9,%ymm9
   DB  196,227,53,74,208,32                ; vblendvps     %ymm2,%ymm0,%ymm9,%ymm2
-  DB  196,226,125,24,5,44,139,2,0         ; vbroadcastss  0x28b2c(%rip),%ymm0        # 2bda0 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
+  DB  196,226,125,24,5,216,161,2,0        ; vbroadcastss  0x2a1d8(%rip),%ymm0        # 2d44c <_sk_srcover_rgba_8888_sse2_lowp+0x398>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,35,139,2,0          ; vbroadcastss  0x28b23(%rip),%ymm8        # 2bda4 <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
+  DB  196,98,125,24,5,207,161,2,0         ; vbroadcastss  0x2a1cf(%rip),%ymm8        # 2d450 <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -2977,10 +2977,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,52                              ; jne           32d6 <_sk_load_a8_hsw+0x4c>
   DB  196,130,121,48,4,24                 ; vpmovzxbw     (%r8,%r11,1),%xmm0
-  DB  197,249,219,5,32,145,2,0            ; vpand         0x29120(%rip),%xmm0,%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,249,219,5,0,168,2,0             ; vpand         0x2a800(%rip),%xmm0,%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,110,138,2,0       ; vbroadcastss  0x28a6e(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,26,161,2,0        ; vbroadcastss  0x2a11a(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
@@ -3051,10 +3051,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,52                              ; jne           33d4 <_sk_load_a8_dst_hsw+0x4c>
   DB  196,130,121,48,36,24                ; vpmovzxbw     (%r8,%r11,1),%xmm4
-  DB  197,217,219,37,34,144,2,0           ; vpand         0x29022(%rip),%xmm4,%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,217,219,37,2,167,2,0            ; vpand         0x2a702(%rip),%xmm4,%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,125,51,228                  ; vpmovzxwd     %xmm4,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,112,137,2,0       ; vbroadcastss  0x28970(%rip),%ymm5        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,45,28,160,2,0        ; vbroadcastss  0x2a01c(%rip),%ymm5        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,220,89,253                      ; vmulps        %ymm5,%ymm4,%ymm7
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
@@ -3102,7 +3102,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,221                 ; callq         ffffffffde00347c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd7a74>
+  DB  232,255,255,255,221                 ; callq         ffffffffde00347c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfd63c8>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,210                             ; callq         *%rdx
@@ -3167,7 +3167,7 @@
   DB  196,227,121,32,192,7                ; vpinsrb       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,183,135,2,0       ; vbroadcastss  0x287b7(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,99,158,2,0        ; vbroadcastss  0x29e63(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
@@ -3187,7 +3187,7 @@
   DB  77,15,175,193                       ; imul          %r9,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,52,135,2,0          ; vbroadcastss  0x28734(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,224,157,2,0         ; vbroadcastss  0x29de0(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,100,89,192                   ; vmulps        %ymm8,%ymm3,%ymm8
   DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
   DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
@@ -3212,13 +3212,13 @@
   DB  196,3,121,20,4,24,0                 ; vpextrb       $0x0,%xmm8,(%r8,%r11,1)
   DB  235,202                             ; jmp           35d4 <_sk_store_a8_hsw+0x40>
   DB  196,3,121,20,68,24,2,4              ; vpextrb       $0x4,%xmm8,0x2(%r8,%r11,1)
-  DB  196,98,57,0,5,197,141,2,0           ; vpshufb       0x28dc5(%rip),%xmm8,%xmm8        # 2c3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x9d8>
+  DB  196,98,57,0,5,165,164,2,0           ; vpshufb       0x2a4a5(%rip),%xmm8,%xmm8        # 2dac0 <_sk_srcover_rgba_8888_sse2_lowp+0xa0c>
   DB  196,3,121,21,4,24,0                 ; vpextrw       $0x0,%xmm8,(%r8,%r11,1)
   DB  235,176                             ; jmp           35d4 <_sk_store_a8_hsw+0x40>
   DB  196,3,121,20,68,24,6,12             ; vpextrb       $0xc,%xmm8,0x6(%r8,%r11,1)
   DB  196,3,121,20,68,24,5,10             ; vpextrb       $0xa,%xmm8,0x5(%r8,%r11,1)
   DB  196,3,121,20,68,24,4,8              ; vpextrb       $0x8,%xmm8,0x4(%r8,%r11,1)
-  DB  196,98,57,0,5,171,141,2,0           ; vpshufb       0x28dab(%rip),%xmm8,%xmm8        # 2c3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x9e8>
+  DB  196,98,57,0,5,139,164,2,0           ; vpshufb       0x2a48b(%rip),%xmm8,%xmm8        # 2dad0 <_sk_srcover_rgba_8888_sse2_lowp+0xa1c>
   DB  196,1,121,126,4,24                  ; vmovd         %xmm8,(%r8,%r11,1)
   DB  235,135                             ; jmp           35d4 <_sk_store_a8_hsw+0x40>
   DB  15,31,0                             ; nopl          (%rax)
@@ -3253,13 +3253,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,57                              ; jne           36bd <_sk_load_g8_hsw+0x51>
   DB  196,130,121,48,4,24                 ; vpmovzxbw     (%r8,%r11,1),%xmm0
-  DB  197,249,219,5,62,141,2,0            ; vpand         0x28d3e(%rip),%xmm0,%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,249,219,5,30,164,2,0            ; vpand         0x2a41e(%rip),%xmm0,%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,140,134,2,0       ; vbroadcastss  0x2868c(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,56,157,2,0        ; vbroadcastss  0x29d38(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,5,134,2,0         ; vbroadcastss  0x28605(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,177,156,2,0       ; vbroadcastss  0x29cb1(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  255,224                             ; jmpq          *%rax
@@ -3304,7 +3304,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,222                 ; jmpq          ffffffffdf003764 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd7d5c>
+  DB  233,255,255,255,222                 ; jmpq          ffffffffdf003764 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd66b0>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,211                             ; callq         *%rbx
@@ -3326,13 +3326,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,57                              ; jne           37c1 <_sk_load_g8_dst_hsw+0x51>
   DB  196,130,121,48,36,24                ; vpmovzxbw     (%r8,%r11,1),%xmm4
-  DB  197,217,219,37,58,140,2,0           ; vpand         0x28c3a(%rip),%xmm4,%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,217,219,37,26,163,2,0           ; vpand         0x2a31a(%rip),%xmm4,%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,125,51,228                  ; vpmovzxwd     %xmm4,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,136,133,2,0       ; vbroadcastss  0x28588(%rip),%ymm5        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,45,52,156,2,0        ; vbroadcastss  0x29c34(%rip),%ymm5        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,61,1,133,2,0         ; vbroadcastss  0x28501(%rip),%ymm7        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,61,173,155,2,0       ; vbroadcastss  0x29bad(%rip),%ymm7        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,236                      ; vmovaps       %ymm4,%ymm5
   DB  197,252,40,244                      ; vmovaps       %ymm4,%ymm6
   DB  255,224                             ; jmpq          *%rax
@@ -3377,7 +3377,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,222                 ; jmpq          ffffffffdf003868 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd7e60>
+  DB  233,255,255,255,222                 ; jmpq          ffffffffdf003868 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefd67b4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,211                             ; callq         *%rbx
@@ -3442,10 +3442,10 @@
   DB  196,227,121,32,192,7                ; vpinsrb       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,49,192                  ; vpmovzxbd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,203,131,2,0       ; vbroadcastss  0x283cb(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,119,154,2,0       ; vbroadcastss  0x29a77(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,68,131,2,0        ; vbroadcastss  0x28344(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,240,153,2,0       ; vbroadcastss  0x299f0(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -3467,23 +3467,23 @@
   DB  117,114                             ; jne           3a12 <_sk_load_565_hsw+0x8d>
   DB  196,129,122,111,4,88                ; vmovdqu       (%r8,%r11,2),%xmm0
   DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
-  DB  196,226,125,88,5,128,131,2,0        ; vpbroadcastd  0x28380(%rip),%ymm0        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,88,5,44,154,2,0         ; vpbroadcastd  0x29a2c(%rip),%ymm0        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,115,131,2,0       ; vbroadcastss  0x28373(%rip),%ymm1        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,13,31,154,2,0        ; vbroadcastss  0x29a1f(%rip),%ymm1        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,106,131,2,0       ; vpbroadcastd  0x2836a(%rip),%ymm1        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,88,13,22,154,2,0        ; vpbroadcastd  0x29a16(%rip),%ymm1        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,93,131,2,0        ; vbroadcastss  0x2835d(%rip),%ymm3        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,29,9,154,2,0         ; vbroadcastss  0x29a09(%rip),%ymm3        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,84,131,2,0        ; vpbroadcastd  0x28354(%rip),%ymm3        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,88,29,0,154,2,0         ; vpbroadcastd  0x29a00(%rip),%ymm3        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,71,131,2,0        ; vbroadcastss  0x28347(%rip),%ymm3        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,29,243,153,2,0       ; vbroadcastss  0x299f3(%rip),%ymm3        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,168,130,2,0       ; vbroadcastss  0x282a8(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,84,153,2,0        ; vbroadcastss  0x29954(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,7                        ; and           $0x7,%r9b
@@ -3544,23 +3544,23 @@
   DB  117,114                             ; jne           3b3d <_sk_load_565_dst_hsw+0x8d>
   DB  196,129,122,111,36,88               ; vmovdqu       (%r8,%r11,2),%xmm4
   DB  196,226,125,51,244                  ; vpmovzxwd     %xmm4,%ymm6
-  DB  196,226,125,88,37,85,130,2,0        ; vpbroadcastd  0x28255(%rip),%ymm4        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,88,37,1,153,2,0         ; vpbroadcastd  0x29901(%rip),%ymm4        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,205,219,228                     ; vpand         %ymm4,%ymm6,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,72,130,2,0        ; vbroadcastss  0x28248(%rip),%ymm5        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,45,244,152,2,0       ; vbroadcastss  0x298f4(%rip),%ymm5        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
-  DB  196,226,125,88,45,63,130,2,0        ; vpbroadcastd  0x2823f(%rip),%ymm5        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,88,45,235,152,2,0       ; vpbroadcastd  0x298eb(%rip),%ymm5        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,205,219,237                     ; vpand         %ymm5,%ymm6,%ymm5
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,226,125,24,61,50,130,2,0        ; vbroadcastss  0x28232(%rip),%ymm7        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,61,222,152,2,0       ; vbroadcastss  0x298de(%rip),%ymm7        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,212,89,239                      ; vmulps        %ymm7,%ymm5,%ymm5
-  DB  196,226,125,88,61,41,130,2,0        ; vpbroadcastd  0x28229(%rip),%ymm7        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,88,61,213,152,2,0       ; vpbroadcastd  0x298d5(%rip),%ymm7        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,205,219,247                     ; vpand         %ymm7,%ymm6,%ymm6
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
-  DB  196,226,125,24,61,28,130,2,0        ; vbroadcastss  0x2821c(%rip),%ymm7        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,61,200,152,2,0       ; vbroadcastss  0x298c8(%rip),%ymm7        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,204,89,247                      ; vmulps        %ymm7,%ymm6,%ymm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,61,125,129,2,0       ; vbroadcastss  0x2817d(%rip),%ymm7        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,61,41,152,2,0        ; vbroadcastss  0x29829(%rip),%ymm7        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,7                        ; and           $0x7,%r9b
@@ -3664,23 +3664,23 @@
   DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
   DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,51,208                  ; vpmovzxwd     %xmm0,%ymm2
-  DB  196,226,125,88,5,112,128,2,0        ; vpbroadcastd  0x28070(%rip),%ymm0        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,88,5,28,151,2,0         ; vpbroadcastd  0x2971c(%rip),%ymm0        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,237,219,192                     ; vpand         %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,99,128,2,0        ; vbroadcastss  0x28063(%rip),%ymm1        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,13,15,151,2,0        ; vbroadcastss  0x2970f(%rip),%ymm1        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,90,128,2,0        ; vpbroadcastd  0x2805a(%rip),%ymm1        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,88,13,6,151,2,0         ; vpbroadcastd  0x29706(%rip),%ymm1        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,77,128,2,0        ; vbroadcastss  0x2804d(%rip),%ymm3        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,29,249,150,2,0       ; vbroadcastss  0x296f9(%rip),%ymm3        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,88,29,68,128,2,0        ; vpbroadcastd  0x28044(%rip),%ymm3        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,88,29,240,150,2,0       ; vpbroadcastd  0x296f0(%rip),%ymm3        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,55,128,2,0        ; vbroadcastss  0x28037(%rip),%ymm3        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,29,227,150,2,0       ; vbroadcastss  0x296e3(%rip),%ymm3        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,152,127,2,0       ; vbroadcastss  0x27f98(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,68,150,2,0        ; vbroadcastss  0x29644(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -3696,11 +3696,11 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,96,128,2,0          ; vbroadcastss  0x28060(%rip),%ymm8        # 2bda8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
+  DB  196,98,125,24,5,12,151,2,0          ; vbroadcastss  0x2970c(%rip),%ymm8        # 2d454 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,193,53,114,241,11               ; vpslld        $0xb,%ymm9,%ymm9
-  DB  196,98,125,24,21,75,128,2,0         ; vbroadcastss  0x2804b(%rip),%ymm10        # 2bdac <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
+  DB  196,98,125,24,21,247,150,2,0        ; vbroadcastss  0x296f7(%rip),%ymm10        # 2d458 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
   DB  196,65,116,89,210                   ; vmulps        %ymm10,%ymm1,%ymm10
   DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
   DB  196,193,45,114,242,5                ; vpslld        $0x5,%ymm10,%ymm10
@@ -3752,7 +3752,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,224                 ; callq         ffffffffe1003e14 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd840c>
+  DB  232,255,255,255,224                 ; callq         ffffffffe1003e14 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd6d60>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -3770,25 +3770,25 @@
   DB  15,133,138,0,0,0                    ; jne           3ec1 <_sk_load_4444_hsw+0xa9>
   DB  196,129,122,111,4,88                ; vmovdqu       (%r8,%r11,2),%xmm0
   DB  196,226,125,51,216                  ; vpmovzxwd     %xmm0,%ymm3
-  DB  196,226,125,88,5,101,127,2,0        ; vpbroadcastd  0x27f65(%rip),%ymm0        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,88,5,17,150,2,0         ; vpbroadcastd  0x29611(%rip),%ymm0        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,88,127,2,0        ; vbroadcastss  0x27f58(%rip),%ymm1        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,13,4,150,2,0         ; vbroadcastss  0x29604(%rip),%ymm1        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,79,127,2,0        ; vpbroadcastd  0x27f4f(%rip),%ymm1        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,88,13,251,149,2,0       ; vpbroadcastd  0x295fb(%rip),%ymm1        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,66,127,2,0        ; vbroadcastss  0x27f42(%rip),%ymm2        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,21,238,149,2,0       ; vbroadcastss  0x295ee(%rip),%ymm2        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,57,127,2,0        ; vpbroadcastd  0x27f39(%rip),%ymm2        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,88,21,229,149,2,0       ; vpbroadcastd  0x295e5(%rip),%ymm2        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,44,127,2,0          ; vbroadcastss  0x27f2c(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,216,149,2,0         ; vbroadcastss  0x295d8(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,34,127,2,0          ; vpbroadcastd  0x27f22(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,88,5,206,149,2,0         ; vpbroadcastd  0x295ce(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,20,127,2,0          ; vbroadcastss  0x27f14(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,192,149,2,0         ; vbroadcastss  0x295c0(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -3854,25 +3854,25 @@
   DB  15,133,138,0,0,0                    ; jne           400d <_sk_load_4444_dst_hsw+0xa9>
   DB  196,129,122,111,36,88               ; vmovdqu       (%r8,%r11,2),%xmm4
   DB  196,226,125,51,252                  ; vpmovzxwd     %xmm4,%ymm7
-  DB  196,226,125,88,37,25,126,2,0        ; vpbroadcastd  0x27e19(%rip),%ymm4        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,88,37,197,148,2,0       ; vpbroadcastd  0x294c5(%rip),%ymm4        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,197,219,228                     ; vpand         %ymm4,%ymm7,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,12,126,2,0        ; vbroadcastss  0x27e0c(%rip),%ymm5        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,45,184,148,2,0       ; vbroadcastss  0x294b8(%rip),%ymm5        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
-  DB  196,226,125,88,45,3,126,2,0         ; vpbroadcastd  0x27e03(%rip),%ymm5        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,88,45,175,148,2,0       ; vpbroadcastd  0x294af(%rip),%ymm5        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,197,219,237                     ; vpand         %ymm5,%ymm7,%ymm5
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,226,125,24,53,246,125,2,0       ; vbroadcastss  0x27df6(%rip),%ymm6        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,53,162,148,2,0       ; vbroadcastss  0x294a2(%rip),%ymm6        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,212,89,238                      ; vmulps        %ymm6,%ymm5,%ymm5
-  DB  196,226,125,88,53,237,125,2,0       ; vpbroadcastd  0x27ded(%rip),%ymm6        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,88,53,153,148,2,0       ; vpbroadcastd  0x29499(%rip),%ymm6        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,197,219,246                     ; vpand         %ymm6,%ymm7,%ymm6
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
-  DB  196,98,125,24,5,224,125,2,0         ; vbroadcastss  0x27de0(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,140,148,2,0         ; vbroadcastss  0x2948c(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,76,89,240                   ; vmulps        %ymm8,%ymm6,%ymm6
-  DB  196,98,125,88,5,214,125,2,0         ; vpbroadcastd  0x27dd6(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,88,5,130,148,2,0         ; vpbroadcastd  0x29482(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,69,219,248                  ; vpand         %ymm8,%ymm7,%ymm7
   DB  197,252,91,255                      ; vcvtdq2ps     %ymm7,%ymm7
-  DB  196,98,125,24,5,200,125,2,0         ; vbroadcastss  0x27dc8(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,116,148,2,0         ; vbroadcastss  0x29474(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,68,89,248                   ; vmulps        %ymm8,%ymm7,%ymm7
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -3978,25 +3978,25 @@
   DB  65,15,183,4,88                      ; movzwl        (%r8,%rbx,2),%eax
   DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
   DB  196,226,125,51,216                  ; vpmovzxwd     %xmm0,%ymm3
-  DB  196,226,125,88,5,24,124,2,0         ; vpbroadcastd  0x27c18(%rip),%ymm0        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,88,5,196,146,2,0        ; vpbroadcastd  0x292c4(%rip),%ymm0        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,229,219,192                     ; vpand         %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,11,124,2,0        ; vbroadcastss  0x27c0b(%rip),%ymm1        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,13,183,146,2,0       ; vbroadcastss  0x292b7(%rip),%ymm1        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,88,13,2,124,2,0         ; vpbroadcastd  0x27c02(%rip),%ymm1        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,88,13,174,146,2,0       ; vpbroadcastd  0x292ae(%rip),%ymm1        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,229,219,201                     ; vpand         %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,245,123,2,0       ; vbroadcastss  0x27bf5(%rip),%ymm2        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,21,161,146,2,0       ; vbroadcastss  0x292a1(%rip),%ymm2        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,88,21,236,123,2,0       ; vpbroadcastd  0x27bec(%rip),%ymm2        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,88,21,152,146,2,0       ; vpbroadcastd  0x29298(%rip),%ymm2        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,229,219,210                     ; vpand         %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,223,123,2,0         ; vbroadcastss  0x27bdf(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,139,146,2,0         ; vbroadcastss  0x2928b(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,88,5,213,123,2,0         ; vpbroadcastd  0x27bd5(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,88,5,129,146,2,0         ; vpbroadcastd  0x29281(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,101,219,216                 ; vpand         %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,199,123,2,0         ; vbroadcastss  0x27bc7(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,115,146,2,0         ; vbroadcastss  0x29273(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -4014,7 +4014,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,156,123,2,0         ; vbroadcastss  0x27b9c(%rip),%ymm8        # 2bdd0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
+  DB  196,98,125,24,5,72,146,2,0          ; vbroadcastss  0x29248(%rip),%ymm8        # 2d47c <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,193,53,114,241,12               ; vpslld        $0xc,%ymm9,%ymm9
@@ -4073,7 +4073,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,224                 ; callq         ffffffffe100430c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd8904>
+  DB  232,255,255,255,224                 ; callq         ffffffffe100430c <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fd7258>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -4090,14 +4090,14 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,86                              ; jne           4382 <_sk_load_8888_hsw+0x72>
   DB  196,129,126,111,28,152              ; vmovdqu       (%r8,%r11,4),%ymm3
-  DB  197,229,219,5,6,124,2,0             ; vpand         0x27c06(%rip),%ymm3,%ymm0        # 2bf40 <_sk_srcover_rgba_8888_sse2_lowp+0x538>
+  DB  197,229,219,5,166,146,2,0           ; vpand         0x292a6(%rip),%ymm3,%ymm0        # 2d5e0 <_sk_srcover_rgba_8888_sse2_lowp+0x52c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,233,121,2,0         ; vbroadcastss  0x279e9(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,149,144,2,0         ; vbroadcastss  0x29095(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,226,101,0,13,11,124,2,0         ; vpshufb       0x27c0b(%rip),%ymm3,%ymm1        # 2bf60 <_sk_srcover_rgba_8888_sse2_lowp+0x558>
+  DB  196,226,101,0,13,171,146,2,0        ; vpshufb       0x292ab(%rip),%ymm3,%ymm1        # 2d600 <_sk_srcover_rgba_8888_sse2_lowp+0x54c>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,226,101,0,21,25,124,2,0         ; vpshufb       0x27c19(%rip),%ymm3,%ymm2        # 2bf80 <_sk_srcover_rgba_8888_sse2_lowp+0x578>
+  DB  196,226,101,0,21,185,146,2,0        ; vpshufb       0x292b9(%rip),%ymm3,%ymm2        # 2d620 <_sk_srcover_rgba_8888_sse2_lowp+0x56c>
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
@@ -4171,14 +4171,14 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,86                              ; jne           44b6 <_sk_load_8888_dst_hsw+0x72>
   DB  196,129,126,111,60,152              ; vmovdqu       (%r8,%r11,4),%ymm7
-  DB  197,197,219,37,50,123,2,0           ; vpand         0x27b32(%rip),%ymm7,%ymm4        # 2bfa0 <_sk_srcover_rgba_8888_sse2_lowp+0x598>
+  DB  197,197,219,37,210,145,2,0          ; vpand         0x291d2(%rip),%ymm7,%ymm4        # 2d640 <_sk_srcover_rgba_8888_sse2_lowp+0x58c>
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,98,125,24,5,181,120,2,0         ; vbroadcastss  0x278b5(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,97,143,2,0          ; vbroadcastss  0x28f61(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,92,89,224                   ; vmulps        %ymm8,%ymm4,%ymm4
-  DB  196,226,69,0,45,55,123,2,0          ; vpshufb       0x27b37(%rip),%ymm7,%ymm5        # 2bfc0 <_sk_srcover_rgba_8888_sse2_lowp+0x5b8>
+  DB  196,226,69,0,45,215,145,2,0         ; vpshufb       0x291d7(%rip),%ymm7,%ymm5        # 2d660 <_sk_srcover_rgba_8888_sse2_lowp+0x5ac>
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
   DB  196,193,84,89,232                   ; vmulps        %ymm8,%ymm5,%ymm5
-  DB  196,226,69,0,53,69,123,2,0          ; vpshufb       0x27b45(%rip),%ymm7,%ymm6        # 2bfe0 <_sk_srcover_rgba_8888_sse2_lowp+0x5d8>
+  DB  196,226,69,0,53,229,145,2,0         ; vpshufb       0x291e5(%rip),%ymm7,%ymm6        # 2d680 <_sk_srcover_rgba_8888_sse2_lowp+0x5cc>
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
   DB  196,193,76,89,240                   ; vmulps        %ymm8,%ymm6,%ymm6
   DB  197,197,114,215,24                  ; vpsrld        $0x18,%ymm7,%ymm7
@@ -4260,14 +4260,14 @@
   DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
   DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
   DB  196,66,101,144,4,128                ; vpgatherdd    %ymm3,(%r8,%ymm0,4),%ymm8
-  DB  197,189,219,5,49,122,2,0            ; vpand         0x27a31(%rip),%ymm8,%ymm0        # 2c000 <_sk_srcover_rgba_8888_sse2_lowp+0x5f8>
+  DB  197,189,219,5,209,144,2,0           ; vpand         0x290d1(%rip),%ymm8,%ymm0        # 2d6a0 <_sk_srcover_rgba_8888_sse2_lowp+0x5ec>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,29,84,119,2,0        ; vbroadcastss  0x27754(%rip),%ymm3        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,29,0,142,2,0         ; vbroadcastss  0x28e00(%rip),%ymm3        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,61,0,13,55,122,2,0          ; vpshufb       0x27a37(%rip),%ymm8,%ymm1        # 2c020 <_sk_srcover_rgba_8888_sse2_lowp+0x618>
+  DB  196,226,61,0,13,215,144,2,0         ; vpshufb       0x290d7(%rip),%ymm8,%ymm1        # 2d6c0 <_sk_srcover_rgba_8888_sse2_lowp+0x60c>
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,61,0,21,70,122,2,0          ; vpshufb       0x27a46(%rip),%ymm8,%ymm2        # 2c040 <_sk_srcover_rgba_8888_sse2_lowp+0x638>
+  DB  196,226,61,0,21,230,144,2,0         ; vpshufb       0x290e6(%rip),%ymm8,%ymm2        # 2d6e0 <_sk_srcover_rgba_8888_sse2_lowp+0x62c>
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  196,193,61,114,208,24               ; vpsrld        $0x18,%ymm8,%ymm8
@@ -4285,7 +4285,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,175,118,2,0         ; vbroadcastss  0x276af(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,91,141,2,0          ; vbroadcastss  0x28d5b(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,65,116,89,208                   ; vmulps        %ymm8,%ymm1,%ymm10
@@ -4339,7 +4339,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,219                 ; jmpq          ffffffffdc004710 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd8d08>
+  DB  233,255,255,255,219                 ; jmpq          ffffffffdc004710 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd765c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,205                             ; dec           %ebp
@@ -4359,14 +4359,14 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,86                              ; jne           478a <_sk_load_bgra_hsw+0x72>
   DB  196,129,126,111,28,152              ; vmovdqu       (%r8,%r11,4),%ymm3
-  DB  197,229,219,5,30,121,2,0            ; vpand         0x2791e(%rip),%ymm3,%ymm0        # 2c060 <_sk_srcover_rgba_8888_sse2_lowp+0x658>
+  DB  197,229,219,5,190,143,2,0           ; vpand         0x28fbe(%rip),%ymm3,%ymm0        # 2d700 <_sk_srcover_rgba_8888_sse2_lowp+0x64c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,225,117,2,0         ; vbroadcastss  0x275e1(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,141,140,2,0         ; vbroadcastss  0x28c8d(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,208                  ; vmulps        %ymm8,%ymm0,%ymm2
-  DB  196,226,101,0,5,35,121,2,0          ; vpshufb       0x27923(%rip),%ymm3,%ymm0        # 2c080 <_sk_srcover_rgba_8888_sse2_lowp+0x678>
+  DB  196,226,101,0,5,195,143,2,0         ; vpshufb       0x28fc3(%rip),%ymm3,%ymm0        # 2d720 <_sk_srcover_rgba_8888_sse2_lowp+0x66c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
   DB  196,193,124,89,200                  ; vmulps        %ymm8,%ymm0,%ymm1
-  DB  196,226,101,0,5,49,121,2,0          ; vpshufb       0x27931(%rip),%ymm3,%ymm0        # 2c0a0 <_sk_srcover_rgba_8888_sse2_lowp+0x698>
+  DB  196,226,101,0,5,209,143,2,0         ; vpshufb       0x28fd1(%rip),%ymm3,%ymm0        # 2d740 <_sk_srcover_rgba_8888_sse2_lowp+0x68c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  197,229,114,211,24                  ; vpsrld        $0x18,%ymm3,%ymm3
@@ -4440,14 +4440,14 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,86                              ; jne           48be <_sk_load_bgra_dst_hsw+0x72>
   DB  196,129,126,111,60,152              ; vmovdqu       (%r8,%r11,4),%ymm7
-  DB  197,197,219,37,74,120,2,0           ; vpand         0x2784a(%rip),%ymm7,%ymm4        # 2c0c0 <_sk_srcover_rgba_8888_sse2_lowp+0x6b8>
+  DB  197,197,219,37,234,142,2,0          ; vpand         0x28eea(%rip),%ymm7,%ymm4        # 2d760 <_sk_srcover_rgba_8888_sse2_lowp+0x6ac>
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,98,125,24,5,173,116,2,0         ; vbroadcastss  0x274ad(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,89,139,2,0          ; vbroadcastss  0x28b59(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,92,89,240                   ; vmulps        %ymm8,%ymm4,%ymm6
-  DB  196,226,69,0,37,79,120,2,0          ; vpshufb       0x2784f(%rip),%ymm7,%ymm4        # 2c0e0 <_sk_srcover_rgba_8888_sse2_lowp+0x6d8>
+  DB  196,226,69,0,37,239,142,2,0         ; vpshufb       0x28eef(%rip),%ymm7,%ymm4        # 2d780 <_sk_srcover_rgba_8888_sse2_lowp+0x6cc>
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
   DB  196,193,92,89,232                   ; vmulps        %ymm8,%ymm4,%ymm5
-  DB  196,226,69,0,37,93,120,2,0          ; vpshufb       0x2785d(%rip),%ymm7,%ymm4        # 2c100 <_sk_srcover_rgba_8888_sse2_lowp+0x6f8>
+  DB  196,226,69,0,37,253,142,2,0         ; vpshufb       0x28efd(%rip),%ymm7,%ymm4        # 2d7a0 <_sk_srcover_rgba_8888_sse2_lowp+0x6ec>
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
   DB  196,193,92,89,224                   ; vmulps        %ymm8,%ymm4,%ymm4
   DB  197,197,114,215,24                  ; vpsrld        $0x18,%ymm7,%ymm7
@@ -4529,14 +4529,14 @@
   DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
   DB  197,245,254,192                     ; vpaddd        %ymm0,%ymm1,%ymm0
   DB  196,66,101,144,4,128                ; vpgatherdd    %ymm3,(%r8,%ymm0,4),%ymm8
-  DB  197,189,219,5,73,119,2,0            ; vpand         0x27749(%rip),%ymm8,%ymm0        # 2c120 <_sk_srcover_rgba_8888_sse2_lowp+0x718>
+  DB  197,189,219,5,233,141,2,0           ; vpand         0x28de9(%rip),%ymm8,%ymm0        # 2d7c0 <_sk_srcover_rgba_8888_sse2_lowp+0x70c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,29,76,115,2,0        ; vbroadcastss  0x2734c(%rip),%ymm3        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,29,248,137,2,0       ; vbroadcastss  0x289f8(%rip),%ymm3        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,211                      ; vmulps        %ymm3,%ymm0,%ymm2
-  DB  196,226,61,0,5,79,119,2,0           ; vpshufb       0x2774f(%rip),%ymm8,%ymm0        # 2c140 <_sk_srcover_rgba_8888_sse2_lowp+0x738>
+  DB  196,226,61,0,5,239,141,2,0          ; vpshufb       0x28def(%rip),%ymm8,%ymm0        # 2d7e0 <_sk_srcover_rgba_8888_sse2_lowp+0x72c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
   DB  197,252,89,203                      ; vmulps        %ymm3,%ymm0,%ymm1
-  DB  196,226,61,0,5,94,119,2,0           ; vpshufb       0x2775e(%rip),%ymm8,%ymm0        # 2c160 <_sk_srcover_rgba_8888_sse2_lowp+0x758>
+  DB  196,226,61,0,5,254,141,2,0          ; vpshufb       0x28dfe(%rip),%ymm8,%ymm0        # 2d800 <_sk_srcover_rgba_8888_sse2_lowp+0x74c>
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
   DB  196,193,61,114,208,24               ; vpsrld        $0x18,%ymm8,%ymm8
@@ -4554,7 +4554,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,167,114,2,0         ; vbroadcastss  0x272a7(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,83,137,2,0          ; vbroadcastss  0x28953(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,108,89,200                   ; vmulps        %ymm8,%ymm2,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,65,116,89,208                   ; vmulps        %ymm8,%ymm1,%ymm10
@@ -4608,7 +4608,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,219                 ; jmpq          ffffffffdc004b18 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd9110>
+  DB  233,255,255,255,219                 ; jmpq          ffffffffdc004b18 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdbfd7a64>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,205                             ; dec           %ebp
@@ -4859,7 +4859,7 @@
   DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
   DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,21,40,110,2,0         ; vbroadcastss  0x26e28(%rip),%ymm10        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,98,125,24,21,212,132,2,0        ; vbroadcastss  0x284d4(%rip),%ymm10        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  196,193,124,89,194                  ; vmulps        %ymm10,%ymm0,%ymm0
   DB  197,185,109,202                     ; vpunpckhqdq   %xmm2,%xmm8,%xmm1
   DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
@@ -4949,7 +4949,7 @@
   DB  197,241,235,192                     ; vpor          %xmm0,%xmm1,%xmm0
   DB  196,226,125,51,192                  ; vpmovzxwd     %xmm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,151,108,2,0        ; vbroadcastss  0x26c97(%rip),%ymm9        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,98,125,24,13,67,131,2,0         ; vbroadcastss  0x28343(%rip),%ymm9        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,185,109,203                     ; vpunpckhqdq   %xmm3,%xmm8,%xmm1
   DB  197,225,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm3
@@ -4965,7 +4965,7 @@
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,183,107,2,0       ; vbroadcastss  0x26bb7(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,99,130,2,0        ; vbroadcastss  0x28263(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
   DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
@@ -5039,7 +5039,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  73,99,193                           ; movslq        %r9d,%rax
-  DB  196,98,125,24,5,128,107,2,0         ; vbroadcastss  0x26b80(%rip),%ymm8        # 2bdd4 <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
+  DB  196,98,125,24,5,44,130,2,0          ; vbroadcastss  0x2822c(%rip),%ymm8        # 2d480 <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,67,125,25,202,1                 ; vextractf128  $0x1,%ymm9,%xmm10
@@ -5296,7 +5296,7 @@
   DB  255,92,255,255                      ; lcall         *-0x1(%rdi,%rdi,8)
   DB  255,70,255                          ; incl          -0x1(%rsi)
   DB  255                                 ; (bad)
-  DB  255,53,255,255,255,40               ; pushq         0x28ffffff(%rip)        # 290056cc <_sk_srcover_rgba_8888_sse2_lowp+0x28fd9cc4>
+  DB  255,53,255,255,255,40               ; pushq         0x28ffffff(%rip)        # 290056cc <_sk_srcover_rgba_8888_sse2_lowp+0x28fd8618>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -5381,7 +5381,7 @@
   DB  196,65,124,92,218                   ; vsubps        %ymm10,%ymm0,%ymm11
   DB  196,193,58,88,192                   ; vaddss        %xmm8,%xmm8,%xmm0
   DB  196,98,125,24,192                   ; vbroadcastss  %xmm0,%ymm8
-  DB  197,178,89,5,173,100,2,0            ; vmulss        0x264ad(%rip),%xmm9,%xmm0        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,178,89,5,89,123,2,0             ; vmulss        0x27b59(%rip),%xmm9,%xmm0        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,226,125,24,192                  ; vbroadcastss  %xmm0,%ymm0
   DB  197,164,89,192                      ; vmulps        %ymm0,%ymm11,%ymm0
   DB  196,227,125,8,192,1                 ; vroundps      $0x1,%ymm0,%ymm0
@@ -5402,7 +5402,7 @@
   DB  196,65,116,92,218                   ; vsubps        %ymm10,%ymm1,%ymm11
   DB  196,193,58,88,200                   ; vaddss        %xmm8,%xmm8,%xmm1
   DB  196,98,125,24,193                   ; vbroadcastss  %xmm1,%ymm8
-  DB  197,178,89,13,92,100,2,0            ; vmulss        0x2645c(%rip),%xmm9,%xmm1        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,178,89,13,8,123,2,0             ; vmulss        0x27b08(%rip),%xmm9,%xmm1        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,226,125,24,201                  ; vbroadcastss  %xmm1,%ymm1
   DB  197,164,89,201                      ; vmulps        %ymm1,%ymm11,%ymm1
   DB  196,227,125,8,201,1                 ; vroundps      $0x1,%ymm1,%ymm1
@@ -5418,7 +5418,7 @@
 _sk_clamp_x_1_hsw LABEL PROC
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,36,100,2,0          ; vbroadcastss  0x26424(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,208,122,2,0         ; vbroadcastss  0x27ad0(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -5429,16 +5429,16 @@
   DB  196,193,124,92,192                  ; vsubps        %ymm8,%ymm0,%ymm0
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,254,99,2,0          ; vbroadcastss  0x263fe(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,170,122,2,0         ; vbroadcastss  0x27aaa(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_mirror_x_1_hsw
 _sk_mirror_x_1_hsw LABEL PROC
-  DB  196,98,125,24,5,4,100,2,0           ; vbroadcastss  0x26404(%rip),%ymm8        # 2bcd0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  DB  196,98,125,24,5,176,122,2,0         ; vbroadcastss  0x27ab0(%rip),%ymm8        # 2d37c <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,13,218,99,2,0         ; vbroadcastss  0x263da(%rip),%ymm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,13,134,122,2,0        ; vbroadcastss  0x27a86(%rip),%ymm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,124,89,201                   ; vmulps        %ymm9,%ymm0,%ymm9
   DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
   DB  196,65,52,88,201                    ; vaddps        %ymm9,%ymm9,%ymm9
@@ -5448,18 +5448,18 @@
   DB  197,60,92,200                       ; vsubps        %ymm0,%ymm8,%ymm9
   DB  197,180,84,192                      ; vandps        %ymm0,%ymm9,%ymm0
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,170,99,2,0          ; vbroadcastss  0x263aa(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,86,122,2,0          ; vbroadcastss  0x27a56(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_luminance_to_alpha_hsw
 _sk_luminance_to_alpha_hsw LABEL PROC
-  DB  196,226,125,24,29,184,100,2,0       ; vbroadcastss  0x264b8(%rip),%ymm3        # 2bdd8 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
-  DB  196,98,125,24,5,179,100,2,0         ; vbroadcastss  0x264b3(%rip),%ymm8        # 2bddc <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
+  DB  196,226,125,24,29,100,123,2,0       ; vbroadcastss  0x27b64(%rip),%ymm3        # 2d484 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
+  DB  196,98,125,24,5,95,123,2,0          ; vbroadcastss  0x27b5f(%rip),%ymm8        # 2d488 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  196,226,125,184,203                 ; vfmadd231ps   %ymm3,%ymm0,%ymm1
-  DB  196,226,125,24,29,164,100,2,0       ; vbroadcastss  0x264a4(%rip),%ymm3        # 2bde0 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
+  DB  196,226,125,24,29,80,123,2,0        ; vbroadcastss  0x27b50(%rip),%ymm3        # 2d48c <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
   DB  196,226,109,168,217                 ; vfmadd213ps   %ymm1,%ymm2,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
@@ -5713,14 +5713,14 @@
 
 PUBLIC _sk_gauss_a_to_rgba_hsw
 _sk_gauss_a_to_rgba_hsw LABEL PROC
-  DB  196,226,125,24,5,134,96,2,0         ; vbroadcastss  0x26086(%rip),%ymm0        # 2bde4 <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
-  DB  196,226,125,24,13,129,96,2,0        ; vbroadcastss  0x26081(%rip),%ymm1        # 2bde8 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
+  DB  196,226,125,24,5,50,119,2,0         ; vbroadcastss  0x27732(%rip),%ymm0        # 2d490 <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
+  DB  196,226,125,24,13,45,119,2,0        ; vbroadcastss  0x2772d(%rip),%ymm1        # 2d494 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
   DB  196,226,101,168,200                 ; vfmadd213ps   %ymm0,%ymm3,%ymm1
-  DB  196,226,125,24,5,119,96,2,0         ; vbroadcastss  0x26077(%rip),%ymm0        # 2bdec <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
+  DB  196,226,125,24,5,35,119,2,0         ; vbroadcastss  0x27723(%rip),%ymm0        # 2d498 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
   DB  196,226,101,184,193                 ; vfmadd231ps   %ymm1,%ymm3,%ymm0
-  DB  196,226,125,24,13,109,96,2,0        ; vbroadcastss  0x2606d(%rip),%ymm1        # 2bdf0 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
+  DB  196,226,125,24,13,25,119,2,0        ; vbroadcastss  0x27719(%rip),%ymm1        # 2d49c <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
   DB  196,226,101,184,200                 ; vfmadd231ps   %ymm0,%ymm3,%ymm1
-  DB  196,226,125,24,5,99,96,2,0          ; vbroadcastss  0x26063(%rip),%ymm0        # 2bdf4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
+  DB  196,226,125,24,5,15,119,2,0         ; vbroadcastss  0x2770f(%rip),%ymm0        # 2d4a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
   DB  196,226,101,184,193                 ; vfmadd231ps   %ymm1,%ymm3,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
@@ -5737,7 +5737,7 @@
   DB  76,139,72,72                        ; mov           0x48(%rax),%r9
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  65,186,1,0,0,0                      ; mov           $0x1,%r10d
-  DB  196,226,125,24,21,240,94,2,0        ; vbroadcastss  0x25ef0(%rip),%ymm2        # 2bcbc <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  DB  196,226,125,24,21,156,117,2,0       ; vbroadcastss  0x2759c(%rip),%ymm2        # 2d368 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   DB  196,65,53,239,201                   ; vpxor         %ymm9,%ymm9,%ymm9
   DB  196,130,125,24,28,145               ; vbroadcastss  (%r9,%r10,4),%ymm3
   DB  197,228,194,216,2                   ; vcmpleps      %ymm0,%ymm3,%ymm3
@@ -5836,24 +5836,24 @@
   DB  196,65,52,95,226                    ; vmaxps        %ymm10,%ymm9,%ymm12
   DB  196,65,36,94,220                    ; vdivps        %ymm12,%ymm11,%ymm11
   DB  196,65,36,89,227                    ; vmulps        %ymm11,%ymm11,%ymm12
-  DB  196,98,125,24,45,127,94,2,0         ; vbroadcastss  0x25e7f(%rip),%ymm13        # 2bdf8 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
-  DB  196,98,125,24,53,122,94,2,0         ; vbroadcastss  0x25e7a(%rip),%ymm14        # 2bdfc <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
+  DB  196,98,125,24,45,43,117,2,0         ; vbroadcastss  0x2752b(%rip),%ymm13        # 2d4a4 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
+  DB  196,98,125,24,53,38,117,2,0         ; vbroadcastss  0x27526(%rip),%ymm14        # 2d4a8 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
   DB  196,66,29,184,245                   ; vfmadd231ps   %ymm13,%ymm12,%ymm14
-  DB  196,98,125,24,45,112,94,2,0         ; vbroadcastss  0x25e70(%rip),%ymm13        # 2be00 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
+  DB  196,98,125,24,45,28,117,2,0         ; vbroadcastss  0x2751c(%rip),%ymm13        # 2d4ac <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
   DB  196,66,29,184,238                   ; vfmadd231ps   %ymm14,%ymm12,%ymm13
-  DB  196,98,125,24,53,102,94,2,0         ; vbroadcastss  0x25e66(%rip),%ymm14        # 2be04 <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
+  DB  196,98,125,24,53,18,117,2,0         ; vbroadcastss  0x27512(%rip),%ymm14        # 2d4b0 <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
   DB  196,66,29,184,245                   ; vfmadd231ps   %ymm13,%ymm12,%ymm14
   DB  196,65,36,89,222                    ; vmulps        %ymm14,%ymm11,%ymm11
   DB  196,65,52,194,202,1                 ; vcmpltps      %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,81,94,2,0          ; vbroadcastss  0x25e51(%rip),%ymm10        # 2be08 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
+  DB  196,98,125,24,21,253,116,2,0        ; vbroadcastss  0x274fd(%rip),%ymm10        # 2d4b4 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
   DB  196,65,44,92,211                    ; vsubps        %ymm11,%ymm10,%ymm10
   DB  196,67,37,74,202,144                ; vblendvps     %ymm9,%ymm10,%ymm11,%ymm9
   DB  196,193,124,194,192,1               ; vcmpltps      %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,21,227,92,2,0         ; vbroadcastss  0x25ce3(%rip),%ymm10        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,21,143,115,2,0        ; vbroadcastss  0x2738f(%rip),%ymm10        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,92,209                    ; vsubps        %ymm9,%ymm10,%ymm10
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  196,65,116,194,200,1                ; vcmpltps      %ymm8,%ymm1,%ymm9
-  DB  196,98,125,24,21,205,92,2,0         ; vbroadcastss  0x25ccd(%rip),%ymm10        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,21,121,115,2,0        ; vbroadcastss  0x27379(%rip),%ymm10        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,44,92,208                       ; vsubps        %ymm0,%ymm10,%ymm10
   DB  196,195,125,74,194,144              ; vblendvps     %ymm9,%ymm10,%ymm0,%ymm0
   DB  196,65,124,194,200,3                ; vcmpunordps   %ymm8,%ymm0,%ymm9
@@ -5877,23 +5877,23 @@
   DB  197,50,89,80,76                     ; vmulss        0x4c(%rax),%xmm9,%xmm10
   DB  196,66,125,24,210                   ; vbroadcastss  %xmm10,%ymm10
   DB  197,44,88,208                       ; vaddps        %ymm0,%ymm10,%ymm10
-  DB  196,98,125,24,29,209,93,2,0         ; vbroadcastss  0x25dd1(%rip),%ymm11        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,29,125,116,2,0        ; vbroadcastss  0x2747d(%rip),%ymm11        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
   DB  197,116,89,217                      ; vmulps        %ymm1,%ymm1,%ymm11
   DB  196,98,125,184,216                  ; vfmadd231ps   %ymm0,%ymm0,%ymm11
   DB  196,193,50,89,193                   ; vmulss        %xmm9,%xmm9,%xmm0
   DB  196,226,125,24,192                  ; vbroadcastss  %xmm0,%ymm0
   DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,13,176,93,2,0         ; vbroadcastss  0x25db0(%rip),%ymm9        # 2be10 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  DB  196,98,125,24,13,92,116,2,0         ; vbroadcastss  0x2745c(%rip),%ymm9        # 2d4bc <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  196,194,45,184,194                  ; vfmadd231ps   %ymm10,%ymm10,%ymm0
   DB  197,252,81,192                      ; vsqrtps       %ymm0,%ymm0
   DB  196,98,125,24,64,68                 ; vbroadcastss  0x44(%rax),%ymm8
-  DB  196,98,125,24,13,147,93,2,0         ; vbroadcastss  0x25d93(%rip),%ymm9        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,13,63,116,2,0         ; vbroadcastss  0x2743f(%rip),%ymm9        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,65,44,87,201                    ; vxorps        %ymm9,%ymm10,%ymm9
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,32,92,2,0          ; vbroadcastss  0x25c20(%rip),%ymm11        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,29,204,114,2,0        ; vbroadcastss  0x272cc(%rip),%ymm11        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  196,65,44,89,208                    ; vmulps        %ymm8,%ymm10,%ymm10
   DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
@@ -5910,23 +5910,23 @@
   DB  197,50,89,80,76                     ; vmulss        0x4c(%rax),%xmm9,%xmm10
   DB  196,66,125,24,210                   ; vbroadcastss  %xmm10,%ymm10
   DB  197,44,88,208                       ; vaddps        %ymm0,%ymm10,%ymm10
-  DB  196,98,125,24,29,57,93,2,0          ; vbroadcastss  0x25d39(%rip),%ymm11        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,29,229,115,2,0        ; vbroadcastss  0x273e5(%rip),%ymm11        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
   DB  197,116,89,217                      ; vmulps        %ymm1,%ymm1,%ymm11
   DB  196,98,125,184,216                  ; vfmadd231ps   %ymm0,%ymm0,%ymm11
   DB  196,193,50,89,193                   ; vmulss        %xmm9,%xmm9,%xmm0
   DB  196,226,125,24,192                  ; vbroadcastss  %xmm0,%ymm0
   DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,13,24,93,2,0          ; vbroadcastss  0x25d18(%rip),%ymm9        # 2be10 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  DB  196,98,125,24,13,196,115,2,0        ; vbroadcastss  0x273c4(%rip),%ymm9        # 2d4bc <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  196,194,45,184,194                  ; vfmadd231ps   %ymm10,%ymm10,%ymm0
   DB  197,252,81,192                      ; vsqrtps       %ymm0,%ymm0
   DB  196,98,125,24,64,68                 ; vbroadcastss  0x44(%rax),%ymm8
-  DB  196,98,125,24,13,251,92,2,0         ; vbroadcastss  0x25cfb(%rip),%ymm9        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,13,167,115,2,0        ; vbroadcastss  0x273a7(%rip),%ymm9        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,65,44,87,201                    ; vxorps        %ymm9,%ymm10,%ymm9
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,136,91,2,0         ; vbroadcastss  0x25b88(%rip),%ymm11        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,29,52,114,2,0         ; vbroadcastss  0x27234(%rip),%ymm11        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  196,65,44,89,208                    ; vmulps        %ymm8,%ymm10,%ymm10
   DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
@@ -5942,14 +5942,14 @@
   DB  197,58,89,72,76                     ; vmulss        0x4c(%rax),%xmm8,%xmm9
   DB  196,66,125,24,201                   ; vbroadcastss  %xmm9,%ymm9
   DB  197,52,88,200                       ; vaddps        %ymm0,%ymm9,%ymm9
-  DB  196,98,125,24,21,167,92,2,0         ; vbroadcastss  0x25ca7(%rip),%ymm10        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,21,83,115,2,0         ; vbroadcastss  0x27353(%rip),%ymm10        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
   DB  197,116,89,209                      ; vmulps        %ymm1,%ymm1,%ymm10
   DB  196,98,125,184,208                  ; vfmadd231ps   %ymm0,%ymm0,%ymm10
   DB  196,193,58,89,192                   ; vmulss        %xmm8,%xmm8,%xmm0
   DB  196,226,125,24,192                  ; vbroadcastss  %xmm0,%ymm0
   DB  197,172,92,192                      ; vsubps        %ymm0,%ymm10,%ymm0
-  DB  196,98,125,24,5,138,92,2,0          ; vbroadcastss  0x25c8a(%rip),%ymm8        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,5,54,115,2,0          ; vbroadcastss  0x27336(%rip),%ymm8        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,193,124,87,192                  ; vxorps        %ymm8,%ymm0,%ymm0
   DB  196,193,124,94,193                  ; vdivps        %ymm9,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -5986,7 +5986,7 @@
 PUBLIC _sk_save_xy_hsw
 _sk_save_xy_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,181,90,2,0          ; vbroadcastss  0x25ab5(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,97,113,2,0          ; vbroadcastss  0x27161(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,124,88,200                   ; vaddps        %ymm8,%ymm0,%ymm9
   DB  196,67,125,8,209,1                  ; vroundps      $0x1,%ymm9,%ymm10
   DB  196,65,52,92,202                    ; vsubps        %ymm10,%ymm9,%ymm9
@@ -6016,9 +6016,9 @@
 PUBLIC _sk_bilinear_nx_hsw
 _sk_bilinear_nx_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,163,91,2,0         ; vbroadcastss  0x25ba3(%rip),%ymm0        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,5,79,114,2,0         ; vbroadcastss  0x2724f(%rip),%ymm0        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,54,90,2,0           ; vbroadcastss  0x25a36(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,226,112,2,0         ; vbroadcastss  0x270e2(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6027,7 +6027,7 @@
 PUBLIC _sk_bilinear_px_hsw
 _sk_bilinear_px_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,19,90,2,0          ; vbroadcastss  0x25a13(%rip),%ymm0        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,5,191,112,2,0        ; vbroadcastss  0x270bf(%rip),%ymm0        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
   DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -6037,9 +6037,9 @@
 PUBLIC _sk_bilinear_ny_hsw
 _sk_bilinear_ny_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,84,91,2,0         ; vbroadcastss  0x25b54(%rip),%ymm1        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,13,0,114,2,0         ; vbroadcastss  0x27200(%rip),%ymm1        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,230,89,2,0          ; vbroadcastss  0x259e6(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,146,112,2,0         ; vbroadcastss  0x27092(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6048,7 +6048,7 @@
 PUBLIC _sk_bilinear_py_hsw
 _sk_bilinear_py_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,195,89,2,0        ; vbroadcastss  0x259c3(%rip),%ymm1        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,13,111,112,2,0       ; vbroadcastss  0x2706f(%rip),%ymm1        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
   DB  197,124,16,128,192,0,0,0            ; vmovups       0xc0(%rax),%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -6058,13 +6058,13 @@
 PUBLIC _sk_bicubic_n3x_hsw
 _sk_bicubic_n3x_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,7,91,2,0           ; vbroadcastss  0x25b07(%rip),%ymm0        # 2be1c <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  DB  196,226,125,24,5,179,113,2,0        ; vbroadcastss  0x271b3(%rip),%ymm0        # 2d4c8 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,150,89,2,0          ; vbroadcastss  0x25996(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,66,112,2,0          ; vbroadcastss  0x27042(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,244,89,2,0         ; vbroadcastss  0x259f4(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  DB  196,98,125,24,29,223,90,2,0         ; vbroadcastss  0x25adf(%rip),%ymm11        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,160,112,2,0        ; vbroadcastss  0x270a0(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,29,139,113,2,0        ; vbroadcastss  0x2718b(%rip),%ymm11        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,66,61,168,218                   ; vfmadd213ps   %ymm10,%ymm8,%ymm11
   DB  196,65,36,89,193                    ; vmulps        %ymm9,%ymm11,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -6074,16 +6074,16 @@
 PUBLIC _sk_bicubic_n1x_hsw
 _sk_bicubic_n1x_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,182,90,2,0         ; vbroadcastss  0x25ab6(%rip),%ymm0        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,5,98,113,2,0         ; vbroadcastss  0x27162(%rip),%ymm0        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,73,89,2,0           ; vbroadcastss  0x25949(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,245,111,2,0         ; vbroadcastss  0x26ff5(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,164,90,2,0         ; vbroadcastss  0x25aa4(%rip),%ymm9        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  DB  196,98,125,24,21,159,90,2,0         ; vbroadcastss  0x25a9f(%rip),%ymm10        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,13,80,113,2,0         ; vbroadcastss  0x27150(%rip),%ymm9        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,21,75,113,2,0         ; vbroadcastss  0x2714b(%rip),%ymm10        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,29,89,2,0          ; vbroadcastss  0x2591d(%rip),%ymm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,13,201,111,2,0        ; vbroadcastss  0x26fc9(%rip),%ymm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,135,90,2,0         ; vbroadcastss  0x25a87(%rip),%ymm10        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,21,51,113,2,0         ; vbroadcastss  0x27133(%rip),%ymm10        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
   DB  197,124,17,144,0,1,0,0              ; vmovups       %ymm10,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6092,14 +6092,14 @@
 PUBLIC _sk_bicubic_p1x_hsw
 _sk_bicubic_p1x_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,243,88,2,0          ; vbroadcastss  0x258f3(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,159,111,2,0         ; vbroadcastss  0x26f9f(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,136,128,0,0,0            ; vmovups       0x80(%rax),%ymm9
-  DB  196,98,125,24,21,78,90,2,0          ; vbroadcastss  0x25a4e(%rip),%ymm10        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  DB  196,98,125,24,29,73,90,2,0          ; vbroadcastss  0x25a49(%rip),%ymm11        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,21,250,112,2,0        ; vbroadcastss  0x270fa(%rip),%ymm10        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,29,245,112,2,0        ; vbroadcastss  0x270f5(%rip),%ymm11        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,58,90,2,0           ; vbroadcastss  0x25a3a(%rip),%ymm8        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,5,230,112,2,0         ; vbroadcastss  0x270e6(%rip),%ymm8        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,66,53,184,195                   ; vfmadd231ps   %ymm11,%ymm9,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6108,12 +6108,12 @@
 PUBLIC _sk_bicubic_p3x_hsw
 _sk_bicubic_p3x_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,22,90,2,0          ; vbroadcastss  0x25a16(%rip),%ymm0        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,226,125,24,5,194,112,2,0        ; vbroadcastss  0x270c2(%rip),%ymm0        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
   DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,4,89,2,0           ; vbroadcastss  0x25904(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  DB  196,98,125,24,29,239,89,2,0         ; vbroadcastss  0x259ef(%rip),%ymm11        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,176,111,2,0        ; vbroadcastss  0x26fb0(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,29,155,112,2,0        ; vbroadcastss  0x2709b(%rip),%ymm11        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,66,61,168,218                   ; vfmadd213ps   %ymm10,%ymm8,%ymm11
   DB  196,65,52,89,195                    ; vmulps        %ymm11,%ymm9,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -6123,13 +6123,13 @@
 PUBLIC _sk_bicubic_n3y_hsw
 _sk_bicubic_n3y_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,202,89,2,0        ; vbroadcastss  0x259ca(%rip),%ymm1        # 2be1c <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  DB  196,226,125,24,13,118,112,2,0       ; vbroadcastss  0x27076(%rip),%ymm1        # 2d4c8 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,88,88,2,0           ; vbroadcastss  0x25858(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,4,111,2,0           ; vbroadcastss  0x26f04(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,182,88,2,0         ; vbroadcastss  0x258b6(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  DB  196,98,125,24,29,161,89,2,0         ; vbroadcastss  0x259a1(%rip),%ymm11        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,98,111,2,0         ; vbroadcastss  0x26f62(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,29,77,112,2,0         ; vbroadcastss  0x2704d(%rip),%ymm11        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,66,61,168,218                   ; vfmadd213ps   %ymm10,%ymm8,%ymm11
   DB  196,65,36,89,193                    ; vmulps        %ymm9,%ymm11,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -6139,16 +6139,16 @@
 PUBLIC _sk_bicubic_n1y_hsw
 _sk_bicubic_n1y_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,120,89,2,0        ; vbroadcastss  0x25978(%rip),%ymm1        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,13,36,112,2,0        ; vbroadcastss  0x27024(%rip),%ymm1        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,10,88,2,0           ; vbroadcastss  0x2580a(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,182,110,2,0         ; vbroadcastss  0x26eb6(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,101,89,2,0         ; vbroadcastss  0x25965(%rip),%ymm9        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  DB  196,98,125,24,21,96,89,2,0          ; vbroadcastss  0x25960(%rip),%ymm10        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,13,17,112,2,0         ; vbroadcastss  0x27011(%rip),%ymm9        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,21,12,112,2,0         ; vbroadcastss  0x2700c(%rip),%ymm10        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,66,61,168,209                   ; vfmadd213ps   %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,13,222,87,2,0         ; vbroadcastss  0x257de(%rip),%ymm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,13,138,110,2,0        ; vbroadcastss  0x26e8a(%rip),%ymm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,66,61,184,202                   ; vfmadd231ps   %ymm10,%ymm8,%ymm9
-  DB  196,98,125,24,21,72,89,2,0          ; vbroadcastss  0x25948(%rip),%ymm10        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,21,244,111,2,0        ; vbroadcastss  0x26ff4(%rip),%ymm10        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,66,61,184,209                   ; vfmadd231ps   %ymm9,%ymm8,%ymm10
   DB  197,124,17,144,64,1,0,0             ; vmovups       %ymm10,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6157,14 +6157,14 @@
 PUBLIC _sk_bicubic_p1y_hsw
 _sk_bicubic_p1y_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,180,87,2,0          ; vbroadcastss  0x257b4(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,96,110,2,0          ; vbroadcastss  0x26e60(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,188,88,72,64                    ; vaddps        0x40(%rax),%ymm8,%ymm1
   DB  197,124,16,136,192,0,0,0            ; vmovups       0xc0(%rax),%ymm9
-  DB  196,98,125,24,21,14,89,2,0          ; vbroadcastss  0x2590e(%rip),%ymm10        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
-  DB  196,98,125,24,29,9,89,2,0           ; vbroadcastss  0x25909(%rip),%ymm11        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,21,186,111,2,0        ; vbroadcastss  0x26fba(%rip),%ymm10        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,29,181,111,2,0        ; vbroadcastss  0x26fb5(%rip),%ymm11        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,66,53,168,218                   ; vfmadd213ps   %ymm10,%ymm9,%ymm11
   DB  196,66,53,168,216                   ; vfmadd213ps   %ymm8,%ymm9,%ymm11
-  DB  196,98,125,24,5,250,88,2,0          ; vbroadcastss  0x258fa(%rip),%ymm8        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,5,166,111,2,0         ; vbroadcastss  0x26fa6(%rip),%ymm8        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,66,53,184,195                   ; vfmadd231ps   %ymm11,%ymm9,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -6173,12 +6173,12 @@
 PUBLIC _sk_bicubic_p3y_hsw
 _sk_bicubic_p3y_hsw LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,214,88,2,0        ; vbroadcastss  0x258d6(%rip),%ymm1        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,226,125,24,13,130,111,2,0       ; vbroadcastss  0x26f82(%rip),%ymm1        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
   DB  197,124,16,128,192,0,0,0            ; vmovups       0xc0(%rax),%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,195,87,2,0         ; vbroadcastss  0x257c3(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
-  DB  196,98,125,24,29,174,88,2,0         ; vbroadcastss  0x258ae(%rip),%ymm11        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,111,110,2,0        ; vbroadcastss  0x26e6f(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,29,90,111,2,0         ; vbroadcastss  0x26f5a(%rip),%ymm11        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,66,61,168,218                   ; vfmadd213ps   %ymm10,%ymm8,%ymm11
   DB  196,65,52,89,195                    ; vmulps        %ymm11,%ymm9,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -6306,7 +6306,7 @@
   DB  196,98,93,64,218                    ; vpmulld       %ymm2,%ymm4,%ymm11
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  197,165,254,193                     ; vpaddd        %ymm1,%ymm11,%ymm0
-  DB  196,98,125,88,61,125,86,2,0         ; vpbroadcastd  0x2567d(%rip),%ymm15        # 2be34 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  DB  196,98,125,88,61,41,109,2,0         ; vpbroadcastd  0x26d29(%rip),%ymm15        # 2d4e0 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   DB  196,194,125,64,199                  ; vpmulld       %ymm15,%ymm0,%ymm0
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
   DB  197,237,118,210                     ; vpcmpeqd      %ymm2,%ymm2,%ymm2
@@ -6318,13 +6318,13 @@
   DB  197,213,118,237                     ; vpcmpeqd      %ymm5,%ymm5,%ymm5
   DB  196,226,85,146,60,144               ; vgatherdps    %ymm5,(%rax,%ymm2,4),%ymm7
   DB  197,252,17,188,36,128,2,0,0         ; vmovups       %ymm7,0x280(%rsp)
-  DB  196,226,125,88,61,199,84,2,0        ; vpbroadcastd  0x254c7(%rip),%ymm7        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,226,125,88,61,115,107,2,0       ; vpbroadcastd  0x26b73(%rip),%ymm7        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  197,253,254,199                     ; vpaddd        %ymm7,%ymm0,%ymm0
   DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
   DB  197,213,118,237                     ; vpcmpeqd      %ymm5,%ymm5,%ymm5
   DB  196,226,85,146,20,128               ; vgatherdps    %ymm5,(%rax,%ymm0,4),%ymm2
   DB  197,252,17,148,36,160,1,0,0         ; vmovups       %ymm2,0x1a0(%rsp)
-  DB  196,226,125,24,5,15,86,2,0          ; vbroadcastss  0x2560f(%rip),%ymm0        # 2be30 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  DB  196,226,125,24,5,187,108,2,0        ; vbroadcastss  0x26cbb(%rip),%ymm0        # 2d4dc <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   DB  197,172,88,232                      ; vaddps        %ymm0,%ymm10,%ymm5
   DB  197,254,91,237                      ; vcvttps2dq    %ymm5,%ymm5
   DB  196,226,93,64,213                   ; vpmulld       %ymm5,%ymm4,%ymm2
@@ -6584,7 +6584,7 @@
   DB  197,254,127,132,36,128,1,0,0        ; vmovdqu       %ymm0,0x180(%rsp)
   DB  196,98,109,64,200                   ; vpmulld       %ymm0,%ymm2,%ymm9
   DB  197,181,254,199                     ; vpaddd        %ymm7,%ymm9,%ymm0
-  DB  196,98,125,88,21,145,80,2,0         ; vpbroadcastd  0x25091(%rip),%ymm10        # 2be34 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  DB  196,98,125,88,21,61,103,2,0         ; vpbroadcastd  0x2673d(%rip),%ymm10        # 2d4e0 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   DB  196,194,125,64,194                  ; vpmulld       %ymm10,%ymm0,%ymm0
   DB  197,213,118,237                     ; vpcmpeqd      %ymm5,%ymm5,%ymm5
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
@@ -6596,13 +6596,13 @@
   DB  196,65,36,87,219                    ; vxorps        %ymm11,%ymm11,%ymm11
   DB  196,98,61,146,28,168                ; vgatherdps    %ymm8,(%rax,%ymm5,4),%ymm11
   DB  197,124,17,156,36,64,1,0,0          ; vmovups       %ymm11,0x140(%rsp)
-  DB  196,98,125,88,29,216,78,2,0         ; vpbroadcastd  0x24ed8(%rip),%ymm11        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,98,125,88,29,132,101,2,0        ; vpbroadcastd  0x26584(%rip),%ymm11        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  196,193,125,254,195                 ; vpaddd        %ymm11,%ymm0,%ymm0
   DB  196,65,61,118,192                   ; vpcmpeqd      %ymm8,%ymm8,%ymm8
   DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
   DB  196,226,61,146,44,128               ; vgatherdps    %ymm8,(%rax,%ymm0,4),%ymm5
   DB  197,252,17,172,36,160,0,0,0         ; vmovups       %ymm5,0xa0(%rsp)
-  DB  196,226,125,24,5,30,80,2,0          ; vbroadcastss  0x2501e(%rip),%ymm0        # 2be30 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  DB  196,226,125,24,5,202,102,2,0        ; vbroadcastss  0x266ca(%rip),%ymm0        # 2d4dc <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   DB  197,116,88,192                      ; vaddps        %ymm0,%ymm1,%ymm8
   DB  196,65,126,91,192                   ; vcvttps2dq    %ymm8,%ymm8
   DB  196,194,109,64,232                  ; vpmulld       %ymm8,%ymm2,%ymm5
@@ -7035,7 +7035,7 @@
   DB  196,193,100,92,210                  ; vsubps        %ymm10,%ymm3,%ymm2
   DB  196,194,77,168,210                  ; vfmadd213ps   %ymm10,%ymm6,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,202,68,2,0        ; vbroadcastss  0x244ca(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,118,91,2,0        ; vbroadcastss  0x25b76(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,16,164,36,224,6,0,0         ; vmovups       0x6e0(%rsp),%ymm4
   DB  197,252,16,172,36,0,7,0,0           ; vmovups       0x700(%rsp),%ymm5
   DB  197,252,16,180,36,32,7,0,0          ; vmovups       0x720(%rsp),%ymm6
@@ -7156,10 +7156,10 @@
   DB  197,249,112,201,0                   ; vpshufd       $0x0,%xmm1,%xmm1
   DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,219,66,2,0        ; vbroadcastss  0x242db(%rip),%ymm2        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,21,135,89,2,0        ; vbroadcastss  0x25987(%rip),%ymm2        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,21,208,66,2,0        ; vbroadcastss  0x242d0(%rip),%ymm2        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,21,124,89,2,0        ; vbroadcastss  0x2597c(%rip),%ymm2        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,228,87,219                      ; vxorps        %ymm3,%ymm3,%ymm3
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
   DB  197,212,87,237                      ; vxorps        %ymm5,%ymm5,%ymm5
@@ -7171,46 +7171,46 @@
 _sk_dither_avx LABEL PROC
   DB  197,121,110,194                     ; vmovd         %edx,%xmm8
   DB  196,65,121,112,192,0                ; vpshufd       $0x0,%xmm8,%xmm8
-  DB  197,57,254,13,240,73,2,0            ; vpaddd        0x249f0(%rip),%xmm8,%xmm9        # 2c400 <_sk_srcover_rgba_8888_sse2_lowp+0x9f8>
-  DB  197,57,254,5,248,73,2,0             ; vpaddd        0x249f8(%rip),%xmm8,%xmm8        # 2c410 <_sk_srcover_rgba_8888_sse2_lowp+0xa08>
+  DB  197,57,254,13,208,96,2,0            ; vpaddd        0x260d0(%rip),%xmm8,%xmm9        # 2dae0 <_sk_srcover_rgba_8888_sse2_lowp+0xa2c>
+  DB  197,57,254,5,216,96,2,0             ; vpaddd        0x260d8(%rip),%xmm8,%xmm8        # 2daf0 <_sk_srcover_rgba_8888_sse2_lowp+0xa3c>
   DB  196,67,53,24,208,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm10
   DB  197,121,110,217                     ; vmovd         %ecx,%xmm11
   DB  196,65,121,112,219,0                ; vpshufd       $0x0,%xmm11,%xmm11
   DB  196,67,37,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm11,%ymm11
   DB  196,65,36,87,218                    ; vxorps        %ymm10,%ymm11,%ymm11
-  DB  196,98,125,24,21,128,66,2,0         ; vbroadcastss  0x24280(%rip),%ymm10        # 2bcbc <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  DB  196,98,125,24,21,44,89,2,0          ; vbroadcastss  0x2592c(%rip),%ymm10        # 2d368 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   DB  196,65,36,84,210                    ; vandps        %ymm10,%ymm11,%ymm10
   DB  196,193,25,114,242,5                ; vpslld        $0x5,%xmm10,%xmm12
   DB  196,67,125,25,210,1                 ; vextractf128  $0x1,%ymm10,%xmm10
   DB  196,193,41,114,242,5                ; vpslld        $0x5,%xmm10,%xmm10
   DB  196,67,29,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm12,%ymm10
-  DB  197,123,18,37,175,81,2,0            ; vmovddup      0x251af(%rip),%xmm12        # 2cc10 <_sk_srcover_rgba_8888_sse2_lowp+0x1208>
+  DB  197,123,18,37,143,104,2,0           ; vmovddup      0x2688f(%rip),%xmm12        # 2e2f0 <_sk_srcover_rgba_8888_sse2_lowp+0x123c>
   DB  196,65,49,219,236                   ; vpand         %xmm12,%xmm9,%xmm13
   DB  196,193,17,114,245,4                ; vpslld        $0x4,%xmm13,%xmm13
   DB  196,65,57,219,228                   ; vpand         %xmm12,%xmm8,%xmm12
   DB  196,193,25,114,244,4                ; vpslld        $0x4,%xmm12,%xmm12
   DB  196,67,21,24,228,1                  ; vinsertf128   $0x1,%xmm12,%ymm13,%ymm12
-  DB  196,98,125,24,45,62,66,2,0          ; vbroadcastss  0x2423e(%rip),%ymm13        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,98,125,24,45,234,88,2,0         ; vbroadcastss  0x258ea(%rip),%ymm13        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  196,65,36,84,237                    ; vandps        %ymm13,%ymm11,%ymm13
   DB  196,193,9,114,245,2                 ; vpslld        $0x2,%xmm13,%xmm14
   DB  196,67,125,25,237,1                 ; vextractf128  $0x1,%ymm13,%xmm13
   DB  196,193,17,114,245,2                ; vpslld        $0x2,%xmm13,%xmm13
   DB  196,67,13,24,237,1                  ; vinsertf128   $0x1,%xmm13,%ymm14,%ymm13
-  DB  197,123,18,53,109,81,2,0            ; vmovddup      0x2516d(%rip),%xmm14        # 2cc18 <_sk_srcover_rgba_8888_sse2_lowp+0x1210>
+  DB  197,123,18,53,77,104,2,0            ; vmovddup      0x2684d(%rip),%xmm14        # 2e2f8 <_sk_srcover_rgba_8888_sse2_lowp+0x1244>
   DB  196,65,49,219,254                   ; vpand         %xmm14,%xmm9,%xmm15
   DB  196,65,1,254,255                    ; vpaddd        %xmm15,%xmm15,%xmm15
   DB  196,65,57,219,246                   ; vpand         %xmm14,%xmm8,%xmm14
   DB  196,65,9,254,246                    ; vpaddd        %xmm14,%xmm14,%xmm14
   DB  196,67,5,24,246,1                   ; vinsertf128   $0x1,%xmm14,%ymm15,%ymm14
   DB  196,65,12,86,228                    ; vorps         %ymm12,%ymm14,%ymm12
-  DB  196,98,125,24,53,237,65,2,0         ; vbroadcastss  0x241ed(%rip),%ymm14        # 2bcc0 <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
+  DB  196,98,125,24,53,153,88,2,0         ; vbroadcastss  0x25899(%rip),%ymm14        # 2d36c <_sk_srcover_rgba_8888_sse2_lowp+0x2b8>
   DB  196,65,36,84,222                    ; vandps        %ymm14,%ymm11,%ymm11
   DB  196,193,9,114,211,1                 ; vpsrld        $0x1,%xmm11,%xmm14
   DB  196,67,125,25,219,1                 ; vextractf128  $0x1,%ymm11,%xmm11
   DB  196,193,33,114,211,1                ; vpsrld        $0x1,%xmm11,%xmm11
   DB  196,67,13,24,219,1                  ; vinsertf128   $0x1,%xmm11,%ymm14,%ymm11
   DB  196,65,20,86,219                    ; vorps         %ymm11,%ymm13,%ymm11
-  DB  197,123,18,45,35,81,2,0             ; vmovddup      0x25123(%rip),%xmm13        # 2cc20 <_sk_srcover_rgba_8888_sse2_lowp+0x1218>
+  DB  197,123,18,45,3,104,2,0             ; vmovddup      0x26803(%rip),%xmm13        # 2e300 <_sk_srcover_rgba_8888_sse2_lowp+0x124c>
   DB  196,65,49,219,205                   ; vpand         %xmm13,%xmm9,%xmm9
   DB  196,65,57,219,197                   ; vpand         %xmm13,%xmm8,%xmm8
   DB  196,193,49,114,209,2                ; vpsrld        $0x2,%xmm9,%xmm9
@@ -7221,9 +7221,9 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  196,65,60,86,195                    ; vorps         %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,144,65,2,0         ; vbroadcastss  0x24190(%rip),%ymm9        # 2bcc8 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
+  DB  196,98,125,24,13,60,88,2,0          ; vbroadcastss  0x2583c(%rip),%ymm9        # 2d374 <_sk_srcover_rgba_8888_sse2_lowp+0x2c0>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,134,65,2,0         ; vbroadcastss  0x24186(%rip),%ymm9        # 2bccc <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
+  DB  196,98,125,24,13,50,88,2,0          ; vbroadcastss  0x25832(%rip),%ymm9        # 2d378 <_sk_srcover_rgba_8888_sse2_lowp+0x2c4>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  196,98,125,24,8                     ; vbroadcastss  (%rax),%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
@@ -7253,7 +7253,7 @@
 PUBLIC _sk_black_color_avx
 _sk_black_color_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,14,65,2,0         ; vbroadcastss  0x2410e(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,186,87,2,0        ; vbroadcastss  0x257ba(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
@@ -7262,7 +7262,7 @@
 PUBLIC _sk_white_color_avx
 _sk_white_color_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,245,64,2,0         ; vbroadcastss  0x240f5(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,161,87,2,0         ; vbroadcastss  0x257a1(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  197,252,40,216                      ; vmovaps       %ymm0,%ymm3
@@ -7300,7 +7300,7 @@
 PUBLIC _sk_srcatop_avx
 _sk_srcatop_avx LABEL PROC
   DB  197,252,89,199                      ; vmulps        %ymm7,%ymm0,%ymm0
-  DB  196,98,125,24,5,148,64,2,0          ; vbroadcastss  0x24094(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,64,87,2,0           ; vbroadcastss  0x25740(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,204                       ; vmulps        %ymm4,%ymm8,%ymm9
   DB  197,180,88,192                      ; vaddps        %ymm0,%ymm9,%ymm0
@@ -7319,7 +7319,7 @@
 PUBLIC _sk_dstatop_avx
 _sk_dstatop_avx LABEL PROC
   DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
-  DB  196,98,125,24,13,82,64,2,0          ; vbroadcastss  0x24052(%rip),%ymm9        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,13,254,86,2,0         ; vbroadcastss  0x256fe(%rip),%ymm9        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,52,92,207                       ; vsubps        %ymm7,%ymm9,%ymm9
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
   DB  197,188,88,192                      ; vaddps        %ymm0,%ymm8,%ymm0
@@ -7355,7 +7355,7 @@
 
 PUBLIC _sk_srcout_avx
 _sk_srcout_avx LABEL PROC
-  DB  196,98,125,24,5,237,63,2,0          ; vbroadcastss  0x23fed(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,153,86,2,0          ; vbroadcastss  0x25699(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
@@ -7366,7 +7366,7 @@
 
 PUBLIC _sk_dstout_avx
 _sk_dstout_avx LABEL PROC
-  DB  196,226,125,24,5,204,63,2,0         ; vbroadcastss  0x23fcc(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,120,86,2,0         ; vbroadcastss  0x25678(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,92,219                      ; vsubps        %ymm3,%ymm0,%ymm3
   DB  197,228,89,196                      ; vmulps        %ymm4,%ymm3,%ymm0
   DB  197,228,89,205                      ; vmulps        %ymm5,%ymm3,%ymm1
@@ -7377,7 +7377,7 @@
 
 PUBLIC _sk_srcover_avx
 _sk_srcover_avx LABEL PROC
-  DB  196,98,125,24,5,171,63,2,0          ; vbroadcastss  0x23fab(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,87,86,2,0           ; vbroadcastss  0x25657(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,204                       ; vmulps        %ymm4,%ymm8,%ymm9
   DB  197,180,88,192                      ; vaddps        %ymm0,%ymm9,%ymm0
@@ -7392,7 +7392,7 @@
 
 PUBLIC _sk_dstover_avx
 _sk_dstover_avx LABEL PROC
-  DB  196,98,125,24,5,122,63,2,0          ; vbroadcastss  0x23f7a(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,38,86,2,0           ; vbroadcastss  0x25626(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,199                       ; vsubps        %ymm7,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
@@ -7416,7 +7416,7 @@
 
 PUBLIC _sk_multiply_avx
 _sk_multiply_avx LABEL PROC
-  DB  196,98,125,24,5,53,63,2,0           ; vbroadcastss  0x23f35(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,225,85,2,0          ; vbroadcastss  0x255e1(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,52,89,208                       ; vmulps        %ymm0,%ymm9,%ymm10
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -7445,7 +7445,7 @@
 PUBLIC _sk_plus__avx
 _sk_plus__avx LABEL PROC
   DB  197,252,88,196                      ; vaddps        %ymm4,%ymm0,%ymm0
-  DB  196,98,125,24,5,196,62,2,0          ; vbroadcastss  0x23ec4(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,112,85,2,0          ; vbroadcastss  0x25570(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  197,244,88,205                      ; vaddps        %ymm5,%ymm1,%ymm1
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
@@ -7475,7 +7475,7 @@
 
 PUBLIC _sk_xor__avx
 _sk_xor__avx LABEL PROC
-  DB  196,98,125,24,5,99,62,2,0           ; vbroadcastss  0x23e63(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,15,85,2,0           ; vbroadcastss  0x2550f(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,180,89,192                      ; vmulps        %ymm0,%ymm9,%ymm0
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -7510,7 +7510,7 @@
   DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
   DB  196,193,108,95,209                  ; vmaxps        %ymm9,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,223,61,2,0          ; vbroadcastss  0x23ddf(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,139,84,2,0          ; vbroadcastss  0x2548b(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
   DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
@@ -7534,7 +7534,7 @@
   DB  197,100,89,206                      ; vmulps        %ymm6,%ymm3,%ymm9
   DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,135,61,2,0          ; vbroadcastss  0x23d87(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,51,84,2,0           ; vbroadcastss  0x25433(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
   DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
@@ -7561,7 +7561,7 @@
   DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
   DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,35,61,2,0           ; vbroadcastss  0x23d23(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,207,83,2,0          ; vbroadcastss  0x253cf(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
   DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
@@ -7582,7 +7582,7 @@
   DB  197,236,89,214                      ; vmulps        %ymm6,%ymm2,%ymm2
   DB  197,236,88,210                      ; vaddps        %ymm2,%ymm2,%ymm2
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
-  DB  196,98,125,24,5,218,60,2,0          ; vbroadcastss  0x23cda(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,134,83,2,0          ; vbroadcastss  0x25386(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
   DB  197,60,89,199                       ; vmulps        %ymm7,%ymm8,%ymm8
   DB  197,188,88,219                      ; vaddps        %ymm3,%ymm8,%ymm3
@@ -7591,7 +7591,7 @@
 
 PUBLIC _sk_colorburn_avx
 _sk_colorburn_avx LABEL PROC
-  DB  196,98,125,24,5,193,60,2,0          ; vbroadcastss  0x23cc1(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,109,83,2,0          ; vbroadcastss  0x2536d(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,52,89,216                       ; vmulps        %ymm0,%ymm9,%ymm11
   DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
@@ -7651,7 +7651,7 @@
 PUBLIC _sk_colordodge_avx
 _sk_colordodge_avx LABEL PROC
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
-  DB  196,98,125,24,13,197,59,2,0         ; vbroadcastss  0x23bc5(%rip),%ymm9        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,13,113,82,2,0         ; vbroadcastss  0x25271(%rip),%ymm9        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,52,92,215                       ; vsubps        %ymm7,%ymm9,%ymm10
   DB  197,44,89,216                       ; vmulps        %ymm0,%ymm10,%ymm11
   DB  197,52,92,203                       ; vsubps        %ymm3,%ymm9,%ymm9
@@ -7706,7 +7706,7 @@
 
 PUBLIC _sk_hardlight_avx
 _sk_hardlight_avx LABEL PROC
-  DB  196,98,125,24,5,208,58,2,0          ; vbroadcastss  0x23ad0(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,124,81,2,0          ; vbroadcastss  0x2517c(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
   DB  197,44,89,200                       ; vmulps        %ymm0,%ymm10,%ymm9
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -7759,7 +7759,7 @@
 
 PUBLIC _sk_overlay_avx
 _sk_overlay_avx LABEL PROC
-  DB  196,98,125,24,5,245,57,2,0          ; vbroadcastss  0x239f5(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,161,80,2,0          ; vbroadcastss  0x250a1(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,215                       ; vsubps        %ymm7,%ymm8,%ymm10
   DB  197,44,89,200                       ; vmulps        %ymm0,%ymm10,%ymm9
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -7824,10 +7824,10 @@
   DB  196,65,60,88,192                    ; vaddps        %ymm8,%ymm8,%ymm8
   DB  196,65,60,89,216                    ; vmulps        %ymm8,%ymm8,%ymm11
   DB  196,65,60,88,195                    ; vaddps        %ymm11,%ymm8,%ymm8
-  DB  196,98,125,24,29,248,56,2,0         ; vbroadcastss  0x238f8(%rip),%ymm11        # 2bcd0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  DB  196,98,125,24,29,164,79,2,0         ; vbroadcastss  0x24fa4(%rip),%ymm11        # 2d37c <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   DB  196,65,28,88,235                    ; vaddps        %ymm11,%ymm12,%ymm13
   DB  196,65,20,89,192                    ; vmulps        %ymm8,%ymm13,%ymm8
-  DB  196,98,125,24,45,233,56,2,0         ; vbroadcastss  0x238e9(%rip),%ymm13        # 2bcd4 <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
+  DB  196,98,125,24,45,149,79,2,0         ; vbroadcastss  0x24f95(%rip),%ymm13        # 2d380 <_sk_srcover_rgba_8888_sse2_lowp+0x2cc>
   DB  196,65,28,89,245                    ; vmulps        %ymm13,%ymm12,%ymm14
   DB  196,65,12,88,192                    ; vaddps        %ymm8,%ymm14,%ymm8
   DB  196,65,124,82,244                   ; vrsqrtps      %ymm12,%ymm14
@@ -7838,7 +7838,7 @@
   DB  197,4,194,255,2                     ; vcmpleps      %ymm7,%ymm15,%ymm15
   DB  196,67,13,74,240,240                ; vblendvps     %ymm15,%ymm8,%ymm14,%ymm14
   DB  197,116,88,249                      ; vaddps        %ymm1,%ymm1,%ymm15
-  DB  196,98,125,24,5,147,56,2,0          ; vbroadcastss  0x23893(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,63,79,2,0           ; vbroadcastss  0x24f3f(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,65,60,92,228                    ; vsubps        %ymm12,%ymm8,%ymm12
   DB  197,132,92,195                      ; vsubps        %ymm3,%ymm15,%ymm0
   DB  196,65,124,89,228                   ; vmulps        %ymm12,%ymm0,%ymm12
@@ -7965,12 +7965,12 @@
   DB  196,65,28,89,219                    ; vmulps        %ymm11,%ymm12,%ymm11
   DB  196,65,36,94,222                    ; vdivps        %ymm14,%ymm11,%ymm11
   DB  196,67,37,74,224,240                ; vblendvps     %ymm15,%ymm8,%ymm11,%ymm12
-  DB  196,98,125,24,53,113,54,2,0         ; vbroadcastss  0x23671(%rip),%ymm14        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,53,29,77,2,0          ; vbroadcastss  0x24d1d(%rip),%ymm14        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,103,54,2,0         ; vbroadcastss  0x23667(%rip),%ymm15        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,61,19,77,2,0          ; vbroadcastss  0x24d13(%rip),%ymm15        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,88,54,2,0          ; vbroadcastss  0x23658(%rip),%ymm0        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,226,125,24,5,4,77,2,0           ; vbroadcastss  0x24d04(%rip),%ymm0        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  197,76,89,232                       ; vmulps        %ymm0,%ymm6,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
   DB  196,65,52,89,238                    ; vmulps        %ymm14,%ymm9,%ymm13
@@ -8031,7 +8031,7 @@
   DB  196,65,36,95,208                    ; vmaxps        %ymm8,%ymm11,%ymm10
   DB  196,195,109,74,209,240              ; vblendvps     %ymm15,%ymm9,%ymm2,%ymm2
   DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,5,53,2,0            ; vbroadcastss  0x23505(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,177,75,2,0          ; vbroadcastss  0x24bb1(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -8088,12 +8088,12 @@
   DB  196,65,28,89,219                    ; vmulps        %ymm11,%ymm12,%ymm11
   DB  196,65,36,94,222                    ; vdivps        %ymm14,%ymm11,%ymm11
   DB  196,67,37,74,224,240                ; vblendvps     %ymm15,%ymm8,%ymm11,%ymm12
-  DB  196,98,125,24,53,41,52,2,0          ; vbroadcastss  0x23429(%rip),%ymm14        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,53,213,74,2,0         ; vbroadcastss  0x24ad5(%rip),%ymm14        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,92,89,222                    ; vmulps        %ymm14,%ymm4,%ymm11
-  DB  196,98,125,24,61,31,52,2,0          ; vbroadcastss  0x2341f(%rip),%ymm15        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,61,203,74,2,0         ; vbroadcastss  0x24acb(%rip),%ymm15        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,239                    ; vmulps        %ymm15,%ymm5,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
-  DB  196,226,125,24,5,16,52,2,0          ; vbroadcastss  0x23410(%rip),%ymm0        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,226,125,24,5,188,74,2,0         ; vbroadcastss  0x24abc(%rip),%ymm0        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  197,76,89,232                       ; vmulps        %ymm0,%ymm6,%ymm13
   DB  196,65,36,88,221                    ; vaddps        %ymm13,%ymm11,%ymm11
   DB  196,65,52,89,238                    ; vmulps        %ymm14,%ymm9,%ymm13
@@ -8154,7 +8154,7 @@
   DB  196,65,36,95,208                    ; vmaxps        %ymm8,%ymm11,%ymm10
   DB  196,195,109,74,209,240              ; vblendvps     %ymm15,%ymm9,%ymm2,%ymm2
   DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,189,50,2,0          ; vbroadcastss  0x232bd(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,105,73,2,0          ; vbroadcastss  0x24969(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,207                       ; vsubps        %ymm7,%ymm8,%ymm9
   DB  197,180,89,201                      ; vmulps        %ymm1,%ymm9,%ymm1
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
@@ -8183,12 +8183,12 @@
   DB  197,252,17,68,36,32                 ; vmovups       %ymm0,0x20(%rsp)
   DB  197,124,89,199                      ; vmulps        %ymm7,%ymm0,%ymm8
   DB  197,116,89,207                      ; vmulps        %ymm7,%ymm1,%ymm9
-  DB  196,98,125,24,45,105,50,2,0         ; vbroadcastss  0x23269(%rip),%ymm13        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,45,21,73,2,0          ; vbroadcastss  0x24915(%rip),%ymm13        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,92,89,213                    ; vmulps        %ymm13,%ymm4,%ymm10
-  DB  196,98,125,24,53,95,50,2,0          ; vbroadcastss  0x2325f(%rip),%ymm14        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,53,11,73,2,0          ; vbroadcastss  0x2490b(%rip),%ymm14        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,84,89,222                    ; vmulps        %ymm14,%ymm5,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,80,50,2,0          ; vbroadcastss  0x23250(%rip),%ymm15        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,98,125,24,61,252,72,2,0         ; vbroadcastss  0x248fc(%rip),%ymm15        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,65,76,89,223                    ; vmulps        %ymm15,%ymm6,%ymm11
   DB  196,193,44,88,195                   ; vaddps        %ymm11,%ymm10,%ymm0
   DB  196,65,60,89,221                    ; vmulps        %ymm13,%ymm8,%ymm11
@@ -8251,7 +8251,7 @@
   DB  196,65,44,95,207                    ; vmaxps        %ymm15,%ymm10,%ymm9
   DB  196,195,37,74,192,0                 ; vblendvps     %ymm0,%ymm8,%ymm11,%ymm0
   DB  196,65,124,95,199                   ; vmaxps        %ymm15,%ymm0,%ymm8
-  DB  196,226,125,24,5,235,48,2,0         ; vbroadcastss  0x230eb(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,151,71,2,0         ; vbroadcastss  0x24797(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,124,92,215                      ; vsubps        %ymm7,%ymm0,%ymm10
   DB  197,172,89,84,36,32                 ; vmulps        0x20(%rsp),%ymm10,%ymm2
   DB  197,124,92,219                      ; vsubps        %ymm3,%ymm0,%ymm11
@@ -8281,12 +8281,12 @@
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  197,100,89,196                      ; vmulps        %ymm4,%ymm3,%ymm8
   DB  197,100,89,205                      ; vmulps        %ymm5,%ymm3,%ymm9
-  DB  196,98,125,24,45,147,48,2,0         ; vbroadcastss  0x23093(%rip),%ymm13        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,45,63,71,2,0          ; vbroadcastss  0x2473f(%rip),%ymm13        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,108,89,213                   ; vmulps        %ymm13,%ymm2,%ymm10
-  DB  196,98,125,24,53,137,48,2,0         ; vbroadcastss  0x23089(%rip),%ymm14        # 2bcdc <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
+  DB  196,98,125,24,53,53,71,2,0          ; vbroadcastss  0x24735(%rip),%ymm14        # 2d388 <_sk_srcover_rgba_8888_sse2_lowp+0x2d4>
   DB  196,65,116,89,222                   ; vmulps        %ymm14,%ymm1,%ymm11
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,61,122,48,2,0         ; vbroadcastss  0x2307a(%rip),%ymm15        # 2bce0 <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
+  DB  196,98,125,24,61,38,71,2,0          ; vbroadcastss  0x24726(%rip),%ymm15        # 2d38c <_sk_srcover_rgba_8888_sse2_lowp+0x2d8>
   DB  196,65,28,89,223                    ; vmulps        %ymm15,%ymm12,%ymm11
   DB  196,193,44,88,195                   ; vaddps        %ymm11,%ymm10,%ymm0
   DB  196,65,60,89,221                    ; vmulps        %ymm13,%ymm8,%ymm11
@@ -8349,7 +8349,7 @@
   DB  196,65,44,95,207                    ; vmaxps        %ymm15,%ymm10,%ymm9
   DB  196,195,37,74,192,0                 ; vblendvps     %ymm0,%ymm8,%ymm11,%ymm0
   DB  196,65,124,95,199                   ; vmaxps        %ymm15,%ymm0,%ymm8
-  DB  196,226,125,24,5,21,47,2,0          ; vbroadcastss  0x22f15(%rip),%ymm0        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,5,193,69,2,0         ; vbroadcastss  0x245c1(%rip),%ymm0        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,124,92,215                      ; vsubps        %ymm7,%ymm0,%ymm10
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  197,124,92,219                      ; vsubps        %ymm3,%ymm0,%ymm11
@@ -8382,7 +8382,7 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,42,1,0,0                     ; jne           8f3b <_sk_srcover_rgba_8888_avx+0x14a>
   DB  196,1,125,16,4,152                  ; vmovupd       (%r8,%r11,4),%ymm8
-  DB  197,125,40,13,97,51,2,0             ; vmovapd       0x23361(%rip),%ymm9        # 2c180 <_sk_srcover_rgba_8888_sse2_lowp+0x778>
+  DB  197,125,40,13,1,74,2,0              ; vmovapd       0x24a01(%rip),%ymm9        # 2d820 <_sk_srcover_rgba_8888_sse2_lowp+0x76c>
   DB  196,193,61,84,225                   ; vandpd        %ymm9,%ymm8,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
   DB  196,193,81,114,208,8                ; vpsrld        $0x8,%xmm8,%xmm5
@@ -8400,9 +8400,9 @@
   DB  197,193,114,215,24                  ; vpsrld        $0x18,%xmm7,%xmm7
   DB  196,227,61,24,255,1                 ; vinsertf128   $0x1,%xmm7,%ymm8,%ymm7
   DB  197,252,91,255                      ; vcvtdq2ps     %ymm7,%ymm7
-  DB  196,98,125,24,5,56,46,2,0           ; vbroadcastss  0x22e38(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,228,68,2,0          ; vbroadcastss  0x244e4(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,195                       ; vsubps        %ymm3,%ymm8,%ymm8
-  DB  196,98,125,24,13,87,46,2,0          ; vbroadcastss  0x22e57(%rip),%ymm9        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,13,3,69,2,0           ; vbroadcastss  0x24503(%rip),%ymm9        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,60,89,212                       ; vmulps        %ymm4,%ymm8,%ymm10
   DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
@@ -8542,7 +8542,7 @@
 
 PUBLIC _sk_clamp_1_avx
 _sk_clamp_1_avx LABEL PROC
-  DB  196,98,125,24,5,246,43,2,0          ; vbroadcastss  0x22bf6(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,162,66,2,0          ; vbroadcastss  0x242a2(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
@@ -8552,7 +8552,7 @@
 
 PUBLIC _sk_clamp_a_avx
 _sk_clamp_a_avx LABEL PROC
-  DB  196,98,125,24,5,213,43,2,0          ; vbroadcastss  0x22bd5(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,129,66,2,0          ; vbroadcastss  0x24281(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
   DB  197,252,93,195                      ; vminps        %ymm3,%ymm0,%ymm0
   DB  197,244,93,203                      ; vminps        %ymm3,%ymm1,%ymm1
@@ -8562,7 +8562,7 @@
 
 PUBLIC _sk_clamp_a_dst_avx
 _sk_clamp_a_dst_avx LABEL PROC
-  DB  196,98,125,24,5,183,43,2,0          ; vbroadcastss  0x22bb7(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,99,66,2,0           ; vbroadcastss  0x24263(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,68,93,248                   ; vminps        %ymm8,%ymm7,%ymm7
   DB  197,220,93,231                      ; vminps        %ymm7,%ymm4,%ymm4
   DB  197,212,93,239                      ; vminps        %ymm7,%ymm5,%ymm5
@@ -8589,7 +8589,7 @@
 
 PUBLIC _sk_invert_avx
 _sk_invert_avx LABEL PROC
-  DB  196,98,125,24,5,114,43,2,0          ; vbroadcastss  0x22b72(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,30,66,2,0           ; vbroadcastss  0x2421e(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,188,92,192                      ; vsubps        %ymm0,%ymm8,%ymm0
   DB  197,188,92,201                      ; vsubps        %ymm1,%ymm8,%ymm1
   DB  197,188,92,210                      ; vsubps        %ymm2,%ymm8,%ymm2
@@ -8633,9 +8633,9 @@
 
 PUBLIC _sk_unpremul_avx
 _sk_unpremul_avx LABEL PROC
-  DB  196,98,125,24,5,13,43,2,0           ; vbroadcastss  0x22b0d(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,185,65,2,0          ; vbroadcastss  0x241b9(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,94,195                       ; vdivps        %ymm3,%ymm8,%ymm8
-  DB  196,98,125,24,13,48,43,2,0          ; vbroadcastss  0x22b30(%rip),%ymm9        # 2bce8 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
+  DB  196,98,125,24,13,220,65,2,0         ; vbroadcastss  0x241dc(%rip),%ymm9        # 2d394 <_sk_srcover_rgba_8888_sse2_lowp+0x2e0>
   DB  196,65,60,194,201,1                 ; vcmpltps      %ymm9,%ymm8,%ymm9
   DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
   DB  196,67,45,74,192,144                ; vblendvps     %ymm9,%ymm8,%ymm10,%ymm8
@@ -8647,17 +8647,17 @@
 
 PUBLIC _sk_from_srgb_avx
 _sk_from_srgb_avx LABEL PROC
-  DB  196,98,125,24,5,10,43,2,0           ; vbroadcastss  0x22b0a(%rip),%ymm8        # 2bcec <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  DB  196,98,125,24,5,182,65,2,0          ; vbroadcastss  0x241b6(%rip),%ymm8        # 2d398 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  197,124,89,208                      ; vmulps        %ymm0,%ymm0,%ymm10
-  DB  196,98,125,24,29,228,42,2,0         ; vbroadcastss  0x22ae4(%rip),%ymm11        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,29,144,65,2,0         ; vbroadcastss  0x24190(%rip),%ymm11        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,124,89,227                   ; vmulps        %ymm11,%ymm0,%ymm12
-  DB  196,98,125,24,45,238,42,2,0         ; vbroadcastss  0x22aee(%rip),%ymm13        # 2bcf0 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  DB  196,98,125,24,45,154,65,2,0         ; vbroadcastss  0x2419a(%rip),%ymm13        # 2d39c <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,223,42,2,0         ; vbroadcastss  0x22adf(%rip),%ymm12        # 2bcf4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  DB  196,98,125,24,37,139,65,2,0         ; vbroadcastss  0x2418b(%rip),%ymm12        # 2d3a0 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,53,213,42,2,0         ; vbroadcastss  0x22ad5(%rip),%ymm14        # 2bcf8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  DB  196,98,125,24,53,129,65,2,0         ; vbroadcastss  0x24181(%rip),%ymm14        # 2d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   DB  196,193,124,194,198,1               ; vcmpltps      %ymm14,%ymm0,%ymm0
   DB  196,195,45,74,193,0                 ; vblendvps     %ymm0,%ymm9,%ymm10,%ymm0
   DB  196,65,116,89,200                   ; vmulps        %ymm8,%ymm1,%ymm9
@@ -8681,17 +8681,17 @@
 
 PUBLIC _sk_from_srgb_dst_avx
 _sk_from_srgb_dst_avx LABEL PROC
-  DB  196,98,125,24,5,94,42,2,0           ; vbroadcastss  0x22a5e(%rip),%ymm8        # 2bcec <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
+  DB  196,98,125,24,5,10,65,2,0           ; vbroadcastss  0x2410a(%rip),%ymm8        # 2d398 <_sk_srcover_rgba_8888_sse2_lowp+0x2e4>
   DB  196,65,92,89,200                    ; vmulps        %ymm8,%ymm4,%ymm9
   DB  197,92,89,212                       ; vmulps        %ymm4,%ymm4,%ymm10
-  DB  196,98,125,24,29,56,42,2,0          ; vbroadcastss  0x22a38(%rip),%ymm11        # 2bcd8 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
+  DB  196,98,125,24,29,228,64,2,0         ; vbroadcastss  0x240e4(%rip),%ymm11        # 2d384 <_sk_srcover_rgba_8888_sse2_lowp+0x2d0>
   DB  196,65,92,89,227                    ; vmulps        %ymm11,%ymm4,%ymm12
-  DB  196,98,125,24,45,66,42,2,0          ; vbroadcastss  0x22a42(%rip),%ymm13        # 2bcf0 <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
+  DB  196,98,125,24,45,238,64,2,0         ; vbroadcastss  0x240ee(%rip),%ymm13        # 2d39c <_sk_srcover_rgba_8888_sse2_lowp+0x2e8>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,44,89,212                    ; vmulps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,37,51,42,2,0          ; vbroadcastss  0x22a33(%rip),%ymm12        # 2bcf4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
+  DB  196,98,125,24,37,223,64,2,0         ; vbroadcastss  0x240df(%rip),%ymm12        # 2d3a0 <_sk_srcover_rgba_8888_sse2_lowp+0x2ec>
   DB  196,65,44,88,212                    ; vaddps        %ymm12,%ymm10,%ymm10
-  DB  196,98,125,24,53,41,42,2,0          ; vbroadcastss  0x22a29(%rip),%ymm14        # 2bcf8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
+  DB  196,98,125,24,53,213,64,2,0         ; vbroadcastss  0x240d5(%rip),%ymm14        # 2d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0x2f0>
   DB  196,193,92,194,230,1                ; vcmpltps      %ymm14,%ymm4,%ymm4
   DB  196,195,45,74,225,64                ; vblendvps     %ymm4,%ymm9,%ymm10,%ymm4
   DB  196,65,84,89,200                    ; vmulps        %ymm8,%ymm5,%ymm9
@@ -8716,20 +8716,20 @@
 PUBLIC _sk_to_srgb_avx
 _sk_to_srgb_avx LABEL PROC
   DB  197,124,82,200                      ; vrsqrtps      %ymm0,%ymm9
-  DB  196,98,125,24,5,190,41,2,0          ; vbroadcastss  0x229be(%rip),%ymm8        # 2bcfc <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
+  DB  196,98,125,24,5,106,64,2,0          ; vbroadcastss  0x2406a(%rip),%ymm8        # 2d3a8 <_sk_srcover_rgba_8888_sse2_lowp+0x2f4>
   DB  196,65,124,89,208                   ; vmulps        %ymm8,%ymm0,%ymm10
-  DB  196,98,125,24,29,184,41,2,0         ; vbroadcastss  0x229b8(%rip),%ymm11        # 2bd04 <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
+  DB  196,98,125,24,29,100,64,2,0         ; vbroadcastss  0x24064(%rip),%ymm11        # 2d3b0 <_sk_srcover_rgba_8888_sse2_lowp+0x2fc>
   DB  196,65,52,89,227                    ; vmulps        %ymm11,%ymm9,%ymm12
-  DB  196,98,125,24,45,166,41,2,0         ; vbroadcastss  0x229a6(%rip),%ymm13        # 2bd00 <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
+  DB  196,98,125,24,45,82,64,2,0          ; vbroadcastss  0x24052(%rip),%ymm13        # 2d3ac <_sk_srcover_rgba_8888_sse2_lowp+0x2f8>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,52,89,228                    ; vmulps        %ymm12,%ymm9,%ymm12
-  DB  196,98,125,24,53,155,41,2,0         ; vbroadcastss  0x2299b(%rip),%ymm14        # 2bd08 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
+  DB  196,98,125,24,53,71,64,2,0          ; vbroadcastss  0x24047(%rip),%ymm14        # 2d3b4 <_sk_srcover_rgba_8888_sse2_lowp+0x300>
   DB  196,65,28,88,230                    ; vaddps        %ymm14,%ymm12,%ymm12
-  DB  196,98,125,24,61,145,41,2,0         ; vbroadcastss  0x22991(%rip),%ymm15        # 2bd0c <_sk_srcover_rgba_8888_sse2_lowp+0x304>
+  DB  196,98,125,24,61,61,64,2,0          ; vbroadcastss  0x2403d(%rip),%ymm15        # 2d3b8 <_sk_srcover_rgba_8888_sse2_lowp+0x304>
   DB  196,65,52,88,207                    ; vaddps        %ymm15,%ymm9,%ymm9
   DB  196,65,124,83,201                   ; vrcpps        %ymm9,%ymm9
   DB  196,65,52,89,204                    ; vmulps        %ymm12,%ymm9,%ymm9
-  DB  196,98,125,24,37,125,41,2,0         ; vbroadcastss  0x2297d(%rip),%ymm12        # 2bd10 <_sk_srcover_rgba_8888_sse2_lowp+0x308>
+  DB  196,98,125,24,37,41,64,2,0          ; vbroadcastss  0x24029(%rip),%ymm12        # 2d3bc <_sk_srcover_rgba_8888_sse2_lowp+0x308>
   DB  196,193,124,194,196,1               ; vcmpltps      %ymm12,%ymm0,%ymm0
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  197,124,82,201                      ; vrsqrtps      %ymm1,%ymm9
@@ -8764,7 +8764,7 @@
   DB  197,116,93,202                      ; vminps        %ymm2,%ymm1,%ymm9
   DB  196,65,124,93,201                   ; vminps        %ymm9,%ymm0,%ymm9
   DB  196,65,60,92,209                    ; vsubps        %ymm9,%ymm8,%ymm10
-  DB  196,98,125,24,29,133,40,2,0         ; vbroadcastss  0x22885(%rip),%ymm11        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,29,49,63,2,0          ; vbroadcastss  0x23f31(%rip),%ymm11        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,65,36,94,218                    ; vdivps        %ymm10,%ymm11,%ymm11
   DB  197,116,92,226                      ; vsubps        %ymm2,%ymm1,%ymm12
   DB  196,65,28,89,227                    ; vmulps        %ymm11,%ymm12,%ymm12
@@ -8774,19 +8774,19 @@
   DB  196,193,108,89,211                  ; vmulps        %ymm11,%ymm2,%ymm2
   DB  197,252,92,201                      ; vsubps        %ymm1,%ymm0,%ymm1
   DB  196,193,116,89,203                  ; vmulps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,182,40,2,0         ; vbroadcastss  0x228b6(%rip),%ymm11        # 2bd1c <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  DB  196,98,125,24,29,98,63,2,0          ; vbroadcastss  0x23f62(%rip),%ymm11        # 2d3c8 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,164,40,2,0         ; vbroadcastss  0x228a4(%rip),%ymm11        # 2bd18 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
+  DB  196,98,125,24,29,80,63,2,0          ; vbroadcastss  0x23f50(%rip),%ymm11        # 2d3c4 <_sk_srcover_rgba_8888_sse2_lowp+0x310>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,227,117,74,202,224              ; vblendvps     %ymm14,%ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,140,40,2,0        ; vbroadcastss  0x2288c(%rip),%ymm2        # 2bd14 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  DB  196,226,125,24,21,56,63,2,0         ; vbroadcastss  0x23f38(%rip),%ymm2        # 2d3c0 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   DB  196,65,12,87,246                    ; vxorps        %ymm14,%ymm14,%ymm14
   DB  196,227,13,74,210,208               ; vblendvps     %ymm13,%ymm2,%ymm14,%ymm2
   DB  197,188,194,192,0                   ; vcmpeqps      %ymm0,%ymm8,%ymm0
   DB  196,193,108,88,212                  ; vaddps        %ymm12,%ymm2,%ymm2
   DB  196,227,117,74,194,0                ; vblendvps     %ymm0,%ymm2,%ymm1,%ymm0
   DB  196,193,60,88,201                   ; vaddps        %ymm9,%ymm8,%ymm1
-  DB  196,98,125,24,37,3,40,2,0           ; vbroadcastss  0x22803(%rip),%ymm12        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,37,175,62,2,0         ; vbroadcastss  0x23eaf(%rip),%ymm12        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,116,89,212                  ; vmulps        %ymm12,%ymm1,%ymm2
   DB  197,28,194,226,1                    ; vcmpltps      %ymm2,%ymm12,%ymm12
   DB  196,65,36,92,216                    ; vsubps        %ymm8,%ymm11,%ymm11
@@ -8796,7 +8796,7 @@
   DB  197,172,94,201                      ; vdivps        %ymm1,%ymm10,%ymm1
   DB  196,195,125,74,198,128              ; vblendvps     %ymm8,%ymm14,%ymm0,%ymm0
   DB  196,195,117,74,206,128              ; vblendvps     %ymm8,%ymm14,%ymm1,%ymm1
-  DB  196,98,125,24,5,54,40,2,0           ; vbroadcastss  0x22836(%rip),%ymm8        # 2bd20 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  DB  196,98,125,24,5,226,62,2,0          ; vbroadcastss  0x23ee2(%rip),%ymm8        # 2d3cc <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -8811,7 +8811,7 @@
   DB  197,252,17,28,36                    ; vmovups       %ymm3,(%rsp)
   DB  197,252,40,225                      ; vmovaps       %ymm1,%ymm4
   DB  197,252,40,216                      ; vmovaps       %ymm0,%ymm3
-  DB  196,98,125,24,5,137,39,2,0          ; vbroadcastss  0x22789(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,53,62,2,0           ; vbroadcastss  0x23e35(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,60,194,202,2                    ; vcmpleps      %ymm2,%ymm8,%ymm9
   DB  197,92,89,210                       ; vmulps        %ymm2,%ymm4,%ymm10
   DB  196,65,92,92,218                    ; vsubps        %ymm10,%ymm4,%ymm11
@@ -8819,23 +8819,23 @@
   DB  197,52,88,210                       ; vaddps        %ymm2,%ymm9,%ymm10
   DB  197,108,88,202                      ; vaddps        %ymm2,%ymm2,%ymm9
   DB  196,65,52,92,202                    ; vsubps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,29,207,39,2,0         ; vbroadcastss  0x227cf(%rip),%ymm11        # 2bd24 <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
+  DB  196,98,125,24,29,123,62,2,0         ; vbroadcastss  0x23e7b(%rip),%ymm11        # 2d3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x31c>
   DB  196,65,100,88,219                   ; vaddps        %ymm11,%ymm3,%ymm11
   DB  196,67,125,8,227,1                  ; vroundps      $0x1,%ymm11,%ymm12
   DB  196,65,36,92,252                    ; vsubps        %ymm12,%ymm11,%ymm15
   DB  196,65,44,92,217                    ; vsubps        %ymm9,%ymm10,%ymm11
-  DB  196,98,125,24,37,161,39,2,0         ; vbroadcastss  0x227a1(%rip),%ymm12        # 2bd14 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
+  DB  196,98,125,24,37,77,62,2,0          ; vbroadcastss  0x23e4d(%rip),%ymm12        # 2d3c0 <_sk_srcover_rgba_8888_sse2_lowp+0x30c>
   DB  196,193,4,89,196                    ; vmulps        %ymm12,%ymm15,%ymm0
-  DB  196,98,125,24,45,155,39,2,0         ; vbroadcastss  0x2279b(%rip),%ymm13        # 2bd1c <_sk_srcover_rgba_8888_sse2_lowp+0x314>
+  DB  196,98,125,24,45,71,62,2,0          ; vbroadcastss  0x23e47(%rip),%ymm13        # 2d3c8 <_sk_srcover_rgba_8888_sse2_lowp+0x314>
   DB  197,20,92,240                       ; vsubps        %ymm0,%ymm13,%ymm14
   DB  196,65,36,89,246                    ; vmulps        %ymm14,%ymm11,%ymm14
   DB  196,65,52,88,246                    ; vaddps        %ymm14,%ymm9,%ymm14
-  DB  196,226,125,24,13,144,39,2,0        ; vbroadcastss  0x22790(%rip),%ymm1        # 2bd28 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
+  DB  196,226,125,24,13,60,62,2,0         ; vbroadcastss  0x23e3c(%rip),%ymm1        # 2d3d4 <_sk_srcover_rgba_8888_sse2_lowp+0x320>
   DB  196,193,116,194,255,2               ; vcmpleps      %ymm15,%ymm1,%ymm7
   DB  196,195,13,74,249,112               ; vblendvps     %ymm7,%ymm9,%ymm14,%ymm7
   DB  196,65,60,194,247,2                 ; vcmpleps      %ymm15,%ymm8,%ymm14
   DB  196,227,45,74,255,224               ; vblendvps     %ymm14,%ymm7,%ymm10,%ymm7
-  DB  196,98,125,24,53,103,39,2,0         ; vbroadcastss  0x22767(%rip),%ymm14        # 2bd20 <_sk_srcover_rgba_8888_sse2_lowp+0x318>
+  DB  196,98,125,24,53,19,62,2,0          ; vbroadcastss  0x23e13(%rip),%ymm14        # 2d3cc <_sk_srcover_rgba_8888_sse2_lowp+0x318>
   DB  196,65,12,194,255,2                 ; vcmpleps      %ymm15,%ymm14,%ymm15
   DB  196,193,124,89,195                  ; vmulps        %ymm11,%ymm0,%ymm0
   DB  197,180,88,192                      ; vaddps        %ymm0,%ymm9,%ymm0
@@ -8854,7 +8854,7 @@
   DB  197,164,89,247                      ; vmulps        %ymm7,%ymm11,%ymm6
   DB  197,180,88,246                      ; vaddps        %ymm6,%ymm9,%ymm6
   DB  196,227,77,74,237,0                 ; vblendvps     %ymm0,%ymm5,%ymm6,%ymm5
-  DB  196,226,125,24,5,17,39,2,0          ; vbroadcastss  0x22711(%rip),%ymm0        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,226,125,24,5,189,61,2,0         ; vbroadcastss  0x23dbd(%rip),%ymm0        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  197,228,88,192                      ; vaddps        %ymm0,%ymm3,%ymm0
   DB  196,227,125,8,216,1                 ; vroundps      $0x1,%ymm0,%ymm3
   DB  197,252,92,195                      ; vsubps        %ymm3,%ymm0,%ymm0
@@ -8906,13 +8906,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,75                              ; jne           9727 <_sk_scale_u8_avx+0x63>
   DB  196,2,121,48,4,24                   ; vpmovzxbw     (%r8,%r11,1),%xmm8
-  DB  197,57,219,5,230,44,2,0             ; vpand         0x22ce6(%rip),%xmm8,%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,57,219,5,198,67,2,0             ; vpand         0x243c6(%rip),%xmm8,%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,66,121,51,200                   ; vpmovzxwd     %xmm8,%xmm9
   DB  196,65,121,112,192,78               ; vpshufd       $0x4e,%xmm8,%xmm8
   DB  196,66,121,51,192                   ; vpmovzxwd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,34,38,2,0          ; vbroadcastss  0x22622(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,206,60,2,0         ; vbroadcastss  0x23cce(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  197,188,89,201                      ; vmulps        %ymm1,%ymm8,%ymm1
@@ -8990,20 +8990,20 @@
   DB  196,65,121,112,192,78               ; vpshufd       $0x4e,%xmm8,%xmm8
   DB  196,66,121,51,192                   ; vpmovzxwd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
-  DB  196,98,125,24,13,20,37,2,0          ; vbroadcastss  0x22514(%rip),%ymm9        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,98,125,24,13,192,59,2,0         ; vbroadcastss  0x23bc0(%rip),%ymm9        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  196,65,60,84,201                    ; vandps        %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,5,37,2,0           ; vbroadcastss  0x22505(%rip),%ymm10        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,98,125,24,21,177,59,2,0         ; vbroadcastss  0x23bb1(%rip),%ymm10        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,251,36,2,0         ; vbroadcastss  0x224fb(%rip),%ymm10        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,98,125,24,21,167,59,2,0         ; vbroadcastss  0x23ba7(%rip),%ymm10        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  196,65,60,84,210                    ; vandps        %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,236,36,2,0         ; vbroadcastss  0x224ec(%rip),%ymm11        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,98,125,24,29,152,59,2,0         ; vbroadcastss  0x23b98(%rip),%ymm11        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,29,226,36,2,0         ; vbroadcastss  0x224e2(%rip),%ymm11        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,98,125,24,29,142,59,2,0         ; vbroadcastss  0x23b8e(%rip),%ymm11        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  196,65,60,84,195                    ; vandps        %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,211,36,2,0         ; vbroadcastss  0x224d3(%rip),%ymm11        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,98,125,24,29,127,59,2,0         ; vbroadcastss  0x23b7f(%rip),%ymm11        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  197,100,194,223,1                   ; vcmpltps      %ymm7,%ymm3,%ymm11
   DB  196,65,44,93,224                    ; vminps        %ymm8,%ymm10,%ymm12
@@ -9092,13 +9092,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,111                             ; jne           9a16 <_sk_lerp_u8_avx+0x87>
   DB  196,2,121,48,4,24                   ; vpmovzxbw     (%r8,%r11,1),%xmm8
-  DB  197,57,219,5,27,42,2,0              ; vpand         0x22a1b(%rip),%xmm8,%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,57,219,5,251,64,2,0             ; vpand         0x240fb(%rip),%xmm8,%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,66,121,51,200                   ; vpmovzxwd     %xmm8,%xmm9
   DB  196,65,121,112,192,78               ; vpshufd       $0x4e,%xmm8,%xmm8
   DB  196,66,121,51,192                   ; vpmovzxwd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,13,87,35,2,0          ; vbroadcastss  0x22357(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,3,58,2,0           ; vbroadcastss  0x23a03(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,252,92,196                      ; vsubps        %ymm4,%ymm0,%ymm0
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
@@ -9154,7 +9154,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,221                 ; callq         ffffffffde009ac4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfde0bc>
+  DB  232,255,255,255,221                 ; callq         ffffffffde009ac4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfdca10>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,210                             ; callq         *%rdx
@@ -9181,20 +9181,20 @@
   DB  196,65,121,112,192,78               ; vpshufd       $0x4e,%xmm8,%xmm8
   DB  196,66,121,51,192                   ; vpmovzxwd     %xmm8,%xmm8
   DB  196,67,53,24,192,1                  ; vinsertf128   $0x1,%xmm8,%ymm9,%ymm8
-  DB  196,98,125,24,13,32,34,2,0          ; vbroadcastss  0x22220(%rip),%ymm9        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,98,125,24,13,204,56,2,0         ; vbroadcastss  0x238cc(%rip),%ymm9        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  196,65,60,84,201                    ; vandps        %ymm9,%ymm8,%ymm9
   DB  196,65,124,91,201                   ; vcvtdq2ps     %ymm9,%ymm9
-  DB  196,98,125,24,21,17,34,2,0          ; vbroadcastss  0x22211(%rip),%ymm10        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,98,125,24,21,189,56,2,0         ; vbroadcastss  0x238bd(%rip),%ymm10        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,7,34,2,0           ; vbroadcastss  0x22207(%rip),%ymm10        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,98,125,24,21,179,56,2,0         ; vbroadcastss  0x238b3(%rip),%ymm10        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  196,65,60,84,210                    ; vandps        %ymm10,%ymm8,%ymm10
   DB  196,65,124,91,210                   ; vcvtdq2ps     %ymm10,%ymm10
-  DB  196,98,125,24,29,248,33,2,0         ; vbroadcastss  0x221f8(%rip),%ymm11        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,98,125,24,29,164,56,2,0         ; vbroadcastss  0x238a4(%rip),%ymm11        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
-  DB  196,98,125,24,29,238,33,2,0         ; vbroadcastss  0x221ee(%rip),%ymm11        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,98,125,24,29,154,56,2,0         ; vbroadcastss  0x2389a(%rip),%ymm11        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  196,65,60,84,195                    ; vandps        %ymm11,%ymm8,%ymm8
   DB  196,65,124,91,192                   ; vcvtdq2ps     %ymm8,%ymm8
-  DB  196,98,125,24,29,223,33,2,0         ; vbroadcastss  0x221df(%rip),%ymm11        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,98,125,24,29,139,56,2,0         ; vbroadcastss  0x2388b(%rip),%ymm11        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  197,100,194,223,1                   ; vcmpltps      %ymm7,%ymm3,%ymm11
   DB  196,65,44,93,224                    ; vminps        %ymm8,%ymm10,%ymm12
@@ -9270,7 +9270,7 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,7,2,0,0                      ; jne           9e86 <_sk_load_tables_avx+0x21e>
   DB  196,65,125,16,20,144                ; vmovupd       (%r8,%rdx,4),%ymm10
-  DB  197,125,40,13,19,37,2,0             ; vmovapd       0x22513(%rip),%ymm9        # 2c1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x798>
+  DB  197,125,40,13,179,59,2,0            ; vmovapd       0x23bb3(%rip),%ymm9        # 2d840 <_sk_srcover_rgba_8888_sse2_lowp+0x78c>
   DB  196,193,45,84,201                   ; vandpd        %ymm9,%ymm10,%ymm1
   DB  196,227,125,25,200,1                ; vextractf128  $0x1,%ymm1,%xmm0
   DB  196,193,249,126,192                 ; vmovq         %xmm0,%r8
@@ -9362,7 +9362,7 @@
   DB  196,193,65,114,213,24               ; vpsrld        $0x18,%xmm13,%xmm7
   DB  196,227,101,24,223,1                ; vinsertf128   $0x1,%xmm7,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,226,125,24,61,187,30,2,0        ; vbroadcastss  0x21ebb(%rip),%ymm7        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,61,103,53,2,0        ; vbroadcastss  0x23567(%rip),%ymm7        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,228,89,223                      ; vmulps        %ymm7,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,16,60,36                    ; vmovups       (%rsp),%ymm7
@@ -9449,7 +9449,7 @@
   DB  197,177,108,208                     ; vpunpcklqdq   %xmm0,%xmm9,%xmm2
   DB  197,49,109,232                      ; vpunpckhqdq   %xmm0,%xmm9,%xmm13
   DB  196,65,57,108,212                   ; vpunpcklqdq   %xmm12,%xmm8,%xmm10
-  DB  197,121,111,29,1,36,2,0             ; vmovdqa       0x22401(%rip),%xmm11        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,121,111,29,225,58,2,0           ; vmovdqa       0x23ae1(%rip),%xmm11        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,193,105,219,195                 ; vpand         %xmm11,%xmm2,%xmm0
   DB  197,249,112,208,78                  ; vpshufd       $0x4e,%xmm0,%xmm2
   DB  196,226,121,51,210                  ; vpmovzxwd     %xmm2,%xmm2
@@ -9543,7 +9543,7 @@
   DB  196,226,121,51,219                  ; vpmovzxwd     %xmm3,%xmm3
   DB  196,227,77,24,219,1                 ; vinsertf128   $0x1,%xmm3,%ymm6,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,226,125,24,53,152,27,2,0        ; vbroadcastss  0x21b98(%rip),%ymm6        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,226,125,24,53,68,50,2,0         ; vbroadcastss  0x23244(%rip),%ymm6        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  197,228,89,222                      ; vmulps        %ymm6,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,16,52,36                    ; vmovups       (%rsp),%ymm6
@@ -9607,7 +9607,7 @@
   DB  197,105,108,216                     ; vpunpcklqdq   %xmm0,%xmm2,%xmm11
   DB  197,241,108,211                     ; vpunpcklqdq   %xmm3,%xmm1,%xmm2
   DB  197,241,109,203                     ; vpunpckhqdq   %xmm3,%xmm1,%xmm1
-  DB  197,121,111,53,251,32,2,0           ; vmovdqa       0x220fb(%rip),%xmm14        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,121,111,53,219,55,2,0           ; vmovdqa       0x237db(%rip),%xmm14        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,193,105,219,214                 ; vpand         %xmm14,%xmm2,%xmm2
   DB  197,249,112,218,78                  ; vpshufd       $0x4e,%xmm2,%xmm3
   DB  196,226,121,51,219                  ; vpmovzxwd     %xmm3,%xmm3
@@ -9693,7 +9693,7 @@
   DB  196,227,105,33,215,48               ; vinsertps     $0x30,%xmm7,%xmm2,%xmm2
   DB  196,227,77,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm6,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,40,24,2,0         ; vbroadcastss  0x21828(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,212,46,2,0        ; vbroadcastss  0x22ed4(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,16,52,36                    ; vmovups       (%rsp),%ymm6
   DB  197,252,16,124,36,32                ; vmovups       0x20(%rsp),%ymm7
   DB  72,131,196,88                       ; add           $0x58,%rsp
@@ -9763,7 +9763,7 @@
 PUBLIC _sk_byte_tables_avx
 _sk_byte_tables_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,14,23,2,0           ; vbroadcastss  0x2170e(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,186,45,2,0          ; vbroadcastss  0x22dba(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  197,125,91,200                      ; vcvtps2dq     %ymm0,%ymm9
   DB  196,65,249,126,200                  ; vmovq         %xmm9,%r8
@@ -9898,7 +9898,7 @@
   DB  196,194,121,49,204                  ; vpmovzxbd     %xmm12,%xmm1
   DB  196,194,121,49,213                  ; vpmovzxbd     %xmm13,%xmm2
   DB  196,227,117,24,202,1                ; vinsertf128   $0x1,%xmm2,%ymm1,%ymm1
-  DB  196,98,125,24,13,202,20,2,0         ; vbroadcastss  0x214ca(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,118,43,2,0         ; vbroadcastss  0x22b76(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
   DB  196,193,116,89,201                  ; vmulps        %ymm9,%ymm1,%ymm1
@@ -10024,7 +10024,7 @@
   DB  196,194,121,49,203                  ; vpmovzxbd     %xmm11,%xmm1
   DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,127,18,2,0         ; vbroadcastss  0x2127f(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,43,41,2,0          ; vbroadcastss  0x2292b(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  196,194,121,49,202                  ; vpmovzxbd     %xmm10,%xmm1
   DB  196,194,121,49,212                  ; vpmovzxbd     %xmm12,%xmm2
@@ -10251,36 +10251,36 @@
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
   DB  196,98,125,24,16                    ; vbroadcastss  (%rax),%ymm10
   DB  197,124,91,216                      ; vcvtdq2ps     %ymm0,%ymm11
-  DB  196,98,125,24,37,16,15,2,0          ; vbroadcastss  0x20f10(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,37,188,37,2,0         ; vbroadcastss  0x225bc(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,6,15,2,0           ; vbroadcastss  0x20f06(%rip),%ymm12        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,178,37,2,0         ; vbroadcastss  0x225b2(%rip),%ymm12        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,193,124,84,196                  ; vandps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,88,14,2,0          ; vbroadcastss  0x20e58(%rip),%ymm12        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,37,4,37,2,0           ; vbroadcastss  0x22504(%rip),%ymm12        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,124,86,196                  ; vorps         %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,238,14,2,0         ; vbroadcastss  0x20eee(%rip),%ymm12        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,37,154,37,2,0         ; vbroadcastss  0x2259a(%rip),%ymm12        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,228,14,2,0         ; vbroadcastss  0x20ee4(%rip),%ymm12        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,37,144,37,2,0         ; vbroadcastss  0x22590(%rip),%ymm12        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,65,124,89,228                   ; vmulps        %ymm12,%ymm0,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,213,14,2,0         ; vbroadcastss  0x20ed5(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,129,37,2,0         ; vbroadcastss  0x22581(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,193,124,88,196                  ; vaddps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,37,203,14,2,0         ; vbroadcastss  0x20ecb(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,119,37,2,0         ; vbroadcastss  0x22577(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  197,156,94,192                      ; vdivps        %ymm0,%ymm12,%ymm0
   DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
   DB  197,172,89,192                      ; vmulps        %ymm0,%ymm10,%ymm0
   DB  196,99,125,8,208,1                  ; vroundps      $0x1,%ymm0,%ymm10
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,175,14,2,0         ; vbroadcastss  0x20eaf(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,91,37,2,0          ; vbroadcastss  0x2255b(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,124,88,195                  ; vaddps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,165,14,2,0         ; vbroadcastss  0x20ea5(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,81,37,2,0          ; vbroadcastss  0x22551(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,124,92,195                  ; vsubps        %ymm11,%ymm0,%ymm0
-  DB  196,98,125,24,29,150,14,2,0         ; vbroadcastss  0x20e96(%rip),%ymm11        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,98,125,24,29,66,37,2,0          ; vbroadcastss  0x22542(%rip),%ymm11        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,140,14,2,0         ; vbroadcastss  0x20e8c(%rip),%ymm11        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,29,56,37,2,0          ; vbroadcastss  0x22538(%rip),%ymm11        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,124,88,194                  ; vaddps        %ymm10,%ymm0,%ymm0
-  DB  196,98,125,24,21,125,14,2,0         ; vbroadcastss  0x20e7d(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,41,37,2,0          ; vbroadcastss  0x22529(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,124,89,194                  ; vmulps        %ymm10,%ymm0,%ymm0
   DB  197,253,91,192                      ; vcvtps2dq     %ymm0,%ymm0
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -10288,7 +10288,7 @@
   DB  196,195,125,74,193,128              ; vblendvps     %ymm8,%ymm9,%ymm0,%ymm0
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,145,13,2,0          ; vbroadcastss  0x20d91(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,61,36,2,0           ; vbroadcastss  0x2243d(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -10308,36 +10308,36 @@
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
   DB  196,98,125,24,16                    ; vbroadcastss  (%rax),%ymm10
   DB  197,124,91,217                      ; vcvtdq2ps     %ymm1,%ymm11
-  DB  196,98,125,24,37,214,13,2,0         ; vbroadcastss  0x20dd6(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,37,130,36,2,0         ; vbroadcastss  0x22482(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,204,13,2,0         ; vbroadcastss  0x20dcc(%rip),%ymm12        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,120,36,2,0         ; vbroadcastss  0x22478(%rip),%ymm12        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,193,116,84,204                  ; vandps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,30,13,2,0          ; vbroadcastss  0x20d1e(%rip),%ymm12        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,37,202,35,2,0         ; vbroadcastss  0x223ca(%rip),%ymm12        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,116,86,204                  ; vorps         %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,180,13,2,0         ; vbroadcastss  0x20db4(%rip),%ymm12        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,37,96,36,2,0          ; vbroadcastss  0x22460(%rip),%ymm12        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,170,13,2,0         ; vbroadcastss  0x20daa(%rip),%ymm12        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,37,86,36,2,0          ; vbroadcastss  0x22456(%rip),%ymm12        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,65,116,89,228                   ; vmulps        %ymm12,%ymm1,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,155,13,2,0         ; vbroadcastss  0x20d9b(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,71,36,2,0          ; vbroadcastss  0x22447(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,193,116,88,204                  ; vaddps        %ymm12,%ymm1,%ymm1
-  DB  196,98,125,24,37,145,13,2,0         ; vbroadcastss  0x20d91(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,61,36,2,0          ; vbroadcastss  0x2243d(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  197,156,94,201                      ; vdivps        %ymm1,%ymm12,%ymm1
   DB  197,164,92,201                      ; vsubps        %ymm1,%ymm11,%ymm1
   DB  197,172,89,201                      ; vmulps        %ymm1,%ymm10,%ymm1
   DB  196,99,125,8,209,1                  ; vroundps      $0x1,%ymm1,%ymm10
   DB  196,65,116,92,210                   ; vsubps        %ymm10,%ymm1,%ymm10
-  DB  196,98,125,24,29,117,13,2,0         ; vbroadcastss  0x20d75(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,33,36,2,0          ; vbroadcastss  0x22421(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,107,13,2,0         ; vbroadcastss  0x20d6b(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,23,36,2,0          ; vbroadcastss  0x22417(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,116,92,203                  ; vsubps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,29,92,13,2,0          ; vbroadcastss  0x20d5c(%rip),%ymm11        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,98,125,24,29,8,36,2,0           ; vbroadcastss  0x22408(%rip),%ymm11        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,82,13,2,0          ; vbroadcastss  0x20d52(%rip),%ymm11        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,29,254,35,2,0         ; vbroadcastss  0x223fe(%rip),%ymm11        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,21,67,13,2,0          ; vbroadcastss  0x20d43(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,239,35,2,0         ; vbroadcastss  0x223ef(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,116,89,202                  ; vmulps        %ymm10,%ymm1,%ymm1
   DB  197,253,91,201                      ; vcvtps2dq     %ymm1,%ymm1
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -10345,7 +10345,7 @@
   DB  196,195,117,74,201,128              ; vblendvps     %ymm8,%ymm9,%ymm1,%ymm1
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,5,87,12,2,0           ; vbroadcastss  0x20c57(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,3,35,2,0            ; vbroadcastss  0x22303(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -10365,36 +10365,36 @@
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,98,125,24,16                    ; vbroadcastss  (%rax),%ymm10
   DB  197,124,91,218                      ; vcvtdq2ps     %ymm2,%ymm11
-  DB  196,98,125,24,37,156,12,2,0         ; vbroadcastss  0x20c9c(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,37,72,35,2,0          ; vbroadcastss  0x22348(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,146,12,2,0         ; vbroadcastss  0x20c92(%rip),%ymm12        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,62,35,2,0          ; vbroadcastss  0x2233e(%rip),%ymm12        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,193,108,84,212                  ; vandps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,228,11,2,0         ; vbroadcastss  0x20be4(%rip),%ymm12        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,37,144,34,2,0         ; vbroadcastss  0x22290(%rip),%ymm12        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,108,86,212                  ; vorps         %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,122,12,2,0         ; vbroadcastss  0x20c7a(%rip),%ymm12        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,37,38,35,2,0          ; vbroadcastss  0x22326(%rip),%ymm12        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,112,12,2,0         ; vbroadcastss  0x20c70(%rip),%ymm12        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,37,28,35,2,0          ; vbroadcastss  0x2231c(%rip),%ymm12        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,65,108,89,228                   ; vmulps        %ymm12,%ymm2,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,97,12,2,0          ; vbroadcastss  0x20c61(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,13,35,2,0          ; vbroadcastss  0x2230d(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,193,108,88,212                  ; vaddps        %ymm12,%ymm2,%ymm2
-  DB  196,98,125,24,37,87,12,2,0          ; vbroadcastss  0x20c57(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,3,35,2,0           ; vbroadcastss  0x22303(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  197,156,94,210                      ; vdivps        %ymm2,%ymm12,%ymm2
   DB  197,164,92,210                      ; vsubps        %ymm2,%ymm11,%ymm2
   DB  197,172,89,210                      ; vmulps        %ymm2,%ymm10,%ymm2
   DB  196,99,125,8,210,1                  ; vroundps      $0x1,%ymm2,%ymm10
   DB  196,65,108,92,210                   ; vsubps        %ymm10,%ymm2,%ymm10
-  DB  196,98,125,24,29,59,12,2,0          ; vbroadcastss  0x20c3b(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,231,34,2,0         ; vbroadcastss  0x222e7(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,49,12,2,0          ; vbroadcastss  0x20c31(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,221,34,2,0         ; vbroadcastss  0x222dd(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,108,92,211                  ; vsubps        %ymm11,%ymm2,%ymm2
-  DB  196,98,125,24,29,34,12,2,0          ; vbroadcastss  0x20c22(%rip),%ymm11        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,98,125,24,29,206,34,2,0         ; vbroadcastss  0x222ce(%rip),%ymm11        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,24,12,2,0          ; vbroadcastss  0x20c18(%rip),%ymm11        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,29,196,34,2,0         ; vbroadcastss  0x222c4(%rip),%ymm11        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,108,88,210                  ; vaddps        %ymm10,%ymm2,%ymm2
-  DB  196,98,125,24,21,9,12,2,0           ; vbroadcastss  0x20c09(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,181,34,2,0         ; vbroadcastss  0x222b5(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,108,89,210                  ; vmulps        %ymm10,%ymm2,%ymm2
   DB  197,253,91,210                      ; vcvtps2dq     %ymm2,%ymm2
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -10402,7 +10402,7 @@
   DB  196,195,109,74,209,128              ; vblendvps     %ymm8,%ymm9,%ymm2,%ymm2
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,29,11,2,0           ; vbroadcastss  0x20b1d(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,201,33,2,0          ; vbroadcastss  0x221c9(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -10422,36 +10422,36 @@
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
   DB  196,98,125,24,16                    ; vbroadcastss  (%rax),%ymm10
   DB  197,124,91,219                      ; vcvtdq2ps     %ymm3,%ymm11
-  DB  196,98,125,24,37,98,11,2,0          ; vbroadcastss  0x20b62(%rip),%ymm12        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,37,14,34,2,0          ; vbroadcastss  0x2220e(%rip),%ymm12        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,88,11,2,0          ; vbroadcastss  0x20b58(%rip),%ymm12        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,37,4,34,2,0           ; vbroadcastss  0x22204(%rip),%ymm12        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,193,100,84,220                  ; vandps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,170,10,2,0         ; vbroadcastss  0x20aaa(%rip),%ymm12        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,37,86,33,2,0          ; vbroadcastss  0x22156(%rip),%ymm12        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,100,86,220                  ; vorps         %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,64,11,2,0          ; vbroadcastss  0x20b40(%rip),%ymm12        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,37,236,33,2,0         ; vbroadcastss  0x221ec(%rip),%ymm12        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,65,36,88,220                    ; vaddps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,54,11,2,0          ; vbroadcastss  0x20b36(%rip),%ymm12        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,37,226,33,2,0         ; vbroadcastss  0x221e2(%rip),%ymm12        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,65,100,89,228                   ; vmulps        %ymm12,%ymm3,%ymm12
   DB  196,65,36,92,220                    ; vsubps        %ymm12,%ymm11,%ymm11
-  DB  196,98,125,24,37,39,11,2,0          ; vbroadcastss  0x20b27(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,211,33,2,0         ; vbroadcastss  0x221d3(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,193,100,88,220                  ; vaddps        %ymm12,%ymm3,%ymm3
-  DB  196,98,125,24,37,29,11,2,0          ; vbroadcastss  0x20b1d(%rip),%ymm12        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,37,201,33,2,0         ; vbroadcastss  0x221c9(%rip),%ymm12        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  197,156,94,219                      ; vdivps        %ymm3,%ymm12,%ymm3
   DB  197,164,92,219                      ; vsubps        %ymm3,%ymm11,%ymm3
   DB  197,172,89,219                      ; vmulps        %ymm3,%ymm10,%ymm3
   DB  196,99,125,8,211,1                  ; vroundps      $0x1,%ymm3,%ymm10
   DB  196,65,100,92,210                   ; vsubps        %ymm10,%ymm3,%ymm10
-  DB  196,98,125,24,29,1,11,2,0           ; vbroadcastss  0x20b01(%rip),%ymm11        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,29,173,33,2,0         ; vbroadcastss  0x221ad(%rip),%ymm11        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,100,88,219                  ; vaddps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,247,10,2,0         ; vbroadcastss  0x20af7(%rip),%ymm11        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,98,125,24,29,163,33,2,0         ; vbroadcastss  0x221a3(%rip),%ymm11        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  196,65,44,89,219                    ; vmulps        %ymm11,%ymm10,%ymm11
   DB  196,193,100,92,219                  ; vsubps        %ymm11,%ymm3,%ymm3
-  DB  196,98,125,24,29,232,10,2,0         ; vbroadcastss  0x20ae8(%rip),%ymm11        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,98,125,24,29,148,33,2,0         ; vbroadcastss  0x22194(%rip),%ymm11        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  196,65,36,92,210                    ; vsubps        %ymm10,%ymm11,%ymm10
-  DB  196,98,125,24,29,222,10,2,0         ; vbroadcastss  0x20ade(%rip),%ymm11        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,98,125,24,29,138,33,2,0         ; vbroadcastss  0x2218a(%rip),%ymm11        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  196,65,36,94,210                    ; vdivps        %ymm10,%ymm11,%ymm10
   DB  196,193,100,88,218                  ; vaddps        %ymm10,%ymm3,%ymm3
-  DB  196,98,125,24,21,207,10,2,0         ; vbroadcastss  0x20acf(%rip),%ymm10        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,98,125,24,21,123,33,2,0         ; vbroadcastss  0x2217b(%rip),%ymm10        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  196,193,100,89,218                  ; vmulps        %ymm10,%ymm3,%ymm3
   DB  197,253,91,219                      ; vcvtps2dq     %ymm3,%ymm3
   DB  196,98,125,24,80,20                 ; vbroadcastss  0x14(%rax),%ymm10
@@ -10459,7 +10459,7 @@
   DB  196,195,101,74,217,128              ; vblendvps     %ymm8,%ymm9,%ymm3,%ymm3
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  196,193,100,95,216                  ; vmaxps        %ymm8,%ymm3,%ymm3
-  DB  196,98,125,24,5,227,9,2,0           ; vbroadcastss  0x209e3(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,143,32,2,0          ; vbroadcastss  0x2208f(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -10476,34 +10476,34 @@
   DB  197,252,40,233                      ; vmovaps       %ymm1,%ymm5
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,91,200                      ; vcvtdq2ps     %ymm0,%ymm1
-  DB  196,98,125,24,5,48,10,2,0           ; vbroadcastss  0x20a30(%rip),%ymm8        # 2bd50 <_sk_srcover_rgba_8888_sse2_lowp+0x348>
+  DB  196,98,125,24,5,220,32,2,0          ; vbroadcastss  0x220dc(%rip),%ymm8        # 2d3fc <_sk_srcover_rgba_8888_sse2_lowp+0x348>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,13,38,10,2,0          ; vbroadcastss  0x20a26(%rip),%ymm9        # 2bd54 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
+  DB  196,98,125,24,13,210,32,2,0         ; vbroadcastss  0x220d2(%rip),%ymm9        # 2d400 <_sk_srcover_rgba_8888_sse2_lowp+0x34c>
   DB  196,193,124,84,193                  ; vandps        %ymm9,%ymm0,%ymm0
-  DB  196,226,125,24,37,120,9,2,0         ; vbroadcastss  0x20978(%rip),%ymm4        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,37,36,32,2,0         ; vbroadcastss  0x22024(%rip),%ymm4        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,252,86,196                      ; vorps         %ymm4,%ymm0,%ymm0
-  DB  196,98,125,24,21,15,10,2,0          ; vbroadcastss  0x20a0f(%rip),%ymm10        # 2bd58 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
+  DB  196,98,125,24,21,187,32,2,0         ; vbroadcastss  0x220bb(%rip),%ymm10        # 2d404 <_sk_srcover_rgba_8888_sse2_lowp+0x350>
   DB  196,193,116,88,202                  ; vaddps        %ymm10,%ymm1,%ymm1
-  DB  196,98,125,24,29,5,10,2,0           ; vbroadcastss  0x20a05(%rip),%ymm11        # 2bd5c <_sk_srcover_rgba_8888_sse2_lowp+0x354>
+  DB  196,98,125,24,29,177,32,2,0         ; vbroadcastss  0x220b1(%rip),%ymm11        # 2d408 <_sk_srcover_rgba_8888_sse2_lowp+0x354>
   DB  196,193,124,89,211                  ; vmulps        %ymm11,%ymm0,%ymm2
   DB  197,244,92,202                      ; vsubps        %ymm2,%ymm1,%ymm1
-  DB  196,98,125,24,37,247,9,2,0          ; vbroadcastss  0x209f7(%rip),%ymm12        # 2bd60 <_sk_srcover_rgba_8888_sse2_lowp+0x358>
+  DB  196,98,125,24,37,163,32,2,0         ; vbroadcastss  0x220a3(%rip),%ymm12        # 2d40c <_sk_srcover_rgba_8888_sse2_lowp+0x358>
   DB  196,193,124,88,196                  ; vaddps        %ymm12,%ymm0,%ymm0
-  DB  196,98,125,24,45,237,9,2,0          ; vbroadcastss  0x209ed(%rip),%ymm13        # 2bd64 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
+  DB  196,98,125,24,45,153,32,2,0         ; vbroadcastss  0x22099(%rip),%ymm13        # 2d410 <_sk_srcover_rgba_8888_sse2_lowp+0x35c>
   DB  197,148,94,192                      ; vdivps        %ymm0,%ymm13,%ymm0
   DB  197,244,92,192                      ; vsubps        %ymm0,%ymm1,%ymm0
   DB  196,98,125,24,48                    ; vbroadcastss  (%rax),%ymm14
   DB  196,193,124,89,198                  ; vmulps        %ymm14,%ymm0,%ymm0
   DB  196,227,125,8,200,1                 ; vroundps      $0x1,%ymm0,%ymm1
   DB  197,252,92,241                      ; vsubps        %ymm1,%ymm0,%ymm6
-  DB  196,98,125,24,61,204,9,2,0          ; vbroadcastss  0x209cc(%rip),%ymm15        # 2bd68 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
+  DB  196,98,125,24,61,120,32,2,0         ; vbroadcastss  0x22078(%rip),%ymm15        # 2d414 <_sk_srcover_rgba_8888_sse2_lowp+0x360>
   DB  196,193,124,88,199                  ; vaddps        %ymm15,%ymm0,%ymm0
-  DB  196,226,125,24,29,194,9,2,0         ; vbroadcastss  0x209c2(%rip),%ymm3        # 2bd6c <_sk_srcover_rgba_8888_sse2_lowp+0x364>
+  DB  196,226,125,24,29,110,32,2,0        ; vbroadcastss  0x2206e(%rip),%ymm3        # 2d418 <_sk_srcover_rgba_8888_sse2_lowp+0x364>
   DB  197,204,89,203                      ; vmulps        %ymm3,%ymm6,%ymm1
   DB  197,252,92,201                      ; vsubps        %ymm1,%ymm0,%ymm1
-  DB  196,226,125,24,21,181,9,2,0         ; vbroadcastss  0x209b5(%rip),%ymm2        # 2bd70 <_sk_srcover_rgba_8888_sse2_lowp+0x368>
+  DB  196,226,125,24,21,97,32,2,0         ; vbroadcastss  0x22061(%rip),%ymm2        # 2d41c <_sk_srcover_rgba_8888_sse2_lowp+0x368>
   DB  197,236,92,198                      ; vsubps        %ymm6,%ymm2,%ymm0
-  DB  196,226,125,24,53,172,9,2,0         ; vbroadcastss  0x209ac(%rip),%ymm6        # 2bd74 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
+  DB  196,226,125,24,53,88,32,2,0         ; vbroadcastss  0x22058(%rip),%ymm6        # 2d420 <_sk_srcover_rgba_8888_sse2_lowp+0x36c>
   DB  197,204,94,192                      ; vdivps        %ymm0,%ymm6,%ymm0
   DB  197,244,88,192                      ; vaddps        %ymm0,%ymm1,%ymm0
   DB  197,252,17,4,36                     ; vmovups       %ymm0,(%rsp)
@@ -10545,7 +10545,7 @@
   DB  197,236,92,212                      ; vsubps        %ymm4,%ymm2,%ymm2
   DB  197,204,94,210                      ; vdivps        %ymm2,%ymm6,%ymm2
   DB  197,244,88,202                      ; vaddps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,240,8,2,0         ; vbroadcastss  0x208f0(%rip),%ymm2        # 2bd78 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
+  DB  196,226,125,24,21,156,31,2,0        ; vbroadcastss  0x21f9c(%rip),%ymm2        # 2d424 <_sk_srcover_rgba_8888_sse2_lowp+0x370>
   DB  197,236,89,28,36                    ; vmulps        (%rsp),%ymm2,%ymm3
   DB  197,252,89,226                      ; vmulps        %ymm2,%ymm0,%ymm4
   DB  197,244,89,210                      ; vmulps        %ymm2,%ymm1,%ymm2
@@ -10563,31 +10563,31 @@
 
 PUBLIC _sk_lab_to_xyz_avx
 _sk_lab_to_xyz_avx LABEL PROC
-  DB  196,98,125,24,5,163,8,2,0           ; vbroadcastss  0x208a3(%rip),%ymm8        # 2bd7c <_sk_srcover_rgba_8888_sse2_lowp+0x374>
+  DB  196,98,125,24,5,79,31,2,0           ; vbroadcastss  0x21f4f(%rip),%ymm8        # 2d428 <_sk_srcover_rgba_8888_sse2_lowp+0x374>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,253,7,2,0           ; vbroadcastss  0x207fd(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,169,30,2,0          ; vbroadcastss  0x21ea9(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
-  DB  196,98,125,24,13,139,8,2,0          ; vbroadcastss  0x2088b(%rip),%ymm9        # 2bd80 <_sk_srcover_rgba_8888_sse2_lowp+0x378>
+  DB  196,98,125,24,13,55,31,2,0          ; vbroadcastss  0x21f37(%rip),%ymm9        # 2d42c <_sk_srcover_rgba_8888_sse2_lowp+0x378>
   DB  196,193,116,88,201                  ; vaddps        %ymm9,%ymm1,%ymm1
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  196,193,108,88,209                  ; vaddps        %ymm9,%ymm2,%ymm2
-  DB  196,98,125,24,5,119,8,2,0           ; vbroadcastss  0x20877(%rip),%ymm8        # 2bd84 <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
+  DB  196,98,125,24,5,35,31,2,0           ; vbroadcastss  0x21f23(%rip),%ymm8        # 2d430 <_sk_srcover_rgba_8888_sse2_lowp+0x37c>
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,109,8,2,0           ; vbroadcastss  0x2086d(%rip),%ymm8        # 2bd88 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
+  DB  196,98,125,24,5,25,31,2,0           ; vbroadcastss  0x21f19(%rip),%ymm8        # 2d434 <_sk_srcover_rgba_8888_sse2_lowp+0x380>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,5,99,8,2,0            ; vbroadcastss  0x20863(%rip),%ymm8        # 2bd8c <_sk_srcover_rgba_8888_sse2_lowp+0x384>
+  DB  196,98,125,24,5,15,31,2,0           ; vbroadcastss  0x21f0f(%rip),%ymm8        # 2d438 <_sk_srcover_rgba_8888_sse2_lowp+0x384>
   DB  196,193,116,89,200                  ; vmulps        %ymm8,%ymm1,%ymm1
   DB  197,252,88,201                      ; vaddps        %ymm1,%ymm0,%ymm1
-  DB  196,98,125,24,5,85,8,2,0            ; vbroadcastss  0x20855(%rip),%ymm8        # 2bd90 <_sk_srcover_rgba_8888_sse2_lowp+0x388>
+  DB  196,98,125,24,5,1,31,2,0            ; vbroadcastss  0x21f01(%rip),%ymm8        # 2d43c <_sk_srcover_rgba_8888_sse2_lowp+0x388>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  197,252,92,210                      ; vsubps        %ymm2,%ymm0,%ymm2
   DB  197,116,89,193                      ; vmulps        %ymm1,%ymm1,%ymm8
   DB  196,65,116,89,192                   ; vmulps        %ymm8,%ymm1,%ymm8
-  DB  196,98,125,24,13,62,8,2,0           ; vbroadcastss  0x2083e(%rip),%ymm9        # 2bd94 <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
+  DB  196,98,125,24,13,234,30,2,0         ; vbroadcastss  0x21eea(%rip),%ymm9        # 2d440 <_sk_srcover_rgba_8888_sse2_lowp+0x38c>
   DB  196,65,52,194,208,1                 ; vcmpltps      %ymm8,%ymm9,%ymm10
-  DB  196,98,125,24,29,51,8,2,0           ; vbroadcastss  0x20833(%rip),%ymm11        # 2bd98 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
+  DB  196,98,125,24,29,223,30,2,0         ; vbroadcastss  0x21edf(%rip),%ymm11        # 2d444 <_sk_srcover_rgba_8888_sse2_lowp+0x390>
   DB  196,193,116,88,203                  ; vaddps        %ymm11,%ymm1,%ymm1
-  DB  196,98,125,24,37,41,8,2,0           ; vbroadcastss  0x20829(%rip),%ymm12        # 2bd9c <_sk_srcover_rgba_8888_sse2_lowp+0x394>
+  DB  196,98,125,24,37,213,30,2,0         ; vbroadcastss  0x21ed5(%rip),%ymm12        # 2d448 <_sk_srcover_rgba_8888_sse2_lowp+0x394>
   DB  196,193,116,89,204                  ; vmulps        %ymm12,%ymm1,%ymm1
   DB  196,67,117,74,192,160               ; vblendvps     %ymm10,%ymm8,%ymm1,%ymm8
   DB  197,252,89,200                      ; vmulps        %ymm0,%ymm0,%ymm1
@@ -10602,9 +10602,9 @@
   DB  196,193,108,88,211                  ; vaddps        %ymm11,%ymm2,%ymm2
   DB  196,193,108,89,212                  ; vmulps        %ymm12,%ymm2,%ymm2
   DB  196,227,109,74,208,144              ; vblendvps     %ymm9,%ymm0,%ymm2,%ymm2
-  DB  196,226,125,24,5,223,7,2,0          ; vbroadcastss  0x207df(%rip),%ymm0        # 2bda0 <_sk_srcover_rgba_8888_sse2_lowp+0x398>
+  DB  196,226,125,24,5,139,30,2,0         ; vbroadcastss  0x21e8b(%rip),%ymm0        # 2d44c <_sk_srcover_rgba_8888_sse2_lowp+0x398>
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,214,7,2,0           ; vbroadcastss  0x207d6(%rip),%ymm8        # 2bda4 <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
+  DB  196,98,125,24,5,130,30,2,0          ; vbroadcastss  0x21e82(%rip),%ymm8        # 2d450 <_sk_srcover_rgba_8888_sse2_lowp+0x39c>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -10620,13 +10620,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,68                              ; jne           b633 <_sk_load_a8_avx+0x5c>
   DB  196,130,121,48,4,24                 ; vpmovzxbw     (%r8,%r11,1),%xmm0
-  DB  197,249,219,5,211,13,2,0            ; vpand         0x20dd3(%rip),%xmm0,%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,249,219,5,179,36,2,0            ; vpand         0x224b3(%rip),%xmm0,%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,121,51,200                  ; vpmovzxwd     %xmm0,%xmm1
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,17,7,2,0          ; vbroadcastss  0x20711(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,189,29,2,0        ; vbroadcastss  0x21dbd(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
@@ -10698,13 +10698,13 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,68                              ; jne           b744 <_sk_load_a8_dst_avx+0x5c>
   DB  196,130,121,48,36,24                ; vpmovzxbw     (%r8,%r11,1),%xmm4
-  DB  197,217,219,37,194,12,2,0           ; vpand         0x20cc2(%rip),%xmm4,%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,217,219,37,162,35,2,0           ; vpand         0x223a2(%rip),%xmm4,%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,121,51,236                  ; vpmovzxwd     %xmm4,%xmm5
   DB  197,249,112,228,78                  ; vpshufd       $0x4e,%xmm4,%xmm4
   DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
   DB  196,227,85,24,228,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,0,6,2,0           ; vbroadcastss  0x20600(%rip),%ymm5        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,45,172,28,2,0        ; vbroadcastss  0x21cac(%rip),%ymm5        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,220,89,253                      ; vmulps        %ymm5,%ymm4,%ymm7
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,220,87,228                      ; vxorps        %ymm4,%ymm4,%ymm4
@@ -10752,7 +10752,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,221                 ; callq         ffffffffde00b7ec <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfdfde4>
+  DB  232,255,255,255,221                 ; callq         ffffffffde00b7ec <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffddfde738>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,210                             ; callq         *%rdx
@@ -10830,7 +10830,7 @@
   DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
   DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,6,4,2,0           ; vbroadcastss  0x20406(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,178,26,2,0        ; vbroadcastss  0x21ab2(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,217                      ; vmulps        %ymm1,%ymm0,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
@@ -10851,7 +10851,7 @@
   DB  77,15,175,193                       ; imul          %r9,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,130,3,2,0           ; vbroadcastss  0x20382(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,46,26,2,0           ; vbroadcastss  0x21a2e(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,100,89,192                   ; vmulps        %ymm8,%ymm3,%ymm8
   DB  196,65,125,91,192                   ; vcvtps2dq     %ymm8,%ymm8
   DB  196,67,125,25,193,1                 ; vextractf128  $0x1,%ymm8,%xmm9
@@ -10876,13 +10876,13 @@
   DB  196,3,121,20,4,24,0                 ; vpextrb       $0x0,%xmm8,(%r8,%r11,1)
   DB  235,202                             ; jmp           b986 <_sk_store_a8_avx+0x40>
   DB  196,3,121,20,68,24,2,4              ; vpextrb       $0x4,%xmm8,0x2(%r8,%r11,1)
-  DB  196,98,57,0,5,19,10,2,0             ; vpshufb       0x20a13(%rip),%xmm8,%xmm8        # 2c3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x9d8>
+  DB  196,98,57,0,5,243,32,2,0            ; vpshufb       0x220f3(%rip),%xmm8,%xmm8        # 2dac0 <_sk_srcover_rgba_8888_sse2_lowp+0xa0c>
   DB  196,3,121,21,4,24,0                 ; vpextrw       $0x0,%xmm8,(%r8,%r11,1)
   DB  235,176                             ; jmp           b986 <_sk_store_a8_avx+0x40>
   DB  196,3,121,20,68,24,6,12             ; vpextrb       $0xc,%xmm8,0x6(%r8,%r11,1)
   DB  196,3,121,20,68,24,5,10             ; vpextrb       $0xa,%xmm8,0x5(%r8,%r11,1)
   DB  196,3,121,20,68,24,4,8              ; vpextrb       $0x8,%xmm8,0x4(%r8,%r11,1)
-  DB  196,98,57,0,5,249,9,2,0             ; vpshufb       0x209f9(%rip),%xmm8,%xmm8        # 2c3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x9e8>
+  DB  196,98,57,0,5,217,32,2,0            ; vpshufb       0x220d9(%rip),%xmm8,%xmm8        # 2dad0 <_sk_srcover_rgba_8888_sse2_lowp+0xa1c>
   DB  196,1,121,126,4,24                  ; vmovd         %xmm8,(%r8,%r11,1)
   DB  235,135                             ; jmp           b986 <_sk_store_a8_avx+0x40>
   DB  144                                 ; nop
@@ -10917,16 +10917,16 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,73                              ; jne           ba7d <_sk_load_g8_avx+0x61>
   DB  196,130,121,48,4,24                 ; vpmovzxbw     (%r8,%r11,1),%xmm0
-  DB  197,249,219,5,142,9,2,0             ; vpand         0x2098e(%rip),%xmm0,%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,249,219,5,110,32,2,0            ; vpand         0x2206e(%rip),%xmm0,%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,121,51,200                  ; vpmovzxwd     %xmm0,%xmm1
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,204,2,2,0         ; vbroadcastss  0x202cc(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,120,25,2,0        ; vbroadcastss  0x21978(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,69,2,2,0          ; vbroadcastss  0x20245(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,241,24,2,0        ; vbroadcastss  0x218f1(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  255,224                             ; jmpq          *%rax
@@ -10971,7 +10971,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,222                 ; jmpq          ffffffffdf00bb24 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefe011c>
+  DB  233,255,255,255,222                 ; jmpq          ffffffffdf00bb24 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefdea70>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,211                             ; callq         *%rbx
@@ -10993,16 +10993,16 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,73                              ; jne           bb91 <_sk_load_g8_dst_avx+0x61>
   DB  196,130,121,48,36,24                ; vpmovzxbw     (%r8,%r11,1),%xmm4
-  DB  197,217,219,37,122,8,2,0            ; vpand         0x2087a(%rip),%xmm4,%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  197,217,219,37,90,31,2,0            ; vpand         0x21f5a(%rip),%xmm4,%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  196,226,121,51,236                  ; vpmovzxwd     %xmm4,%xmm5
   DB  197,249,112,228,78                  ; vpshufd       $0x4e,%xmm4,%xmm4
   DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
   DB  196,227,85,24,228,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,184,1,2,0         ; vbroadcastss  0x201b8(%rip),%ymm5        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,45,100,24,2,0        ; vbroadcastss  0x21864(%rip),%ymm5        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,61,49,1,2,0          ; vbroadcastss  0x20131(%rip),%ymm7        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,61,221,23,2,0        ; vbroadcastss  0x217dd(%rip),%ymm7        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,236                      ; vmovaps       %ymm4,%ymm5
   DB  197,252,40,244                      ; vmovaps       %ymm4,%ymm6
   DB  255,224                             ; jmpq          *%rax
@@ -11047,7 +11047,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  233,255,255,255,222                 ; jmpq          ffffffffdf00bc38 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefe0230>
+  DB  233,255,255,255,222                 ; jmpq          ffffffffdf00bc38 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdefdeb84>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,211                             ; callq         *%rbx
@@ -11125,10 +11125,10 @@
   DB  196,226,121,49,201                  ; vpmovzxbd     %xmm1,%xmm1
   DB  196,227,125,24,193,1                ; vinsertf128   $0x1,%xmm1,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,186,255,1,0       ; vbroadcastss  0x1ffba(%rip),%ymm1        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,226,125,24,13,102,22,2,0        ; vbroadcastss  0x21666(%rip),%ymm1        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,51,255,1,0        ; vbroadcastss  0x1ff33(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,223,21,2,0        ; vbroadcastss  0x215df(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
   DB  197,252,40,208                      ; vmovaps       %ymm0,%ymm2
   DB  91                                  ; pop           %rbx
@@ -11154,23 +11154,23 @@
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,208,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm2
-  DB  196,226,125,24,5,90,255,1,0         ; vbroadcastss  0x1ff5a(%rip),%ymm0        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,24,5,6,22,2,0           ; vbroadcastss  0x21606(%rip),%ymm0        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,77,255,1,0        ; vbroadcastss  0x1ff4d(%rip),%ymm1        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,13,249,21,2,0        ; vbroadcastss  0x215f9(%rip),%ymm1        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,68,255,1,0        ; vbroadcastss  0x1ff44(%rip),%ymm1        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,24,13,240,21,2,0        ; vbroadcastss  0x215f0(%rip),%ymm1        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,55,255,1,0        ; vbroadcastss  0x1ff37(%rip),%ymm3        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,29,227,21,2,0        ; vbroadcastss  0x215e3(%rip),%ymm3        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,46,255,1,0        ; vbroadcastss  0x1ff2e(%rip),%ymm3        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,24,29,218,21,2,0        ; vbroadcastss  0x215da(%rip),%ymm3        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,33,255,1,0        ; vbroadcastss  0x1ff21(%rip),%ymm3        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,29,205,21,2,0        ; vbroadcastss  0x215cd(%rip),%ymm3        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,130,254,1,0       ; vbroadcastss  0x1fe82(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,46,21,2,0         ; vbroadcastss  0x2152e(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,7                        ; and           $0x7,%r9b
@@ -11233,23 +11233,23 @@
   DB  197,249,112,228,78                  ; vpshufd       $0x4e,%xmm4,%xmm4
   DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
   DB  196,227,85,24,244,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm6
-  DB  196,226,125,24,37,25,254,1,0        ; vbroadcastss  0x1fe19(%rip),%ymm4        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,24,37,197,20,2,0        ; vbroadcastss  0x214c5(%rip),%ymm4        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,204,84,228                      ; vandps        %ymm4,%ymm6,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,12,254,1,0        ; vbroadcastss  0x1fe0c(%rip),%ymm5        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,45,184,20,2,0        ; vbroadcastss  0x214b8(%rip),%ymm5        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
-  DB  196,226,125,24,45,3,254,1,0         ; vbroadcastss  0x1fe03(%rip),%ymm5        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,24,45,175,20,2,0        ; vbroadcastss  0x214af(%rip),%ymm5        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,204,84,237                      ; vandps        %ymm5,%ymm6,%ymm5
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,226,125,24,61,246,253,1,0       ; vbroadcastss  0x1fdf6(%rip),%ymm7        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,61,162,20,2,0        ; vbroadcastss  0x214a2(%rip),%ymm7        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,212,89,239                      ; vmulps        %ymm7,%ymm5,%ymm5
-  DB  196,226,125,24,61,237,253,1,0       ; vbroadcastss  0x1fded(%rip),%ymm7        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,24,61,153,20,2,0        ; vbroadcastss  0x21499(%rip),%ymm7        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,204,84,247                      ; vandps        %ymm7,%ymm6,%ymm6
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
-  DB  196,226,125,24,61,224,253,1,0       ; vbroadcastss  0x1fde0(%rip),%ymm7        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,61,140,20,2,0        ; vbroadcastss  0x2148c(%rip),%ymm7        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,204,89,247                      ; vmulps        %ymm7,%ymm6,%ymm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,61,65,253,1,0        ; vbroadcastss  0x1fd41(%rip),%ymm7        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,61,237,19,2,0        ; vbroadcastss  0x213ed(%rip),%ymm7        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,7                        ; and           $0x7,%r9b
@@ -11367,23 +11367,23 @@
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,208,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm2
-  DB  196,226,125,24,5,233,251,1,0        ; vbroadcastss  0x1fbe9(%rip),%ymm0        # 2bd34 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
+  DB  196,226,125,24,5,149,18,2,0         ; vbroadcastss  0x21295(%rip),%ymm0        # 2d3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x32c>
   DB  197,236,84,192                      ; vandps        %ymm0,%ymm2,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,220,251,1,0       ; vbroadcastss  0x1fbdc(%rip),%ymm1        # 2bd38 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
+  DB  196,226,125,24,13,136,18,2,0        ; vbroadcastss  0x21288(%rip),%ymm1        # 2d3e4 <_sk_srcover_rgba_8888_sse2_lowp+0x330>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,211,251,1,0       ; vbroadcastss  0x1fbd3(%rip),%ymm1        # 2bd3c <_sk_srcover_rgba_8888_sse2_lowp+0x334>
+  DB  196,226,125,24,13,127,18,2,0        ; vbroadcastss  0x2127f(%rip),%ymm1        # 2d3e8 <_sk_srcover_rgba_8888_sse2_lowp+0x334>
   DB  197,236,84,201                      ; vandps        %ymm1,%ymm2,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,29,198,251,1,0       ; vbroadcastss  0x1fbc6(%rip),%ymm3        # 2bd40 <_sk_srcover_rgba_8888_sse2_lowp+0x338>
+  DB  196,226,125,24,29,114,18,2,0        ; vbroadcastss  0x21272(%rip),%ymm3        # 2d3ec <_sk_srcover_rgba_8888_sse2_lowp+0x338>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
-  DB  196,226,125,24,29,189,251,1,0       ; vbroadcastss  0x1fbbd(%rip),%ymm3        # 2bd44 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
+  DB  196,226,125,24,29,105,18,2,0        ; vbroadcastss  0x21269(%rip),%ymm3        # 2d3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x33c>
   DB  197,236,84,211                      ; vandps        %ymm3,%ymm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,176,251,1,0       ; vbroadcastss  0x1fbb0(%rip),%ymm3        # 2bd48 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
+  DB  196,226,125,24,29,92,18,2,0         ; vbroadcastss  0x2125c(%rip),%ymm3        # 2d3f4 <_sk_srcover_rgba_8888_sse2_lowp+0x340>
   DB  197,236,89,211                      ; vmulps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,17,251,1,0        ; vbroadcastss  0x1fb11(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,189,17,2,0        ; vbroadcastss  0x211bd(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  91                                  ; pop           %rbx
   DB  65,92                               ; pop           %r12
   DB  65,94                               ; pop           %r14
@@ -11400,14 +11400,14 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,216,251,1,0         ; vbroadcastss  0x1fbd8(%rip),%ymm8        # 2bda8 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
+  DB  196,98,125,24,5,132,18,2,0          ; vbroadcastss  0x21284(%rip),%ymm8        # 2d454 <_sk_srcover_rgba_8888_sse2_lowp+0x3a0>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,193,41,114,241,11               ; vpslld        $0xb,%xmm9,%xmm10
   DB  196,67,125,25,201,1                 ; vextractf128  $0x1,%ymm9,%xmm9
   DB  196,193,49,114,241,11               ; vpslld        $0xb,%xmm9,%xmm9
   DB  196,67,45,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm10,%ymm9
-  DB  196,98,125,24,21,177,251,1,0        ; vbroadcastss  0x1fbb1(%rip),%ymm10        # 2bdac <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
+  DB  196,98,125,24,21,93,18,2,0          ; vbroadcastss  0x2125d(%rip),%ymm10        # 2d458 <_sk_srcover_rgba_8888_sse2_lowp+0x3a4>
   DB  196,65,116,89,210                   ; vmulps        %ymm10,%ymm1,%ymm10
   DB  196,65,125,91,210                   ; vcvtps2dq     %ymm10,%ymm10
   DB  196,193,33,114,242,5                ; vpslld        $0x5,%xmm10,%xmm11
@@ -11462,7 +11462,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,224                 ; callq         ffffffffe100c2c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fe08b8>
+  DB  232,255,255,255,224                 ; callq         ffffffffe100c2c0 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe0fdf20c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -11483,25 +11483,25 @@
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,216,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm3
-  DB  196,226,125,24,5,169,250,1,0        ; vbroadcastss  0x1faa9(%rip),%ymm0        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,24,5,85,17,2,0          ; vbroadcastss  0x21155(%rip),%ymm0        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,156,250,1,0       ; vbroadcastss  0x1fa9c(%rip),%ymm1        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,13,72,17,2,0         ; vbroadcastss  0x21148(%rip),%ymm1        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,147,250,1,0       ; vbroadcastss  0x1fa93(%rip),%ymm1        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,24,13,63,17,2,0         ; vbroadcastss  0x2113f(%rip),%ymm1        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,134,250,1,0       ; vbroadcastss  0x1fa86(%rip),%ymm2        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,21,50,17,2,0         ; vbroadcastss  0x21132(%rip),%ymm2        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,125,250,1,0       ; vbroadcastss  0x1fa7d(%rip),%ymm2        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,24,21,41,17,2,0         ; vbroadcastss  0x21129(%rip),%ymm2        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,112,250,1,0         ; vbroadcastss  0x1fa70(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,28,17,2,0           ; vbroadcastss  0x2111c(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,102,250,1,0         ; vbroadcastss  0x1fa66(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,24,5,18,17,2,0           ; vbroadcastss  0x21112(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,88,250,1,0          ; vbroadcastss  0x1fa58(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,4,17,2,0            ; vbroadcastss  0x21104(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -11570,25 +11570,25 @@
   DB  197,249,112,228,78                  ; vpshufd       $0x4e,%xmm4,%xmm4
   DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
   DB  196,227,85,24,252,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm7
-  DB  196,226,125,24,37,77,249,1,0        ; vbroadcastss  0x1f94d(%rip),%ymm4        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,24,37,249,15,2,0        ; vbroadcastss  0x20ff9(%rip),%ymm4        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,196,84,228                      ; vandps        %ymm4,%ymm7,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,226,125,24,45,64,249,1,0        ; vbroadcastss  0x1f940(%rip),%ymm5        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,45,236,15,2,0        ; vbroadcastss  0x20fec(%rip),%ymm5        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,220,89,229                      ; vmulps        %ymm5,%ymm4,%ymm4
-  DB  196,226,125,24,45,55,249,1,0        ; vbroadcastss  0x1f937(%rip),%ymm5        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,24,45,227,15,2,0        ; vbroadcastss  0x20fe3(%rip),%ymm5        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,196,84,237                      ; vandps        %ymm5,%ymm7,%ymm5
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,226,125,24,53,42,249,1,0        ; vbroadcastss  0x1f92a(%rip),%ymm6        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,53,214,15,2,0        ; vbroadcastss  0x20fd6(%rip),%ymm6        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,212,89,238                      ; vmulps        %ymm6,%ymm5,%ymm5
-  DB  196,226,125,24,53,33,249,1,0        ; vbroadcastss  0x1f921(%rip),%ymm6        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,24,53,205,15,2,0        ; vbroadcastss  0x20fcd(%rip),%ymm6        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,196,84,246                      ; vandps        %ymm6,%ymm7,%ymm6
   DB  197,252,91,246                      ; vcvtdq2ps     %ymm6,%ymm6
-  DB  196,98,125,24,5,20,249,1,0          ; vbroadcastss  0x1f914(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,192,15,2,0          ; vbroadcastss  0x20fc0(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,76,89,240                   ; vmulps        %ymm8,%ymm6,%ymm6
-  DB  196,98,125,24,5,10,249,1,0          ; vbroadcastss  0x1f90a(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,24,5,182,15,2,0          ; vbroadcastss  0x20fb6(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,68,84,248                   ; vandps        %ymm8,%ymm7,%ymm7
   DB  197,252,91,255                      ; vcvtdq2ps     %ymm7,%ymm7
-  DB  196,98,125,24,5,252,248,1,0         ; vbroadcastss  0x1f8fc(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,168,15,2,0          ; vbroadcastss  0x20fa8(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,68,89,248                   ; vmulps        %ymm8,%ymm7,%ymm7
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -11708,25 +11708,25 @@
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,216,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm3
-  DB  196,226,125,24,5,5,247,1,0          ; vbroadcastss  0x1f705(%rip),%ymm0        # 2bdb0 <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
+  DB  196,226,125,24,5,177,13,2,0         ; vbroadcastss  0x20db1(%rip),%ymm0        # 2d45c <_sk_srcover_rgba_8888_sse2_lowp+0x3a8>
   DB  197,228,84,192                      ; vandps        %ymm0,%ymm3,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,226,125,24,13,248,246,1,0       ; vbroadcastss  0x1f6f8(%rip),%ymm1        # 2bdb4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
+  DB  196,226,125,24,13,164,13,2,0        ; vbroadcastss  0x20da4(%rip),%ymm1        # 2d460 <_sk_srcover_rgba_8888_sse2_lowp+0x3ac>
   DB  197,252,89,193                      ; vmulps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,239,246,1,0       ; vbroadcastss  0x1f6ef(%rip),%ymm1        # 2bdb8 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
+  DB  196,226,125,24,13,155,13,2,0        ; vbroadcastss  0x20d9b(%rip),%ymm1        # 2d464 <_sk_srcover_rgba_8888_sse2_lowp+0x3b0>
   DB  197,228,84,201                      ; vandps        %ymm1,%ymm3,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,226,125,24,21,226,246,1,0       ; vbroadcastss  0x1f6e2(%rip),%ymm2        # 2bdbc <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
+  DB  196,226,125,24,21,142,13,2,0        ; vbroadcastss  0x20d8e(%rip),%ymm2        # 2d468 <_sk_srcover_rgba_8888_sse2_lowp+0x3b4>
   DB  197,244,89,202                      ; vmulps        %ymm2,%ymm1,%ymm1
-  DB  196,226,125,24,21,217,246,1,0       ; vbroadcastss  0x1f6d9(%rip),%ymm2        # 2bdc0 <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
+  DB  196,226,125,24,21,133,13,2,0        ; vbroadcastss  0x20d85(%rip),%ymm2        # 2d46c <_sk_srcover_rgba_8888_sse2_lowp+0x3b8>
   DB  197,228,84,210                      ; vandps        %ymm2,%ymm3,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,98,125,24,5,204,246,1,0         ; vbroadcastss  0x1f6cc(%rip),%ymm8        # 2bdc4 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
+  DB  196,98,125,24,5,120,13,2,0          ; vbroadcastss  0x20d78(%rip),%ymm8        # 2d470 <_sk_srcover_rgba_8888_sse2_lowp+0x3bc>
   DB  196,193,108,89,208                  ; vmulps        %ymm8,%ymm2,%ymm2
-  DB  196,98,125,24,5,194,246,1,0         ; vbroadcastss  0x1f6c2(%rip),%ymm8        # 2bdc8 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
+  DB  196,98,125,24,5,110,13,2,0          ; vbroadcastss  0x20d6e(%rip),%ymm8        # 2d474 <_sk_srcover_rgba_8888_sse2_lowp+0x3c0>
   DB  196,193,100,84,216                  ; vandps        %ymm8,%ymm3,%ymm3
   DB  197,252,91,219                      ; vcvtdq2ps     %ymm3,%ymm3
-  DB  196,98,125,24,5,180,246,1,0         ; vbroadcastss  0x1f6b4(%rip),%ymm8        # 2bdcc <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
+  DB  196,98,125,24,5,96,13,2,0           ; vbroadcastss  0x20d60(%rip),%ymm8        # 2d478 <_sk_srcover_rgba_8888_sse2_lowp+0x3c4>
   DB  196,193,100,89,216                  ; vmulps        %ymm8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
@@ -11745,7 +11745,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,136,246,1,0         ; vbroadcastss  0x1f688(%rip),%ymm8        # 2bdd0 <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
+  DB  196,98,125,24,5,52,13,2,0           ; vbroadcastss  0x20d34(%rip),%ymm8        # 2d47c <_sk_srcover_rgba_8888_sse2_lowp+0x3c8>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,193,41,114,241,12               ; vpslld        $0xc,%xmm9,%xmm10
@@ -11831,10 +11831,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,135,0,0,0                    ; jne           c903 <_sk_load_8888_avx+0xa7>
   DB  196,1,125,16,12,152                 ; vmovupd       (%r8,%r11,4),%ymm9
-  DB  197,125,40,21,54,249,1,0            ; vmovapd       0x1f936(%rip),%ymm10        # 2c1c0 <_sk_srcover_rgba_8888_sse2_lowp+0x7b8>
+  DB  197,125,40,21,214,15,2,0            ; vmovapd       0x20fd6(%rip),%ymm10        # 2d860 <_sk_srcover_rgba_8888_sse2_lowp+0x7ac>
   DB  196,193,53,84,194                   ; vandpd        %ymm10,%ymm9,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,5,148,244,1,0         ; vbroadcastss  0x1f494(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,64,11,2,0           ; vbroadcastss  0x20b40(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,192                  ; vmulps        %ymm8,%ymm0,%ymm0
   DB  196,193,113,114,209,8               ; vpsrld        $0x8,%xmm9,%xmm1
   DB  196,99,125,25,203,1                 ; vextractf128  $0x1,%ymm9,%xmm3
@@ -11924,10 +11924,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,135,0,0,0                    ; jne           ca7b <_sk_load_8888_dst_avx+0xa7>
   DB  196,1,125,16,12,152                 ; vmovupd       (%r8,%r11,4),%ymm9
-  DB  197,125,40,21,222,247,1,0           ; vmovapd       0x1f7de(%rip),%ymm10        # 2c1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x7d8>
+  DB  197,125,40,21,126,14,2,0            ; vmovapd       0x20e7e(%rip),%ymm10        # 2d880 <_sk_srcover_rgba_8888_sse2_lowp+0x7cc>
   DB  196,193,53,84,226                   ; vandpd        %ymm10,%ymm9,%ymm4
   DB  197,252,91,228                      ; vcvtdq2ps     %ymm4,%ymm4
-  DB  196,98,125,24,5,28,243,1,0          ; vbroadcastss  0x1f31c(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,200,9,2,0           ; vbroadcastss  0x209c8(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,92,89,224                   ; vmulps        %ymm8,%ymm4,%ymm4
   DB  196,193,81,114,209,8                ; vpsrld        $0x8,%xmm9,%xmm5
   DB  196,99,125,25,207,1                 ; vextractf128  $0x1,%ymm9,%xmm7
@@ -12060,10 +12060,10 @@
   DB  196,131,121,34,4,176,2              ; vpinsrd       $0x2,(%r8,%r14,4),%xmm0,%xmm0
   DB  196,195,121,34,28,152,3             ; vpinsrd       $0x3,(%r8,%rbx,4),%xmm0,%xmm3
   DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  DB  197,124,40,21,176,245,1,0           ; vmovaps       0x1f5b0(%rip),%ymm10        # 2c200 <_sk_srcover_rgba_8888_sse2_lowp+0x7f8>
+  DB  197,124,40,21,80,12,2,0             ; vmovaps       0x20c50(%rip),%ymm10        # 2d8a0 <_sk_srcover_rgba_8888_sse2_lowp+0x7ec>
   DB  196,193,124,84,194                  ; vandps        %ymm10,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,206,240,1,0        ; vbroadcastss  0x1f0ce(%rip),%ymm9        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,13,122,7,2,0          ; vbroadcastss  0x2077a(%rip),%ymm9        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  196,193,113,114,208,8               ; vpsrld        $0x8,%xmm8,%xmm1
   DB  197,233,114,211,8                   ; vpsrld        $0x8,%xmm3,%xmm2
@@ -12098,7 +12098,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,250,239,1,0         ; vbroadcastss  0x1effa(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,166,6,2,0           ; vbroadcastss  0x206a6(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,65,116,89,208                   ; vmulps        %ymm8,%ymm1,%ymm10
@@ -12161,7 +12161,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,218                 ; callq         ffffffffdb00cdfc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe13f4>
+  DB  232,255,255,255,218                 ; callq         ffffffffdb00cdfc <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafdfd48>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,204                             ; dec           %esp
@@ -12181,10 +12181,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,135,0,0,0                    ; jne           ceab <_sk_load_bgra_avx+0xa7>
   DB  196,1,125,16,12,152                 ; vmovupd       (%r8,%r11,4),%ymm9
-  DB  197,125,40,21,238,243,1,0           ; vmovapd       0x1f3ee(%rip),%ymm10        # 2c220 <_sk_srcover_rgba_8888_sse2_lowp+0x818>
+  DB  197,125,40,21,142,10,2,0            ; vmovapd       0x20a8e(%rip),%ymm10        # 2d8c0 <_sk_srcover_rgba_8888_sse2_lowp+0x80c>
   DB  196,193,53,84,202                   ; vandpd        %ymm10,%ymm9,%ymm1
   DB  197,252,91,201                      ; vcvtdq2ps     %ymm1,%ymm1
-  DB  196,98,125,24,5,236,238,1,0         ; vbroadcastss  0x1eeec(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,152,5,2,0           ; vbroadcastss  0x20598(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,116,89,208                  ; vmulps        %ymm8,%ymm1,%ymm2
   DB  196,193,113,114,209,8               ; vpsrld        $0x8,%xmm9,%xmm1
   DB  196,99,125,25,203,1                 ; vextractf128  $0x1,%ymm9,%xmm3
@@ -12274,10 +12274,10 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,135,0,0,0                    ; jne           d023 <_sk_load_bgra_dst_avx+0xa7>
   DB  196,1,125,16,12,152                 ; vmovupd       (%r8,%r11,4),%ymm9
-  DB  197,125,40,21,150,242,1,0           ; vmovapd       0x1f296(%rip),%ymm10        # 2c240 <_sk_srcover_rgba_8888_sse2_lowp+0x838>
+  DB  197,125,40,21,54,9,2,0              ; vmovapd       0x20936(%rip),%ymm10        # 2d8e0 <_sk_srcover_rgba_8888_sse2_lowp+0x82c>
   DB  196,193,53,84,234                   ; vandpd        %ymm10,%ymm9,%ymm5
   DB  197,252,91,237                      ; vcvtdq2ps     %ymm5,%ymm5
-  DB  196,98,125,24,5,116,237,1,0         ; vbroadcastss  0x1ed74(%rip),%ymm8        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,5,32,4,2,0            ; vbroadcastss  0x20420(%rip),%ymm8        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,84,89,240                   ; vmulps        %ymm8,%ymm5,%ymm6
   DB  196,193,81,114,209,8                ; vpsrld        $0x8,%xmm9,%xmm5
   DB  196,99,125,25,207,1                 ; vextractf128  $0x1,%ymm9,%xmm7
@@ -12410,10 +12410,10 @@
   DB  196,131,121,34,4,176,2              ; vpinsrd       $0x2,(%r8,%r14,4),%xmm0,%xmm0
   DB  196,195,121,34,28,152,3             ; vpinsrd       $0x3,(%r8,%rbx,4),%xmm0,%xmm3
   DB  196,227,61,24,195,1                 ; vinsertf128   $0x1,%xmm3,%ymm8,%ymm0
-  DB  197,124,40,13,104,240,1,0           ; vmovaps       0x1f068(%rip),%ymm9        # 2c260 <_sk_srcover_rgba_8888_sse2_lowp+0x858>
+  DB  197,124,40,13,8,7,2,0               ; vmovaps       0x20708(%rip),%ymm9        # 2d900 <_sk_srcover_rgba_8888_sse2_lowp+0x84c>
   DB  196,193,124,84,193                  ; vandps        %ymm9,%ymm0,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,21,38,235,1,0         ; vbroadcastss  0x1eb26(%rip),%ymm10        # 2bd30 <_sk_srcover_rgba_8888_sse2_lowp+0x328>
+  DB  196,98,125,24,21,210,1,2,0          ; vbroadcastss  0x201d2(%rip),%ymm10        # 2d3dc <_sk_srcover_rgba_8888_sse2_lowp+0x328>
   DB  196,193,124,89,210                  ; vmulps        %ymm10,%ymm0,%ymm2
   DB  196,193,121,114,208,8               ; vpsrld        $0x8,%xmm8,%xmm0
   DB  197,241,114,211,8                   ; vpsrld        $0x8,%xmm3,%xmm1
@@ -12448,7 +12448,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,218                           ; movslq        %edx,%r11
-  DB  196,98,125,24,5,82,234,1,0          ; vbroadcastss  0x1ea52(%rip),%ymm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,125,24,5,254,0,2,0           ; vbroadcastss  0x200fe(%rip),%ymm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  196,65,108,89,200                   ; vmulps        %ymm8,%ymm2,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,65,116,89,208                   ; vmulps        %ymm8,%ymm1,%ymm10
@@ -12511,7 +12511,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,255,255,255,218                 ; callq         ffffffffdb00d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe199c>
+  DB  232,255,255,255,218                 ; callq         ffffffffdb00d3a4 <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffdafe02f0>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,204                             ; dec           %esp
@@ -12551,14 +12551,14 @@
   DB  196,226,121,51,200                  ; vpmovzxwd     %xmm0,%xmm1
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
-  DB  197,123,18,21,229,247,1,0           ; vmovddup      0x1f7e5(%rip),%xmm10        # 2cc28 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  DB  197,123,18,21,197,14,2,0            ; vmovddup      0x20ec5(%rip),%xmm10        # 2e308 <_sk_srcover_rgba_8888_sse2_lowp+0x1254>
   DB  196,193,121,219,218                 ; vpand         %xmm10,%xmm0,%xmm3
-  DB  196,98,121,24,29,231,233,1,0        ; vbroadcastss  0x1e9e7(%rip),%xmm11        # 2be38 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
+  DB  196,98,121,24,29,147,0,2,0          ; vbroadcastss  0x20093(%rip),%xmm11        # 2d4e4 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
   DB  197,33,102,227                      ; vpcmpgtd      %xmm3,%xmm11,%xmm12
   DB  196,193,113,219,210                 ; vpand         %xmm10,%xmm1,%xmm2
   DB  197,33,102,234                      ; vpcmpgtd      %xmm2,%xmm11,%xmm13
   DB  196,195,21,24,252,1                 ; vinsertf128   $0x1,%xmm12,%ymm13,%ymm7
-  DB  197,123,18,37,196,247,1,0           ; vmovddup      0x1f7c4(%rip),%xmm12        # 2cc30 <_sk_srcover_rgba_8888_sse2_lowp+0x1228>
+  DB  197,123,18,37,164,14,2,0            ; vmovddup      0x20ea4(%rip),%xmm12        # 2e310 <_sk_srcover_rgba_8888_sse2_lowp+0x125c>
   DB  196,193,113,219,204                 ; vpand         %xmm12,%xmm1,%xmm1
   DB  197,241,114,241,16                  ; vpslld        $0x10,%xmm1,%xmm1
   DB  196,193,121,219,196                 ; vpand         %xmm12,%xmm0,%xmm0
@@ -12567,7 +12567,7 @@
   DB  197,241,235,202                     ; vpor          %xmm2,%xmm1,%xmm1
   DB  197,233,114,243,13                  ; vpslld        $0xd,%xmm3,%xmm2
   DB  197,249,235,194                     ; vpor          %xmm2,%xmm0,%xmm0
-  DB  196,226,121,24,29,161,233,1,0       ; vbroadcastss  0x1e9a1(%rip),%xmm3        # 2be3c <_sk_srcover_rgba_8888_sse2_lowp+0x434>
+  DB  196,226,121,24,29,77,0,2,0          ; vbroadcastss  0x2004d(%rip),%xmm3        # 2d4e8 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
   DB  197,249,254,195                     ; vpaddd        %xmm3,%xmm0,%xmm0
   DB  197,241,254,203                     ; vpaddd        %xmm3,%xmm1,%xmm1
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
@@ -12702,14 +12702,14 @@
   DB  196,226,121,51,236                  ; vpmovzxwd     %xmm4,%xmm5
   DB  197,249,112,228,78                  ; vpshufd       $0x4e,%xmm4,%xmm4
   DB  196,226,121,51,228                  ; vpmovzxwd     %xmm4,%xmm4
-  DB  197,123,18,21,10,245,1,0            ; vmovddup      0x1f50a(%rip),%xmm10        # 2cc28 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  DB  197,123,18,21,234,11,2,0            ; vmovddup      0x20bea(%rip),%xmm10        # 2e308 <_sk_srcover_rgba_8888_sse2_lowp+0x1254>
   DB  196,193,89,219,250                  ; vpand         %xmm10,%xmm4,%xmm7
-  DB  196,98,121,24,29,12,231,1,0         ; vbroadcastss  0x1e70c(%rip),%xmm11        # 2be38 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
+  DB  196,98,121,24,29,184,253,1,0        ; vbroadcastss  0x1fdb8(%rip),%xmm11        # 2d4e4 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
   DB  197,33,102,231                      ; vpcmpgtd      %xmm7,%xmm11,%xmm12
   DB  196,193,81,219,242                  ; vpand         %xmm10,%xmm5,%xmm6
   DB  197,33,102,238                      ; vpcmpgtd      %xmm6,%xmm11,%xmm13
   DB  196,195,21,24,220,1                 ; vinsertf128   $0x1,%xmm12,%ymm13,%ymm3
-  DB  197,123,18,37,233,244,1,0           ; vmovddup      0x1f4e9(%rip),%xmm12        # 2cc30 <_sk_srcover_rgba_8888_sse2_lowp+0x1228>
+  DB  197,123,18,37,201,11,2,0            ; vmovddup      0x20bc9(%rip),%xmm12        # 2e310 <_sk_srcover_rgba_8888_sse2_lowp+0x125c>
   DB  196,193,81,219,236                  ; vpand         %xmm12,%xmm5,%xmm5
   DB  197,209,114,245,16                  ; vpslld        $0x10,%xmm5,%xmm5
   DB  196,193,89,219,228                  ; vpand         %xmm12,%xmm4,%xmm4
@@ -12718,7 +12718,7 @@
   DB  197,209,235,238                     ; vpor          %xmm6,%xmm5,%xmm5
   DB  197,201,114,247,13                  ; vpslld        $0xd,%xmm7,%xmm6
   DB  197,217,235,230                     ; vpor          %xmm6,%xmm4,%xmm4
-  DB  196,226,121,24,61,198,230,1,0       ; vbroadcastss  0x1e6c6(%rip),%xmm7        # 2be3c <_sk_srcover_rgba_8888_sse2_lowp+0x434>
+  DB  196,226,121,24,61,114,253,1,0       ; vbroadcastss  0x1fd72(%rip),%xmm7        # 2d4e8 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
   DB  197,217,254,231                     ; vpaddd        %xmm7,%xmm4,%xmm4
   DB  197,209,254,239                     ; vpaddd        %xmm7,%xmm5,%xmm5
   DB  196,227,85,24,228,1                 ; vinsertf128   $0x1,%xmm4,%ymm5,%ymm4
@@ -12896,14 +12896,14 @@
   DB  196,226,121,51,208                  ; vpmovzxwd     %xmm0,%xmm2
   DB  197,249,112,192,78                  ; vpshufd       $0x4e,%xmm0,%xmm0
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
-  DB  197,123,18,29,108,241,1,0           ; vmovddup      0x1f16c(%rip),%xmm11        # 2cc28 <_sk_srcover_rgba_8888_sse2_lowp+0x1220>
+  DB  197,123,18,29,76,8,2,0              ; vmovddup      0x2084c(%rip),%xmm11        # 2e308 <_sk_srcover_rgba_8888_sse2_lowp+0x1254>
   DB  196,193,121,219,219                 ; vpand         %xmm11,%xmm0,%xmm3
-  DB  196,98,121,24,37,110,227,1,0        ; vbroadcastss  0x1e36e(%rip),%xmm12        # 2be38 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
+  DB  196,98,121,24,37,26,250,1,0         ; vbroadcastss  0x1fa1a(%rip),%xmm12        # 2d4e4 <_sk_srcover_rgba_8888_sse2_lowp+0x430>
   DB  197,25,102,235                      ; vpcmpgtd      %xmm3,%xmm12,%xmm13
   DB  196,193,105,219,203                 ; vpand         %xmm11,%xmm2,%xmm1
   DB  197,25,102,193                      ; vpcmpgtd      %xmm1,%xmm12,%xmm8
   DB  196,67,61,24,197,1                  ; vinsertf128   $0x1,%xmm13,%ymm8,%ymm8
-  DB  197,123,18,45,75,241,1,0            ; vmovddup      0x1f14b(%rip),%xmm13        # 2cc30 <_sk_srcover_rgba_8888_sse2_lowp+0x1228>
+  DB  197,123,18,45,43,8,2,0              ; vmovddup      0x2082b(%rip),%xmm13        # 2e310 <_sk_srcover_rgba_8888_sse2_lowp+0x125c>
   DB  196,193,105,219,213                 ; vpand         %xmm13,%xmm2,%xmm2
   DB  197,233,114,242,16                  ; vpslld        $0x10,%xmm2,%xmm2
   DB  197,241,114,241,13                  ; vpslld        $0xd,%xmm1,%xmm1
@@ -12912,7 +12912,7 @@
   DB  197,249,114,240,16                  ; vpslld        $0x10,%xmm0,%xmm0
   DB  197,233,114,243,13                  ; vpslld        $0xd,%xmm3,%xmm2
   DB  197,249,235,194                     ; vpor          %xmm2,%xmm0,%xmm0
-  DB  196,226,121,24,29,40,227,1,0        ; vbroadcastss  0x1e328(%rip),%xmm3        # 2be3c <_sk_srcover_rgba_8888_sse2_lowp+0x434>
+  DB  196,226,121,24,29,212,249,1,0       ; vbroadcastss  0x1f9d4(%rip),%xmm3        # 2d4e8 <_sk_srcover_rgba_8888_sse2_lowp+0x434>
   DB  197,249,254,195                     ; vpaddd        %xmm3,%xmm0,%xmm0
   DB  197,241,254,203                     ; vpaddd        %xmm3,%xmm1,%xmm1
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
@@ -13000,12 +13000,12 @@
   DB  197,252,17,108,36,64                ; vmovups       %ymm5,0x40(%rsp)
   DB  197,252,17,100,36,32                ; vmovups       %ymm4,0x20(%rsp)
   DB  197,252,40,225                      ; vmovaps       %ymm1,%ymm4
-  DB  196,98,125,24,5,93,225,1,0          ; vbroadcastss  0x1e15d(%rip),%ymm8        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,5,9,248,1,0           ; vbroadcastss  0x1f809(%rip),%ymm8        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,65,124,84,216                   ; vandps        %ymm8,%ymm0,%ymm11
-  DB  196,98,125,24,21,123,225,1,0        ; vbroadcastss  0x1e17b(%rip),%ymm10        # 2be40 <_sk_srcover_rgba_8888_sse2_lowp+0x438>
+  DB  196,98,125,24,21,39,248,1,0         ; vbroadcastss  0x1f827(%rip),%ymm10        # 2d4ec <_sk_srcover_rgba_8888_sse2_lowp+0x438>
   DB  196,65,124,84,226                   ; vandps        %ymm10,%ymm0,%ymm12
   DB  196,67,125,25,229,1                 ; vextractf128  $0x1,%ymm12,%xmm13
-  DB  196,98,121,24,13,107,225,1,0        ; vbroadcastss  0x1e16b(%rip),%xmm9        # 2be44 <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
+  DB  196,98,121,24,13,23,248,1,0         ; vbroadcastss  0x1f817(%rip),%xmm9        # 2d4f0 <_sk_srcover_rgba_8888_sse2_lowp+0x43c>
   DB  196,65,49,102,245                   ; vpcmpgtd      %xmm13,%xmm9,%xmm14
   DB  196,65,49,102,252                   ; vpcmpgtd      %xmm12,%xmm9,%xmm15
   DB  196,67,5,24,246,1                   ; vinsertf128   $0x1,%xmm14,%ymm15,%ymm14
@@ -13016,7 +13016,7 @@
   DB  196,65,17,254,239                   ; vpaddd        %xmm15,%xmm13,%xmm13
   DB  196,193,25,114,212,13               ; vpsrld        $0xd,%xmm12,%xmm12
   DB  196,65,25,254,227                   ; vpaddd        %xmm11,%xmm12,%xmm12
-  DB  196,98,121,24,29,46,225,1,0         ; vbroadcastss  0x1e12e(%rip),%xmm11        # 2be48 <_sk_srcover_rgba_8888_sse2_lowp+0x440>
+  DB  196,98,121,24,29,218,247,1,0        ; vbroadcastss  0x1f7da(%rip),%xmm11        # 2d4f4 <_sk_srcover_rgba_8888_sse2_lowp+0x440>
   DB  196,65,17,254,235                   ; vpaddd        %xmm11,%xmm13,%xmm13
   DB  196,65,25,254,227                   ; vpaddd        %xmm11,%xmm12,%xmm12
   DB  196,67,29,24,237,1                  ; vinsertf128   $0x1,%xmm13,%ymm12,%ymm13
@@ -13165,7 +13165,7 @@
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,112,221,1,0        ; vbroadcastss  0x1dd70(%rip),%ymm9        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,98,125,24,13,28,244,1,0         ; vbroadcastss  0x1f41c(%rip),%ymm9        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,169,109,202                     ; vpunpckhqdq   %xmm2,%xmm10,%xmm1
   DB  197,233,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm2
@@ -13267,7 +13267,7 @@
   DB  196,226,121,51,192                  ; vpmovzxwd     %xmm0,%xmm0
   DB  196,227,117,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm1,%ymm0
   DB  197,252,91,192                      ; vcvtdq2ps     %ymm0,%ymm0
-  DB  196,98,125,24,13,159,219,1,0        ; vbroadcastss  0x1db9f(%rip),%ymm9        # 2bd4c <_sk_srcover_rgba_8888_sse2_lowp+0x344>
+  DB  196,98,125,24,13,75,242,1,0         ; vbroadcastss  0x1f24b(%rip),%ymm9        # 2d3f8 <_sk_srcover_rgba_8888_sse2_lowp+0x344>
   DB  196,193,124,89,193                  ; vmulps        %ymm9,%ymm0,%ymm0
   DB  197,185,109,203                     ; vpunpckhqdq   %xmm3,%xmm8,%xmm1
   DB  197,225,113,241,8                   ; vpsllw        $0x8,%xmm1,%xmm3
@@ -13289,7 +13289,7 @@
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
   DB  196,193,108,89,209                  ; vmulps        %ymm9,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,159,218,1,0       ; vbroadcastss  0x1da9f(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,75,241,1,0        ; vbroadcastss  0x1f14b(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,4,64                ; vmovd         (%r8,%rax,2),%xmm0
   DB  196,65,49,239,201                   ; vpxor         %xmm9,%xmm9,%xmm9
@@ -13363,7 +13363,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  73,99,193                           ; movslq        %r9d,%rax
-  DB  196,98,125,24,5,104,218,1,0         ; vbroadcastss  0x1da68(%rip),%ymm8        # 2bdd4 <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
+  DB  196,98,125,24,5,20,241,1,0          ; vbroadcastss  0x1f114(%rip),%ymm8        # 2d480 <_sk_srcover_rgba_8888_sse2_lowp+0x3cc>
   DB  196,65,124,89,200                   ; vmulps        %ymm8,%ymm0,%ymm9
   DB  196,65,125,91,201                   ; vcvtps2dq     %ymm9,%ymm9
   DB  196,67,125,25,202,1                 ; vextractf128  $0x1,%ymm9,%xmm10
@@ -13620,7 +13620,7 @@
   DB  255,92,255,255                      ; lcall         *-0x1(%rdi,%rdi,8)
   DB  255,70,255                          ; incl          -0x1(%rsi)
   DB  255                                 ; (bad)
-  DB  255,53,255,255,255,40               ; pushq         0x28ffffff(%rip)        # 2900e7e4 <_sk_srcover_rgba_8888_sse2_lowp+0x28fe2ddc>
+  DB  255,53,255,255,255,40               ; pushq         0x28ffffff(%rip)        # 2900e7e4 <_sk_srcover_rgba_8888_sse2_lowp+0x28fe1730>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -13709,7 +13709,7 @@
   DB  196,193,58,88,192                   ; vaddss        %xmm8,%xmm8,%xmm0
   DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
   DB  196,99,125,24,192,1                 ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm8
-  DB  197,178,89,5,125,211,1,0            ; vmulss        0x1d37d(%rip),%xmm9,%xmm0        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,178,89,5,41,234,1,0             ; vmulss        0x1ea29(%rip),%xmm9,%xmm0        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
   DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   DB  197,164,89,192                      ; vmulps        %ymm0,%ymm11,%ymm0
@@ -13734,7 +13734,7 @@
   DB  196,193,58,88,200                   ; vaddss        %xmm8,%xmm8,%xmm1
   DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
   DB  196,99,117,24,193,1                 ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm8
-  DB  197,178,89,13,19,211,1,0            ; vmulss        0x1d313(%rip),%xmm9,%xmm1        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,178,89,13,191,233,1,0           ; vmulss        0x1e9bf(%rip),%xmm9,%xmm1        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,227,121,4,201,0                 ; vpermilps     $0x0,%xmm1,%xmm1
   DB  196,227,117,24,201,1                ; vinsertf128   $0x1,%xmm1,%ymm1,%ymm1
   DB  197,164,89,201                      ; vmulps        %ymm1,%ymm11,%ymm1
@@ -13752,7 +13752,7 @@
 _sk_clamp_x_1_avx LABEL PROC
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,208,210,1,0         ; vbroadcastss  0x1d2d0(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,124,233,1,0         ; vbroadcastss  0x1e97c(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -13763,16 +13763,16 @@
   DB  196,193,124,92,192                  ; vsubps        %ymm8,%ymm0,%ymm0
   DB  196,65,60,87,192                    ; vxorps        %ymm8,%ymm8,%ymm8
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,170,210,1,0         ; vbroadcastss  0x1d2aa(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,86,233,1,0          ; vbroadcastss  0x1e956(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_mirror_x_1_avx
 _sk_mirror_x_1_avx LABEL PROC
-  DB  196,98,125,24,5,176,210,1,0         ; vbroadcastss  0x1d2b0(%rip),%ymm8        # 2bcd0 <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
+  DB  196,98,125,24,5,92,233,1,0          ; vbroadcastss  0x1e95c(%rip),%ymm8        # 2d37c <_sk_srcover_rgba_8888_sse2_lowp+0x2c8>
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,13,134,210,1,0        ; vbroadcastss  0x1d286(%rip),%ymm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,13,50,233,1,0         ; vbroadcastss  0x1e932(%rip),%ymm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,124,89,201                   ; vmulps        %ymm9,%ymm0,%ymm9
   DB  196,67,125,8,201,1                  ; vroundps      $0x1,%ymm9,%ymm9
   DB  196,65,52,88,201                    ; vaddps        %ymm9,%ymm9,%ymm9
@@ -13782,19 +13782,19 @@
   DB  197,60,92,200                       ; vsubps        %ymm0,%ymm8,%ymm9
   DB  197,180,84,192                      ; vandps        %ymm0,%ymm9,%ymm0
   DB  197,188,95,192                      ; vmaxps        %ymm0,%ymm8,%ymm0
-  DB  196,98,125,24,5,86,210,1,0          ; vbroadcastss  0x1d256(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,2,233,1,0           ; vbroadcastss  0x1e902(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_luminance_to_alpha_avx
 _sk_luminance_to_alpha_avx LABEL PROC
-  DB  196,226,125,24,29,100,211,1,0       ; vbroadcastss  0x1d364(%rip),%ymm3        # 2bdd8 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
+  DB  196,226,125,24,29,16,234,1,0        ; vbroadcastss  0x1ea10(%rip),%ymm3        # 2d484 <_sk_srcover_rgba_8888_sse2_lowp+0x3d0>
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,29,91,211,1,0        ; vbroadcastss  0x1d35b(%rip),%ymm3        # 2bddc <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
+  DB  196,226,125,24,29,7,234,1,0         ; vbroadcastss  0x1ea07(%rip),%ymm3        # 2d488 <_sk_srcover_rgba_8888_sse2_lowp+0x3d4>
   DB  197,244,89,203                      ; vmulps        %ymm3,%ymm1,%ymm1
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
-  DB  196,226,125,24,13,78,211,1,0        ; vbroadcastss  0x1d34e(%rip),%ymm1        # 2bde0 <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
+  DB  196,226,125,24,13,250,233,1,0       ; vbroadcastss  0x1e9fa(%rip),%ymm1        # 2d48c <_sk_srcover_rgba_8888_sse2_lowp+0x3d8>
   DB  197,236,89,201                      ; vmulps        %ymm1,%ymm2,%ymm1
   DB  197,252,88,217                      ; vaddps        %ymm1,%ymm0,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14186,18 +14186,18 @@
 
 PUBLIC _sk_gauss_a_to_rgba_avx
 _sk_gauss_a_to_rgba_avx LABEL PROC
-  DB  196,226,125,24,5,99,204,1,0         ; vbroadcastss  0x1cc63(%rip),%ymm0        # 2bde8 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
+  DB  196,226,125,24,5,15,227,1,0         ; vbroadcastss  0x1e30f(%rip),%ymm0        # 2d494 <_sk_srcover_rgba_8888_sse2_lowp+0x3e0>
   DB  197,228,89,192                      ; vmulps        %ymm0,%ymm3,%ymm0
-  DB  196,226,125,24,13,82,204,1,0        ; vbroadcastss  0x1cc52(%rip),%ymm1        # 2bde4 <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
+  DB  196,226,125,24,13,254,226,1,0       ; vbroadcastss  0x1e2fe(%rip),%ymm1        # 2d490 <_sk_srcover_rgba_8888_sse2_lowp+0x3dc>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,13,73,204,1,0        ; vbroadcastss  0x1cc49(%rip),%ymm1        # 2bdec <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
+  DB  196,226,125,24,13,245,226,1,0       ; vbroadcastss  0x1e2f5(%rip),%ymm1        # 2d498 <_sk_srcover_rgba_8888_sse2_lowp+0x3e4>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,13,60,204,1,0        ; vbroadcastss  0x1cc3c(%rip),%ymm1        # 2bdf0 <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
+  DB  196,226,125,24,13,232,226,1,0       ; vbroadcastss  0x1e2e8(%rip),%ymm1        # 2d49c <_sk_srcover_rgba_8888_sse2_lowp+0x3e8>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  197,252,89,195                      ; vmulps        %ymm3,%ymm0,%ymm0
-  DB  196,226,125,24,13,47,204,1,0        ; vbroadcastss  0x1cc2f(%rip),%ymm1        # 2bdf4 <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
+  DB  196,226,125,24,13,219,226,1,0       ; vbroadcastss  0x1e2db(%rip),%ymm1        # 2d4a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3ec>
   DB  197,252,88,193                      ; vaddps        %ymm1,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,40,200                      ; vmovaps       %ymm0,%ymm1
@@ -14223,7 +14223,7 @@
   DB  73,255,200                          ; dec           %r8
   DB  72,131,195,4                        ; add           $0x4,%rbx
   DB  196,65,52,87,201                    ; vxorps        %ymm9,%ymm9,%ymm9
-  DB  196,98,125,24,21,169,202,1,0        ; vbroadcastss  0x1caa9(%rip),%ymm10        # 2bcbc <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
+  DB  196,98,125,24,21,85,225,1,0         ; vbroadcastss  0x1e155(%rip),%ymm10        # 2d368 <_sk_srcover_rgba_8888_sse2_lowp+0x2b4>
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  196,98,125,24,3                     ; vbroadcastss  (%rbx),%ymm8
   DB  197,60,194,192,2                    ; vcmpleps      %ymm0,%ymm8,%ymm8
@@ -14414,27 +14414,27 @@
   DB  196,65,52,95,226                    ; vmaxps        %ymm10,%ymm9,%ymm12
   DB  196,65,36,94,220                    ; vdivps        %ymm12,%ymm11,%ymm11
   DB  196,65,36,89,227                    ; vmulps        %ymm11,%ymm11,%ymm12
-  DB  196,98,125,24,45,212,199,1,0        ; vbroadcastss  0x1c7d4(%rip),%ymm13        # 2bdf8 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
+  DB  196,98,125,24,45,128,222,1,0        ; vbroadcastss  0x1de80(%rip),%ymm13        # 2d4a4 <_sk_srcover_rgba_8888_sse2_lowp+0x3f0>
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,202,199,1,0        ; vbroadcastss  0x1c7ca(%rip),%ymm14        # 2bdfc <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
+  DB  196,98,125,24,53,118,222,1,0        ; vbroadcastss  0x1de76(%rip),%ymm14        # 2d4a8 <_sk_srcover_rgba_8888_sse2_lowp+0x3f4>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,237                    ; vmulps        %ymm13,%ymm12,%ymm13
-  DB  196,98,125,24,53,187,199,1,0        ; vbroadcastss  0x1c7bb(%rip),%ymm14        # 2be00 <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
+  DB  196,98,125,24,53,103,222,1,0        ; vbroadcastss  0x1de67(%rip),%ymm14        # 2d4ac <_sk_srcover_rgba_8888_sse2_lowp+0x3f8>
   DB  196,65,20,88,238                    ; vaddps        %ymm14,%ymm13,%ymm13
   DB  196,65,28,89,229                    ; vmulps        %ymm13,%ymm12,%ymm12
-  DB  196,98,125,24,45,172,199,1,0        ; vbroadcastss  0x1c7ac(%rip),%ymm13        # 2be04 <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
+  DB  196,98,125,24,45,88,222,1,0         ; vbroadcastss  0x1de58(%rip),%ymm13        # 2d4b0 <_sk_srcover_rgba_8888_sse2_lowp+0x3fc>
   DB  196,65,28,88,229                    ; vaddps        %ymm13,%ymm12,%ymm12
   DB  196,65,36,89,220                    ; vmulps        %ymm12,%ymm11,%ymm11
   DB  196,65,52,194,202,1                 ; vcmpltps      %ymm10,%ymm9,%ymm9
-  DB  196,98,125,24,21,151,199,1,0        ; vbroadcastss  0x1c797(%rip),%ymm10        # 2be08 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
+  DB  196,98,125,24,21,67,222,1,0         ; vbroadcastss  0x1de43(%rip),%ymm10        # 2d4b4 <_sk_srcover_rgba_8888_sse2_lowp+0x400>
   DB  196,65,44,92,211                    ; vsubps        %ymm11,%ymm10,%ymm10
   DB  196,67,37,74,202,144                ; vblendvps     %ymm9,%ymm10,%ymm11,%ymm9
   DB  196,193,124,194,192,1               ; vcmpltps      %ymm8,%ymm0,%ymm0
-  DB  196,98,125,24,21,41,198,1,0         ; vbroadcastss  0x1c629(%rip),%ymm10        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,21,213,220,1,0        ; vbroadcastss  0x1dcd5(%rip),%ymm10        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,44,92,209                    ; vsubps        %ymm9,%ymm10,%ymm10
   DB  196,195,53,74,194,0                 ; vblendvps     %ymm0,%ymm10,%ymm9,%ymm0
   DB  196,65,116,194,200,1                ; vcmpltps      %ymm8,%ymm1,%ymm9
-  DB  196,98,125,24,21,19,198,1,0         ; vbroadcastss  0x1c613(%rip),%ymm10        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,21,191,220,1,0        ; vbroadcastss  0x1dcbf(%rip),%ymm10        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,44,92,208                       ; vsubps        %ymm0,%ymm10,%ymm10
   DB  196,195,125,74,194,144              ; vblendvps     %ymm9,%ymm10,%ymm0,%ymm0
   DB  196,65,124,194,200,3                ; vcmpunordps   %ymm8,%ymm0,%ymm9
@@ -14460,7 +14460,7 @@
   DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
   DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
   DB  197,44,88,208                       ; vaddps        %ymm0,%ymm10,%ymm10
-  DB  196,98,125,24,29,13,199,1,0         ; vbroadcastss  0x1c70d(%rip),%ymm11        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,29,185,221,1,0        ; vbroadcastss  0x1ddb9(%rip),%ymm11        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
   DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
   DB  197,116,89,217                      ; vmulps        %ymm1,%ymm1,%ymm11
@@ -14469,17 +14469,17 @@
   DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
   DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,13,225,198,1,0        ; vbroadcastss  0x1c6e1(%rip),%ymm9        # 2be10 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  DB  196,98,125,24,13,141,221,1,0        ; vbroadcastss  0x1dd8d(%rip),%ymm9        # 2d4bc <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  196,65,44,89,194                    ; vmulps        %ymm10,%ymm10,%ymm8
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
   DB  197,252,81,192                      ; vsqrtps       %ymm0,%ymm0
   DB  196,98,125,24,64,68                 ; vbroadcastss  0x44(%rax),%ymm8
-  DB  196,98,125,24,13,191,198,1,0        ; vbroadcastss  0x1c6bf(%rip),%ymm9        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,13,107,221,1,0        ; vbroadcastss  0x1dd6b(%rip),%ymm9        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,65,44,87,201                    ; vxorps        %ymm9,%ymm10,%ymm9
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,76,197,1,0         ; vbroadcastss  0x1c54c(%rip),%ymm11        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,29,248,219,1,0        ; vbroadcastss  0x1dbf8(%rip),%ymm11        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  196,65,60,89,210                    ; vmulps        %ymm10,%ymm8,%ymm10
   DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
@@ -14497,7 +14497,7 @@
   DB  196,67,121,4,210,0                  ; vpermilps     $0x0,%xmm10,%xmm10
   DB  196,67,45,24,210,1                  ; vinsertf128   $0x1,%xmm10,%ymm10,%ymm10
   DB  197,44,88,208                       ; vaddps        %ymm0,%ymm10,%ymm10
-  DB  196,98,125,24,29,95,198,1,0         ; vbroadcastss  0x1c65f(%rip),%ymm11        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,29,11,221,1,0         ; vbroadcastss  0x1dd0b(%rip),%ymm11        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,44,89,211                    ; vmulps        %ymm11,%ymm10,%ymm10
   DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
   DB  197,116,89,217                      ; vmulps        %ymm1,%ymm1,%ymm11
@@ -14506,17 +14506,17 @@
   DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
   DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   DB  197,164,92,192                      ; vsubps        %ymm0,%ymm11,%ymm0
-  DB  196,98,125,24,13,51,198,1,0         ; vbroadcastss  0x1c633(%rip),%ymm9        # 2be10 <_sk_srcover_rgba_8888_sse2_lowp+0x408>
+  DB  196,98,125,24,13,223,220,1,0        ; vbroadcastss  0x1dcdf(%rip),%ymm9        # 2d4bc <_sk_srcover_rgba_8888_sse2_lowp+0x408>
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
   DB  197,188,89,192                      ; vmulps        %ymm0,%ymm8,%ymm0
   DB  196,65,44,89,194                    ; vmulps        %ymm10,%ymm10,%ymm8
   DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
   DB  197,252,81,192                      ; vsqrtps       %ymm0,%ymm0
   DB  196,98,125,24,64,68                 ; vbroadcastss  0x44(%rax),%ymm8
-  DB  196,98,125,24,13,17,198,1,0         ; vbroadcastss  0x1c611(%rip),%ymm9        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,13,189,220,1,0        ; vbroadcastss  0x1dcbd(%rip),%ymm9        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,65,44,87,201                    ; vxorps        %ymm9,%ymm10,%ymm9
   DB  196,65,124,92,210                   ; vsubps        %ymm10,%ymm0,%ymm10
-  DB  196,98,125,24,29,158,196,1,0        ; vbroadcastss  0x1c49e(%rip),%ymm11        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,29,74,219,1,0         ; vbroadcastss  0x1db4a(%rip),%ymm11        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,60,89,195                    ; vmulps        %ymm11,%ymm8,%ymm8
   DB  196,65,60,89,210                    ; vmulps        %ymm10,%ymm8,%ymm10
   DB  197,180,92,192                      ; vsubps        %ymm0,%ymm9,%ymm0
@@ -14533,7 +14533,7 @@
   DB  196,67,121,4,201,0                  ; vpermilps     $0x0,%xmm9,%xmm9
   DB  196,67,53,24,201,1                  ; vinsertf128   $0x1,%xmm9,%ymm9,%ymm9
   DB  197,52,88,200                       ; vaddps        %ymm0,%ymm9,%ymm9
-  DB  196,98,125,24,21,183,197,1,0        ; vbroadcastss  0x1c5b7(%rip),%ymm10        # 2be0c <_sk_srcover_rgba_8888_sse2_lowp+0x404>
+  DB  196,98,125,24,21,99,220,1,0         ; vbroadcastss  0x1dc63(%rip),%ymm10        # 2d4b8 <_sk_srcover_rgba_8888_sse2_lowp+0x404>
   DB  196,65,52,89,202                    ; vmulps        %ymm10,%ymm9,%ymm9
   DB  197,252,89,192                      ; vmulps        %ymm0,%ymm0,%ymm0
   DB  197,116,89,209                      ; vmulps        %ymm1,%ymm1,%ymm10
@@ -14542,7 +14542,7 @@
   DB  196,227,121,4,192,0                 ; vpermilps     $0x0,%xmm0,%xmm0
   DB  196,227,125,24,192,1                ; vinsertf128   $0x1,%xmm0,%ymm0,%ymm0
   DB  197,172,92,192                      ; vsubps        %ymm0,%ymm10,%ymm0
-  DB  196,98,125,24,5,143,197,1,0         ; vbroadcastss  0x1c58f(%rip),%ymm8        # 2be14 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
+  DB  196,98,125,24,5,59,220,1,0          ; vbroadcastss  0x1dc3b(%rip),%ymm8        # 2d4c0 <_sk_srcover_rgba_8888_sse2_lowp+0x40c>
   DB  196,193,124,87,192                  ; vxorps        %ymm8,%ymm0,%ymm0
   DB  196,193,124,94,193                  ; vdivps        %ymm9,%ymm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14580,7 +14580,7 @@
 PUBLIC _sk_save_xy_avx
 _sk_save_xy_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,181,195,1,0         ; vbroadcastss  0x1c3b5(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,97,218,1,0          ; vbroadcastss  0x1da61(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,124,88,200                   ; vaddps        %ymm8,%ymm0,%ymm9
   DB  196,67,125,8,209,1                  ; vroundps      $0x1,%ymm9,%ymm10
   DB  196,65,52,92,202                    ; vsubps        %ymm10,%ymm9,%ymm9
@@ -14613,9 +14613,9 @@
 PUBLIC _sk_bilinear_nx_avx
 _sk_bilinear_nx_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,155,196,1,0        ; vbroadcastss  0x1c49b(%rip),%ymm0        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,5,71,219,1,0         ; vbroadcastss  0x1db47(%rip),%ymm0        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,46,195,1,0          ; vbroadcastss  0x1c32e(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,218,217,1,0         ; vbroadcastss  0x1d9da(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14624,7 +14624,7 @@
 PUBLIC _sk_bilinear_px_avx
 _sk_bilinear_px_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,11,195,1,0         ; vbroadcastss  0x1c30b(%rip),%ymm0        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,5,183,217,1,0        ; vbroadcastss  0x1d9b7(%rip),%ymm0        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
   DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -14634,9 +14634,9 @@
 PUBLIC _sk_bilinear_ny_avx
 _sk_bilinear_ny_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,76,196,1,0        ; vbroadcastss  0x1c44c(%rip),%ymm1        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,13,248,218,1,0       ; vbroadcastss  0x1daf8(%rip),%ymm1        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,222,194,1,0         ; vbroadcastss  0x1c2de(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,138,217,1,0         ; vbroadcastss  0x1d98a(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14645,7 +14645,7 @@
 PUBLIC _sk_bilinear_py_avx
 _sk_bilinear_py_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,187,194,1,0       ; vbroadcastss  0x1c2bb(%rip),%ymm1        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,13,103,217,1,0       ; vbroadcastss  0x1d967(%rip),%ymm1        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
   DB  197,124,16,128,192,0,0,0            ; vmovups       0xc0(%rax),%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -14655,14 +14655,14 @@
 PUBLIC _sk_bicubic_n3x_avx
 _sk_bicubic_n3x_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,255,195,1,0        ; vbroadcastss  0x1c3ff(%rip),%ymm0        # 2be1c <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  DB  196,226,125,24,5,171,218,1,0        ; vbroadcastss  0x1daab(%rip),%ymm0        # 2d4c8 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,142,194,1,0         ; vbroadcastss  0x1c28e(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,58,217,1,0          ; vbroadcastss  0x1d93a(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,224,195,1,0        ; vbroadcastss  0x1c3e0(%rip),%ymm10        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,140,218,1,0        ; vbroadcastss  0x1da8c(%rip),%ymm10        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,222,194,1,0        ; vbroadcastss  0x1c2de(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,21,138,217,1,0        ; vbroadcastss  0x1d98a(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -14672,19 +14672,19 @@
 PUBLIC _sk_bicubic_n1x_avx
 _sk_bicubic_n1x_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,169,195,1,0        ; vbroadcastss  0x1c3a9(%rip),%ymm0        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,5,85,218,1,0         ; vbroadcastss  0x1da55(%rip),%ymm0        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
-  DB  196,98,125,24,5,60,194,1,0          ; vbroadcastss  0x1c23c(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,232,216,1,0         ; vbroadcastss  0x1d8e8(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,128,0,0,0             ; vsubps        0x80(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,155,195,1,0        ; vbroadcastss  0x1c39b(%rip),%ymm9        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,13,71,218,1,0         ; vbroadcastss  0x1da47(%rip),%ymm9        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,137,195,1,0        ; vbroadcastss  0x1c389(%rip),%ymm10        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,21,53,218,1,0         ; vbroadcastss  0x1da35(%rip),%ymm10        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,6,194,1,0          ; vbroadcastss  0x1c206(%rip),%ymm10        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,21,178,216,1,0        ; vbroadcastss  0x1d8b2(%rip),%ymm10        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,107,195,1,0        ; vbroadcastss  0x1c36b(%rip),%ymm9        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,13,23,218,1,0         ; vbroadcastss  0x1da17(%rip),%ymm9        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14693,17 +14693,17 @@
 PUBLIC _sk_bicubic_p1x_avx
 _sk_bicubic_p1x_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,215,193,1,0         ; vbroadcastss  0x1c1d7(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,131,216,1,0         ; vbroadcastss  0x1d883(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,188,88,0                        ; vaddps        (%rax),%ymm8,%ymm0
   DB  197,124,16,136,128,0,0,0            ; vmovups       0x80(%rax),%ymm9
-  DB  196,98,125,24,21,54,195,1,0         ; vbroadcastss  0x1c336(%rip),%ymm10        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,21,226,217,1,0        ; vbroadcastss  0x1d9e2(%rip),%ymm10        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,36,195,1,0         ; vbroadcastss  0x1c324(%rip),%ymm11        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,29,208,217,1,0        ; vbroadcastss  0x1d9d0(%rip),%ymm11        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
   DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
-  DB  196,98,125,24,13,15,195,1,0         ; vbroadcastss  0x1c30f(%rip),%ymm9        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,13,187,217,1,0        ; vbroadcastss  0x1d9bb(%rip),%ymm9        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14712,13 +14712,13 @@
 PUBLIC _sk_bicubic_p3x_avx
 _sk_bicubic_p3x_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,5,235,194,1,0        ; vbroadcastss  0x1c2eb(%rip),%ymm0        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,226,125,24,5,151,217,1,0        ; vbroadcastss  0x1d997(%rip),%ymm0        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  197,252,88,0                        ; vaddps        (%rax),%ymm0,%ymm0
   DB  197,124,16,128,128,0,0,0            ; vmovups       0x80(%rax),%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,205,194,1,0        ; vbroadcastss  0x1c2cd(%rip),%ymm10        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,121,217,1,0        ; vbroadcastss  0x1d979(%rip),%ymm10        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,203,193,1,0        ; vbroadcastss  0x1c1cb(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,21,119,216,1,0        ; vbroadcastss  0x1d877(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,0,1,0,0              ; vmovups       %ymm8,0x100(%rax)
@@ -14728,14 +14728,14 @@
 PUBLIC _sk_bicubic_n3y_avx
 _sk_bicubic_n3y_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,154,194,1,0       ; vbroadcastss  0x1c29a(%rip),%ymm1        # 2be1c <_sk_srcover_rgba_8888_sse2_lowp+0x414>
+  DB  196,226,125,24,13,70,217,1,0        ; vbroadcastss  0x1d946(%rip),%ymm1        # 2d4c8 <_sk_srcover_rgba_8888_sse2_lowp+0x414>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,40,193,1,0          ; vbroadcastss  0x1c128(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,212,215,1,0         ; vbroadcastss  0x1d7d4(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,122,194,1,0        ; vbroadcastss  0x1c27a(%rip),%ymm10        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,38,217,1,0         ; vbroadcastss  0x1d926(%rip),%ymm10        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,120,193,1,0        ; vbroadcastss  0x1c178(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,21,36,216,1,0         ; vbroadcastss  0x1d824(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -14745,19 +14745,19 @@
 PUBLIC _sk_bicubic_n1y_avx
 _sk_bicubic_n1y_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,67,194,1,0        ; vbroadcastss  0x1c243(%rip),%ymm1        # 2be18 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
+  DB  196,226,125,24,13,239,216,1,0       ; vbroadcastss  0x1d8ef(%rip),%ymm1        # 2d4c4 <_sk_srcover_rgba_8888_sse2_lowp+0x410>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
-  DB  196,98,125,24,5,213,192,1,0         ; vbroadcastss  0x1c0d5(%rip),%ymm8        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,98,125,24,5,129,215,1,0         ; vbroadcastss  0x1d781(%rip),%ymm8        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,60,92,128,192,0,0,0             ; vsubps        0xc0(%rax),%ymm8,%ymm8
-  DB  196,98,125,24,13,52,194,1,0         ; vbroadcastss  0x1c234(%rip),%ymm9        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,13,224,216,1,0        ; vbroadcastss  0x1d8e0(%rip),%ymm9        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,34,194,1,0         ; vbroadcastss  0x1c222(%rip),%ymm10        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,21,206,216,1,0        ; vbroadcastss  0x1d8ce(%rip),%ymm10        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,201                    ; vmulps        %ymm9,%ymm8,%ymm9
-  DB  196,98,125,24,21,159,192,1,0        ; vbroadcastss  0x1c09f(%rip),%ymm10        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,21,75,215,1,0         ; vbroadcastss  0x1d74b(%rip),%ymm10        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,65,52,88,202                    ; vaddps        %ymm10,%ymm9,%ymm9
   DB  196,65,60,89,193                    ; vmulps        %ymm9,%ymm8,%ymm8
-  DB  196,98,125,24,13,4,194,1,0          ; vbroadcastss  0x1c204(%rip),%ymm9        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,13,176,216,1,0        ; vbroadcastss  0x1d8b0(%rip),%ymm9        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14766,17 +14766,17 @@
 PUBLIC _sk_bicubic_p1y_avx
 _sk_bicubic_p1y_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,98,125,24,5,112,192,1,0         ; vbroadcastss  0x1c070(%rip),%ymm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,98,125,24,5,28,215,1,0          ; vbroadcastss  0x1d71c(%rip),%ymm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,188,88,72,64                    ; vaddps        0x40(%rax),%ymm8,%ymm1
   DB  197,124,16,136,192,0,0,0            ; vmovups       0xc0(%rax),%ymm9
-  DB  196,98,125,24,21,206,193,1,0        ; vbroadcastss  0x1c1ce(%rip),%ymm10        # 2be28 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
+  DB  196,98,125,24,21,122,216,1,0        ; vbroadcastss  0x1d87a(%rip),%ymm10        # 2d4d4 <_sk_srcover_rgba_8888_sse2_lowp+0x420>
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
-  DB  196,98,125,24,29,188,193,1,0        ; vbroadcastss  0x1c1bc(%rip),%ymm11        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,98,125,24,29,104,216,1,0        ; vbroadcastss  0x1d868(%rip),%ymm11        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  196,65,44,88,211                    ; vaddps        %ymm11,%ymm10,%ymm10
   DB  196,65,52,89,210                    ; vmulps        %ymm10,%ymm9,%ymm10
   DB  196,65,44,88,192                    ; vaddps        %ymm8,%ymm10,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
-  DB  196,98,125,24,13,167,193,1,0        ; vbroadcastss  0x1c1a7(%rip),%ymm9        # 2be2c <_sk_srcover_rgba_8888_sse2_lowp+0x424>
+  DB  196,98,125,24,13,83,216,1,0         ; vbroadcastss  0x1d853(%rip),%ymm9        # 2d4d8 <_sk_srcover_rgba_8888_sse2_lowp+0x424>
   DB  196,65,60,88,193                    ; vaddps        %ymm9,%ymm8,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -14785,13 +14785,13 @@
 PUBLIC _sk_bicubic_p3y_avx
 _sk_bicubic_p3y_avx LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,13,131,193,1,0       ; vbroadcastss  0x1c183(%rip),%ymm1        # 2be24 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
+  DB  196,226,125,24,13,47,216,1,0        ; vbroadcastss  0x1d82f(%rip),%ymm1        # 2d4d0 <_sk_srcover_rgba_8888_sse2_lowp+0x41c>
   DB  197,244,88,72,64                    ; vaddps        0x40(%rax),%ymm1,%ymm1
   DB  197,124,16,128,192,0,0,0            ; vmovups       0xc0(%rax),%ymm8
   DB  196,65,60,89,200                    ; vmulps        %ymm8,%ymm8,%ymm9
-  DB  196,98,125,24,21,100,193,1,0        ; vbroadcastss  0x1c164(%rip),%ymm10        # 2be20 <_sk_srcover_rgba_8888_sse2_lowp+0x418>
+  DB  196,98,125,24,21,16,216,1,0         ; vbroadcastss  0x1d810(%rip),%ymm10        # 2d4cc <_sk_srcover_rgba_8888_sse2_lowp+0x418>
   DB  196,65,60,89,194                    ; vmulps        %ymm10,%ymm8,%ymm8
-  DB  196,98,125,24,21,98,192,1,0         ; vbroadcastss  0x1c062(%rip),%ymm10        # 2bd2c <_sk_srcover_rgba_8888_sse2_lowp+0x324>
+  DB  196,98,125,24,21,14,215,1,0         ; vbroadcastss  0x1d70e(%rip),%ymm10        # 2d3d8 <_sk_srcover_rgba_8888_sse2_lowp+0x324>
   DB  196,65,60,88,194                    ; vaddps        %ymm10,%ymm8,%ymm8
   DB  196,65,52,89,192                    ; vmulps        %ymm8,%ymm9,%ymm8
   DB  197,124,17,128,64,1,0,0             ; vmovups       %ymm8,0x140(%rax)
@@ -14927,7 +14927,7 @@
   DB  72,139,0                            ; mov           (%rax),%rax
   DB  197,217,254,207                     ; vpaddd        %xmm7,%xmm4,%xmm1
   DB  197,249,254,193                     ; vpaddd        %xmm1,%xmm0,%xmm0
-  DB  196,226,121,24,37,6,191,1,0         ; vbroadcastss  0x1bf06(%rip),%xmm4        # 2be34 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  DB  196,226,121,24,37,178,213,1,0       ; vbroadcastss  0x1d5b2(%rip),%xmm4        # 2d4e0 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   DB  196,98,121,64,228                   ; vpmulld       %xmm4,%xmm0,%xmm12
   DB  196,226,49,64,198                   ; vpmulld       %xmm6,%xmm9,%xmm0
   DB  197,249,127,132,36,0,1,0,0          ; vmovdqa       %xmm0,0x100(%rsp)
@@ -14981,7 +14981,7 @@
   DB  196,161,122,16,60,128               ; vmovss        (%rax,%r8,4),%xmm7
   DB  196,227,73,33,247,48                ; vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   DB  196,227,77,24,237,1                 ; vinsertf128   $0x1,%xmm5,%ymm6,%ymm5
-  DB  196,98,121,24,21,107,188,1,0        ; vbroadcastss  0x1bc6b(%rip),%xmm10        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,98,121,24,21,23,211,1,0         ; vbroadcastss  0x1d317(%rip),%xmm10        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  196,193,105,254,210                 ; vpaddd        %xmm10,%xmm2,%xmm2
   DB  196,195,249,22,208,1                ; vpextrq       $0x1,%xmm2,%r8
   DB  196,193,249,126,209                 ; vmovq         %xmm2,%r9
@@ -15009,7 +15009,7 @@
   DB  196,161,122,16,60,128               ; vmovss        (%rax,%r8,4),%xmm7
   DB  196,227,73,33,247,48                ; vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   DB  196,227,77,24,210,1                 ; vinsertf128   $0x1,%xmm2,%ymm6,%ymm2
-  DB  196,226,125,24,61,67,189,1,0        ; vbroadcastss  0x1bd43(%rip),%ymm7        # 2be30 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  DB  196,226,125,24,61,239,211,1,0       ; vbroadcastss  0x1d3ef(%rip),%ymm7        # 2d4dc <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   DB  197,148,88,247                      ; vaddps        %ymm7,%ymm13,%ymm6
   DB  197,124,40,231                      ; vmovaps       %ymm7,%ymm12
   DB  197,124,17,164,36,128,0,0,0         ; vmovups       %ymm12,0x80(%rsp)
@@ -15750,7 +15750,7 @@
   DB  197,249,127,132,36,160,0,0,0        ; vmovdqa       %xmm0,0xa0(%rsp)
   DB  197,225,254,216                     ; vpaddd        %xmm0,%xmm3,%xmm3
   DB  197,233,254,195                     ; vpaddd        %xmm3,%xmm2,%xmm0
-  DB  196,98,121,24,13,212,173,1,0        ; vbroadcastss  0x1add4(%rip),%xmm9        # 2be34 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
+  DB  196,98,121,24,13,128,196,1,0        ; vbroadcastss  0x1c480(%rip),%xmm9        # 2d4e0 <_sk_srcover_rgba_8888_sse2_lowp+0x42c>
   DB  196,194,121,64,209                  ; vpmulld       %xmm9,%xmm0,%xmm2
   DB  196,226,81,64,199                   ; vpmulld       %xmm7,%xmm5,%xmm0
   DB  197,249,127,132,36,224,1,0,0        ; vmovdqa       %xmm0,0x1e0(%rsp)
@@ -15807,7 +15807,7 @@
   DB  196,161,122,16,60,128               ; vmovss        (%rax,%r8,4),%xmm7
   DB  196,227,73,33,247,48                ; vinsertps     $0x30,%xmm7,%xmm6,%xmm6
   DB  196,227,77,24,237,1                 ; vinsertf128   $0x1,%xmm5,%ymm6,%ymm5
-  DB  196,98,121,24,21,41,171,1,0         ; vbroadcastss  0x1ab29(%rip),%xmm10        # 2bcc4 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
+  DB  196,98,121,24,21,213,193,1,0        ; vbroadcastss  0x1c1d5(%rip),%xmm10        # 2d370 <_sk_srcover_rgba_8888_sse2_lowp+0x2bc>
   DB  196,193,121,254,194                 ; vpaddd        %xmm10,%xmm0,%xmm0
   DB  196,195,249,22,192,1                ; vpextrq       $0x1,%xmm0,%r8
   DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
@@ -15835,7 +15835,7 @@
   DB  196,161,122,16,52,128               ; vmovss        (%rax,%r8,4),%xmm6
   DB  196,227,105,33,214,48               ; vinsertps     $0x30,%xmm6,%xmm2,%xmm2
   DB  196,227,109,24,208,1                ; vinsertf128   $0x1,%xmm0,%ymm2,%ymm2
-  DB  196,98,125,24,37,1,172,1,0          ; vbroadcastss  0x1ac01(%rip),%ymm12        # 2be30 <_sk_srcover_rgba_8888_sse2_lowp+0x428>
+  DB  196,98,125,24,37,173,194,1,0        ; vbroadcastss  0x1c2ad(%rip),%ymm12        # 2d4dc <_sk_srcover_rgba_8888_sse2_lowp+0x428>
   DB  196,193,4,88,196                    ; vaddps        %ymm12,%ymm15,%ymm0
   DB  197,124,17,164,36,128,0,0,0         ; vmovups       %ymm12,0x80(%rsp)
   DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
@@ -17233,7 +17233,7 @@
   DB  197,228,89,210                      ; vmulps        %ymm2,%ymm3,%ymm2
   DB  197,220,88,210                      ; vaddps        %ymm2,%ymm4,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,24,29,3,141,1,0         ; vbroadcastss  0x18d03(%rip),%ymm3        # 2bcb8 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
+  DB  196,226,125,24,29,175,163,1,0       ; vbroadcastss  0x1a3af(%rip),%ymm3        # 2d364 <_sk_srcover_rgba_8888_sse2_lowp+0x2b0>
   DB  197,252,16,164,36,192,2,0,0         ; vmovups       0x2c0(%rsp),%ymm4
   DB  197,252,16,172,36,224,2,0,0         ; vmovups       0x2e0(%rsp),%ymm5
   DB  197,252,16,180,36,0,3,0,0           ; vmovups       0x300(%rsp),%ymm6
@@ -17355,9 +17355,9 @@
   DB  102,15,110,201                      ; movd          %ecx,%xmm1
   DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,88,13,159,146,1,0                ; addps         0x1929f(%rip),%xmm1        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,13,127,169,1,0                ; addps         0x1a97f(%rip),%xmm1        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,21,166,146,1,0                ; movaps        0x192a6(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,134,169,1,0                ; movaps        0x1a986(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -17370,18 +17370,18 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  102,68,15,110,194                   ; movd          %edx,%xmm8
   DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
-  DB  102,68,15,254,5,79,146,1,0          ; paddd         0x1924f(%rip),%xmm8        # 2c400 <_sk_srcover_rgba_8888_sse2_lowp+0x9f8>
+  DB  102,68,15,254,5,47,169,1,0          ; paddd         0x1a92f(%rip),%xmm8        # 2dae0 <_sk_srcover_rgba_8888_sse2_lowp+0xa2c>
   DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
   DB  102,69,15,112,201,0                 ; pshufd        $0x0,%xmm9,%xmm9
   DB  102,69,15,239,200                   ; pxor          %xmm8,%xmm9
-  DB  102,68,15,111,21,118,146,1,0        ; movdqa        0x19276(%rip),%xmm10        # 2c440 <_sk_srcover_rgba_8888_sse2_lowp+0xa38>
+  DB  102,68,15,111,21,86,169,1,0         ; movdqa        0x1a956(%rip),%xmm10        # 2db20 <_sk_srcover_rgba_8888_sse2_lowp+0xa6c>
   DB  102,69,15,111,217                   ; movdqa        %xmm9,%xmm11
   DB  102,69,15,219,218                   ; pand          %xmm10,%xmm11
   DB  102,65,15,114,243,5                 ; pslld         $0x5,%xmm11
   DB  102,69,15,219,208                   ; pand          %xmm8,%xmm10
   DB  102,65,15,114,242,4                 ; pslld         $0x4,%xmm10
-  DB  102,68,15,111,37,98,146,1,0         ; movdqa        0x19262(%rip),%xmm12        # 2c450 <_sk_srcover_rgba_8888_sse2_lowp+0xa48>
-  DB  102,68,15,111,45,105,146,1,0        ; movdqa        0x19269(%rip),%xmm13        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,37,66,169,1,0         ; movdqa        0x1a942(%rip),%xmm12        # 2db30 <_sk_srcover_rgba_8888_sse2_lowp+0xa7c>
+  DB  102,68,15,111,45,73,169,1,0         ; movdqa        0x1a949(%rip),%xmm13        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,69,15,111,241                   ; movdqa        %xmm9,%xmm14
   DB  102,69,15,219,245                   ; pand          %xmm13,%xmm14
   DB  102,65,15,114,246,2                 ; pslld         $0x2,%xmm14
@@ -17397,8 +17397,8 @@
   DB  102,69,15,235,198                   ; por           %xmm14,%xmm8
   DB  102,69,15,235,193                   ; por           %xmm9,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,36,146,1,0               ; mulps         0x19224(%rip),%xmm8        # 2c470 <_sk_srcover_rgba_8888_sse2_lowp+0xa68>
-  DB  68,15,88,5,44,146,1,0               ; addps         0x1922c(%rip),%xmm8        # 2c480 <_sk_srcover_rgba_8888_sse2_lowp+0xa78>
+  DB  68,15,89,5,4,169,1,0                ; mulps         0x1a904(%rip),%xmm8        # 2db50 <_sk_srcover_rgba_8888_sse2_lowp+0xa9c>
+  DB  68,15,88,5,12,169,1,0               ; addps         0x1a90c(%rip),%xmm8        # 2db60 <_sk_srcover_rgba_8888_sse2_lowp+0xaac>
   DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
@@ -17436,7 +17436,7 @@
 PUBLIC _sk_black_color_sse41
 _sk_black_color_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,98,145,1,0                 ; movaps        0x19162(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,66,168,1,0                 ; movaps        0x1a842(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -17445,7 +17445,7 @@
 PUBLIC _sk_white_color_sse41
 _sk_white_color_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,5,78,145,1,0                  ; movaps        0x1914e(%rip),%xmm0        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,5,46,168,1,0                  ; movaps        0x1a82e(%rip),%xmm0        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -17483,7 +17483,7 @@
 PUBLIC _sk_srcatop_sse41
 _sk_srcatop_sse41 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,254,144,1,0              ; movaps        0x190fe(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,222,167,1,0              ; movaps        0x1a7de(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -17506,7 +17506,7 @@
 _sk_dstatop_sse41 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
   DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
-  DB  68,15,40,13,177,144,1,0             ; movaps        0x190b1(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,145,167,1,0             ; movaps        0x1a791(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -17547,7 +17547,7 @@
 
 PUBLIC _sk_srcout_sse41
 _sk_srcout_sse41 LABEL PROC
-  DB  68,15,40,5,69,144,1,0               ; movaps        0x19045(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,37,167,1,0               ; movaps        0x1a725(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -17558,7 +17558,7 @@
 
 PUBLIC _sk_dstout_sse41
 _sk_dstout_sse41 LABEL PROC
-  DB  68,15,40,5,37,144,1,0               ; movaps        0x19025(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,5,167,1,0                ; movaps        0x1a705(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -17573,7 +17573,7 @@
 
 PUBLIC _sk_srcover_sse41
 _sk_srcover_sse41 LABEL PROC
-  DB  68,15,40,5,248,143,1,0              ; movaps        0x18ff8(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,216,166,1,0              ; movaps        0x1a6d8(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -17591,7 +17591,7 @@
 
 PUBLIC _sk_dstover_sse41
 _sk_dstover_sse41 LABEL PROC
-  DB  68,15,40,5,188,143,1,0              ; movaps        0x18fbc(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,156,166,1,0              ; movaps        0x1a69c(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -17615,7 +17615,7 @@
 
 PUBLIC _sk_multiply_sse41
 _sk_multiply_sse41 LABEL PROC
-  DB  68,15,40,5,128,143,1,0              ; movaps        0x18f80(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,96,166,1,0               ; movaps        0x1a660(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -17651,7 +17651,7 @@
 PUBLIC _sk_plus__sse41
 _sk_plus__sse41 LABEL PROC
   DB  15,88,196                           ; addps         %xmm4,%xmm0
-  DB  68,15,40,5,1,143,1,0                ; movaps        0x18f01(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,225,165,1,0              ; movaps        0x1a5e1(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  15,88,205                           ; addps         %xmm5,%xmm1
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
@@ -17689,7 +17689,7 @@
 PUBLIC _sk_xor__sse41
 _sk_xor__sse41 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,141,142,1,0                ; movaps        0x18e8d(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,109,165,1,0                ; movaps        0x1a56d(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -17735,7 +17735,7 @@
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,232,141,1,0                ; movaps        0x18de8(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,200,164,1,0                ; movaps        0x1a4c8(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -17767,7 +17767,7 @@
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,125,141,1,0                ; movaps        0x18d7d(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,93,164,1,0                 ; movaps        0x1a45d(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -17802,7 +17802,7 @@
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,7,141,1,0                  ; movaps        0x18d07(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,231,163,1,0                ; movaps        0x1a3e7(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -17828,7 +17828,7 @@
   DB  15,89,214                           ; mulps         %xmm6,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,183,140,1,0                ; movaps        0x18cb7(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,151,163,1,0                ; movaps        0x1a397(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -17838,7 +17838,7 @@
 
 PUBLIC _sk_colorburn_sse41
 _sk_colorburn_sse41 LABEL PROC
-  DB  68,15,40,29,158,140,1,0             ; movaps        0x18c9e(%rip),%xmm11        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,29,126,163,1,0             ; movaps        0x1a37e(%rip),%xmm11        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
   DB  68,15,92,231                        ; subps         %xmm7,%xmm12
   DB  69,15,40,204                        ; movaps        %xmm12,%xmm9
@@ -17920,7 +17920,7 @@
 PUBLIC _sk_colordodge_sse41
 _sk_colordodge_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,21,104,139,1,0             ; movaps        0x18b68(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,72,162,1,0              ; movaps        0x1a248(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -18001,7 +18001,7 @@
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
-  DB  68,15,40,21,48,138,1,0              ; movaps        0x18a30(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,16,161,1,0              ; movaps        0x1a110(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,40,234                        ; movaps        %xmm10,%xmm5
   DB  15,92,239                           ; subps         %xmm7,%xmm5
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
@@ -18083,7 +18083,7 @@
 _sk_overlay_sse41 LABEL PROC
   DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
-  DB  68,15,40,21,2,137,1,0               ; movaps        0x18902(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,226,159,1,0             ; movaps        0x19fe2(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
@@ -18167,7 +18167,7 @@
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  65,15,84,193                        ; andps         %xmm9,%xmm0
-  DB  15,40,13,197,135,1,0                ; movaps        0x187c5(%rip),%xmm1        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,13,165,158,1,0                ; movaps        0x19ea5(%rip),%xmm1        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  68,15,92,208                        ; subps         %xmm0,%xmm10
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
@@ -18180,10 +18180,10 @@
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,89,210                           ; mulps         %xmm2,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
-  DB  68,15,40,45,243,135,1,0             ; movaps        0x187f3(%rip),%xmm13        # 2c490 <_sk_srcover_rgba_8888_sse2_lowp+0xa88>
+  DB  68,15,40,45,211,158,1,0             ; movaps        0x19ed3(%rip),%xmm13        # 2db70 <_sk_srcover_rgba_8888_sse2_lowp+0xabc>
   DB  69,15,88,245                        ; addps         %xmm13,%xmm14
   DB  68,15,89,242                        ; mulps         %xmm2,%xmm14
-  DB  68,15,40,37,243,135,1,0             ; movaps        0x187f3(%rip),%xmm12        # 2c4a0 <_sk_srcover_rgba_8888_sse2_lowp+0xa98>
+  DB  68,15,40,37,211,158,1,0             ; movaps        0x19ed3(%rip),%xmm12        # 2db80 <_sk_srcover_rgba_8888_sse2_lowp+0xacc>
   DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
   DB  69,15,88,254                        ; addps         %xmm14,%xmm15
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
@@ -18369,12 +18369,12 @@
   DB  68,15,84,208                        ; andps         %xmm0,%xmm10
   DB  15,84,200                           ; andps         %xmm0,%xmm1
   DB  68,15,84,232                        ; andps         %xmm0,%xmm13
-  DB  15,40,5,89,133,1,0                  ; movaps        0x18559(%rip),%xmm0        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  15,40,5,57,156,1,0                  ; movaps        0x19c39(%rip),%xmm0        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,89,224                        ; mulps         %xmm0,%xmm12
-  DB  15,40,21,94,133,1,0                 ; movaps        0x1855e(%rip),%xmm2        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  15,40,21,62,156,1,0                 ; movaps        0x19c3e(%rip),%xmm2        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  15,89,250                           ; mulps         %xmm2,%xmm7
   DB  65,15,88,252                        ; addps         %xmm12,%xmm7
-  DB  68,15,40,53,95,133,1,0              ; movaps        0x1855f(%rip),%xmm14        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,53,63,156,1,0              ; movaps        0x19c3f(%rip),%xmm14        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  68,15,40,252                        ; movaps        %xmm4,%xmm15
   DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
   DB  68,15,88,255                        ; addps         %xmm7,%xmm15
@@ -18457,7 +18457,7 @@
   DB  65,15,88,214                        ; addps         %xmm14,%xmm2
   DB  15,40,196                           ; movaps        %xmm4,%xmm0
   DB  102,15,56,20,202                    ; blendvps      %xmm0,%xmm2,%xmm1
-  DB  68,15,40,13,116,131,1,0             ; movaps        0x18374(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,84,154,1,0              ; movaps        0x19a54(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,40,225                        ; movaps        %xmm9,%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  15,40,68,36,48                      ; movaps        0x30(%rsp),%xmm0
@@ -18551,14 +18551,14 @@
   DB  68,15,84,215                        ; andps         %xmm7,%xmm10
   DB  68,15,84,223                        ; andps         %xmm7,%xmm11
   DB  68,15,84,199                        ; andps         %xmm7,%xmm8
-  DB  15,40,21,151,130,1,0                ; movaps        0x18297(%rip),%xmm2        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  15,40,21,119,153,1,0                ; movaps        0x19977(%rip),%xmm2        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,218                           ; mulps         %xmm2,%xmm3
-  DB  15,40,13,154,130,1,0                ; movaps        0x1829a(%rip),%xmm1        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  15,40,13,122,153,1,0                ; movaps        0x1997a(%rip),%xmm1        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
   DB  15,89,249                           ; mulps         %xmm1,%xmm7
   DB  15,88,251                           ; addps         %xmm3,%xmm7
-  DB  68,15,40,45,153,130,1,0             ; movaps        0x18299(%rip),%xmm13        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,45,121,153,1,0             ; movaps        0x19979(%rip),%xmm13        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
   DB  68,15,88,247                        ; addps         %xmm7,%xmm14
   DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
@@ -18639,7 +18639,7 @@
   DB  65,15,88,253                        ; addps         %xmm13,%xmm7
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  102,68,15,56,20,223                 ; blendvps      %xmm0,%xmm7,%xmm11
-  DB  68,15,40,13,175,128,1,0             ; movaps        0x180af(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,143,151,1,0             ; movaps        0x1978f(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  68,15,92,204                        ; subps         %xmm4,%xmm9
   DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
@@ -18694,14 +18694,14 @@
   DB  15,40,231                           ; movaps        %xmm7,%xmm4
   DB  68,15,89,244                        ; mulps         %xmm4,%xmm14
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
-  DB  68,15,40,13,100,128,1,0             ; movaps        0x18064(%rip),%xmm9        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,13,68,151,1,0              ; movaps        0x19744(%rip),%xmm9        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  65,15,40,250                        ; movaps        %xmm10,%xmm7
   DB  65,15,89,249                        ; mulps         %xmm9,%xmm7
-  DB  68,15,40,21,100,128,1,0             ; movaps        0x18064(%rip),%xmm10        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  68,15,40,21,68,151,1,0              ; movaps        0x19744(%rip),%xmm10        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
   DB  65,15,89,218                        ; mulps         %xmm10,%xmm3
   DB  15,88,223                           ; addps         %xmm7,%xmm3
-  DB  68,15,40,29,97,128,1,0              ; movaps        0x18061(%rip),%xmm11        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,29,65,151,1,0              ; movaps        0x19741(%rip),%xmm11        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
@@ -18786,7 +18786,7 @@
   DB  65,15,88,251                        ; addps         %xmm11,%xmm7
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  102,15,56,20,207                    ; blendvps      %xmm0,%xmm7,%xmm1
-  DB  68,15,40,13,109,126,1,0             ; movaps        0x17e6d(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,77,149,1,0              ; movaps        0x1954d(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  68,15,89,192                        ; mulps         %xmm0,%xmm8
@@ -18838,13 +18838,13 @@
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,89,205                        ; mulps         %xmm5,%xmm9
-  DB  68,15,40,5,47,126,1,0               ; movaps        0x17e2f(%rip),%xmm8        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,5,15,149,1,0               ; movaps        0x1950f(%rip),%xmm8        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
-  DB  68,15,40,21,51,126,1,0              ; movaps        0x17e33(%rip),%xmm10        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  68,15,40,21,19,149,1,0              ; movaps        0x19513(%rip),%xmm10        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  65,15,89,234                        ; mulps         %xmm10,%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  68,15,40,37,49,126,1,0              ; movaps        0x17e31(%rip),%xmm12        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,37,17,149,1,0              ; movaps        0x19511(%rip),%xmm12        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  68,15,40,242                        ; movaps        %xmm2,%xmm14
   DB  69,15,89,244                        ; mulps         %xmm12,%xmm14
   DB  68,15,88,245                        ; addps         %xmm5,%xmm14
@@ -18929,7 +18929,7 @@
   DB  65,15,88,244                        ; addps         %xmm12,%xmm6
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  102,68,15,56,20,206                 ; blendvps      %xmm0,%xmm6,%xmm9
-  DB  15,40,5,55,124,1,0                  ; movaps        0x17c37(%rip),%xmm0        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,5,23,147,1,0                  ; movaps        0x19317(%rip),%xmm0        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,92,215                           ; subps         %xmm7,%xmm2
   DB  15,89,226                           ; mulps         %xmm2,%xmm4
@@ -18976,20 +18976,20 @@
   DB  15,133,219,0,0,0                    ; jne           14964 <_sk_srcover_rgba_8888_sse41+0xfb>
   DB  243,65,15,111,60,128                ; movdqu        (%r8,%rax,4),%xmm7
   DB  72,133,255                          ; test          %rdi,%rdi
-  DB  102,15,111,37,70,124,1,0            ; movdqa        0x17c46(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,37,38,147,1,0            ; movdqa        0x19326(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,231                      ; pand          %xmm7,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
   DB  102,15,111,239                      ; movdqa        %xmm7,%xmm5
-  DB  102,15,56,0,45,66,124,1,0           ; pshufb        0x17c42(%rip),%xmm5        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,45,34,147,1,0           ; pshufb        0x19322(%rip),%xmm5        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
   DB  102,15,111,247                      ; movdqa        %xmm7,%xmm6
-  DB  102,15,56,0,53,66,124,1,0           ; pshufb        0x17c42(%rip),%xmm6        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,53,34,147,1,0           ; pshufb        0x19322(%rip),%xmm6        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  102,15,114,215,24                   ; psrld         $0x18,%xmm7
   DB  15,91,255                           ; cvtdq2ps      %xmm7,%xmm7
-  DB  68,15,40,5,95,123,1,0               ; movaps        0x17b5f(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,63,146,1,0               ; movaps        0x1923f(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
-  DB  68,15,40,37,51,124,1,0              ; movaps        0x17c33(%rip),%xmm12        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,37,19,147,1,0              ; movaps        0x19313(%rip),%xmm12        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -19065,7 +19065,7 @@
 
 PUBLIC _sk_clamp_1_sse41
 _sk_clamp_1_sse41 LABEL PROC
-  DB  68,15,40,5,46,122,1,0               ; movaps        0x17a2e(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,14,145,1,0               ; movaps        0x1910e(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -19075,7 +19075,7 @@
 
 PUBLIC _sk_clamp_a_sse41
 _sk_clamp_a_sse41 LABEL PROC
-  DB  15,93,29,19,122,1,0                 ; minps         0x17a13(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,29,243,144,1,0                ; minps         0x190f3(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -19084,7 +19084,7 @@
 
 PUBLIC _sk_clamp_a_dst_sse41
 _sk_clamp_a_dst_sse41 LABEL PROC
-  DB  15,93,61,255,121,1,0                ; minps         0x179ff(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,61,223,144,1,0                ; minps         0x190df(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,93,231                           ; minps         %xmm7,%xmm4
   DB  15,93,239                           ; minps         %xmm7,%xmm5
   DB  15,93,247                           ; minps         %xmm7,%xmm6
@@ -19113,7 +19113,7 @@
 
 PUBLIC _sk_invert_sse41
 _sk_invert_sse41 LABEL PROC
-  DB  68,15,40,5,187,121,1,0              ; movaps        0x179bb(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,155,144,1,0              ; movaps        0x1909b(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,200                        ; subps         %xmm0,%xmm9
   DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
@@ -19164,10 +19164,10 @@
 
 PUBLIC _sk_unpremul_sse41
 _sk_unpremul_sse41 LABEL PROC
-  DB  68,15,40,5,73,121,1,0               ; movaps        0x17949(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,41,144,1,0               ; movaps        0x19029(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,94,195                        ; divps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
-  DB  68,15,194,13,40,122,1,0,1           ; cmpltps       0x17a28(%rip),%xmm9        # 2c520 <_sk_srcover_rgba_8888_sse2_lowp+0xb18>
+  DB  68,15,194,13,8,145,1,0,1            ; cmpltps       0x19108(%rip),%xmm9        # 2dc00 <_sk_srcover_rgba_8888_sse2_lowp+0xb4c>
   DB  69,15,84,200                        ; andps         %xmm8,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
@@ -19177,20 +19177,20 @@
 
 PUBLIC _sk_from_srgb_sse41
 _sk_from_srgb_sse41 LABEL PROC
-  DB  68,15,40,29,28,122,1,0              ; movaps        0x17a1c(%rip),%xmm11        # 2c530 <_sk_srcover_rgba_8888_sse2_lowp+0xb28>
+  DB  68,15,40,29,252,144,1,0             ; movaps        0x190fc(%rip),%xmm11        # 2dc10 <_sk_srcover_rgba_8888_sse2_lowp+0xb5c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
   DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
   DB  69,15,89,210                        ; mulps         %xmm10,%xmm10
-  DB  68,15,40,37,132,121,1,0             ; movaps        0x17984(%rip),%xmm12        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,37,100,144,1,0             ; movaps        0x19064(%rip),%xmm12        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,4,122,1,0               ; movaps        0x17a04(%rip),%xmm13        # 2c540 <_sk_srcover_rgba_8888_sse2_lowp+0xb38>
+  DB  68,15,40,45,228,144,1,0             ; movaps        0x190e4(%rip),%xmm13        # 2dc20 <_sk_srcover_rgba_8888_sse2_lowp+0xb6c>
   DB  69,15,88,197                        ; addps         %xmm13,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
-  DB  68,15,40,53,4,122,1,0               ; movaps        0x17a04(%rip),%xmm14        # 2c550 <_sk_srcover_rgba_8888_sse2_lowp+0xb48>
+  DB  68,15,40,53,228,144,1,0             ; movaps        0x190e4(%rip),%xmm14        # 2dc30 <_sk_srcover_rgba_8888_sse2_lowp+0xb7c>
   DB  69,15,88,198                        ; addps         %xmm14,%xmm8
-  DB  68,15,40,61,8,122,1,0               ; movaps        0x17a08(%rip),%xmm15        # 2c560 <_sk_srcover_rgba_8888_sse2_lowp+0xb58>
+  DB  68,15,40,61,232,144,1,0             ; movaps        0x190e8(%rip),%xmm15        # 2dc40 <_sk_srcover_rgba_8888_sse2_lowp+0xb8c>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
@@ -19225,19 +19225,19 @@
 _sk_from_srgb_dst_sse41 LABEL PROC
   DB  68,15,40,204                        ; movaps        %xmm4,%xmm9
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,29,83,121,1,0              ; movaps        0x17953(%rip),%xmm11        # 2c530 <_sk_srcover_rgba_8888_sse2_lowp+0xb28>
+  DB  68,15,40,29,51,144,1,0              ; movaps        0x19033(%rip),%xmm11        # 2dc10 <_sk_srcover_rgba_8888_sse2_lowp+0xb5c>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  69,15,89,211                        ; mulps         %xmm11,%xmm10
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
-  DB  68,15,40,37,188,120,1,0             ; movaps        0x178bc(%rip),%xmm12        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,37,156,143,1,0             ; movaps        0x18f9c(%rip),%xmm12        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  65,15,89,228                        ; mulps         %xmm12,%xmm4
-  DB  68,15,40,45,64,121,1,0              ; movaps        0x17940(%rip),%xmm13        # 2c540 <_sk_srcover_rgba_8888_sse2_lowp+0xb38>
+  DB  68,15,40,45,32,144,1,0              ; movaps        0x19020(%rip),%xmm13        # 2dc20 <_sk_srcover_rgba_8888_sse2_lowp+0xb6c>
   DB  65,15,88,229                        ; addps         %xmm13,%xmm4
   DB  15,89,224                           ; mulps         %xmm0,%xmm4
-  DB  68,15,40,53,65,121,1,0              ; movaps        0x17941(%rip),%xmm14        # 2c550 <_sk_srcover_rgba_8888_sse2_lowp+0xb48>
+  DB  68,15,40,53,33,144,1,0              ; movaps        0x19021(%rip),%xmm14        # 2dc30 <_sk_srcover_rgba_8888_sse2_lowp+0xb7c>
   DB  65,15,88,230                        ; addps         %xmm14,%xmm4
-  DB  68,15,40,61,69,121,1,0              ; movaps        0x17945(%rip),%xmm15        # 2c560 <_sk_srcover_rgba_8888_sse2_lowp+0xb58>
+  DB  68,15,40,61,37,144,1,0              ; movaps        0x19025(%rip),%xmm15        # 2dc40 <_sk_srcover_rgba_8888_sse2_lowp+0xb8c>
   DB  69,15,194,207,1                     ; cmpltps       %xmm15,%xmm9
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  102,65,15,56,20,226                 ; blendvps      %xmm0,%xmm10,%xmm4
@@ -19280,22 +19280,22 @@
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  15,40,209                           ; movaps        %xmm1,%xmm2
   DB  68,15,82,192                        ; rsqrtps       %xmm0,%xmm8
-  DB  68,15,40,29,182,120,1,0             ; movaps        0x178b6(%rip),%xmm11        # 2c570 <_sk_srcover_rgba_8888_sse2_lowp+0xb68>
+  DB  68,15,40,29,150,143,1,0             ; movaps        0x18f96(%rip),%xmm11        # 2dc50 <_sk_srcover_rgba_8888_sse2_lowp+0xb9c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
-  DB  68,15,40,37,182,120,1,0             ; movaps        0x178b6(%rip),%xmm12        # 2c580 <_sk_srcover_rgba_8888_sse2_lowp+0xb78>
+  DB  68,15,40,37,150,143,1,0             ; movaps        0x18f96(%rip),%xmm12        # 2dc60 <_sk_srcover_rgba_8888_sse2_lowp+0xbac>
   DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
   DB  69,15,89,252                        ; mulps         %xmm12,%xmm15
-  DB  68,15,40,21,182,120,1,0             ; movaps        0x178b6(%rip),%xmm10        # 2c590 <_sk_srcover_rgba_8888_sse2_lowp+0xb88>
+  DB  68,15,40,21,150,143,1,0             ; movaps        0x18f96(%rip),%xmm10        # 2dc70 <_sk_srcover_rgba_8888_sse2_lowp+0xbbc>
   DB  69,15,88,250                        ; addps         %xmm10,%xmm15
   DB  69,15,89,248                        ; mulps         %xmm8,%xmm15
-  DB  68,15,40,45,182,120,1,0             ; movaps        0x178b6(%rip),%xmm13        # 2c5a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb98>
+  DB  68,15,40,45,150,143,1,0             ; movaps        0x18f96(%rip),%xmm13        # 2dc80 <_sk_srcover_rgba_8888_sse2_lowp+0xbcc>
   DB  69,15,88,253                        ; addps         %xmm13,%xmm15
-  DB  68,15,40,53,186,120,1,0             ; movaps        0x178ba(%rip),%xmm14        # 2c5b0 <_sk_srcover_rgba_8888_sse2_lowp+0xba8>
+  DB  68,15,40,53,154,143,1,0             ; movaps        0x18f9a(%rip),%xmm14        # 2dc90 <_sk_srcover_rgba_8888_sse2_lowp+0xbdc>
   DB  69,15,88,198                        ; addps         %xmm14,%xmm8
   DB  69,15,83,192                        ; rcpps         %xmm8,%xmm8
   DB  69,15,89,199                        ; mulps         %xmm15,%xmm8
-  DB  68,15,40,61,182,120,1,0             ; movaps        0x178b6(%rip),%xmm15        # 2c5c0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb8>
+  DB  68,15,40,61,150,143,1,0             ; movaps        0x18f96(%rip),%xmm15        # 2dca0 <_sk_srcover_rgba_8888_sse2_lowp+0xbec>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  68,15,82,202                        ; rsqrtps       %xmm2,%xmm9
@@ -19349,7 +19349,7 @@
   DB  68,15,93,224                        ; minps         %xmm0,%xmm12
   DB  65,15,40,203                        ; movaps        %xmm11,%xmm1
   DB  65,15,92,204                        ; subps         %xmm12,%xmm1
-  DB  68,15,40,53,90,118,1,0              ; movaps        0x1765a(%rip),%xmm14        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,53,58,141,1,0              ; movaps        0x18d3a(%rip),%xmm14        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,94,241                        ; divps         %xmm1,%xmm14
   DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
   DB  69,15,194,208,0                     ; cmpeqps       %xmm8,%xmm10
@@ -19358,27 +19358,27 @@
   DB  65,15,89,198                        ; mulps         %xmm14,%xmm0
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  68,15,194,250,1                     ; cmpltps       %xmm2,%xmm15
-  DB  68,15,84,61,209,119,1,0             ; andps         0x177d1(%rip),%xmm15        # 2c5d0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc8>
+  DB  68,15,84,61,177,142,1,0             ; andps         0x18eb1(%rip),%xmm15        # 2dcb0 <_sk_srcover_rgba_8888_sse2_lowp+0xbfc>
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,194,193,0                     ; cmpeqps       %xmm9,%xmm0
   DB  65,15,92,208                        ; subps         %xmm8,%xmm2
   DB  65,15,89,214                        ; mulps         %xmm14,%xmm2
-  DB  68,15,40,45,196,119,1,0             ; movaps        0x177c4(%rip),%xmm13        # 2c5e0 <_sk_srcover_rgba_8888_sse2_lowp+0xbd8>
+  DB  68,15,40,45,164,142,1,0             ; movaps        0x18ea4(%rip),%xmm13        # 2dcc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc0c>
   DB  65,15,88,213                        ; addps         %xmm13,%xmm2
   DB  69,15,92,193                        ; subps         %xmm9,%xmm8
   DB  69,15,89,198                        ; mulps         %xmm14,%xmm8
-  DB  68,15,88,5,192,119,1,0              ; addps         0x177c0(%rip),%xmm8        # 2c5f0 <_sk_srcover_rgba_8888_sse2_lowp+0xbe8>
+  DB  68,15,88,5,160,142,1,0              ; addps         0x18ea0(%rip),%xmm8        # 2dcd0 <_sk_srcover_rgba_8888_sse2_lowp+0xc1c>
   DB  102,68,15,56,20,194                 ; blendvps      %xmm0,%xmm2,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  102,69,15,56,20,199                 ; blendvps      %xmm0,%xmm15,%xmm8
-  DB  68,15,89,5,184,119,1,0              ; mulps         0x177b8(%rip),%xmm8        # 2c600 <_sk_srcover_rgba_8888_sse2_lowp+0xbf8>
+  DB  68,15,89,5,152,142,1,0              ; mulps         0x18e98(%rip),%xmm8        # 2dce0 <_sk_srcover_rgba_8888_sse2_lowp+0xc2c>
   DB  69,15,40,203                        ; movaps        %xmm11,%xmm9
   DB  69,15,194,204,4                     ; cmpneqps      %xmm12,%xmm9
   DB  69,15,84,193                        ; andps         %xmm9,%xmm8
   DB  69,15,92,235                        ; subps         %xmm11,%xmm13
   DB  69,15,88,220                        ; addps         %xmm12,%xmm11
-  DB  15,40,5,188,117,1,0                 ; movaps        0x175bc(%rip),%xmm0        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,40,5,156,140,1,0                 ; movaps        0x18c9c(%rip),%xmm0        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,40,211                        ; movaps        %xmm11,%xmm2
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
   DB  15,194,194,1                        ; cmpltps       %xmm2,%xmm0
@@ -19399,7 +19399,7 @@
   DB  15,41,100,36,32                     ; movaps        %xmm4,0x20(%rsp)
   DB  15,41,92,36,16                      ; movaps        %xmm3,0x10(%rsp)
   DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
-  DB  68,15,40,13,110,117,1,0             ; movaps        0x1756e(%rip),%xmm9        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,13,78,140,1,0              ; movaps        0x18c4e(%rip),%xmm9        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,194,194,2                        ; cmpleps       %xmm2,%xmm0
   DB  15,40,217                           ; movaps        %xmm1,%xmm3
@@ -19412,19 +19412,19 @@
   DB  15,41,20,36                         ; movaps        %xmm2,(%rsp)
   DB  69,15,88,192                        ; addps         %xmm8,%xmm8
   DB  68,15,92,197                        ; subps         %xmm5,%xmm8
-  DB  68,15,40,53,42,119,1,0              ; movaps        0x1772a(%rip),%xmm14        # 2c610 <_sk_srcover_rgba_8888_sse2_lowp+0xc08>
+  DB  68,15,40,53,10,142,1,0              ; movaps        0x18e0a(%rip),%xmm14        # 2dcf0 <_sk_srcover_rgba_8888_sse2_lowp+0xc3c>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  102,65,15,58,8,198,1                ; roundps       $0x1,%xmm14,%xmm0
   DB  68,15,92,240                        ; subps         %xmm0,%xmm14
-  DB  68,15,40,29,35,119,1,0              ; movaps        0x17723(%rip),%xmm11        # 2c620 <_sk_srcover_rgba_8888_sse2_lowp+0xc18>
+  DB  68,15,40,29,3,142,1,0               ; movaps        0x18e03(%rip),%xmm11        # 2dd00 <_sk_srcover_rgba_8888_sse2_lowp+0xc4c>
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  15,40,245                           ; movaps        %xmm5,%xmm6
   DB  65,15,92,240                        ; subps         %xmm8,%xmm6
-  DB  15,40,61,188,118,1,0                ; movaps        0x176bc(%rip),%xmm7        # 2c5d0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc8>
+  DB  15,40,61,156,141,1,0                ; movaps        0x18d9c(%rip),%xmm7        # 2dcb0 <_sk_srcover_rgba_8888_sse2_lowp+0xbfc>
   DB  69,15,40,238                        ; movaps        %xmm14,%xmm13
   DB  68,15,89,239                        ; mulps         %xmm7,%xmm13
-  DB  15,40,29,205,118,1,0                ; movaps        0x176cd(%rip),%xmm3        # 2c5f0 <_sk_srcover_rgba_8888_sse2_lowp+0xbe8>
+  DB  15,40,29,173,141,1,0                ; movaps        0x18dad(%rip),%xmm3        # 2dcd0 <_sk_srcover_rgba_8888_sse2_lowp+0xc1c>
   DB  68,15,40,227                        ; movaps        %xmm3,%xmm12
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  68,15,89,230                        ; mulps         %xmm6,%xmm12
@@ -19434,7 +19434,7 @@
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  68,15,40,253                        ; movaps        %xmm5,%xmm15
   DB  102,69,15,56,20,252                 ; blendvps      %xmm0,%xmm12,%xmm15
-  DB  68,15,40,37,172,118,1,0             ; movaps        0x176ac(%rip),%xmm12        # 2c600 <_sk_srcover_rgba_8888_sse2_lowp+0xbf8>
+  DB  68,15,40,37,140,141,1,0             ; movaps        0x18d8c(%rip),%xmm12        # 2dce0 <_sk_srcover_rgba_8888_sse2_lowp+0xc2c>
   DB  65,15,40,196                        ; movaps        %xmm12,%xmm0
   DB  65,15,194,198,2                     ; cmpleps       %xmm14,%xmm0
   DB  68,15,89,238                        ; mulps         %xmm6,%xmm13
@@ -19468,7 +19468,7 @@
   DB  65,15,40,198                        ; movaps        %xmm14,%xmm0
   DB  15,40,20,36                         ; movaps        (%rsp),%xmm2
   DB  102,15,56,20,202                    ; blendvps      %xmm0,%xmm2,%xmm1
-  DB  68,15,88,21,69,118,1,0              ; addps         0x17645(%rip),%xmm10        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,88,21,37,141,1,0              ; addps         0x18d25(%rip),%xmm10        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  102,65,15,58,8,194,1                ; roundps       $0x1,%xmm10,%xmm0
   DB  68,15,92,208                        ; subps         %xmm0,%xmm10
   DB  69,15,194,218,2                     ; cmpleps       %xmm10,%xmm11
@@ -19522,9 +19522,9 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,52                              ; jne           150d6 <_sk_scale_u8_sse41+0x4c>
   DB  102,71,15,56,49,4,16                ; pmovzxbd      (%r8,%r10,1),%xmm8
-  DB  102,68,15,219,5,46,116,1,0          ; pand          0x1742e(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,219,5,14,139,1,0          ; pand          0x18b0e(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,130,117,1,0              ; mulps         0x17582(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,89,5,98,140,1,0               ; mulps         0x18c62(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -19566,17 +19566,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,159,0,0,0                    ; jne           151f0 <_sk_scale_565_sse41+0xc3>
   DB  102,71,15,56,51,28,80               ; pmovzxwd      (%r8,%r10,2),%xmm11
-  DB  102,15,111,5,240,116,1,0            ; movdqa        0x174f0(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,208,139,1,0            ; movdqa        0x18bd0(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,65,15,219,195                   ; pand          %xmm11,%xmm0
   DB  68,15,91,200                        ; cvtdq2ps      %xmm0,%xmm9
-  DB  68,15,89,13,239,116,1,0             ; mulps         0x174ef(%rip),%xmm9        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,5,247,116,1,0            ; movdqa        0x174f7(%rip),%xmm0        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  68,15,89,13,207,139,1,0             ; mulps         0x18bcf(%rip),%xmm9        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,5,215,139,1,0            ; movdqa        0x18bd7(%rip),%xmm0        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,65,15,219,195                   ; pand          %xmm11,%xmm0
   DB  68,15,91,208                        ; cvtdq2ps      %xmm0,%xmm10
-  DB  68,15,89,21,246,116,1,0             ; mulps         0x174f6(%rip),%xmm10        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,68,15,219,29,253,116,1,0        ; pand          0x174fd(%rip),%xmm11        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  68,15,89,21,214,139,1,0             ; mulps         0x18bd6(%rip),%xmm10        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,68,15,219,29,221,139,1,0        ; pand          0x18bdd(%rip),%xmm11        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
-  DB  68,15,89,29,1,117,1,0               ; mulps         0x17501(%rip),%xmm11        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  68,15,89,29,225,139,1,0             ; mulps         0x18be1(%rip),%xmm11        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  15,40,195                           ; movaps        %xmm3,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
   DB  69,15,40,226                        ; movaps        %xmm10,%xmm12
@@ -19649,9 +19649,9 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,72                              ; jne           152e0 <_sk_lerp_u8_sse41+0x60>
   DB  102,71,15,56,49,4,16                ; pmovzxbd      (%r8,%r10,1),%xmm8
-  DB  102,68,15,219,5,56,114,1,0          ; pand          0x17238(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,219,5,24,137,1,0          ; pand          0x18918(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,140,115,1,0              ; mulps         0x1738c(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,89,5,108,138,1,0              ; mulps         0x18a6c(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -19700,17 +19700,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,165,0,0,0                    ; jne           15403 <_sk_lerp_565_sse41+0xc9>
   DB  102,71,15,56,51,20,80               ; pmovzxwd      (%r8,%r10,2),%xmm10
-  DB  102,15,111,5,227,114,1,0            ; movdqa        0x172e3(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,195,137,1,0            ; movdqa        0x189c3(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,65,15,219,194                   ; pand          %xmm10,%xmm0
   DB  68,15,91,200                        ; cvtdq2ps      %xmm0,%xmm9
-  DB  68,15,89,13,226,114,1,0             ; mulps         0x172e2(%rip),%xmm9        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,5,234,114,1,0            ; movdqa        0x172ea(%rip),%xmm0        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  68,15,89,13,194,137,1,0             ; mulps         0x189c2(%rip),%xmm9        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,5,202,137,1,0            ; movdqa        0x189ca(%rip),%xmm0        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,65,15,219,194                   ; pand          %xmm10,%xmm0
   DB  68,15,91,216                        ; cvtdq2ps      %xmm0,%xmm11
-  DB  68,15,89,29,233,114,1,0             ; mulps         0x172e9(%rip),%xmm11        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,68,15,219,21,240,114,1,0        ; pand          0x172f0(%rip),%xmm10        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  68,15,89,29,201,137,1,0             ; mulps         0x189c9(%rip),%xmm11        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,68,15,219,21,208,137,1,0        ; pand          0x189d0(%rip),%xmm10        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,244,114,1,0             ; mulps         0x172f4(%rip),%xmm10        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  68,15,89,21,212,137,1,0             ; mulps         0x189d4(%rip),%xmm10        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  15,40,195                           ; movaps        %xmm3,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -19764,7 +19764,7 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,17,1,0,0                     ; jne           1557d <_sk_load_tables_sse41+0x122>
   DB  243,69,15,111,4,144                 ; movdqu        (%r8,%rdx,4),%xmm8
-  DB  102,15,111,5,102,112,1,0            ; movdqa        0x17066(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,70,135,1,0             ; movdqa        0x18746(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,73,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%r8
   DB  102,73,15,126,193                   ; movq          %xmm0,%r9
@@ -19779,7 +19779,7 @@
   DB  102,66,15,58,33,4,155,32            ; insertps      $0x20,(%rbx,%r11,4),%xmm0
   DB  102,66,15,58,33,4,3,48              ; insertps      $0x30,(%rbx,%r8,1),%xmm0
   DB  102,65,15,111,200                   ; movdqa        %xmm8,%xmm1
-  DB  102,15,56,0,13,33,112,1,0           ; pshufb        0x17021(%rip),%xmm1        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,13,1,135,1,0            ; pshufb        0x18701(%rip),%xmm1        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  102,73,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%r8
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
   DB  68,15,182,203                       ; movzbl        %bl,%r9d
@@ -19794,7 +19794,7 @@
   DB  102,15,58,33,202,48                 ; insertps      $0x30,%xmm2,%xmm1
   DB  76,139,64,24                        ; mov           0x18(%rax),%r8
   DB  102,65,15,111,208                   ; movdqa        %xmm8,%xmm2
-  DB  102,15,56,0,21,221,111,1,0          ; pshufb        0x16fdd(%rip),%xmm2        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,21,189,134,1,0          ; pshufb        0x186bd(%rip),%xmm2        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  102,72,15,58,22,211,1               ; pextrq        $0x1,%xmm2,%rbx
   DB  102,72,15,126,208                   ; movq          %xmm2,%rax
   DB  68,15,182,200                       ; movzbl        %al,%r9d
@@ -19809,7 +19809,7 @@
   DB  102,15,58,33,211,48                 ; insertps      $0x30,%xmm3,%xmm2
   DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
   DB  65,15,91,216                        ; cvtdq2ps      %xmm8,%xmm3
-  DB  15,89,29,202,112,1,0                ; mulps         0x170ca(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,170,135,1,0                ; mulps         0x187aa(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
@@ -19848,7 +19848,7 @@
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
-  DB  102,68,15,111,5,156,112,1,0         ; movdqa        0x1709c(%rip),%xmm8        # 2c6b0 <_sk_srcover_rgba_8888_sse2_lowp+0xca8>
+  DB  102,68,15,111,5,124,135,1,0         ; movdqa        0x1877c(%rip),%xmm8        # 2dd90 <_sk_srcover_rgba_8888_sse2_lowp+0xcdc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
@@ -19866,7 +19866,7 @@
   DB  102,15,58,33,194,32                 ; insertps      $0x20,%xmm2,%xmm0
   DB  243,66,15,16,20,3                   ; movss         (%rbx,%r8,1),%xmm2
   DB  102,15,58,33,194,48                 ; insertps      $0x30,%xmm2,%xmm0
-  DB  102,15,56,0,13,75,112,1,0           ; pshufb        0x1704b(%rip),%xmm1        # 2c6c0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb8>
+  DB  102,15,56,0,13,43,135,1,0           ; pshufb        0x1872b(%rip),%xmm1        # 2dda0 <_sk_srcover_rgba_8888_sse2_lowp+0xcec>
   DB  102,15,56,51,201                    ; pmovzxwd      %xmm1,%xmm1
   DB  102,73,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%r8
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
@@ -19902,7 +19902,7 @@
   DB  102,65,15,235,216                   ; por           %xmm8,%xmm3
   DB  102,15,56,51,219                    ; pmovzxwd      %xmm3,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,153,111,1,0                ; mulps         0x16f99(%rip),%xmm3        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  15,89,29,121,134,1,0                ; mulps         0x18679(%rip),%xmm3        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
@@ -19940,7 +19940,7 @@
   DB  102,15,97,209                       ; punpcklwd     %xmm1,%xmm2
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
-  DB  102,68,15,111,5,226,110,1,0         ; movdqa        0x16ee2(%rip),%xmm8        # 2c6b0 <_sk_srcover_rgba_8888_sse2_lowp+0xca8>
+  DB  102,68,15,111,5,194,133,1,0         ; movdqa        0x185c2(%rip),%xmm8        # 2dd90 <_sk_srcover_rgba_8888_sse2_lowp+0xcdc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,56,51,192                    ; pmovzxwd      %xmm0,%xmm0
@@ -19958,7 +19958,7 @@
   DB  102,15,58,33,195,32                 ; insertps      $0x20,%xmm3,%xmm0
   DB  243,66,15,16,28,3                   ; movss         (%rbx,%r8,1),%xmm3
   DB  102,15,58,33,195,48                 ; insertps      $0x30,%xmm3,%xmm0
-  DB  102,15,56,0,13,145,110,1,0          ; pshufb        0x16e91(%rip),%xmm1        # 2c6c0 <_sk_srcover_rgba_8888_sse2_lowp+0xcb8>
+  DB  102,15,56,0,13,113,133,1,0          ; pshufb        0x18571(%rip),%xmm1        # 2dda0 <_sk_srcover_rgba_8888_sse2_lowp+0xcec>
   DB  102,15,56,51,201                    ; pmovzxwd      %xmm1,%xmm1
   DB  102,73,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%r8
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
@@ -19989,7 +19989,7 @@
   DB  243,65,15,16,28,24                  ; movss         (%r8,%rbx,1),%xmm3
   DB  102,15,58,33,211,48                 ; insertps      $0x30,%xmm3,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,92,107,1,0                 ; movaps        0x16b5c(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,60,130,1,0                 ; movaps        0x1823c(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
   DB  255,224                             ; jmpq          *%rax
@@ -20019,7 +20019,7 @@
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,201,107,1,0              ; movaps        0x16bc9(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,169,130,1,0              ; movaps        0x182a9(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  102,73,15,58,22,192,1               ; pextrq        $0x1,%xmm0,%r8
@@ -20042,7 +20042,7 @@
   DB  102,15,58,32,197,3                  ; pinsrb        $0x3,%ebp,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,13,143,108,1,0             ; movaps        0x16c8f(%rip),%xmm9        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,13,111,131,1,0             ; movaps        0x1836f(%rip),%xmm9        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20142,7 +20142,7 @@
   DB  102,15,58,32,197,3                  ; pinsrb        $0x3,%ebp,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,13,235,106,1,0             ; movaps        0x16aeb(%rip),%xmm9        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,13,203,129,1,0             ; movaps        0x181cb(%rip),%xmm9        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -20314,31 +20314,31 @@
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,194                        ; cvtdq2ps      %xmm10,%xmm8
-  DB  68,15,89,5,198,104,1,0              ; mulps         0x168c6(%rip),%xmm8        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,21,206,104,1,0             ; andps         0x168ce(%rip),%xmm10        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,21,246,101,1,0             ; orps          0x165f6(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,5,206,104,1,0              ; addps         0x168ce(%rip),%xmm8        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  68,15,40,37,214,104,1,0             ; movaps        0x168d6(%rip),%xmm12        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,5,166,127,1,0              ; mulps         0x17fa6(%rip),%xmm8        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,21,174,127,1,0             ; andps         0x17fae(%rip),%xmm10        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,21,214,124,1,0             ; orps          0x17cd6(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,5,174,127,1,0              ; addps         0x17fae(%rip),%xmm8        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  68,15,40,37,182,127,1,0             ; movaps        0x17fb6(%rip),%xmm12        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,196                        ; subps         %xmm12,%xmm8
-  DB  68,15,88,21,214,104,1,0             ; addps         0x168d6(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  68,15,40,37,222,104,1,0             ; movaps        0x168de(%rip),%xmm12        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,21,182,127,1,0             ; addps         0x17fb6(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  68,15,40,37,190,127,1,0             ; movaps        0x17fbe(%rip),%xmm12        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,196                        ; subps         %xmm12,%xmm8
   DB  69,15,89,195                        ; mulps         %xmm11,%xmm8
   DB  102,69,15,58,8,208,1                ; roundps       $0x1,%xmm8,%xmm10
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,5,203,104,1,0              ; addps         0x168cb(%rip),%xmm8        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  68,15,40,21,211,104,1,0             ; movaps        0x168d3(%rip),%xmm10        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,5,171,127,1,0              ; addps         0x17fab(%rip),%xmm8        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  68,15,40,21,179,127,1,0             ; movaps        0x17fb3(%rip),%xmm10        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  69,15,89,211                        ; mulps         %xmm11,%xmm10
   DB  69,15,92,194                        ; subps         %xmm10,%xmm8
-  DB  68,15,40,21,211,104,1,0             ; movaps        0x168d3(%rip),%xmm10        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,21,179,127,1,0             ; movaps        0x17fb3(%rip),%xmm10        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  68,15,40,29,215,104,1,0             ; movaps        0x168d7(%rip),%xmm11        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,29,183,127,1,0             ; movaps        0x17fb7(%rip),%xmm11        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  69,15,94,218                        ; divps         %xmm10,%xmm11
   DB  69,15,88,216                        ; addps         %xmm8,%xmm11
-  DB  68,15,89,29,215,104,1,0             ; mulps         0x168d7(%rip),%xmm11        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  68,15,89,29,183,127,1,0             ; mulps         0x17fb7(%rip),%xmm11        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,69,15,91,211                    ; cvtps2dq      %xmm11,%xmm10
   DB  243,68,15,16,64,20                  ; movss         0x14(%rax),%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
@@ -20346,7 +20346,7 @@
   DB  102,69,15,56,20,193                 ; blendvps      %xmm0,%xmm9,%xmm8
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
-  DB  68,15,93,5,94,101,1,0               ; minps         0x1655e(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,93,5,62,124,1,0               ; minps         0x17c3e(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20374,31 +20374,31 @@
   DB  68,15,88,217                        ; addps         %xmm1,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,159,103,1,0             ; mulps         0x1679f(%rip),%xmm12        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,29,167,103,1,0             ; andps         0x167a7(%rip),%xmm11        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,29,207,100,1,0             ; orps          0x164cf(%rip),%xmm11        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,37,167,103,1,0             ; addps         0x167a7(%rip),%xmm12        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  15,40,13,176,103,1,0                ; movaps        0x167b0(%rip),%xmm1        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,37,127,126,1,0             ; mulps         0x17e7f(%rip),%xmm12        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,29,135,126,1,0             ; andps         0x17e87(%rip),%xmm11        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,29,175,123,1,0             ; orps          0x17baf(%rip),%xmm11        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,37,135,126,1,0             ; addps         0x17e87(%rip),%xmm12        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  15,40,13,144,126,1,0                ; movaps        0x17e90(%rip),%xmm1        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,88,29,176,103,1,0             ; addps         0x167b0(%rip),%xmm11        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  15,40,13,185,103,1,0                ; movaps        0x167b9(%rip),%xmm1        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,29,144,126,1,0             ; addps         0x17e90(%rip),%xmm11        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  15,40,13,153,126,1,0                ; movaps        0x17e99(%rip),%xmm1        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  65,15,94,203                        ; divps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,166,103,1,0             ; addps         0x167a6(%rip),%xmm12        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  15,40,13,175,103,1,0                ; movaps        0x167af(%rip),%xmm1        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,37,134,126,1,0             ; addps         0x17e86(%rip),%xmm12        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  15,40,13,143,126,1,0                ; movaps        0x17e8f(%rip),%xmm1        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  68,15,92,225                        ; subps         %xmm1,%xmm12
-  DB  68,15,40,21,175,103,1,0             ; movaps        0x167af(%rip),%xmm10        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,21,143,126,1,0             ; movaps        0x17e8f(%rip),%xmm10        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,13,180,103,1,0                ; movaps        0x167b4(%rip),%xmm1        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  15,40,13,148,126,1,0                ; movaps        0x17e94(%rip),%xmm1        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  65,15,94,202                        ; divps         %xmm10,%xmm1
   DB  65,15,88,204                        ; addps         %xmm12,%xmm1
-  DB  15,89,13,181,103,1,0                ; mulps         0x167b5(%rip),%xmm1        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  15,89,13,149,126,1,0                ; mulps         0x17e95(%rip),%xmm1        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,68,15,91,209                    ; cvtps2dq      %xmm1,%xmm10
   DB  243,15,16,72,20                     ; movss         0x14(%rax),%xmm1
   DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
@@ -20406,7 +20406,7 @@
   DB  102,65,15,56,20,201                 ; blendvps      %xmm0,%xmm9,%xmm1
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,200                           ; maxps         %xmm0,%xmm1
-  DB  15,93,13,64,100,1,0                 ; minps         0x16440(%rip),%xmm1        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,13,32,123,1,0                 ; minps         0x17b20(%rip),%xmm1        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20434,31 +20434,31 @@
   DB  68,15,88,218                        ; addps         %xmm2,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,129,102,1,0             ; mulps         0x16681(%rip),%xmm12        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,29,137,102,1,0             ; andps         0x16689(%rip),%xmm11        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,29,177,99,1,0              ; orps          0x163b1(%rip),%xmm11        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,37,137,102,1,0             ; addps         0x16689(%rip),%xmm12        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  15,40,21,146,102,1,0                ; movaps        0x16692(%rip),%xmm2        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,37,97,125,1,0              ; mulps         0x17d61(%rip),%xmm12        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,29,105,125,1,0             ; andps         0x17d69(%rip),%xmm11        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,29,145,122,1,0             ; orps          0x17a91(%rip),%xmm11        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,37,105,125,1,0             ; addps         0x17d69(%rip),%xmm12        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  15,40,21,114,125,1,0                ; movaps        0x17d72(%rip),%xmm2        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,88,29,146,102,1,0             ; addps         0x16692(%rip),%xmm11        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  15,40,21,155,102,1,0                ; movaps        0x1669b(%rip),%xmm2        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,29,114,125,1,0             ; addps         0x17d72(%rip),%xmm11        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  15,40,21,123,125,1,0                ; movaps        0x17d7b(%rip),%xmm2        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  65,15,94,211                        ; divps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,136,102,1,0             ; addps         0x16688(%rip),%xmm12        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  15,40,21,145,102,1,0                ; movaps        0x16691(%rip),%xmm2        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,37,104,125,1,0             ; addps         0x17d68(%rip),%xmm12        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  15,40,21,113,125,1,0                ; movaps        0x17d71(%rip),%xmm2        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  65,15,89,211                        ; mulps         %xmm11,%xmm2
   DB  68,15,92,226                        ; subps         %xmm2,%xmm12
-  DB  68,15,40,21,145,102,1,0             ; movaps        0x16691(%rip),%xmm10        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,21,113,125,1,0             ; movaps        0x17d71(%rip),%xmm10        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,21,150,102,1,0                ; movaps        0x16696(%rip),%xmm2        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  15,40,21,118,125,1,0                ; movaps        0x17d76(%rip),%xmm2        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  65,15,94,210                        ; divps         %xmm10,%xmm2
   DB  65,15,88,212                        ; addps         %xmm12,%xmm2
-  DB  15,89,21,151,102,1,0                ; mulps         0x16697(%rip),%xmm2        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  15,89,21,119,125,1,0                ; mulps         0x17d77(%rip),%xmm2        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,68,15,91,210                    ; cvtps2dq      %xmm2,%xmm10
   DB  243,15,16,80,20                     ; movss         0x14(%rax),%xmm2
   DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
@@ -20466,7 +20466,7 @@
   DB  102,65,15,56,20,209                 ; blendvps      %xmm0,%xmm9,%xmm2
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,208                           ; maxps         %xmm0,%xmm2
-  DB  15,93,21,34,99,1,0                  ; minps         0x16322(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,21,2,122,1,0                  ; minps         0x17a02(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20494,31 +20494,31 @@
   DB  68,15,88,219                        ; addps         %xmm3,%xmm11
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,91,227                        ; cvtdq2ps      %xmm11,%xmm12
-  DB  68,15,89,37,99,101,1,0              ; mulps         0x16563(%rip),%xmm12        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,29,107,101,1,0             ; andps         0x1656b(%rip),%xmm11        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,29,147,98,1,0              ; orps          0x16293(%rip),%xmm11        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,37,107,101,1,0             ; addps         0x1656b(%rip),%xmm12        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  15,40,29,116,101,1,0                ; movaps        0x16574(%rip),%xmm3        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,37,67,124,1,0              ; mulps         0x17c43(%rip),%xmm12        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,29,75,124,1,0              ; andps         0x17c4b(%rip),%xmm11        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,29,115,121,1,0             ; orps          0x17973(%rip),%xmm11        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,37,75,124,1,0              ; addps         0x17c4b(%rip),%xmm12        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  15,40,29,84,124,1,0                 ; movaps        0x17c54(%rip),%xmm3        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,88,29,116,101,1,0             ; addps         0x16574(%rip),%xmm11        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  15,40,29,125,101,1,0                ; movaps        0x1657d(%rip),%xmm3        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,29,84,124,1,0              ; addps         0x17c54(%rip),%xmm11        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  15,40,29,93,124,1,0                 ; movaps        0x17c5d(%rip),%xmm3        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  65,15,94,219                        ; divps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  102,69,15,58,8,212,1                ; roundps       $0x1,%xmm12,%xmm10
   DB  69,15,40,220                        ; movaps        %xmm12,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  68,15,88,37,106,101,1,0             ; addps         0x1656a(%rip),%xmm12        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  15,40,29,115,101,1,0                ; movaps        0x16573(%rip),%xmm3        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,37,74,124,1,0              ; addps         0x17c4a(%rip),%xmm12        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  15,40,29,83,124,1,0                 ; movaps        0x17c53(%rip),%xmm3        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  65,15,89,219                        ; mulps         %xmm11,%xmm3
   DB  68,15,92,227                        ; subps         %xmm3,%xmm12
-  DB  68,15,40,21,115,101,1,0             ; movaps        0x16573(%rip),%xmm10        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,21,83,124,1,0              ; movaps        0x17c53(%rip),%xmm10        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
-  DB  15,40,29,120,101,1,0                ; movaps        0x16578(%rip),%xmm3        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  15,40,29,88,124,1,0                 ; movaps        0x17c58(%rip),%xmm3        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  65,15,94,218                        ; divps         %xmm10,%xmm3
   DB  65,15,88,220                        ; addps         %xmm12,%xmm3
-  DB  15,89,29,121,101,1,0                ; mulps         0x16579(%rip),%xmm3        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  15,89,29,89,124,1,0                 ; mulps         0x17c59(%rip),%xmm3        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,68,15,91,211                    ; cvtps2dq      %xmm3,%xmm10
   DB  243,15,16,88,20                     ; movss         0x14(%rax),%xmm3
   DB  15,198,219,0                        ; shufps        $0x0,%xmm3,%xmm3
@@ -20526,7 +20526,7 @@
   DB  102,65,15,56,20,217                 ; blendvps      %xmm0,%xmm9,%xmm3
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,95,216                           ; maxps         %xmm0,%xmm3
-  DB  15,93,29,4,98,1,0                   ; minps         0x16204(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,29,228,120,1,0                ; minps         0x178e4(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -20542,20 +20542,20 @@
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,91,194                           ; cvtdq2ps      %xmm2,%xmm0
-  DB  15,89,5,128,100,1,0                 ; mulps         0x16480(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,40,53,136,100,1,0             ; movaps        0x16488(%rip),%xmm14        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
+  DB  15,89,5,96,123,1,0                  ; mulps         0x17b60(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,40,53,104,123,1,0             ; movaps        0x17b68(%rip),%xmm14        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
   DB  65,15,84,214                        ; andps         %xmm14,%xmm2
-  DB  68,15,40,37,172,97,1,0              ; movaps        0x161ac(%rip),%xmm12        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,37,140,120,1,0             ; movaps        0x1788c(%rip),%xmm12        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,86,212                        ; orps          %xmm12,%xmm2
-  DB  68,15,40,21,128,100,1,0             ; movaps        0x16480(%rip),%xmm10        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
+  DB  68,15,40,21,96,123,1,0              ; movaps        0x17b60(%rip),%xmm10        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
   DB  65,15,88,194                        ; addps         %xmm10,%xmm0
-  DB  68,15,40,29,132,100,1,0             ; movaps        0x16484(%rip),%xmm11        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,40,29,100,123,1,0             ; movaps        0x17b64(%rip),%xmm11        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  15,40,226                           ; movaps        %xmm2,%xmm4
   DB  65,15,89,227                        ; mulps         %xmm11,%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  68,15,40,13,130,100,1,0             ; movaps        0x16482(%rip),%xmm9        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
+  DB  68,15,40,13,98,123,1,0              ; movaps        0x17b62(%rip),%xmm9        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
   DB  65,15,88,209                        ; addps         %xmm9,%xmm2
-  DB  68,15,40,45,134,100,1,0             ; movaps        0x16486(%rip),%xmm13        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,40,45,102,123,1,0             ; movaps        0x17b66(%rip),%xmm13        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  65,15,40,229                        ; movaps        %xmm13,%xmm4
   DB  15,94,226                           ; divps         %xmm2,%xmm4
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -20566,21 +20566,21 @@
   DB  102,15,58,8,208,1                   ; roundps       $0x1,%xmm0,%xmm2
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
   DB  15,92,226                           ; subps         %xmm2,%xmm4
-  DB  15,40,53,105,100,1,0                ; movaps        0x16469(%rip),%xmm6        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
+  DB  15,40,53,73,123,1,0                 ; movaps        0x17b49(%rip),%xmm6        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
   DB  15,88,198                           ; addps         %xmm6,%xmm0
-  DB  15,40,61,127,100,1,0                ; movaps        0x1647f(%rip),%xmm7        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  15,40,61,95,123,1,0                 ; movaps        0x17b5f(%rip),%xmm7        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  15,40,239                           ; movaps        %xmm7,%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,40,212                           ; movaps        %xmm4,%xmm2
-  DB  15,40,37,95,100,1,0                 ; movaps        0x1645f(%rip),%xmm4        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  15,40,37,63,123,1,0                 ; movaps        0x17b3f(%rip),%xmm4        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  15,92,194                           ; subps         %xmm2,%xmm0
-  DB  68,15,40,61,113,100,1,0             ; movaps        0x16471(%rip),%xmm15        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,61,81,123,1,0              ; movaps        0x17b51(%rip),%xmm15        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  65,15,40,215                        ; movaps        %xmm15,%xmm2
   DB  15,94,213                           ; divps         %xmm5,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  15,89,5,205,99,1,0                  ; mulps         0x163cd(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
+  DB  15,89,5,173,122,1,0                 ; mulps         0x17aad(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
   DB  65,15,84,206                        ; andps         %xmm14,%xmm1
   DB  65,15,86,204                        ; orps          %xmm12,%xmm1
   DB  65,15,88,194                        ; addps         %xmm10,%xmm0
@@ -20606,8 +20606,8 @@
   DB  15,94,206                           ; divps         %xmm6,%xmm1
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  15,91,195                           ; cvtdq2ps      %xmm3,%xmm0
-  DB  15,89,5,110,99,1,0                  ; mulps         0x1636e(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  15,84,29,119,99,1,0                 ; andps         0x16377(%rip),%xmm3        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
+  DB  15,89,5,78,122,1,0                  ; mulps         0x17a4e(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  15,84,29,87,122,1,0                 ; andps         0x17a57(%rip),%xmm3        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
   DB  65,15,86,220                        ; orps          %xmm12,%xmm3
   DB  65,15,88,194                        ; addps         %xmm10,%xmm0
   DB  68,15,89,219                        ; mulps         %xmm3,%xmm11
@@ -20625,7 +20625,7 @@
   DB  15,92,253                           ; subps         %xmm5,%xmm7
   DB  68,15,94,255                        ; divps         %xmm7,%xmm15
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
-  DB  15,40,5,191,99,1,0                  ; movaps        0x163bf(%rip),%xmm0        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  15,40,5,159,122,1,0                 ; movaps        0x17a9f(%rip),%xmm0        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  68,15,89,248                        ; mulps         %xmm0,%xmm15
@@ -20644,29 +20644,29 @@
 PUBLIC _sk_lab_to_xyz_sse41
 _sk_lab_to_xyz_sse41 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,89,5,140,99,1,0               ; mulps         0x1638c(%rip),%xmm8        # 2c790 <_sk_srcover_rgba_8888_sse2_lowp+0xd88>
-  DB  68,15,40,13,4,97,1,0                ; movaps        0x16104(%rip),%xmm9        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,89,5,108,122,1,0              ; mulps         0x17a6c(%rip),%xmm8        # 2de70 <_sk_srcover_rgba_8888_sse2_lowp+0xdbc>
+  DB  68,15,40,13,228,119,1,0             ; movaps        0x177e4(%rip),%xmm9        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
-  DB  15,40,5,137,99,1,0                  ; movaps        0x16389(%rip),%xmm0        # 2c7a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd98>
+  DB  15,40,5,105,122,1,0                 ; movaps        0x17a69(%rip),%xmm0        # 2de80 <_sk_srcover_rgba_8888_sse2_lowp+0xdcc>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
-  DB  68,15,88,5,135,99,1,0               ; addps         0x16387(%rip),%xmm8        # 2c7b0 <_sk_srcover_rgba_8888_sse2_lowp+0xda8>
-  DB  68,15,89,5,143,99,1,0               ; mulps         0x1638f(%rip),%xmm8        # 2c7c0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb8>
-  DB  15,89,13,152,99,1,0                 ; mulps         0x16398(%rip),%xmm1        # 2c7d0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc8>
+  DB  68,15,88,5,103,122,1,0              ; addps         0x17a67(%rip),%xmm8        # 2de90 <_sk_srcover_rgba_8888_sse2_lowp+0xddc>
+  DB  68,15,89,5,111,122,1,0              ; mulps         0x17a6f(%rip),%xmm8        # 2dea0 <_sk_srcover_rgba_8888_sse2_lowp+0xdec>
+  DB  15,89,13,120,122,1,0                ; mulps         0x17a78(%rip),%xmm1        # 2deb0 <_sk_srcover_rgba_8888_sse2_lowp+0xdfc>
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  15,89,21,157,99,1,0                 ; mulps         0x1639d(%rip),%xmm2        # 2c7e0 <_sk_srcover_rgba_8888_sse2_lowp+0xdd8>
+  DB  15,89,21,125,122,1,0                ; mulps         0x17a7d(%rip),%xmm2        # 2dec0 <_sk_srcover_rgba_8888_sse2_lowp+0xe0c>
   DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
   DB  68,15,92,210                        ; subps         %xmm2,%xmm10
   DB  68,15,40,217                        ; movaps        %xmm1,%xmm11
   DB  69,15,89,219                        ; mulps         %xmm11,%xmm11
   DB  68,15,89,217                        ; mulps         %xmm1,%xmm11
-  DB  68,15,40,13,145,99,1,0              ; movaps        0x16391(%rip),%xmm9        # 2c7f0 <_sk_srcover_rgba_8888_sse2_lowp+0xde8>
+  DB  68,15,40,13,113,122,1,0             ; movaps        0x17a71(%rip),%xmm9        # 2ded0 <_sk_srcover_rgba_8888_sse2_lowp+0xe1c>
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  15,40,21,145,99,1,0                 ; movaps        0x16391(%rip),%xmm2        # 2c800 <_sk_srcover_rgba_8888_sse2_lowp+0xdf8>
+  DB  15,40,21,113,122,1,0                ; movaps        0x17a71(%rip),%xmm2        # 2dee0 <_sk_srcover_rgba_8888_sse2_lowp+0xe2c>
   DB  15,88,202                           ; addps         %xmm2,%xmm1
-  DB  68,15,40,37,150,99,1,0              ; movaps        0x16396(%rip),%xmm12        # 2c810 <_sk_srcover_rgba_8888_sse2_lowp+0xe08>
+  DB  68,15,40,37,118,122,1,0             ; movaps        0x17a76(%rip),%xmm12        # 2def0 <_sk_srcover_rgba_8888_sse2_lowp+0xe3c>
   DB  65,15,89,204                        ; mulps         %xmm12,%xmm1
   DB  102,65,15,56,20,203                 ; blendvps      %xmm0,%xmm11,%xmm1
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
@@ -20685,8 +20685,8 @@
   DB  65,15,89,212                        ; mulps         %xmm12,%xmm2
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  102,65,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm2
-  DB  15,89,13,79,99,1,0                  ; mulps         0x1634f(%rip),%xmm1        # 2c820 <_sk_srcover_rgba_8888_sse2_lowp+0xe18>
-  DB  15,89,21,88,99,1,0                  ; mulps         0x16358(%rip),%xmm2        # 2c830 <_sk_srcover_rgba_8888_sse2_lowp+0xe28>
+  DB  15,89,13,47,122,1,0                 ; mulps         0x17a2f(%rip),%xmm1        # 2df00 <_sk_srcover_rgba_8888_sse2_lowp+0xe4c>
+  DB  15,89,21,56,122,1,0                 ; mulps         0x17a38(%rip),%xmm2        # 2df10 <_sk_srcover_rgba_8888_sse2_lowp+0xe5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  65,15,40,200                        ; movaps        %xmm8,%xmm1
@@ -20703,9 +20703,9 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,39                              ; jne           16522 <_sk_load_a8_sse41+0x3f>
   DB  102,67,15,56,49,4,16                ; pmovzxbd      (%r8,%r10,1),%xmm0
-  DB  102,15,219,5,214,95,1,0             ; pand          0x15fd6(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,5,182,118,1,0            ; pand          0x176b6(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,44,97,1,0                  ; mulps         0x1612c(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,12,120,1,0                 ; mulps         0x1780c(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20743,9 +20743,9 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,39                              ; jne           165ae <_sk_load_a8_dst_sse41+0x3f>
   DB  102,67,15,56,49,36,16               ; pmovzxbd      (%r8,%r10,1),%xmm4
-  DB  102,15,219,37,74,95,1,0             ; pand          0x15f4a(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,37,42,118,1,0            ; pand          0x1762a(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,252                           ; cvtdq2ps      %xmm4,%xmm7
-  DB  15,89,61,160,96,1,0                 ; mulps         0x160a0(%rip),%xmm7        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,61,128,119,1,0                ; mulps         0x17780(%rip),%xmm7        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
@@ -20811,7 +20811,7 @@
   DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,152,95,1,0                 ; mulps         0x15f98(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,120,118,1,0                ; mulps         0x17678(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -20827,7 +20827,7 @@
   DB  77,15,175,193                       ; imul          %r9,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,210                           ; movslq        %edx,%r10
-  DB  68,15,40,5,59,94,1,0                ; movaps        0x15e3b(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,27,117,1,0               ; movaps        0x1751b(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
   DB  102,68,15,56,43,192                 ; packusdw      %xmm0,%xmm8
@@ -20849,7 +20849,7 @@
   DB  65,128,249,3                        ; cmp           $0x3,%r9b
   DB  117,217                             ; jne           166f7 <_sk_store_a8_sse41+0x41>
   DB  102,71,15,58,20,68,16,2,8           ; pextrb        $0x8,%xmm8,0x2(%r8,%r10,1)
-  DB  102,68,15,56,0,5,15,97,1,0          ; pshufb        0x1610f(%rip),%xmm8        # 2c840 <_sk_srcover_rgba_8888_sse2_lowp+0xe38>
+  DB  102,68,15,56,0,5,239,119,1,0        ; pshufb        0x177ef(%rip),%xmm8        # 2df20 <_sk_srcover_rgba_8888_sse2_lowp+0xe6c>
   DB  102,71,15,58,21,4,16,0              ; pextrw        $0x0,%xmm8,(%r8,%r10,1)
   DB  235,188                             ; jmp           166f7 <_sk_store_a8_sse41+0x41>
   DB  102,71,15,58,20,4,16,0              ; pextrb        $0x0,%xmm8,(%r8,%r10,1)
@@ -20866,11 +20866,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,42                              ; jne           16787 <_sk_load_g8_sse41+0x42>
   DB  102,67,15,56,49,4,16                ; pmovzxbd      (%r8,%r10,1),%xmm0
-  DB  102,15,219,5,116,93,1,0             ; pand          0x15d74(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,5,84,116,1,0             ; pand          0x17454(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,202,94,1,0                  ; mulps         0x15eca(%rip),%xmm0        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,5,170,117,1,0                 ; mulps         0x175aa(%rip),%xmm0        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,177,92,1,0                 ; movaps        0x15cb1(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,145,115,1,0                ; movaps        0x17391(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -20906,11 +20906,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,42                              ; jne           16816 <_sk_load_g8_dst_sse41+0x42>
   DB  102,67,15,56,49,36,16               ; pmovzxbd      (%r8,%r10,1),%xmm4
-  DB  102,15,219,37,229,92,1,0            ; pand          0x15ce5(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,37,197,115,1,0           ; pand          0x173c5(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,59,94,1,0                  ; mulps         0x15e3b(%rip),%xmm4        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,37,27,117,1,0                 ; mulps         0x1751b(%rip),%xmm4        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,34,92,1,0                  ; movaps        0x15c22(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,61,2,115,1,0                  ; movaps        0x17302(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  255,224                             ; jmpq          *%rax
@@ -20974,9 +20974,9 @@
   DB  102,15,58,32,192,3                  ; pinsrb        $0x3,%eax,%xmm0
   DB  102,15,56,49,192                    ; pmovzxbd      %xmm0,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,48,93,1,0                   ; mulps         0x15d30(%rip),%xmm0        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,5,16,116,1,0                  ; mulps         0x17410(%rip),%xmm0        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,23,91,1,0                  ; movaps        0x15b17(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,247,113,1,0                ; movaps        0x171f7(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -20993,19 +20993,19 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,80                              ; jne           1698c <_sk_load_565_sse41+0x6b>
   DB  102,67,15,56,51,20,80               ; pmovzxwd      (%r8,%r10,2),%xmm2
-  DB  102,15,111,5,5,93,1,0               ; movdqa        0x15d05(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,229,115,1,0            ; movdqa        0x173e5(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,7,93,1,0                    ; mulps         0x15d07(%rip),%xmm0        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,13,15,93,1,0             ; movdqa        0x15d0f(%rip),%xmm1        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,5,231,115,1,0                 ; mulps         0x173e7(%rip),%xmm0        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,13,239,115,1,0           ; movdqa        0x173ef(%rip),%xmm1        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,17,93,1,0                  ; mulps         0x15d11(%rip),%xmm1        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,21,25,93,1,0             ; pand          0x15d19(%rip),%xmm2        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,13,241,115,1,0                ; mulps         0x173f1(%rip),%xmm1        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,21,249,115,1,0           ; pand          0x173f9(%rip),%xmm2        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,31,93,1,0                  ; mulps         0x15d1f(%rip),%xmm2        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,21,255,115,1,0                ; mulps         0x173ff(%rip),%xmm2        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,166,90,1,0                 ; movaps        0x15aa6(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,134,113,1,0                ; movaps        0x17186(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,3                        ; and           $0x3,%r9b
@@ -21039,19 +21039,19 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,80                              ; jne           16a47 <_sk_load_565_dst_sse41+0x6b>
   DB  102,67,15,56,51,52,80               ; pmovzxwd      (%r8,%r10,2),%xmm6
-  DB  102,15,111,37,74,92,1,0             ; movdqa        0x15c4a(%rip),%xmm4        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,37,42,115,1,0            ; movdqa        0x1732a(%rip),%xmm4        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,76,92,1,0                  ; mulps         0x15c4c(%rip),%xmm4        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,45,84,92,1,0             ; movdqa        0x15c54(%rip),%xmm5        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,37,44,115,1,0                 ; mulps         0x1732c(%rip),%xmm4        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,45,52,115,1,0            ; movdqa        0x17334(%rip),%xmm5        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,238                      ; pand          %xmm6,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,45,86,92,1,0                  ; mulps         0x15c56(%rip),%xmm5        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,53,94,92,1,0             ; pand          0x15c5e(%rip),%xmm6        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,45,54,115,1,0                 ; mulps         0x17336(%rip),%xmm5        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,53,62,115,1,0            ; pand          0x1733e(%rip),%xmm6        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,53,100,92,1,0                 ; mulps         0x15c64(%rip),%xmm6        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,53,68,115,1,0                 ; mulps         0x17344(%rip),%xmm6        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,235,89,1,0                 ; movaps        0x159eb(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,61,203,112,1,0                ; movaps        0x170cb(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,3                        ; and           $0x3,%r9b
@@ -21111,19 +21111,19 @@
   DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
   DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
   DB  102,15,56,51,208                    ; pmovzxwd      %xmm0,%xmm2
-  DB  102,15,111,5,17,91,1,0              ; movdqa        0x15b11(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,241,113,1,0            ; movdqa        0x171f1(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,19,91,1,0                   ; mulps         0x15b13(%rip),%xmm0        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,13,27,91,1,0             ; movdqa        0x15b1b(%rip),%xmm1        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,5,243,113,1,0                 ; mulps         0x171f3(%rip),%xmm0        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,13,251,113,1,0           ; movdqa        0x171fb(%rip),%xmm1        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,29,91,1,0                  ; mulps         0x15b1d(%rip),%xmm1        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,21,37,91,1,0             ; pand          0x15b25(%rip),%xmm2        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,13,253,113,1,0                ; mulps         0x171fd(%rip),%xmm1        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,21,5,114,1,0             ; pand          0x17205(%rip),%xmm2        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,43,91,1,0                  ; mulps         0x15b2b(%rip),%xmm2        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,21,11,114,1,0                 ; mulps         0x1720b(%rip),%xmm2        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,178,88,1,0                 ; movaps        0x158b2(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,146,111,1,0                ; movaps        0x16f92(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_565_sse41
@@ -21135,12 +21135,12 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,178,92,1,0               ; movaps        0x15cb2(%rip),%xmm8        # 2c850 <_sk_srcover_rgba_8888_sse2_lowp+0xe48>
+  DB  68,15,40,5,146,115,1,0              ; movaps        0x17392(%rip),%xmm8        # 2df30 <_sk_srcover_rgba_8888_sse2_lowp+0xe7c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
   DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
-  DB  68,15,40,21,167,92,1,0              ; movaps        0x15ca7(%rip),%xmm10        # 2c860 <_sk_srcover_rgba_8888_sse2_lowp+0xe58>
+  DB  68,15,40,21,135,115,1,0             ; movaps        0x17387(%rip),%xmm10        # 2df40 <_sk_srcover_rgba_8888_sse2_lowp+0xe8c>
   DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
   DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
   DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
@@ -21182,21 +21182,21 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,95                              ; jne           16caa <_sk_load_4444_sse41+0x7a>
   DB  102,67,15,56,51,28,80               ; pmovzxwd      (%r8,%r10,2),%xmm3
-  DB  102,15,111,5,22,92,1,0              ; movdqa        0x15c16(%rip),%xmm0        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,5,246,114,1,0            ; movdqa        0x172f6(%rip),%xmm0        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,24,92,1,0                   ; mulps         0x15c18(%rip),%xmm0        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,13,32,92,1,0             ; movdqa        0x15c20(%rip),%xmm1        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,5,248,114,1,0                 ; mulps         0x172f8(%rip),%xmm0        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,13,0,115,1,0             ; movdqa        0x17300(%rip),%xmm1        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,34,92,1,0                  ; mulps         0x15c22(%rip),%xmm1        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,21,42,92,1,0             ; movdqa        0x15c2a(%rip),%xmm2        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,13,2,115,1,0                  ; mulps         0x17302(%rip),%xmm1        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,21,10,115,1,0            ; movdqa        0x1730a(%rip),%xmm2        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,44,92,1,0                  ; mulps         0x15c2c(%rip),%xmm2        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,29,52,92,1,0             ; pand          0x15c34(%rip),%xmm3        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,21,12,115,1,0                 ; mulps         0x1730c(%rip),%xmm2        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,29,20,115,1,0            ; pand          0x17314(%rip),%xmm3        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,58,92,1,0                  ; mulps         0x15c3a(%rip),%xmm3        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,29,26,115,1,0                 ; mulps         0x1731a(%rip),%xmm3        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
@@ -21231,21 +21231,21 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,95                              ; jne           16d74 <_sk_load_4444_dst_sse41+0x7a>
   DB  102,67,15,56,51,60,80               ; pmovzxwd      (%r8,%r10,2),%xmm7
-  DB  102,15,111,37,76,91,1,0             ; movdqa        0x15b4c(%rip),%xmm4        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,37,44,114,1,0            ; movdqa        0x1722c(%rip),%xmm4        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,231                      ; pand          %xmm7,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,78,91,1,0                  ; mulps         0x15b4e(%rip),%xmm4        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,45,86,91,1,0             ; movdqa        0x15b56(%rip),%xmm5        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,37,46,114,1,0                 ; mulps         0x1722e(%rip),%xmm4        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,45,54,114,1,0            ; movdqa        0x17236(%rip),%xmm5        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,239                      ; pand          %xmm7,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,45,88,91,1,0                  ; mulps         0x15b58(%rip),%xmm5        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,53,96,91,1,0             ; movdqa        0x15b60(%rip),%xmm6        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,45,56,114,1,0                 ; mulps         0x17238(%rip),%xmm5        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,53,64,114,1,0            ; movdqa        0x17240(%rip),%xmm6        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,247                      ; pand          %xmm7,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,53,98,91,1,0                  ; mulps         0x15b62(%rip),%xmm6        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,61,106,91,1,0            ; pand          0x15b6a(%rip),%xmm7        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,53,66,114,1,0                 ; mulps         0x17242(%rip),%xmm6        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,61,74,114,1,0            ; pand          0x1724a(%rip),%xmm7        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,255                           ; cvtdq2ps      %xmm7,%xmm7
-  DB  15,89,61,112,91,1,0                 ; mulps         0x15b70(%rip),%xmm7        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,61,80,114,1,0                 ; mulps         0x17250(%rip),%xmm7        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
@@ -21306,21 +21306,21 @@
   DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
   DB  102,15,196,192,3                    ; pinsrw        $0x3,%eax,%xmm0
   DB  102,15,56,51,216                    ; pmovzxwd      %xmm0,%xmm3
-  DB  102,15,111,5,4,90,1,0               ; movdqa        0x15a04(%rip),%xmm0        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,5,228,112,1,0            ; movdqa        0x170e4(%rip),%xmm0        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,6,90,1,0                    ; mulps         0x15a06(%rip),%xmm0        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,13,14,90,1,0             ; movdqa        0x15a0e(%rip),%xmm1        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,5,230,112,1,0                 ; mulps         0x170e6(%rip),%xmm0        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,13,238,112,1,0           ; movdqa        0x170ee(%rip),%xmm1        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,16,90,1,0                  ; mulps         0x15a10(%rip),%xmm1        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,21,24,90,1,0             ; movdqa        0x15a18(%rip),%xmm2        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,13,240,112,1,0                ; mulps         0x170f0(%rip),%xmm1        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,21,248,112,1,0           ; movdqa        0x170f8(%rip),%xmm2        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,26,90,1,0                  ; mulps         0x15a1a(%rip),%xmm2        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,29,34,90,1,0             ; pand          0x15a22(%rip),%xmm3        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,21,250,112,1,0                ; mulps         0x170fa(%rip),%xmm2        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,29,2,113,1,0             ; pand          0x17102(%rip),%xmm3        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,40,90,1,0                  ; mulps         0x15a28(%rip),%xmm3        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,29,8,113,1,0                  ; mulps         0x17108(%rip),%xmm3        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -21333,7 +21333,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,22,90,1,0                ; movaps        0x15a16(%rip),%xmm8        # 2c8f0 <_sk_srcover_rgba_8888_sse2_lowp+0xee8>
+  DB  68,15,40,5,246,112,1,0              ; movaps        0x170f6(%rip),%xmm8        # 2dfd0 <_sk_srcover_rgba_8888_sse2_lowp+0xf1c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -21385,17 +21385,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,89                              ; jne           16ff5 <_sk_load_8888_sse41+0x75>
   DB  243,65,15,111,28,128                ; movdqu        (%r8,%rax,4),%xmm3
-  DB  102,15,111,5,54,85,1,0              ; movdqa        0x15536(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,22,108,1,0             ; movdqa        0x16c16(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,135,86,1,0               ; movaps        0x15687(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,103,109,1,0              ; movaps        0x16d67(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,38,85,1,0            ; pshufb        0x15526(%rip),%xmm1        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,13,6,108,1,0            ; pshufb        0x16c06(%rip),%xmm1        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,34,85,1,0            ; pshufb        0x15522(%rip),%xmm2        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,21,2,108,1,0            ; pshufb        0x16c02(%rip),%xmm2        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -21432,17 +21432,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,89                              ; jne           170af <_sk_load_8888_dst_sse41+0x75>
   DB  243,65,15,111,60,128                ; movdqu        (%r8,%rax,4),%xmm7
-  DB  102,15,111,37,124,84,1,0            ; movdqa        0x1547c(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,37,92,107,1,0            ; movdqa        0x16b5c(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,231                      ; pand          %xmm7,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  68,15,40,5,205,85,1,0               ; movaps        0x155cd(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,173,108,1,0              ; movaps        0x16cad(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,224                        ; mulps         %xmm8,%xmm4
   DB  102,15,111,239                      ; movdqa        %xmm7,%xmm5
-  DB  102,15,56,0,45,108,84,1,0           ; pshufb        0x1546c(%rip),%xmm5        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,45,76,107,1,0           ; pshufb        0x16b4c(%rip),%xmm5        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
   DB  65,15,89,232                        ; mulps         %xmm8,%xmm5
   DB  102,15,111,247                      ; movdqa        %xmm7,%xmm6
-  DB  102,15,56,0,53,104,84,1,0           ; pshufb        0x15468(%rip),%xmm6        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,53,72,107,1,0           ; pshufb        0x16b48(%rip),%xmm6        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  65,15,89,240                        ; mulps         %xmm8,%xmm6
   DB  102,15,114,215,24                   ; psrld         $0x18,%xmm7
@@ -21500,17 +21500,17 @@
   DB  102,65,15,58,34,28,128,1            ; pinsrd        $0x1,(%r8,%rax,4),%xmm3
   DB  102,67,15,58,34,28,152,2            ; pinsrd        $0x2,(%r8,%r11,4),%xmm3
   DB  102,67,15,58,34,28,144,3            ; pinsrd        $0x3,(%r8,%r10,4),%xmm3
-  DB  102,15,111,5,85,83,1,0              ; movdqa        0x15355(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,53,106,1,0             ; movdqa        0x16a35(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,166,84,1,0               ; movaps        0x154a6(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,134,107,1,0              ; movaps        0x16b86(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
-  DB  102,15,56,0,13,69,83,1,0            ; pshufb        0x15345(%rip),%xmm1        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,13,37,106,1,0           ; pshufb        0x16a25(%rip),%xmm1        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,211                      ; movdqa        %xmm3,%xmm2
-  DB  102,15,56,0,21,65,83,1,0            ; pshufb        0x15341(%rip),%xmm2        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,21,33,106,1,0           ; pshufb        0x16a21(%rip),%xmm2        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -21528,7 +21528,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,27,83,1,0                ; movaps        0x1531b(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,251,105,1,0              ; movaps        0x169fb(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -21577,17 +21577,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,89                              ; jne           172fc <_sk_load_bgra_sse41+0x75>
   DB  243,65,15,111,28,128                ; movdqu        (%r8,%rax,4),%xmm3
-  DB  102,15,111,5,47,82,1,0              ; movdqa        0x1522f(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,15,105,1,0             ; movdqa        0x1690f(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,208                           ; cvtdq2ps      %xmm0,%xmm2
-  DB  68,15,40,5,128,83,1,0               ; movaps        0x15380(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,96,106,1,0               ; movaps        0x16a60(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,15,56,0,5,31,82,1,0             ; pshufb        0x1521f(%rip),%xmm0        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,5,255,104,1,0           ; pshufb        0x168ff(%rip),%xmm0        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,15,56,0,5,27,82,1,0             ; pshufb        0x1521b(%rip),%xmm0        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,5,251,104,1,0           ; pshufb        0x168fb(%rip),%xmm0        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -21624,17 +21624,17 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,89                              ; jne           173b6 <_sk_load_bgra_dst_sse41+0x75>
   DB  243,65,15,111,60,128                ; movdqu        (%r8,%rax,4),%xmm7
-  DB  102,15,111,37,117,81,1,0            ; movdqa        0x15175(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,37,85,104,1,0            ; movdqa        0x16855(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,231                      ; pand          %xmm7,%xmm4
   DB  15,91,244                           ; cvtdq2ps      %xmm4,%xmm6
-  DB  68,15,40,5,198,82,1,0               ; movaps        0x152c6(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,166,105,1,0              ; movaps        0x169a6(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,240                        ; mulps         %xmm8,%xmm6
   DB  102,15,111,231                      ; movdqa        %xmm7,%xmm4
-  DB  102,15,56,0,37,101,81,1,0           ; pshufb        0x15165(%rip),%xmm4        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,37,69,104,1,0           ; pshufb        0x16845(%rip),%xmm4        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,236                           ; cvtdq2ps      %xmm4,%xmm5
   DB  65,15,89,232                        ; mulps         %xmm8,%xmm5
   DB  102,15,111,231                      ; movdqa        %xmm7,%xmm4
-  DB  102,15,56,0,37,97,81,1,0            ; pshufb        0x15161(%rip),%xmm4        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,37,65,104,1,0           ; pshufb        0x16841(%rip),%xmm4        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
   DB  65,15,89,224                        ; mulps         %xmm8,%xmm4
   DB  102,15,114,215,24                   ; psrld         $0x18,%xmm7
@@ -21692,17 +21692,17 @@
   DB  102,65,15,58,34,28,128,1            ; pinsrd        $0x1,(%r8,%rax,4),%xmm3
   DB  102,67,15,58,34,28,152,2            ; pinsrd        $0x2,(%r8,%r11,4),%xmm3
   DB  102,67,15,58,34,28,144,3            ; pinsrd        $0x3,(%r8,%r10,4),%xmm3
-  DB  102,15,111,5,78,80,1,0              ; movdqa        0x1504e(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,46,103,1,0             ; movdqa        0x1672e(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,208                           ; cvtdq2ps      %xmm0,%xmm2
-  DB  68,15,40,5,159,81,1,0               ; movaps        0x1519f(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,127,104,1,0              ; movaps        0x1687f(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,15,56,0,5,62,80,1,0             ; pshufb        0x1503e(%rip),%xmm0        # 2c4f0 <_sk_srcover_rgba_8888_sse2_lowp+0xae8>
+  DB  102,15,56,0,5,30,103,1,0            ; pshufb        0x1671e(%rip),%xmm0        # 2dbd0 <_sk_srcover_rgba_8888_sse2_lowp+0xb1c>
   DB  15,91,200                           ; cvtdq2ps      %xmm0,%xmm1
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,15,56,0,5,58,80,1,0             ; pshufb        0x1503a(%rip),%xmm0        # 2c500 <_sk_srcover_rgba_8888_sse2_lowp+0xaf8>
+  DB  102,15,56,0,5,26,103,1,0            ; pshufb        0x1671a(%rip),%xmm0        # 2dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0xb2c>
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
@@ -21720,7 +21720,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,20,80,1,0                ; movaps        0x15014(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,244,102,1,0              ; movaps        0x166f4(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,40,202                        ; movaps        %xmm2,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -21777,19 +21777,19 @@
   DB  102,68,15,97,224                    ; punpcklwd     %xmm0,%xmm12
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
   DB  102,65,15,56,51,212                 ; pmovzxwd      %xmm12,%xmm2
-  DB  102,68,15,111,5,25,83,1,0           ; movdqa        0x15319(%rip),%xmm8        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,5,249,105,1,0         ; movdqa        0x169f9(%rip),%xmm8        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,219,200                   ; pand          %xmm8,%xmm1
-  DB  102,68,15,111,21,23,83,1,0          ; movdqa        0x15317(%rip),%xmm10        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,21,247,105,1,0        ; movdqa        0x169f7(%rip),%xmm10        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,210                   ; pand          %xmm10,%xmm2
-  DB  102,15,111,29,26,83,1,0             ; movdqa        0x1531a(%rip),%xmm3        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,29,250,105,1,0           ; movdqa        0x169fa(%rip),%xmm3        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
   DB  102,15,56,63,195                    ; pmaxud        %xmm3,%xmm0
   DB  102,15,118,194                      ; pcmpeqd       %xmm2,%xmm0
   DB  102,15,114,242,13                   ; pslld         $0xd,%xmm2
   DB  102,15,235,209                      ; por           %xmm1,%xmm2
-  DB  102,68,15,111,29,6,83,1,0           ; movdqa        0x15306(%rip),%xmm11        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,29,230,105,1,0        ; movdqa        0x169e6(%rip),%xmm11        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  102,65,15,112,204,238               ; pshufd        $0xee,%xmm12,%xmm1
@@ -21863,19 +21863,19 @@
   DB  102,68,15,97,228                    ; punpcklwd     %xmm4,%xmm12
   DB  102,68,15,105,204                   ; punpckhwd     %xmm4,%xmm9
   DB  102,65,15,56,51,244                 ; pmovzxwd      %xmm12,%xmm6
-  DB  102,68,15,111,5,123,81,1,0          ; movdqa        0x1517b(%rip),%xmm8        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,5,91,104,1,0          ; movdqa        0x1685b(%rip),%xmm8        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,15,111,238                      ; movdqa        %xmm6,%xmm5
   DB  102,65,15,219,232                   ; pand          %xmm8,%xmm5
-  DB  102,68,15,111,21,121,81,1,0         ; movdqa        0x15179(%rip),%xmm10        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,21,89,104,1,0         ; movdqa        0x16859(%rip),%xmm10        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,242                   ; pand          %xmm10,%xmm6
-  DB  102,15,111,61,124,81,1,0            ; movdqa        0x1517c(%rip),%xmm7        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,61,92,104,1,0            ; movdqa        0x1685c(%rip),%xmm7        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,15,114,245,16                   ; pslld         $0x10,%xmm5
   DB  102,15,111,230                      ; movdqa        %xmm6,%xmm4
   DB  102,15,56,63,231                    ; pmaxud        %xmm7,%xmm4
   DB  102,15,118,230                      ; pcmpeqd       %xmm6,%xmm4
   DB  102,15,114,246,13                   ; pslld         $0xd,%xmm6
   DB  102,15,235,245                      ; por           %xmm5,%xmm6
-  DB  102,68,15,111,29,104,81,1,0         ; movdqa        0x15168(%rip),%xmm11        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,29,72,104,1,0         ; movdqa        0x16848(%rip),%xmm11        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,243                   ; paddd         %xmm11,%xmm6
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  102,65,15,112,236,238               ; pshufd        $0xee,%xmm12,%xmm5
@@ -21971,19 +21971,19 @@
   DB  102,68,15,97,226                    ; punpcklwd     %xmm2,%xmm12
   DB  102,68,15,105,202                   ; punpckhwd     %xmm2,%xmm9
   DB  102,65,15,56,51,212                 ; pmovzxwd      %xmm12,%xmm2
-  DB  102,68,15,111,5,121,79,1,0          ; movdqa        0x14f79(%rip),%xmm8        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,5,89,102,1,0          ; movdqa        0x16659(%rip),%xmm8        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,219,200                   ; pand          %xmm8,%xmm1
-  DB  102,68,15,111,21,119,79,1,0         ; movdqa        0x14f77(%rip),%xmm10        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,21,87,102,1,0         ; movdqa        0x16657(%rip),%xmm10        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,210                   ; pand          %xmm10,%xmm2
-  DB  102,15,111,29,122,79,1,0            ; movdqa        0x14f7a(%rip),%xmm3        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,29,90,102,1,0            ; movdqa        0x1665a(%rip),%xmm3        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
   DB  102,15,56,63,195                    ; pmaxud        %xmm3,%xmm0
   DB  102,15,118,194                      ; pcmpeqd       %xmm2,%xmm0
   DB  102,15,114,242,13                   ; pslld         $0xd,%xmm2
   DB  102,15,235,209                      ; por           %xmm1,%xmm2
-  DB  102,68,15,111,29,102,79,1,0         ; movdqa        0x14f66(%rip),%xmm11        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,29,70,102,1,0         ; movdqa        0x16646(%rip),%xmm11        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,211                   ; paddd         %xmm11,%xmm2
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  102,65,15,112,204,238               ; pshufd        $0xee,%xmm12,%xmm1
@@ -22034,20 +22034,20 @@
   DB  73,193,224,3                        ; shl           $0x3,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  102,68,15,111,13,142,78,1,0         ; movdqa        0x14e8e(%rip),%xmm9        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
+  DB  102,68,15,111,13,110,101,1,0        ; movdqa        0x1656e(%rip),%xmm9        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
   DB  102,68,15,111,224                   ; movdqa        %xmm0,%xmm12
   DB  102,69,15,219,225                   ; pand          %xmm9,%xmm12
-  DB  102,68,15,111,29,139,78,1,0         ; movdqa        0x14e8b(%rip),%xmm11        # 2c950 <_sk_srcover_rgba_8888_sse2_lowp+0xf48>
+  DB  102,68,15,111,29,107,101,1,0        ; movdqa        0x1656b(%rip),%xmm11        # 2e030 <_sk_srcover_rgba_8888_sse2_lowp+0xf7c>
   DB  102,68,15,111,232                   ; movdqa        %xmm0,%xmm13
   DB  102,69,15,219,235                   ; pand          %xmm11,%xmm13
-  DB  102,68,15,111,21,136,78,1,0         ; movdqa        0x14e88(%rip),%xmm10        # 2c960 <_sk_srcover_rgba_8888_sse2_lowp+0xf58>
+  DB  102,68,15,111,21,104,101,1,0        ; movdqa        0x16568(%rip),%xmm10        # 2e040 <_sk_srcover_rgba_8888_sse2_lowp+0xf8c>
   DB  102,65,15,114,212,16                ; psrld         $0x10,%xmm12
   DB  102,69,15,111,197                   ; movdqa        %xmm13,%xmm8
   DB  102,69,15,56,63,194                 ; pmaxud        %xmm10,%xmm8
   DB  102,69,15,118,197                   ; pcmpeqd       %xmm13,%xmm8
   DB  102,65,15,114,213,13                ; psrld         $0xd,%xmm13
   DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
-  DB  102,68,15,111,37,110,78,1,0         ; movdqa        0x14e6e(%rip),%xmm12        # 2c970 <_sk_srcover_rgba_8888_sse2_lowp+0xf68>
+  DB  102,68,15,111,37,78,101,1,0         ; movdqa        0x1654e(%rip),%xmm12        # 2e050 <_sk_srcover_rgba_8888_sse2_lowp+0xf9c>
   DB  102,69,15,254,236                   ; paddd         %xmm12,%xmm13
   DB  102,69,15,219,197                   ; pand          %xmm13,%xmm8
   DB  102,68,15,56,43,192                 ; packusdw      %xmm0,%xmm8
@@ -22135,7 +22135,7 @@
   DB  102,15,235,200                      ; por           %xmm0,%xmm1
   DB  102,15,56,51,193                    ; pmovzxwd      %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,54,74,1,0                ; movaps        0x14a36(%rip),%xmm8        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  68,15,40,5,22,97,1,0                ; movaps        0x16116(%rip),%xmm8        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -22207,7 +22207,7 @@
   DB  102,15,235,200                      ; por           %xmm0,%xmm1
   DB  102,15,56,51,193                    ; pmovzxwd      %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,5,73,1,0                 ; movaps        0x14905(%rip),%xmm8        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  68,15,40,5,229,95,1,0               ; movaps        0x15fe5(%rip),%xmm8        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -22224,7 +22224,7 @@
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,28,70,1,0                  ; movaps        0x1461c(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,252,92,1,0                 ; movaps        0x15cfc(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,20,64                 ; movd          (%r8,%rax,2),%xmm2
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -22256,7 +22256,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  73,99,193                           ; movslq        %r9d,%rax
-  DB  68,15,40,21,229,74,1,0              ; movaps        0x14ae5(%rip),%xmm10        # 2c980 <_sk_srcover_rgba_8888_sse2_lowp+0xf78>
+  DB  68,15,40,21,197,97,1,0              ; movaps        0x161c5(%rip),%xmm10        # 2e060 <_sk_srcover_rgba_8888_sse2_lowp+0xfac>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
@@ -22475,7 +22475,7 @@
   DB  65,15,92,194                        ; subps         %xmm10,%xmm0
   DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
-  DB  243,68,15,89,13,215,58,1,0          ; mulss         0x13ad7(%rip),%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,13,131,81,1,0          ; mulss         0x15183(%rip),%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
   DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
@@ -22498,7 +22498,7 @@
   DB  65,15,92,202                        ; subps         %xmm10,%xmm1
   DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
-  DB  243,68,15,89,13,126,58,1,0          ; mulss         0x13a7e(%rip),%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,13,42,81,1,0           ; mulss         0x1512a(%rip),%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
   DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
@@ -22515,7 +22515,7 @@
 _sk_clamp_x_1_sse41 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
   DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
-  DB  68,15,93,5,190,65,1,0               ; minps         0x141be(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,93,5,158,88,1,0               ; minps         0x1589e(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -22526,16 +22526,16 @@
   DB  65,15,92,192                        ; subps         %xmm8,%xmm0
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
   DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
-  DB  68,15,93,5,155,65,1,0               ; minps         0x1419b(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,93,5,123,88,1,0               ; minps         0x1587b(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_mirror_x_1_sse41
 _sk_mirror_x_1_sse41 LABEL PROC
-  DB  68,15,40,5,235,65,1,0               ; movaps        0x141eb(%rip),%xmm8        # 2c490 <_sk_srcover_rgba_8888_sse2_lowp+0xa88>
+  DB  68,15,40,5,203,88,1,0               ; movaps        0x158cb(%rip),%xmm8        # 2db70 <_sk_srcover_rgba_8888_sse2_lowp+0xabc>
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,13,111,65,1,0              ; movaps        0x1416f(%rip),%xmm9        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,13,79,88,1,0               ; movaps        0x1584f(%rip),%xmm9        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
   DB  102,69,15,58,8,201,1                ; roundps       $0x1,%xmm9,%xmm9
   DB  69,15,88,201                        ; addps         %xmm9,%xmm9
@@ -22546,7 +22546,7 @@
   DB  68,15,92,200                        ; subps         %xmm0,%xmm9
   DB  68,15,84,200                        ; andps         %xmm0,%xmm9
   DB  69,15,95,193                        ; maxps         %xmm9,%xmm8
-  DB  68,15,93,5,76,65,1,0                ; minps         0x1414c(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,93,5,44,88,1,0                ; minps         0x1582c(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -22554,10 +22554,10 @@
 PUBLIC _sk_luminance_to_alpha_sse41
 _sk_luminance_to_alpha_sse41 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,154,70,1,0                  ; mulps         0x1469a(%rip),%xmm0        # 2c990 <_sk_srcover_rgba_8888_sse2_lowp+0xf88>
-  DB  15,89,13,163,70,1,0                 ; mulps         0x146a3(%rip),%xmm1        # 2c9a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf98>
+  DB  15,89,5,122,93,1,0                  ; mulps         0x15d7a(%rip),%xmm0        # 2e070 <_sk_srcover_rgba_8888_sse2_lowp+0xfbc>
+  DB  15,89,13,131,93,1,0                 ; mulps         0x15d83(%rip),%xmm1        # 2e080 <_sk_srcover_rgba_8888_sse2_lowp+0xfcc>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,169,70,1,0                 ; mulps         0x146a9(%rip),%xmm3        # 2c9b0 <_sk_srcover_rgba_8888_sse2_lowp+0xfa8>
+  DB  15,89,29,137,93,1,0                 ; mulps         0x15d89(%rip),%xmm3        # 2e090 <_sk_srcover_rgba_8888_sse2_lowp+0xfdc>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -22942,15 +22942,15 @@
 
 PUBLIC _sk_gauss_a_to_rgba_sse41
 _sk_gauss_a_to_rgba_sse41 LABEL PROC
-  DB  15,40,5,88,64,1,0                   ; movaps        0x14058(%rip),%xmm0        # 2c9c0 <_sk_srcover_rgba_8888_sse2_lowp+0xfb8>
+  DB  15,40,5,56,87,1,0                   ; movaps        0x15738(%rip),%xmm0        # 2e0a0 <_sk_srcover_rgba_8888_sse2_lowp+0xfec>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,94,64,1,0                   ; addps         0x1405e(%rip),%xmm0        # 2c9d0 <_sk_srcover_rgba_8888_sse2_lowp+0xfc8>
+  DB  15,88,5,62,87,1,0                   ; addps         0x1573e(%rip),%xmm0        # 2e0b0 <_sk_srcover_rgba_8888_sse2_lowp+0xffc>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,100,64,1,0                  ; addps         0x14064(%rip),%xmm0        # 2c9e0 <_sk_srcover_rgba_8888_sse2_lowp+0xfd8>
+  DB  15,88,5,68,87,1,0                   ; addps         0x15744(%rip),%xmm0        # 2e0c0 <_sk_srcover_rgba_8888_sse2_lowp+0x100c>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,106,64,1,0                  ; addps         0x1406a(%rip),%xmm0        # 2c9f0 <_sk_srcover_rgba_8888_sse2_lowp+0xfe8>
+  DB  15,88,5,74,87,1,0                   ; addps         0x1574a(%rip),%xmm0        # 2e0d0 <_sk_srcover_rgba_8888_sse2_lowp+0x101c>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,112,64,1,0                  ; addps         0x14070(%rip),%xmm0        # 2ca00 <_sk_srcover_rgba_8888_sse2_lowp+0xff8>
+  DB  15,88,5,80,87,1,0                   ; addps         0x15750(%rip),%xmm0        # 2e0e0 <_sk_srcover_rgba_8888_sse2_lowp+0x102c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
@@ -23101,26 +23101,26 @@
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,40,236                        ; movaps        %xmm12,%xmm13
   DB  69,15,89,237                        ; mulps         %xmm13,%xmm13
-  DB  68,15,40,21,231,61,1,0              ; movaps        0x13de7(%rip),%xmm10        # 2ca10 <_sk_srcover_rgba_8888_sse2_lowp+0x1008>
+  DB  68,15,40,21,199,84,1,0              ; movaps        0x154c7(%rip),%xmm10        # 2e0f0 <_sk_srcover_rgba_8888_sse2_lowp+0x103c>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,235,61,1,0              ; addps         0x13deb(%rip),%xmm10        # 2ca20 <_sk_srcover_rgba_8888_sse2_lowp+0x1018>
+  DB  68,15,88,21,203,84,1,0              ; addps         0x154cb(%rip),%xmm10        # 2e100 <_sk_srcover_rgba_8888_sse2_lowp+0x104c>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,239,61,1,0              ; addps         0x13def(%rip),%xmm10        # 2ca30 <_sk_srcover_rgba_8888_sse2_lowp+0x1028>
+  DB  68,15,88,21,207,84,1,0              ; addps         0x154cf(%rip),%xmm10        # 2e110 <_sk_srcover_rgba_8888_sse2_lowp+0x105c>
   DB  69,15,89,213                        ; mulps         %xmm13,%xmm10
-  DB  68,15,88,21,243,61,1,0              ; addps         0x13df3(%rip),%xmm10        # 2ca40 <_sk_srcover_rgba_8888_sse2_lowp+0x1038>
+  DB  68,15,88,21,211,84,1,0              ; addps         0x154d3(%rip),%xmm10        # 2e120 <_sk_srcover_rgba_8888_sse2_lowp+0x106c>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,194,195,1                     ; cmpltps       %xmm11,%xmm0
-  DB  68,15,40,29,242,61,1,0              ; movaps        0x13df2(%rip),%xmm11        # 2ca50 <_sk_srcover_rgba_8888_sse2_lowp+0x1048>
+  DB  68,15,40,29,210,84,1,0              ; movaps        0x154d2(%rip),%xmm11        # 2e130 <_sk_srcover_rgba_8888_sse2_lowp+0x107c>
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  102,69,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm10
   DB  69,15,194,200,1                     ; cmpltps       %xmm8,%xmm9
-  DB  68,15,40,29,171,55,1,0              ; movaps        0x137ab(%rip),%xmm11        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,29,139,78,1,0              ; movaps        0x14e8b(%rip),%xmm11        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  102,69,15,56,20,211                 ; blendvps      %xmm0,%xmm11,%xmm10
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  65,15,194,192,1                     ; cmpltps       %xmm8,%xmm0
-  DB  68,15,40,13,157,55,1,0              ; movaps        0x1379d(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,125,78,1,0              ; movaps        0x14e7d(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  102,69,15,56,20,209                 ; blendvps      %xmm0,%xmm9,%xmm10
   DB  69,15,194,194,7                     ; cmpordps      %xmm10,%xmm8
@@ -23150,7 +23150,7 @@
   DB  243,69,15,89,203                    ; mulss         %xmm11,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,102,61,1,0              ; mulps         0x13d66(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,70,84,1,0               ; mulps         0x15446(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
@@ -23158,7 +23158,7 @@
   DB  243,69,15,89,219                    ; mulss         %xmm11,%xmm11
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,92,227                        ; subps         %xmm11,%xmm12
-  DB  68,15,89,21,81,61,1,0               ; mulps         0x13d51(%rip),%xmm10        # 2ca70 <_sk_srcover_rgba_8888_sse2_lowp+0x1068>
+  DB  68,15,89,21,49,84,1,0               ; mulps         0x15431(%rip),%xmm10        # 2e150 <_sk_srcover_rgba_8888_sse2_lowp+0x109c>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
@@ -23167,8 +23167,8 @@
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,92,193                        ; subps         %xmm9,%xmm0
-  DB  68,15,87,13,249,59,1,0              ; xorps         0x13bf9(%rip),%xmm9        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
-  DB  68,15,89,5,209,54,1,0               ; mulps         0x136d1(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,87,13,217,82,1,0              ; xorps         0x152d9(%rip),%xmm9        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
+  DB  68,15,89,5,177,77,1,0               ; mulps         0x14db1(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -23187,7 +23187,7 @@
   DB  243,69,15,89,203                    ; mulss         %xmm11,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,200,60,1,0              ; mulps         0x13cc8(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,168,83,1,0              ; mulps         0x153a8(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
@@ -23195,7 +23195,7 @@
   DB  243,69,15,89,219                    ; mulss         %xmm11,%xmm11
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,92,227                        ; subps         %xmm11,%xmm12
-  DB  68,15,89,21,179,60,1,0              ; mulps         0x13cb3(%rip),%xmm10        # 2ca70 <_sk_srcover_rgba_8888_sse2_lowp+0x1068>
+  DB  68,15,89,21,147,83,1,0              ; mulps         0x15393(%rip),%xmm10        # 2e150 <_sk_srcover_rgba_8888_sse2_lowp+0x109c>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
@@ -23204,8 +23204,8 @@
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,92,193                        ; subps         %xmm9,%xmm0
-  DB  68,15,87,13,91,59,1,0               ; xorps         0x13b5b(%rip),%xmm9        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
-  DB  68,15,89,5,51,54,1,0                ; mulps         0x13633(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,87,13,59,82,1,0               ; xorps         0x1523b(%rip),%xmm9        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
+  DB  68,15,89,5,19,77,1,0                ; mulps         0x14d13(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -23221,7 +23221,7 @@
   DB  243,69,15,89,200                    ; mulss         %xmm8,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,59,60,1,0               ; mulps         0x13c3b(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,27,83,1,0               ; mulps         0x1531b(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  69,15,89,210                        ; mulps         %xmm10,%xmm10
@@ -23229,7 +23229,7 @@
   DB  243,69,15,89,192                    ; mulss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,92,192                        ; subps         %xmm8,%xmm0
-  DB  15,87,5,247,58,1,0                  ; xorps         0x13af7(%rip),%xmm0        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
+  DB  15,87,5,215,81,1,0                  ; xorps         0x151d7(%rip),%xmm0        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
   DB  65,15,94,193                        ; divps         %xmm9,%xmm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23271,7 +23271,7 @@
 PUBLIC _sk_save_xy_sse41
 _sk_save_xy_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,86,53,1,0                ; movaps        0x13556(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,54,76,1,0                ; movaps        0x14c36(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -23311,8 +23311,8 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,31,59,1,0                   ; addps         0x13b1f(%rip),%xmm0        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,199,52,1,0              ; movaps        0x134c7(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,255,81,1,0                  ; addps         0x151ff(%rip),%xmm0        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,167,75,1,0              ; movaps        0x14ba7(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,0,1,0,0                ; movups        %xmm9,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23323,7 +23323,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,147,52,1,0                  ; addps         0x13493(%rip),%xmm0        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,5,115,75,1,0                  ; addps         0x14b73(%rip),%xmm0        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23333,8 +23333,8 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,210,58,1,0                 ; addps         0x13ad2(%rip),%xmm1        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,122,52,1,0              ; movaps        0x1347a(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,178,81,1,0                 ; addps         0x151b2(%rip),%xmm1        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,90,75,1,0               ; movaps        0x14b5a(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,64,1,0,0               ; movups        %xmm9,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23345,7 +23345,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,69,52,1,0                  ; addps         0x13445(%rip),%xmm1        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,13,37,75,1,0                  ; addps         0x14b25(%rip),%xmm1        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23355,13 +23355,13 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,149,58,1,0                  ; addps         0x13a95(%rip),%xmm0        # 2ca90 <_sk_srcover_rgba_8888_sse2_lowp+0x1088>
-  DB  68,15,40,13,45,52,1,0               ; movaps        0x1342d(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,117,81,1,0                  ; addps         0x15175(%rip),%xmm0        # 2e170 <_sk_srcover_rgba_8888_sse2_lowp+0x10bc>
+  DB  68,15,40,13,13,75,1,0               ; movaps        0x14b0d(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,137,58,1,0              ; mulps         0x13a89(%rip),%xmm9        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,13,17,54,1,0               ; addps         0x13611(%rip),%xmm9        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,13,105,81,1,0              ; mulps         0x15169(%rip),%xmm9        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,13,241,76,1,0              ; addps         0x14cf1(%rip),%xmm9        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,0,1,0,0                ; movups        %xmm9,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23372,16 +23372,16 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,61,58,1,0                   ; addps         0x13a3d(%rip),%xmm0        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,229,51,1,0              ; movaps        0x133e5(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,29,81,1,0                   ; addps         0x1511d(%rip),%xmm0        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,197,74,1,0              ; movaps        0x14ac5(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,89,58,1,0                ; movaps        0x13a59(%rip),%xmm8        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,5,57,81,1,0                ; movaps        0x15139(%rip),%xmm8        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,93,58,1,0                ; addps         0x13a5d(%rip),%xmm8        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,5,61,81,1,0                ; addps         0x1513d(%rip),%xmm8        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,177,51,1,0               ; addps         0x133b1(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,88,5,145,74,1,0               ; addps         0x14a91(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,85,58,1,0                ; addps         0x13a55(%rip),%xmm8        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,5,53,81,1,0                ; addps         0x15135(%rip),%xmm8        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23389,17 +23389,17 @@
 PUBLIC _sk_bicubic_p1x_sse41
 _sk_bicubic_p1x_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,143,51,1,0               ; movaps        0x1338f(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,111,74,1,0               ; movaps        0x14a6f(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,136,128,0,0,0              ; movups        0x80(%rax),%xmm9
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,21,8,58,1,0                ; movaps        0x13a08(%rip),%xmm10        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,21,232,80,1,0              ; movaps        0x150e8(%rip),%xmm10        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,12,58,1,0               ; addps         0x13a0c(%rip),%xmm10        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,21,236,80,1,0              ; addps         0x150ec(%rip),%xmm10        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,8,58,1,0                ; addps         0x13a08(%rip),%xmm10        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,21,232,80,1,0              ; addps         0x150e8(%rip),%xmm10        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,144,0,1,0,0                ; movups        %xmm10,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23409,11 +23409,11 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,216,57,1,0                  ; addps         0x139d8(%rip),%xmm0        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  15,88,5,184,80,1,0                  ; addps         0x150b8(%rip),%xmm0        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,168,57,1,0               ; mulps         0x139a8(%rip),%xmm8        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,5,48,53,1,0                ; addps         0x13530(%rip),%xmm8        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,5,136,80,1,0               ; mulps         0x15088(%rip),%xmm8        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,5,16,76,1,0                ; addps         0x14c10(%rip),%xmm8        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23424,13 +23424,13 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,107,57,1,0                 ; addps         0x1396b(%rip),%xmm1        # 2ca90 <_sk_srcover_rgba_8888_sse2_lowp+0x1088>
-  DB  68,15,40,13,3,51,1,0                ; movaps        0x13303(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,75,80,1,0                  ; addps         0x1504b(%rip),%xmm1        # 2e170 <_sk_srcover_rgba_8888_sse2_lowp+0x10bc>
+  DB  68,15,40,13,227,73,1,0              ; movaps        0x149e3(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,95,57,1,0               ; mulps         0x1395f(%rip),%xmm9        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,13,231,52,1,0              ; addps         0x134e7(%rip),%xmm9        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,13,63,80,1,0               ; mulps         0x1503f(%rip),%xmm9        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,13,199,75,1,0              ; addps         0x14bc7(%rip),%xmm9        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,64,1,0,0               ; movups        %xmm9,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23441,16 +23441,16 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,18,57,1,0                  ; addps         0x13912(%rip),%xmm1        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,186,50,1,0              ; movaps        0x132ba(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,242,79,1,0                 ; addps         0x14ff2(%rip),%xmm1        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,154,73,1,0              ; movaps        0x1499a(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,46,57,1,0                ; movaps        0x1392e(%rip),%xmm8        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,5,14,80,1,0                ; movaps        0x1500e(%rip),%xmm8        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,50,57,1,0                ; addps         0x13932(%rip),%xmm8        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,5,18,80,1,0                ; addps         0x15012(%rip),%xmm8        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,134,50,1,0               ; addps         0x13286(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,88,5,102,73,1,0               ; addps         0x14966(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,42,57,1,0                ; addps         0x1392a(%rip),%xmm8        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,5,10,80,1,0                ; addps         0x1500a(%rip),%xmm8        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23458,17 +23458,17 @@
 PUBLIC _sk_bicubic_p1y_sse41
 _sk_bicubic_p1y_sse41 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,100,50,1,0               ; movaps        0x13264(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,68,73,1,0                ; movaps        0x14944(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,136,192,0,0,0              ; movups        0xc0(%rax),%xmm9
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  68,15,40,21,220,56,1,0              ; movaps        0x138dc(%rip),%xmm10        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,21,188,79,1,0              ; movaps        0x14fbc(%rip),%xmm10        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,224,56,1,0              ; addps         0x138e0(%rip),%xmm10        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,21,192,79,1,0              ; addps         0x14fc0(%rip),%xmm10        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,220,56,1,0              ; addps         0x138dc(%rip),%xmm10        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,21,188,79,1,0              ; addps         0x14fbc(%rip),%xmm10        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,144,64,1,0,0               ; movups        %xmm10,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -23478,11 +23478,11 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,171,56,1,0                 ; addps         0x138ab(%rip),%xmm1        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  15,88,13,139,79,1,0                 ; addps         0x14f8b(%rip),%xmm1        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,123,56,1,0               ; mulps         0x1387b(%rip),%xmm8        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,5,3,52,1,0                 ; addps         0x13403(%rip),%xmm8        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,5,91,79,1,0                ; mulps         0x14f5b(%rip),%xmm8        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,5,227,74,1,0               ; addps         0x14ae3(%rip),%xmm8        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -23577,7 +23577,7 @@
   DB  15,89,194                           ; mulps         %xmm2,%xmm0
   DB  15,41,68,36,80                      ; movaps        %xmm0,0x50(%rsp)
   DB  243,15,91,240                       ; cvttps2dq     %xmm0,%xmm6
-  DB  15,40,37,126,55,1,0                 ; movaps        0x1377e(%rip),%xmm4        # 2cae0 <_sk_srcover_rgba_8888_sse2_lowp+0x10d8>
+  DB  15,40,37,94,78,1,0                  ; movaps        0x14e5e(%rip),%xmm4        # 2e1c0 <_sk_srcover_rgba_8888_sse2_lowp+0x110c>
   DB  15,88,196                           ; addps         %xmm4,%xmm0
   DB  15,41,68,36,48                      ; movaps        %xmm0,0x30(%rsp)
   DB  102,65,15,110,208                   ; movd          %r8d,%xmm2
@@ -23616,7 +23616,7 @@
   DB  102,68,15,56,64,192                 ; pmulld        %xmm0,%xmm8
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  102,15,254,218                      ; paddd         %xmm2,%xmm3
-  DB  102,68,15,111,37,213,54,1,0         ; movdqa        0x136d5(%rip),%xmm12        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,68,15,111,37,181,77,1,0         ; movdqa        0x14db5(%rip),%xmm12        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,65,15,56,64,220                 ; pmulld        %xmm12,%xmm3
   DB  102,15,118,228                      ; pcmpeqd       %xmm4,%xmm4
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
@@ -23639,7 +23639,7 @@
   DB  102,65,15,58,22,217,1               ; pextrd        $0x1,%xmm3,%r9d
   DB  102,65,15,58,22,218,2               ; pextrd        $0x2,%xmm3,%r10d
   DB  102,65,15,58,22,219,3               ; pextrd        $0x3,%xmm3,%r11d
-  DB  102,15,111,37,201,47,1,0            ; movdqa        0x12fc9(%rip),%xmm4        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,37,169,70,1,0            ; movdqa        0x146a9(%rip),%xmm4        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,220                      ; paddd         %xmm4,%xmm3
   DB  102,15,111,252                      ; movdqa        %xmm4,%xmm7
   DB  102,73,15,58,22,222,1               ; pextrq        $0x1,%xmm3,%r14
@@ -23812,7 +23812,7 @@
   DB  102,68,15,254,202                   ; paddd         %xmm2,%xmm9
   DB  102,65,15,111,192                   ; movdqa        %xmm8,%xmm0
   DB  102,65,15,254,193                   ; paddd         %xmm9,%xmm0
-  DB  102,15,111,45,228,50,1,0            ; movdqa        0x132e4(%rip),%xmm5        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,45,196,73,1,0            ; movdqa        0x149c4(%rip),%xmm5        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,56,64,197                    ; pmulld        %xmm5,%xmm0
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,118,246                      ; pcmpeqd       %xmm6,%xmm6
@@ -23833,7 +23833,7 @@
   DB  102,65,15,58,22,193,1               ; pextrd        $0x1,%xmm0,%r9d
   DB  102,65,15,58,22,194,2               ; pextrd        $0x2,%xmm0,%r10d
   DB  102,65,15,58,22,195,3               ; pextrd        $0x3,%xmm0,%r11d
-  DB  102,15,111,61,229,43,1,0            ; movdqa        0x12be5(%rip),%xmm7        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,61,197,66,1,0            ; movdqa        0x142c5(%rip),%xmm7        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,199                      ; paddd         %xmm7,%xmm0
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
   DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
@@ -23899,7 +23899,7 @@
   DB  15,88,248                           ; addps         %xmm0,%xmm7
   DB  102,68,15,254,210                   ; paddd         %xmm2,%xmm10
   DB  102,69,15,254,194                   ; paddd         %xmm10,%xmm8
-  DB  102,15,111,13,32,49,1,0             ; movdqa        0x13120(%rip),%xmm1        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,13,0,72,1,0              ; movdqa        0x14800(%rip),%xmm1        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,68,15,56,64,193                 ; pmulld        %xmm1,%xmm8
   DB  102,65,15,111,192                   ; movdqa        %xmm8,%xmm0
   DB  102,15,118,237                      ; pcmpeqd       %xmm5,%xmm5
@@ -23920,7 +23920,7 @@
   DB  102,69,15,58,22,193,1               ; pextrd        $0x1,%xmm8,%r9d
   DB  102,69,15,58,22,194,2               ; pextrd        $0x2,%xmm8,%r10d
   DB  102,69,15,58,22,195,3               ; pextrd        $0x3,%xmm8,%r11d
-  DB  102,15,111,21,27,42,1,0             ; movdqa        0x12a1b(%rip),%xmm2        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,21,251,64,1,0            ; movdqa        0x140fb(%rip),%xmm2        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,68,15,254,194                   ; paddd         %xmm2,%xmm8
   DB  102,76,15,58,22,193,1               ; pextrq        $0x1,%xmm8,%rcx
   DB  102,76,15,126,195                   ; movq          %xmm8,%rbx
@@ -24045,7 +24045,7 @@
   DB  15,89,203                           ; mulps         %xmm3,%xmm1
   DB  15,41,140,36,176,0,0,0              ; movaps        %xmm1,0xb0(%rsp)
   DB  243,15,91,249                       ; cvttps2dq     %xmm1,%xmm7
-  DB  15,40,29,98,46,1,0                  ; movaps        0x12e62(%rip),%xmm3        # 2cae0 <_sk_srcover_rgba_8888_sse2_lowp+0x10d8>
+  DB  15,40,29,66,69,1,0                  ; movaps        0x14542(%rip),%xmm3        # 2e1c0 <_sk_srcover_rgba_8888_sse2_lowp+0x110c>
   DB  15,88,203                           ; addps         %xmm3,%xmm1
   DB  15,41,76,36,32                      ; movaps        %xmm1,0x20(%rsp)
   DB  102,65,15,110,232                   ; movd          %r8d,%xmm5
@@ -24102,7 +24102,7 @@
   DB  102,68,15,56,64,253                 ; pmulld        %xmm5,%xmm15
   DB  102,65,15,111,223                   ; movdqa        %xmm15,%xmm3
   DB  102,15,254,217                      ; paddd         %xmm1,%xmm3
-  DB  102,68,15,111,29,92,45,1,0          ; movdqa        0x12d5c(%rip),%xmm11        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,68,15,111,29,60,68,1,0          ; movdqa        0x1443c(%rip),%xmm11        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,65,15,56,64,219                 ; pmulld        %xmm11,%xmm3
   DB  102,15,118,192                      ; pcmpeqd       %xmm0,%xmm0
   DB  102,15,111,243                      ; movdqa        %xmm3,%xmm6
@@ -24124,7 +24124,7 @@
   DB  102,65,15,58,22,217,1               ; pextrd        $0x1,%xmm3,%r9d
   DB  102,65,15,58,22,218,2               ; pextrd        $0x2,%xmm3,%r10d
   DB  102,65,15,58,22,219,3               ; pextrd        $0x3,%xmm3,%r11d
-  DB  102,68,15,111,21,84,38,1,0          ; movdqa        0x12654(%rip),%xmm10        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,21,52,61,1,0          ; movdqa        0x13d34(%rip),%xmm10        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,218                   ; paddd         %xmm10,%xmm3
   DB  102,73,15,58,22,222,1               ; pextrq        $0x1,%xmm3,%r14
   DB  102,72,15,126,219                   ; movq          %xmm3,%rbx
@@ -24142,7 +24142,7 @@
   DB  102,65,15,254,201                   ; paddd         %xmm9,%xmm1
   DB  102,65,15,56,64,203                 ; pmulld        %xmm11,%xmm1
   DB  102,15,111,217                      ; movdqa        %xmm1,%xmm3
-  DB  102,15,250,29,146,44,1,0            ; psubd         0x12c92(%rip),%xmm3        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,29,114,67,1,0            ; psubd         0x14372(%rip),%xmm3        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,73,15,58,22,222,1               ; pextrq        $0x1,%xmm3,%r14
   DB  102,72,15,126,219                   ; movq          %xmm3,%rbx
   DB  65,137,223                          ; mov           %ebx,%r15d
@@ -24197,7 +24197,7 @@
   DB  102,65,15,254,199                   ; paddd         %xmm15,%xmm0
   DB  102,65,15,56,64,195                 ; pmulld        %xmm11,%xmm0
   DB  102,15,111,232                      ; movdqa        %xmm0,%xmm5
-  DB  102,15,250,45,114,43,1,0            ; psubd         0x12b72(%rip),%xmm5        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,45,82,66,1,0             ; psubd         0x14252(%rip),%xmm5        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,69,15,118,192                   ; pcmpeqd       %xmm8,%xmm8
   DB  102,72,15,58,22,233,1               ; pextrq        $0x1,%xmm5,%rcx
   DB  102,72,15,126,237                   ; movq          %xmm5,%rbp
@@ -24322,7 +24322,7 @@
   DB  72,193,233,32                       ; shr           $0x20,%rcx
   DB  243,15,16,36,136                    ; movss         (%rax,%rcx,4),%xmm4
   DB  102,68,15,58,33,220,48              ; insertps      $0x30,%xmm4,%xmm11
-  DB  102,15,111,61,69,34,1,0             ; movdqa        0x12245(%rip),%xmm7        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,61,37,57,1,0             ; movdqa        0x13925(%rip),%xmm7        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,223                      ; paddd         %xmm7,%xmm3
   DB  102,72,15,58,22,217,1               ; pextrq        $0x1,%xmm3,%rcx
   DB  102,72,15,126,221                   ; movq          %xmm3,%rbp
@@ -24408,7 +24408,7 @@
   DB  102,65,15,58,22,201,1               ; pextrd        $0x1,%xmm1,%r9d
   DB  102,65,15,58,22,202,2               ; pextrd        $0x2,%xmm1,%r10d
   DB  102,65,15,58,22,203,3               ; pextrd        $0x3,%xmm1,%r11d
-  DB  102,68,15,111,21,133,32,1,0         ; movdqa        0x12085(%rip),%xmm10        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,21,101,55,1,0         ; movdqa        0x13765(%rip),%xmm10        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,202                   ; paddd         %xmm10,%xmm1
   DB  102,72,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%rcx
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
@@ -24504,7 +24504,7 @@
   DB  102,15,111,124,36,64                ; movdqa        0x40(%rsp),%xmm7
   DB  102,15,111,199                      ; movdqa        %xmm7,%xmm0
   DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
-  DB  102,15,111,21,41,37,1,0             ; movdqa        0x12529(%rip),%xmm2        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,21,9,60,1,0              ; movdqa        0x13c09(%rip),%xmm2        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,56,64,194                    ; pmulld        %xmm2,%xmm0
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,69,15,118,192                   ; pcmpeqd       %xmm8,%xmm8
@@ -24525,7 +24525,7 @@
   DB  102,65,15,58,22,193,1               ; pextrd        $0x1,%xmm0,%r9d
   DB  102,65,15,58,22,194,2               ; pextrd        $0x2,%xmm0,%r10d
   DB  102,65,15,58,22,195,3               ; pextrd        $0x3,%xmm0,%r11d
-  DB  102,15,111,37,40,30,1,0             ; movdqa        0x11e28(%rip),%xmm4        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,37,8,53,1,0              ; movdqa        0x13508(%rip),%xmm4        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,196                      ; paddd         %xmm4,%xmm0
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
   DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
@@ -24597,7 +24597,7 @@
   DB  102,68,15,111,226                   ; movdqa        %xmm2,%xmm12
   DB  102,65,15,56,64,204                 ; pmulld        %xmm12,%xmm1
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
-  DB  102,15,250,21,84,35,1,0             ; psubd         0x12354(%rip),%xmm2        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,21,52,58,1,0             ; psubd         0x13a34(%rip),%xmm2        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,72,15,58,22,209,1               ; pextrq        $0x1,%xmm2,%rcx
   DB  102,72,15,126,213                   ; movq          %xmm2,%rbp
   DB  137,235                             ; mov           %ebp,%ebx
@@ -24631,7 +24631,7 @@
   DB  102,65,15,111,249                   ; movdqa        %xmm9,%xmm7
   DB  102,69,15,56,64,236                 ; pmulld        %xmm12,%xmm13
   DB  102,65,15,111,205                   ; movdqa        %xmm13,%xmm1
-  DB  102,15,250,13,160,34,1,0            ; psubd         0x122a0(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,128,57,1,0            ; psubd         0x13980(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,72,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%rcx
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
   DB  137,221                             ; mov           %ebx,%ebp
@@ -24696,7 +24696,7 @@
   DB  102,65,15,111,206                   ; movdqa        %xmm14,%xmm1
   DB  102,15,111,108,36,64                ; movdqa        0x40(%rsp),%xmm5
   DB  102,15,254,205                      ; paddd         %xmm5,%xmm1
-  DB  102,15,111,37,52,33,1,0             ; movdqa        0x12134(%rip),%xmm4        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,37,20,56,1,0             ; movdqa        0x13814(%rip),%xmm4        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,56,64,204                    ; pmulld        %xmm4,%xmm1
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,118,246                      ; pcmpeqd       %xmm6,%xmm6
@@ -24717,7 +24717,7 @@
   DB  102,65,15,58,22,201,1               ; pextrd        $0x1,%xmm1,%r9d
   DB  102,65,15,58,22,202,2               ; pextrd        $0x2,%xmm1,%r10d
   DB  102,65,15,58,22,203,3               ; pextrd        $0x3,%xmm1,%r11d
-  DB  102,15,111,29,53,26,1,0             ; movdqa        0x11a35(%rip),%xmm3        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,29,21,49,1,0             ; movdqa        0x13115(%rip),%xmm3        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
   DB  102,72,15,58,22,201,1               ; pextrq        $0x1,%xmm1,%rcx
   DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
@@ -24789,7 +24789,7 @@
   DB  102,15,56,64,239                    ; pmulld        %xmm7,%xmm5
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
-  DB  102,15,250,5,101,31,1,0             ; psubd         0x11f65(%rip),%xmm0        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,5,69,54,1,0              ; psubd         0x13645(%rip),%xmm0        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
   DB  102,72,15,126,197                   ; movq          %xmm0,%rbp
   DB  137,235                             ; mov           %ebp,%ebx
@@ -24807,7 +24807,7 @@
   DB  102,65,15,58,22,193,1               ; pextrd        $0x1,%xmm0,%r9d
   DB  102,65,15,58,22,194,2               ; pextrd        $0x2,%xmm0,%r10d
   DB  102,65,15,58,22,195,3               ; pextrd        $0x3,%xmm0,%r11d
-  DB  102,15,254,5,99,24,1,0              ; paddd         0x11863(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,5,67,47,1,0              ; paddd         0x12f43(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
   DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
   DB  137,221                             ; mov           %ebx,%ebp
@@ -24823,7 +24823,7 @@
   DB  102,15,254,84,36,48                 ; paddd         0x30(%rsp),%xmm2
   DB  102,15,56,64,215                    ; pmulld        %xmm7,%xmm2
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
-  DB  102,15,250,5,172,30,1,0             ; psubd         0x11eac(%rip),%xmm0        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,5,140,53,1,0             ; psubd         0x1358c(%rip),%xmm0        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,72,15,58,22,193,1               ; pextrq        $0x1,%xmm0,%rcx
   DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
   DB  137,221                             ; mov           %ebx,%ebp
@@ -24841,7 +24841,7 @@
   DB  102,65,15,58,22,199,1               ; pextrd        $0x1,%xmm0,%r15d
   DB  102,65,15,58,22,196,2               ; pextrd        $0x2,%xmm0,%r12d
   DB  102,65,15,58,22,198,3               ; pextrd        $0x3,%xmm0,%r14d
-  DB  102,15,254,5,170,23,1,0             ; paddd         0x117aa(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,5,138,46,1,0             ; paddd         0x12e8a(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,72,15,58,22,195,1               ; pextrq        $0x1,%xmm0,%rbx
   DB  102,72,15,126,197                   ; movq          %xmm0,%rbp
   DB  137,233                             ; mov           %ebp,%ecx
@@ -24904,7 +24904,7 @@
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  65,15,88,211                        ; addps         %xmm11,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,98,22,1,0                  ; movaps        0x11662(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,66,45,1,0                  ; movaps        0x12d42(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,139,76,36,120                    ; mov           0x78(%rsp),%rcx
   DB  15,40,164,36,192,0,0,0              ; movaps        0xc0(%rsp),%xmm4
   DB  15,40,172,36,208,0,0,0              ; movaps        0xd0(%rsp),%xmm5
@@ -25029,9 +25029,9 @@
   DB  102,15,110,201                      ; movd          %ecx,%xmm1
   DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,88,13,130,20,1,0                 ; addps         0x11482(%rip),%xmm1        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,13,98,43,1,0                  ; addps         0x12b62(%rip),%xmm1        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,21,137,20,1,0                 ; movaps        0x11489(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,105,43,1,0                 ; movaps        0x12b69(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
@@ -25044,18 +25044,18 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  102,68,15,110,194                   ; movd          %edx,%xmm8
   DB  102,69,15,112,192,0                 ; pshufd        $0x0,%xmm8,%xmm8
-  DB  102,68,15,254,5,50,20,1,0           ; paddd         0x11432(%rip),%xmm8        # 2c400 <_sk_srcover_rgba_8888_sse2_lowp+0x9f8>
+  DB  102,68,15,254,5,18,43,1,0           ; paddd         0x12b12(%rip),%xmm8        # 2dae0 <_sk_srcover_rgba_8888_sse2_lowp+0xa2c>
   DB  102,68,15,110,201                   ; movd          %ecx,%xmm9
   DB  102,69,15,112,201,0                 ; pshufd        $0x0,%xmm9,%xmm9
   DB  102,69,15,239,200                   ; pxor          %xmm8,%xmm9
-  DB  102,68,15,111,21,89,20,1,0          ; movdqa        0x11459(%rip),%xmm10        # 2c440 <_sk_srcover_rgba_8888_sse2_lowp+0xa38>
+  DB  102,68,15,111,21,57,43,1,0          ; movdqa        0x12b39(%rip),%xmm10        # 2db20 <_sk_srcover_rgba_8888_sse2_lowp+0xa6c>
   DB  102,69,15,111,217                   ; movdqa        %xmm9,%xmm11
   DB  102,69,15,219,218                   ; pand          %xmm10,%xmm11
   DB  102,65,15,114,243,5                 ; pslld         $0x5,%xmm11
   DB  102,69,15,219,208                   ; pand          %xmm8,%xmm10
   DB  102,65,15,114,242,4                 ; pslld         $0x4,%xmm10
-  DB  102,68,15,111,37,69,20,1,0          ; movdqa        0x11445(%rip),%xmm12        # 2c450 <_sk_srcover_rgba_8888_sse2_lowp+0xa48>
-  DB  102,68,15,111,45,76,20,1,0          ; movdqa        0x1144c(%rip),%xmm13        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,37,37,43,1,0          ; movdqa        0x12b25(%rip),%xmm12        # 2db30 <_sk_srcover_rgba_8888_sse2_lowp+0xa7c>
+  DB  102,68,15,111,45,44,43,1,0          ; movdqa        0x12b2c(%rip),%xmm13        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,69,15,111,241                   ; movdqa        %xmm9,%xmm14
   DB  102,69,15,219,245                   ; pand          %xmm13,%xmm14
   DB  102,65,15,114,246,2                 ; pslld         $0x2,%xmm14
@@ -25071,8 +25071,8 @@
   DB  102,69,15,235,198                   ; por           %xmm14,%xmm8
   DB  102,69,15,235,193                   ; por           %xmm9,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,7,20,1,0                 ; mulps         0x11407(%rip),%xmm8        # 2c470 <_sk_srcover_rgba_8888_sse2_lowp+0xa68>
-  DB  68,15,88,5,15,20,1,0                ; addps         0x1140f(%rip),%xmm8        # 2c480 <_sk_srcover_rgba_8888_sse2_lowp+0xa78>
+  DB  68,15,89,5,231,42,1,0               ; mulps         0x12ae7(%rip),%xmm8        # 2db50 <_sk_srcover_rgba_8888_sse2_lowp+0xa9c>
+  DB  68,15,88,5,239,42,1,0               ; addps         0x12aef(%rip),%xmm8        # 2db60 <_sk_srcover_rgba_8888_sse2_lowp+0xaac>
   DB  243,68,15,16,16                     ; movss         (%rax),%xmm10
   DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
@@ -25110,7 +25110,7 @@
 PUBLIC _sk_black_color_sse2
 _sk_black_color_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,69,19,1,0                  ; movaps        0x11345(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,37,42,1,0                  ; movaps        0x12a25(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -25119,7 +25119,7 @@
 PUBLIC _sk_white_color_sse2
 _sk_white_color_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,5,49,19,1,0                   ; movaps        0x11331(%rip),%xmm0        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,5,17,42,1,0                   ; movaps        0x12a11(%rip),%xmm0        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -25157,7 +25157,7 @@
 PUBLIC _sk_srcatop_sse2
 _sk_srcatop_sse2 LABEL PROC
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  68,15,40,5,225,18,1,0               ; movaps        0x112e1(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,193,41,1,0               ; movaps        0x129c1(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -25180,7 +25180,7 @@
 _sk_dstatop_sse2 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
   DB  68,15,89,196                        ; mulps         %xmm4,%xmm8
-  DB  68,15,40,13,148,18,1,0              ; movaps        0x11294(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,116,41,1,0              ; movaps        0x12974(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
@@ -25221,7 +25221,7 @@
 
 PUBLIC _sk_srcout_sse2
 _sk_srcout_sse2 LABEL PROC
-  DB  68,15,40,5,40,18,1,0                ; movaps        0x11228(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,8,41,1,0                 ; movaps        0x12908(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
@@ -25232,7 +25232,7 @@
 
 PUBLIC _sk_dstout_sse2
 _sk_dstout_sse2 LABEL PROC
-  DB  68,15,40,5,8,18,1,0                 ; movaps        0x11208(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,232,40,1,0               ; movaps        0x128e8(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -25247,7 +25247,7 @@
 
 PUBLIC _sk_srcover_sse2
 _sk_srcover_sse2 LABEL PROC
-  DB  68,15,40,5,219,17,1,0               ; movaps        0x111db(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,187,40,1,0               ; movaps        0x128bb(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -25265,7 +25265,7 @@
 
 PUBLIC _sk_dstover_sse2
 _sk_dstover_sse2 LABEL PROC
-  DB  68,15,40,5,159,17,1,0               ; movaps        0x1119f(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,127,40,1,0               ; movaps        0x1287f(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -25289,7 +25289,7 @@
 
 PUBLIC _sk_multiply_sse2
 _sk_multiply_sse2 LABEL PROC
-  DB  68,15,40,5,99,17,1,0                ; movaps        0x11163(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,67,40,1,0                ; movaps        0x12843(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
@@ -25325,7 +25325,7 @@
 PUBLIC _sk_plus__sse2
 _sk_plus__sse2 LABEL PROC
   DB  15,88,196                           ; addps         %xmm4,%xmm0
-  DB  68,15,40,5,228,16,1,0               ; movaps        0x110e4(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,196,39,1,0               ; movaps        0x127c4(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  15,88,205                           ; addps         %xmm5,%xmm1
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
@@ -25363,7 +25363,7 @@
 PUBLIC _sk_xor__sse2
 _sk_xor__sse2 LABEL PROC
   DB  68,15,40,195                        ; movaps        %xmm3,%xmm8
-  DB  15,40,29,112,16,1,0                 ; movaps        0x11070(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,80,39,1,0                  ; movaps        0x12750(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,40,203                        ; movaps        %xmm3,%xmm9
   DB  68,15,92,207                        ; subps         %xmm7,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
@@ -25409,7 +25409,7 @@
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,95,209                        ; maxps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,203,15,1,0                 ; movaps        0x10fcb(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,171,38,1,0                 ; movaps        0x126ab(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -25441,7 +25441,7 @@
   DB  68,15,89,206                        ; mulps         %xmm6,%xmm9
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,96,15,1,0                  ; movaps        0x10f60(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,64,38,1,0                  ; movaps        0x12640(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -25476,7 +25476,7 @@
   DB  65,15,93,209                        ; minps         %xmm9,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,234,14,1,0                 ; movaps        0x10eea(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,202,37,1,0                 ; movaps        0x125ca(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -25502,7 +25502,7 @@
   DB  15,89,214                           ; mulps         %xmm6,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  68,15,92,194                        ; subps         %xmm2,%xmm8
-  DB  15,40,21,154,14,1,0                 ; movaps        0x10e9a(%rip),%xmm2        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,21,122,37,1,0                 ; movaps        0x1257a(%rip),%xmm2        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,215                           ; mulps         %xmm7,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -25513,7 +25513,7 @@
 PUBLIC _sk_colorburn_sse2
 _sk_colorburn_sse2 LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
-  DB  68,15,40,13,125,14,1,0              ; movaps        0x10e7d(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,93,37,1,0               ; movaps        0x1255d(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -25606,7 +25606,7 @@
 PUBLIC _sk_colordodge_sse2
 _sk_colordodge_sse2 LABEL PROC
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
-  DB  68,15,40,21,37,13,1,0               ; movaps        0x10d25(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,5,36,1,0                ; movaps        0x12405(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
   DB  68,15,92,223                        ; subps         %xmm7,%xmm11
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
@@ -25697,7 +25697,7 @@
   DB  15,41,52,36                         ; movaps        %xmm6,(%rsp)
   DB  15,40,245                           ; movaps        %xmm5,%xmm6
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
-  DB  68,15,40,29,203,11,1,0              ; movaps        0x10bcb(%rip),%xmm11        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,29,171,34,1,0              ; movaps        0x122ab(%rip),%xmm11        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,211                        ; movaps        %xmm11,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
@@ -25784,7 +25784,7 @@
 _sk_overlay_sse2 LABEL PROC
   DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
   DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
-  DB  68,15,40,13,134,10,1,0              ; movaps        0x10a86(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,102,33,1,0              ; movaps        0x12166(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  68,15,92,215                        ; subps         %xmm7,%xmm10
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -25874,7 +25874,7 @@
   DB  68,15,40,213                        ; movaps        %xmm5,%xmm10
   DB  68,15,94,215                        ; divps         %xmm7,%xmm10
   DB  69,15,84,212                        ; andps         %xmm12,%xmm10
-  DB  68,15,40,13,48,9,1,0                ; movaps        0x10930(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,16,32,1,0               ; movaps        0x12010(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,249                        ; movaps        %xmm9,%xmm15
   DB  69,15,92,250                        ; subps         %xmm10,%xmm15
   DB  69,15,40,218                        ; movaps        %xmm10,%xmm11
@@ -25887,10 +25887,10 @@
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  65,15,88,194                        ; addps         %xmm10,%xmm0
-  DB  68,15,40,53,90,9,1,0                ; movaps        0x1095a(%rip),%xmm14        # 2c490 <_sk_srcover_rgba_8888_sse2_lowp+0xa88>
+  DB  68,15,40,53,58,32,1,0               ; movaps        0x1203a(%rip),%xmm14        # 2db70 <_sk_srcover_rgba_8888_sse2_lowp+0xabc>
   DB  69,15,88,222                        ; addps         %xmm14,%xmm11
   DB  68,15,89,216                        ; mulps         %xmm0,%xmm11
-  DB  68,15,40,21,90,9,1,0                ; movaps        0x1095a(%rip),%xmm10        # 2c4a0 <_sk_srcover_rgba_8888_sse2_lowp+0xa98>
+  DB  68,15,40,21,58,32,1,0               ; movaps        0x1203a(%rip),%xmm10        # 2db80 <_sk_srcover_rgba_8888_sse2_lowp+0xacc>
   DB  69,15,89,234                        ; mulps         %xmm10,%xmm13
   DB  69,15,88,235                        ; addps         %xmm11,%xmm13
   DB  15,88,228                           ; addps         %xmm4,%xmm4
@@ -26035,7 +26035,7 @@
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  68,15,89,211                        ; mulps         %xmm3,%xmm10
-  DB  68,15,40,5,230,6,1,0                ; movaps        0x106e6(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,198,29,1,0               ; movaps        0x11dc6(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,216                        ; movaps        %xmm8,%xmm11
   DB  15,40,207                           ; movaps        %xmm7,%xmm1
   DB  68,15,92,217                        ; subps         %xmm1,%xmm11
@@ -26083,12 +26083,12 @@
   DB  69,15,84,206                        ; andps         %xmm14,%xmm9
   DB  69,15,84,214                        ; andps         %xmm14,%xmm10
   DB  65,15,84,214                        ; andps         %xmm14,%xmm2
-  DB  68,15,40,61,163,6,1,0               ; movaps        0x106a3(%rip),%xmm15        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,61,131,29,1,0              ; movaps        0x11d83(%rip),%xmm15        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  65,15,89,231                        ; mulps         %xmm15,%xmm4
-  DB  15,40,5,168,6,1,0                   ; movaps        0x106a8(%rip),%xmm0        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  15,40,5,136,29,1,0                  ; movaps        0x11d88(%rip),%xmm0        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  15,89,240                           ; mulps         %xmm0,%xmm6
   DB  15,88,244                           ; addps         %xmm4,%xmm6
-  DB  68,15,40,53,170,6,1,0               ; movaps        0x106aa(%rip),%xmm14        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,53,138,29,1,0              ; movaps        0x11d8a(%rip),%xmm14        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  68,15,40,239                        ; movaps        %xmm7,%xmm13
   DB  69,15,89,238                        ; mulps         %xmm14,%xmm13
   DB  68,15,88,238                        ; addps         %xmm6,%xmm13
@@ -26266,14 +26266,14 @@
   DB  68,15,84,211                        ; andps         %xmm3,%xmm10
   DB  68,15,84,203                        ; andps         %xmm3,%xmm9
   DB  15,84,195                           ; andps         %xmm3,%xmm0
-  DB  68,15,40,5,247,3,1,0                ; movaps        0x103f7(%rip),%xmm8        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,5,215,26,1,0               ; movaps        0x11ad7(%rip),%xmm8        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  15,40,214                           ; movaps        %xmm6,%xmm2
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
-  DB  15,40,13,249,3,1,0                  ; movaps        0x103f9(%rip),%xmm1        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  15,40,13,217,26,1,0                 ; movaps        0x11ad9(%rip),%xmm1        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  15,40,221                           ; movaps        %xmm5,%xmm3
   DB  15,89,217                           ; mulps         %xmm1,%xmm3
   DB  15,88,218                           ; addps         %xmm2,%xmm3
-  DB  68,15,40,37,248,3,1,0               ; movaps        0x103f8(%rip),%xmm12        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,37,216,26,1,0              ; movaps        0x11ad8(%rip),%xmm12        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
   DB  68,15,88,235                        ; addps         %xmm3,%xmm13
   DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
@@ -26318,7 +26318,7 @@
   DB  15,40,223                           ; movaps        %xmm7,%xmm3
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,221                           ; mulps         %xmm5,%xmm3
-  DB  68,15,40,5,173,2,1,0                ; movaps        0x102ad(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,141,25,1,0               ; movaps        0x1198d(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  68,15,92,199                        ; subps         %xmm7,%xmm8
   DB  15,88,253                           ; addps         %xmm5,%xmm7
@@ -26419,14 +26419,14 @@
   DB  68,15,40,213                        ; movaps        %xmm5,%xmm10
   DB  69,15,89,208                        ; mulps         %xmm8,%xmm10
   DB  65,15,40,208                        ; movaps        %xmm8,%xmm2
-  DB  68,15,40,45,181,1,1,0               ; movaps        0x101b5(%rip),%xmm13        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,45,149,24,1,0              ; movaps        0x11895(%rip),%xmm13        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,40,198                        ; movaps        %xmm6,%xmm8
   DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
-  DB  68,15,40,53,181,1,1,0               ; movaps        0x101b5(%rip),%xmm14        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  68,15,40,53,149,24,1,0              ; movaps        0x11895(%rip),%xmm14        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  65,15,40,195                        ; movaps        %xmm11,%xmm0
   DB  65,15,89,198                        ; mulps         %xmm14,%xmm0
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,29,177,1,1,0               ; movaps        0x101b1(%rip),%xmm11        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
+  DB  68,15,40,29,145,24,1,0              ; movaps        0x11891(%rip),%xmm11        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
   DB  69,15,89,227                        ; mulps         %xmm11,%xmm12
   DB  68,15,88,224                        ; addps         %xmm0,%xmm12
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
@@ -26434,7 +26434,7 @@
   DB  69,15,40,250                        ; movaps        %xmm10,%xmm15
   DB  69,15,89,254                        ; mulps         %xmm14,%xmm15
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
-  DB  68,15,40,5,237,0,1,0                ; movaps        0x100ed(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,205,23,1,0               ; movaps        0x117cd(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,40,224                        ; movaps        %xmm8,%xmm4
   DB  15,92,226                           ; subps         %xmm2,%xmm4
   DB  15,89,252                           ; mulps         %xmm4,%xmm7
@@ -26570,15 +26570,15 @@
   DB  68,15,40,205                        ; movaps        %xmm5,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
   DB  15,89,222                           ; mulps         %xmm6,%xmm3
-  DB  68,15,40,37,111,255,0,0             ; movaps        0xff6f(%rip),%xmm12        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,37,79,22,1,0               ; movaps        0x1164f(%rip),%xmm12        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,40,199                        ; movaps        %xmm7,%xmm8
   DB  69,15,89,196                        ; mulps         %xmm12,%xmm8
-  DB  68,15,40,45,111,255,0,0             ; movaps        0xff6f(%rip),%xmm13        # 2c4c0 <_sk_srcover_rgba_8888_sse2_lowp+0xab8>
+  DB  68,15,40,45,79,22,1,0               ; movaps        0x1164f(%rip),%xmm13        # 2dba0 <_sk_srcover_rgba_8888_sse2_lowp+0xaec>
   DB  68,15,40,241                        ; movaps        %xmm1,%xmm14
   DB  69,15,89,245                        ; mulps         %xmm13,%xmm14
   DB  69,15,88,240                        ; addps         %xmm8,%xmm14
-  DB  68,15,40,29,107,255,0,0             ; movaps        0xff6b(%rip),%xmm11        # 2c4d0 <_sk_srcover_rgba_8888_sse2_lowp+0xac8>
-  DB  68,15,40,5,195,254,0,0              ; movaps        0xfec3(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,29,75,22,1,0               ; movaps        0x1164b(%rip),%xmm11        # 2dbb0 <_sk_srcover_rgba_8888_sse2_lowp+0xafc>
+  DB  68,15,40,5,163,21,1,0               ; movaps        0x115a3(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,248                        ; movaps        %xmm8,%xmm15
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  68,15,92,248                        ; subps         %xmm0,%xmm15
@@ -26721,7 +26721,7 @@
   DB  15,133,228,0,0,0                    ; jne           1c85a <_sk_srcover_rgba_8888_sse2+0x104>
   DB  243,69,15,111,4,128                 ; movdqu        (%r8,%rax,4),%xmm8
   DB  72,133,255                          ; test          %rdi,%rdi
-  DB  102,15,111,53,89,253,0,0            ; movdqa        0xfd59(%rip),%xmm6        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,53,57,20,1,0             ; movdqa        0x11439(%rip),%xmm6        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,224                   ; movdqa        %xmm8,%xmm4
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
@@ -26735,9 +26735,9 @@
   DB  15,91,247                           ; cvtdq2ps      %xmm7,%xmm6
   DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
   DB  65,15,91,248                        ; cvtdq2ps      %xmm8,%xmm7
-  DB  68,15,40,5,105,252,0,0              ; movaps        0xfc69(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,73,19,1,0                ; movaps        0x11349(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,92,195                        ; subps         %xmm3,%xmm8
-  DB  68,15,40,37,61,253,0,0              ; movaps        0xfd3d(%rip),%xmm12        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,37,29,20,1,0               ; movaps        0x1141d(%rip),%xmm12        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,89,204                        ; mulps         %xmm4,%xmm9
@@ -26813,7 +26813,7 @@
 
 PUBLIC _sk_clamp_1_sse2
 _sk_clamp_1_sse2 LABEL PROC
-  DB  68,15,40,5,57,251,0,0               ; movaps        0xfb39(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,25,18,1,0                ; movaps        0x11219(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,93,192                        ; minps         %xmm8,%xmm0
   DB  65,15,93,200                        ; minps         %xmm8,%xmm1
   DB  65,15,93,208                        ; minps         %xmm8,%xmm2
@@ -26823,7 +26823,7 @@
 
 PUBLIC _sk_clamp_a_sse2
 _sk_clamp_a_sse2 LABEL PROC
-  DB  15,93,29,30,251,0,0                 ; minps         0xfb1e(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,29,254,17,1,0                 ; minps         0x111fe(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -26832,7 +26832,7 @@
 
 PUBLIC _sk_clamp_a_dst_sse2
 _sk_clamp_a_dst_sse2 LABEL PROC
-  DB  15,93,61,10,251,0,0                 ; minps         0xfb0a(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,93,61,234,17,1,0                 ; minps         0x111ea(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,93,231                           ; minps         %xmm7,%xmm4
   DB  15,93,239                           ; minps         %xmm7,%xmm5
   DB  15,93,247                           ; minps         %xmm7,%xmm6
@@ -26861,7 +26861,7 @@
 
 PUBLIC _sk_invert_sse2
 _sk_invert_sse2 LABEL PROC
-  DB  68,15,40,5,198,250,0,0              ; movaps        0xfac6(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,166,17,1,0               ; movaps        0x111a6(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  68,15,92,200                        ; subps         %xmm0,%xmm9
   DB  69,15,40,208                        ; movaps        %xmm8,%xmm10
@@ -26912,10 +26912,10 @@
 
 PUBLIC _sk_unpremul_sse2
 _sk_unpremul_sse2 LABEL PROC
-  DB  68,15,40,5,84,250,0,0               ; movaps        0xfa54(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,5,52,17,1,0                ; movaps        0x11134(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,94,195                        ; divps         %xmm3,%xmm8
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
-  DB  68,15,194,13,51,251,0,0,1           ; cmpltps       0xfb33(%rip),%xmm9        # 2c520 <_sk_srcover_rgba_8888_sse2_lowp+0xb18>
+  DB  68,15,194,13,19,18,1,0,1            ; cmpltps       0x11213(%rip),%xmm9        # 2dc00 <_sk_srcover_rgba_8888_sse2_lowp+0xb4c>
   DB  69,15,84,200                        ; andps         %xmm8,%xmm9
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  65,15,89,201                        ; mulps         %xmm9,%xmm1
@@ -26925,20 +26925,20 @@
 
 PUBLIC _sk_from_srgb_sse2
 _sk_from_srgb_sse2 LABEL PROC
-  DB  68,15,40,5,39,251,0,0               ; movaps        0xfb27(%rip),%xmm8        # 2c530 <_sk_srcover_rgba_8888_sse2_lowp+0xb28>
+  DB  68,15,40,5,7,18,1,0                 ; movaps        0x11207(%rip),%xmm8        # 2dc10 <_sk_srcover_rgba_8888_sse2_lowp+0xb5c>
   DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
   DB  69,15,89,232                        ; mulps         %xmm8,%xmm13
   DB  68,15,40,216                        ; movaps        %xmm0,%xmm11
   DB  69,15,89,219                        ; mulps         %xmm11,%xmm11
-  DB  68,15,40,13,143,250,0,0             ; movaps        0xfa8f(%rip),%xmm9        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,13,111,17,1,0              ; movaps        0x1116f(%rip),%xmm9        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,241                        ; mulps         %xmm9,%xmm14
-  DB  68,15,40,21,15,251,0,0              ; movaps        0xfb0f(%rip),%xmm10        # 2c540 <_sk_srcover_rgba_8888_sse2_lowp+0xb38>
+  DB  68,15,40,21,239,17,1,0              ; movaps        0x111ef(%rip),%xmm10        # 2dc20 <_sk_srcover_rgba_8888_sse2_lowp+0xb6c>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  69,15,89,243                        ; mulps         %xmm11,%xmm14
-  DB  68,15,40,29,15,251,0,0              ; movaps        0xfb0f(%rip),%xmm11        # 2c550 <_sk_srcover_rgba_8888_sse2_lowp+0xb48>
+  DB  68,15,40,29,239,17,1,0              ; movaps        0x111ef(%rip),%xmm11        # 2dc30 <_sk_srcover_rgba_8888_sse2_lowp+0xb7c>
   DB  69,15,88,243                        ; addps         %xmm11,%xmm14
-  DB  68,15,40,37,19,251,0,0              ; movaps        0xfb13(%rip),%xmm12        # 2c560 <_sk_srcover_rgba_8888_sse2_lowp+0xb58>
+  DB  68,15,40,37,243,17,1,0              ; movaps        0x111f3(%rip),%xmm12        # 2dc40 <_sk_srcover_rgba_8888_sse2_lowp+0xb8c>
   DB  65,15,194,196,1                     ; cmpltps       %xmm12,%xmm0
   DB  68,15,84,232                        ; andps         %xmm0,%xmm13
   DB  65,15,85,198                        ; andnps        %xmm14,%xmm0
@@ -26972,20 +26972,20 @@
 
 PUBLIC _sk_from_srgb_dst_sse2
 _sk_from_srgb_dst_sse2 LABEL PROC
-  DB  68,15,40,5,100,250,0,0              ; movaps        0xfa64(%rip),%xmm8        # 2c530 <_sk_srcover_rgba_8888_sse2_lowp+0xb28>
+  DB  68,15,40,5,68,17,1,0                ; movaps        0x11144(%rip),%xmm8        # 2dc10 <_sk_srcover_rgba_8888_sse2_lowp+0xb5c>
   DB  68,15,40,236                        ; movaps        %xmm4,%xmm13
   DB  69,15,89,232                        ; mulps         %xmm8,%xmm13
   DB  68,15,40,220                        ; movaps        %xmm4,%xmm11
   DB  69,15,89,219                        ; mulps         %xmm11,%xmm11
-  DB  68,15,40,13,204,249,0,0             ; movaps        0xf9cc(%rip),%xmm9        # 2c4b0 <_sk_srcover_rgba_8888_sse2_lowp+0xaa8>
+  DB  68,15,40,13,172,16,1,0              ; movaps        0x110ac(%rip),%xmm9        # 2db90 <_sk_srcover_rgba_8888_sse2_lowp+0xadc>
   DB  68,15,40,244                        ; movaps        %xmm4,%xmm14
   DB  69,15,89,241                        ; mulps         %xmm9,%xmm14
-  DB  68,15,40,21,76,250,0,0              ; movaps        0xfa4c(%rip),%xmm10        # 2c540 <_sk_srcover_rgba_8888_sse2_lowp+0xb38>
+  DB  68,15,40,21,44,17,1,0               ; movaps        0x1112c(%rip),%xmm10        # 2dc20 <_sk_srcover_rgba_8888_sse2_lowp+0xb6c>
   DB  69,15,88,242                        ; addps         %xmm10,%xmm14
   DB  69,15,89,243                        ; mulps         %xmm11,%xmm14
-  DB  68,15,40,29,76,250,0,0              ; movaps        0xfa4c(%rip),%xmm11        # 2c550 <_sk_srcover_rgba_8888_sse2_lowp+0xb48>
+  DB  68,15,40,29,44,17,1,0               ; movaps        0x1112c(%rip),%xmm11        # 2dc30 <_sk_srcover_rgba_8888_sse2_lowp+0xb7c>
   DB  69,15,88,243                        ; addps         %xmm11,%xmm14
-  DB  68,15,40,37,80,250,0,0              ; movaps        0xfa50(%rip),%xmm12        # 2c560 <_sk_srcover_rgba_8888_sse2_lowp+0xb58>
+  DB  68,15,40,37,48,17,1,0               ; movaps        0x11130(%rip),%xmm12        # 2dc40 <_sk_srcover_rgba_8888_sse2_lowp+0xb8c>
   DB  65,15,194,228,1                     ; cmpltps       %xmm12,%xmm4
   DB  68,15,84,236                        ; andps         %xmm4,%xmm13
   DB  65,15,85,230                        ; andnps        %xmm14,%xmm4
@@ -27020,22 +27020,22 @@
 PUBLIC _sk_to_srgb_sse2
 _sk_to_srgb_sse2 LABEL PROC
   DB  68,15,82,232                        ; rsqrtps       %xmm0,%xmm13
-  DB  68,15,40,5,221,249,0,0              ; movaps        0xf9dd(%rip),%xmm8        # 2c570 <_sk_srcover_rgba_8888_sse2_lowp+0xb68>
+  DB  68,15,40,5,189,16,1,0               ; movaps        0x110bd(%rip),%xmm8        # 2dc50 <_sk_srcover_rgba_8888_sse2_lowp+0xb9c>
   DB  68,15,40,240                        ; movaps        %xmm0,%xmm14
   DB  69,15,89,240                        ; mulps         %xmm8,%xmm14
-  DB  68,15,40,13,221,249,0,0             ; movaps        0xf9dd(%rip),%xmm9        # 2c580 <_sk_srcover_rgba_8888_sse2_lowp+0xb78>
+  DB  68,15,40,13,189,16,1,0              ; movaps        0x110bd(%rip),%xmm9        # 2dc60 <_sk_srcover_rgba_8888_sse2_lowp+0xbac>
   DB  69,15,40,253                        ; movaps        %xmm13,%xmm15
   DB  69,15,89,249                        ; mulps         %xmm9,%xmm15
-  DB  68,15,40,21,221,249,0,0             ; movaps        0xf9dd(%rip),%xmm10        # 2c590 <_sk_srcover_rgba_8888_sse2_lowp+0xb88>
+  DB  68,15,40,21,189,16,1,0              ; movaps        0x110bd(%rip),%xmm10        # 2dc70 <_sk_srcover_rgba_8888_sse2_lowp+0xbbc>
   DB  69,15,88,250                        ; addps         %xmm10,%xmm15
   DB  69,15,89,253                        ; mulps         %xmm13,%xmm15
-  DB  68,15,40,29,221,249,0,0             ; movaps        0xf9dd(%rip),%xmm11        # 2c5a0 <_sk_srcover_rgba_8888_sse2_lowp+0xb98>
+  DB  68,15,40,29,189,16,1,0              ; movaps        0x110bd(%rip),%xmm11        # 2dc80 <_sk_srcover_rgba_8888_sse2_lowp+0xbcc>
   DB  69,15,88,251                        ; addps         %xmm11,%xmm15
-  DB  68,15,40,37,225,249,0,0             ; movaps        0xf9e1(%rip),%xmm12        # 2c5b0 <_sk_srcover_rgba_8888_sse2_lowp+0xba8>
+  DB  68,15,40,37,193,16,1,0              ; movaps        0x110c1(%rip),%xmm12        # 2dc90 <_sk_srcover_rgba_8888_sse2_lowp+0xbdc>
   DB  69,15,88,236                        ; addps         %xmm12,%xmm13
   DB  69,15,83,237                        ; rcpps         %xmm13,%xmm13
   DB  69,15,89,239                        ; mulps         %xmm15,%xmm13
-  DB  68,15,40,61,221,249,0,0             ; movaps        0xf9dd(%rip),%xmm15        # 2c5c0 <_sk_srcover_rgba_8888_sse2_lowp+0xbb8>
+  DB  68,15,40,61,189,16,1,0              ; movaps        0x110bd(%rip),%xmm15        # 2dca0 <_sk_srcover_rgba_8888_sse2_lowp+0xbec>
   DB  65,15,194,199,1                     ; cmpltps       %xmm15,%xmm0
   DB  68,15,84,240                        ; andps         %xmm0,%xmm14
   DB  65,15,85,197                        ; andnps        %xmm13,%xmm0
@@ -27085,7 +27085,7 @@
   DB  68,15,93,216                        ; minps         %xmm0,%xmm11
   DB  65,15,40,202                        ; movaps        %xmm10,%xmm1
   DB  65,15,92,203                        ; subps         %xmm11,%xmm1
-  DB  68,15,40,45,144,247,0,0             ; movaps        0xf790(%rip),%xmm13        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,45,112,14,1,0              ; movaps        0x10e70(%rip),%xmm13        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  68,15,94,233                        ; divps         %xmm1,%xmm13
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,194,192,0                     ; cmpeqps       %xmm8,%xmm0
@@ -27094,30 +27094,30 @@
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,40,241                        ; movaps        %xmm9,%xmm14
   DB  68,15,194,242,1                     ; cmpltps       %xmm2,%xmm14
-  DB  68,15,84,53,6,249,0,0               ; andps         0xf906(%rip),%xmm14        # 2c5d0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc8>
+  DB  68,15,84,53,230,15,1,0              ; andps         0x10fe6(%rip),%xmm14        # 2dcb0 <_sk_srcover_rgba_8888_sse2_lowp+0xbfc>
   DB  69,15,88,244                        ; addps         %xmm12,%xmm14
   DB  69,15,40,250                        ; movaps        %xmm10,%xmm15
   DB  69,15,194,249,0                     ; cmpeqps       %xmm9,%xmm15
   DB  65,15,92,208                        ; subps         %xmm8,%xmm2
   DB  65,15,89,213                        ; mulps         %xmm13,%xmm2
-  DB  68,15,40,37,249,248,0,0             ; movaps        0xf8f9(%rip),%xmm12        # 2c5e0 <_sk_srcover_rgba_8888_sse2_lowp+0xbd8>
+  DB  68,15,40,37,217,15,1,0              ; movaps        0x10fd9(%rip),%xmm12        # 2dcc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc0c>
   DB  65,15,88,212                        ; addps         %xmm12,%xmm2
   DB  69,15,92,193                        ; subps         %xmm9,%xmm8
   DB  69,15,89,197                        ; mulps         %xmm13,%xmm8
-  DB  68,15,88,5,245,248,0,0              ; addps         0xf8f5(%rip),%xmm8        # 2c5f0 <_sk_srcover_rgba_8888_sse2_lowp+0xbe8>
+  DB  68,15,88,5,213,15,1,0               ; addps         0x10fd5(%rip),%xmm8        # 2dcd0 <_sk_srcover_rgba_8888_sse2_lowp+0xc1c>
   DB  65,15,84,215                        ; andps         %xmm15,%xmm2
   DB  69,15,85,248                        ; andnps        %xmm8,%xmm15
   DB  68,15,86,250                        ; orps          %xmm2,%xmm15
   DB  68,15,84,240                        ; andps         %xmm0,%xmm14
   DB  65,15,85,199                        ; andnps        %xmm15,%xmm0
   DB  65,15,86,198                        ; orps          %xmm14,%xmm0
-  DB  15,89,5,230,248,0,0                 ; mulps         0xf8e6(%rip),%xmm0        # 2c600 <_sk_srcover_rgba_8888_sse2_lowp+0xbf8>
+  DB  15,89,5,198,15,1,0                  ; mulps         0x10fc6(%rip),%xmm0        # 2dce0 <_sk_srcover_rgba_8888_sse2_lowp+0xc2c>
   DB  69,15,40,194                        ; movaps        %xmm10,%xmm8
   DB  69,15,194,195,4                     ; cmpneqps      %xmm11,%xmm8
   DB  65,15,84,192                        ; andps         %xmm8,%xmm0
   DB  69,15,92,226                        ; subps         %xmm10,%xmm12
   DB  69,15,88,211                        ; addps         %xmm11,%xmm10
-  DB  68,15,40,13,233,246,0,0             ; movaps        0xf6e9(%rip),%xmm9        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,13,201,13,1,0              ; movaps        0x10dc9(%rip),%xmm9        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  68,15,194,202,1                     ; cmpltps       %xmm2,%xmm9
@@ -27140,7 +27140,7 @@
   DB  15,41,92,36,32                      ; movaps        %xmm3,0x20(%rsp)
   DB  68,15,40,218                        ; movaps        %xmm2,%xmm11
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
-  DB  68,15,40,13,148,246,0,0             ; movaps        0xf694(%rip),%xmm9        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,13,116,13,1,0              ; movaps        0x10d74(%rip),%xmm9        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,40,209                        ; movaps        %xmm9,%xmm10
   DB  69,15,194,211,2                     ; cmpleps       %xmm11,%xmm10
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
@@ -27157,28 +27157,28 @@
   DB  69,15,88,211                        ; addps         %xmm11,%xmm10
   DB  69,15,88,219                        ; addps         %xmm11,%xmm11
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
-  DB  15,40,5,62,248,0,0                  ; movaps        0xf83e(%rip),%xmm0        # 2c610 <_sk_srcover_rgba_8888_sse2_lowp+0xc08>
+  DB  15,40,5,30,15,1,0                   ; movaps        0x10f1e(%rip),%xmm0        # 2dcf0 <_sk_srcover_rgba_8888_sse2_lowp+0xc3c>
   DB  15,88,198                           ; addps         %xmm6,%xmm0
   DB  243,15,91,200                       ; cvttps2dq     %xmm0,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
   DB  15,194,217,1                        ; cmpltps       %xmm1,%xmm3
-  DB  15,84,29,70,246,0,0                 ; andps         0xf646(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,84,29,38,13,1,0                  ; andps         0x10d26(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,92,193                           ; subps         %xmm1,%xmm0
-  DB  68,15,40,45,40,248,0,0              ; movaps        0xf828(%rip),%xmm13        # 2c620 <_sk_srcover_rgba_8888_sse2_lowp+0xc18>
+  DB  68,15,40,45,8,15,1,0                ; movaps        0x10f08(%rip),%xmm13        # 2dd00 <_sk_srcover_rgba_8888_sse2_lowp+0xc4c>
   DB  69,15,40,197                        ; movaps        %xmm13,%xmm8
   DB  68,15,194,192,2                     ; cmpleps       %xmm0,%xmm8
   DB  69,15,40,242                        ; movaps        %xmm10,%xmm14
   DB  69,15,92,243                        ; subps         %xmm11,%xmm14
   DB  65,15,40,217                        ; movaps        %xmm9,%xmm3
   DB  15,194,216,2                        ; cmpleps       %xmm0,%xmm3
-  DB  15,40,21,232,247,0,0                ; movaps        0xf7e8(%rip),%xmm2        # 2c600 <_sk_srcover_rgba_8888_sse2_lowp+0xbf8>
+  DB  15,40,21,200,14,1,0                 ; movaps        0x10ec8(%rip),%xmm2        # 2dce0 <_sk_srcover_rgba_8888_sse2_lowp+0xc2c>
   DB  68,15,40,250                        ; movaps        %xmm2,%xmm15
   DB  68,15,194,248,2                     ; cmpleps       %xmm0,%xmm15
-  DB  15,40,13,168,247,0,0                ; movaps        0xf7a8(%rip),%xmm1        # 2c5d0 <_sk_srcover_rgba_8888_sse2_lowp+0xbc8>
+  DB  15,40,13,136,14,1,0                 ; movaps        0x10e88(%rip),%xmm1        # 2dcb0 <_sk_srcover_rgba_8888_sse2_lowp+0xbfc>
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,45,190,247,0,0                ; movaps        0xf7be(%rip),%xmm5        # 2c5f0 <_sk_srcover_rgba_8888_sse2_lowp+0xbe8>
+  DB  15,40,45,158,14,1,0                 ; movaps        0x10e9e(%rip),%xmm5        # 2dcd0 <_sk_srcover_rgba_8888_sse2_lowp+0xc1c>
   DB  15,40,229                           ; movaps        %xmm5,%xmm4
   DB  15,92,224                           ; subps         %xmm0,%xmm4
   DB  65,15,89,230                        ; mulps         %xmm14,%xmm4
@@ -27201,7 +27201,7 @@
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,222                           ; movaps        %xmm6,%xmm3
   DB  15,194,216,1                        ; cmpltps       %xmm0,%xmm3
-  DB  15,84,29,163,245,0,0                ; andps         0xf5a3(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,84,29,131,12,1,0                 ; andps         0x10c83(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  68,15,40,230                        ; movaps        %xmm6,%xmm12
   DB  68,15,92,224                        ; subps         %xmm0,%xmm12
@@ -27231,12 +27231,12 @@
   DB  15,40,60,36                         ; movaps        (%rsp),%xmm7
   DB  15,40,231                           ; movaps        %xmm7,%xmm4
   DB  15,85,227                           ; andnps        %xmm3,%xmm4
-  DB  15,88,53,44,247,0,0                 ; addps         0xf72c(%rip),%xmm6        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  15,88,53,12,14,1,0                  ; addps         0x10e0c(%rip),%xmm6        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  243,15,91,198                       ; cvttps2dq     %xmm6,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,222                           ; movaps        %xmm6,%xmm3
   DB  15,194,216,1                        ; cmpltps       %xmm0,%xmm3
-  DB  15,84,29,23,245,0,0                 ; andps         0xf517(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,84,29,247,11,1,0                 ; andps         0x10bf7(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,92,195                           ; subps         %xmm3,%xmm0
   DB  15,92,240                           ; subps         %xmm0,%xmm6
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
@@ -27301,9 +27301,9 @@
   DB  102,71,15,110,4,16                  ; movd          (%r8,%r10,1),%xmm8
   DB  102,68,15,96,192                    ; punpcklbw     %xmm0,%xmm8
   DB  102,68,15,97,192                    ; punpcklwd     %xmm0,%xmm8
-  DB  102,68,15,219,5,230,244,0,0         ; pand          0xf4e6(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,219,5,198,11,1,0          ; pand          0x10bc6(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,58,246,0,0               ; mulps         0xf63a(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,89,5,26,13,1,0                ; mulps         0x10d1a(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
@@ -27346,17 +27346,17 @@
   DB  15,133,173,0,0,0                    ; jne           1d143 <_sk_scale_565_sse2+0xcc>
   DB  243,71,15,126,20,80                 ; movq          (%r8,%r10,2),%xmm10
   DB  102,68,15,97,208                    ; punpcklwd     %xmm0,%xmm10
-  DB  102,68,15,111,5,166,245,0,0         ; movdqa        0xf5a6(%rip),%xmm8        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,68,15,111,5,134,12,1,0          ; movdqa        0x10c86(%rip),%xmm8        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,69,15,219,194                   ; pand          %xmm10,%xmm8
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,165,245,0,0              ; mulps         0xf5a5(%rip),%xmm8        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,68,15,111,13,172,245,0,0        ; movdqa        0xf5ac(%rip),%xmm9        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  68,15,89,5,133,12,1,0               ; mulps         0x10c85(%rip),%xmm8        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,68,15,111,13,140,12,1,0         ; movdqa        0x10c8c(%rip),%xmm9        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,69,15,219,202                   ; pand          %xmm10,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,171,245,0,0             ; mulps         0xf5ab(%rip),%xmm9        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,68,15,219,21,178,245,0,0        ; pand          0xf5b2(%rip),%xmm10        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  68,15,89,13,139,12,1,0              ; mulps         0x10c8b(%rip),%xmm9        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,68,15,219,21,146,12,1,0         ; pand          0x10c92(%rip),%xmm10        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
-  DB  68,15,89,21,182,245,0,0             ; mulps         0xf5b6(%rip),%xmm10        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  68,15,89,21,150,12,1,0              ; mulps         0x10c96(%rip),%xmm10        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
   DB  68,15,194,223,1                     ; cmpltps       %xmm7,%xmm11
   DB  69,15,40,225                        ; movaps        %xmm9,%xmm12
@@ -27433,9 +27433,9 @@
   DB  102,71,15,110,4,16                  ; movd          (%r8,%r10,1),%xmm8
   DB  102,68,15,96,192                    ; punpcklbw     %xmm0,%xmm8
   DB  102,68,15,97,192                    ; punpcklwd     %xmm0,%xmm8
-  DB  102,68,15,219,5,220,242,0,0         ; pand          0xf2dc(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,219,5,188,9,1,0           ; pand          0x109bc(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,48,244,0,0               ; mulps         0xf430(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,89,5,16,11,1,0                ; mulps         0x10b10(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  15,88,196                           ; addps         %xmm4,%xmm0
@@ -27485,17 +27485,17 @@
   DB  15,133,173,0,0,0                    ; jne           1d364 <_sk_lerp_565_sse2+0xcc>
   DB  243,71,15,126,4,80                  ; movq          (%r8,%r10,2),%xmm8
   DB  102,68,15,97,192                    ; punpcklwd     %xmm0,%xmm8
-  DB  102,68,15,111,13,133,243,0,0        ; movdqa        0xf385(%rip),%xmm9        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,68,15,111,13,101,10,1,0         ; movdqa        0x10a65(%rip),%xmm9        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,69,15,219,200                   ; pand          %xmm8,%xmm9
   DB  69,15,91,201                        ; cvtdq2ps      %xmm9,%xmm9
-  DB  68,15,89,13,132,243,0,0             ; mulps         0xf384(%rip),%xmm9        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,68,15,111,21,139,243,0,0        ; movdqa        0xf38b(%rip),%xmm10        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  68,15,89,13,100,10,1,0              ; mulps         0x10a64(%rip),%xmm9        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,68,15,111,21,107,10,1,0         ; movdqa        0x10a6b(%rip),%xmm10        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,69,15,219,208                   ; pand          %xmm8,%xmm10
   DB  69,15,91,218                        ; cvtdq2ps      %xmm10,%xmm11
-  DB  68,15,89,29,138,243,0,0             ; mulps         0xf38a(%rip),%xmm11        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,68,15,219,5,145,243,0,0         ; pand          0xf391(%rip),%xmm8        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  68,15,89,29,106,10,1,0              ; mulps         0x10a6a(%rip),%xmm11        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,68,15,219,5,113,10,1,0          ; pand          0x10a71(%rip),%xmm8        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  69,15,91,192                        ; cvtdq2ps      %xmm8,%xmm8
-  DB  68,15,89,5,149,243,0,0              ; mulps         0xf395(%rip),%xmm8        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  68,15,89,5,117,10,1,0               ; mulps         0x10a75(%rip),%xmm8        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  68,15,40,211                        ; movaps        %xmm3,%xmm10
   DB  68,15,194,215,1                     ; cmpltps       %xmm7,%xmm10
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
@@ -27550,7 +27550,7 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  15,133,31,1,0,0                     ; jne           1d4ec <_sk_load_tables_sse2+0x130>
   DB  243,69,15,111,12,144                ; movdqu        (%r8,%rdx,4),%xmm9
-  DB  102,68,15,111,5,4,241,0,0           ; movdqa        0xf104(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,111,5,228,7,1,0           ; movdqa        0x107e4(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
@@ -27606,7 +27606,7 @@
   DB  102,15,20,211                       ; unpcklpd      %xmm3,%xmm2
   DB  102,65,15,114,209,24                ; psrld         $0x18,%xmm9
   DB  65,15,91,217                        ; cvtdq2ps      %xmm9,%xmm3
-  DB  15,89,29,91,241,0,0                 ; mulps         0xf15b(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,59,8,1,0                   ; mulps         0x1083b(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
@@ -27644,7 +27644,7 @@
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
   DB  102,68,15,105,200                   ; punpckhwd     %xmm0,%xmm9
-  DB  102,15,111,21,53,241,0,0            ; movdqa        0xf135(%rip),%xmm2        # 2c6b0 <_sk_srcover_rgba_8888_sse2_lowp+0xca8>
+  DB  102,15,111,21,21,8,1,0              ; movdqa        0x10815(%rip),%xmm2        # 2dd90 <_sk_srcover_rgba_8888_sse2_lowp+0xcdc>
   DB  102,15,112,217,238                  ; pshufd        $0xee,%xmm1,%xmm3
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
@@ -27705,7 +27705,7 @@
   DB  102,65,15,235,217                   ; por           %xmm9,%xmm3
   DB  102,65,15,97,216                    ; punpcklwd     %xmm8,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,51,240,0,0                 ; mulps         0xf033(%rip),%xmm3        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  15,89,29,19,7,1,0                   ; mulps         0x10713(%rip),%xmm3        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
@@ -27744,7 +27744,7 @@
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
   DB  102,65,15,97,194                    ; punpcklwd     %xmm10,%xmm0
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
-  DB  102,68,15,111,5,119,239,0,0         ; movdqa        0xef77(%rip),%xmm8        # 2c6b0 <_sk_srcover_rgba_8888_sse2_lowp+0xca8>
+  DB  102,68,15,111,5,87,6,1,0            ; movdqa        0x10657(%rip),%xmm8        # 2dd90 <_sk_srcover_rgba_8888_sse2_lowp+0xcdc>
   DB  102,65,15,219,192                   ; pand          %xmm8,%xmm0
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
@@ -27799,7 +27799,7 @@
   DB  65,15,20,216                        ; unpcklps      %xmm8,%xmm3
   DB  102,15,20,211                       ; unpcklpd      %xmm3,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,244,235,0,0                ; movaps        0xebf4(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,212,2,1,0                  ; movaps        0x102d4(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  91                                  ; pop           %rbx
   DB  65,94                               ; pop           %r14
   DB  255,224                             ; jmpq          *%rax
@@ -27829,7 +27829,7 @@
   DB  65,86                               ; push          %r14
   DB  83                                  ; push          %rbx
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,97,236,0,0               ; movaps        0xec61(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,65,3,1,0                 ; movaps        0x10341(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  102,73,15,126,192                   ; movq          %xmm0,%r8
@@ -27857,7 +27857,7 @@
   DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
   DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,21,28,237,0,0              ; movaps        0xed1c(%rip),%xmm10        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,21,252,3,1,0               ; movaps        0x103fc(%rip),%xmm10        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -27974,7 +27974,7 @@
   DB  102,65,15,96,193                    ; punpcklbw     %xmm9,%xmm0
   DB  102,65,15,97,193                    ; punpcklwd     %xmm9,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,21,92,235,0,0              ; movaps        0xeb5c(%rip),%xmm10        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,21,60,2,1,0                ; movaps        0x1023c(%rip),%xmm10        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,194                        ; mulps         %xmm10,%xmm0
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
@@ -28162,15 +28162,15 @@
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,40,233,0,0              ; mulps         0xe928(%rip),%xmm9        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,21,48,233,0,0              ; andps         0xe930(%rip),%xmm10        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,21,88,230,0,0              ; orps          0xe658(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,13,48,233,0,0              ; addps         0xe930(%rip),%xmm9        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  68,15,40,37,56,233,0,0              ; movaps        0xe938(%rip),%xmm12        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,13,8,0,1,0                 ; mulps         0x10008(%rip),%xmm9        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,21,16,0,1,0                ; andps         0x10010(%rip),%xmm10        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,21,56,253,0,0              ; orps          0xfd38(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,13,16,0,1,0                ; addps         0x10010(%rip),%xmm9        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  68,15,40,37,24,0,1,0                ; movaps        0x10018(%rip),%xmm12        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,56,233,0,0              ; addps         0xe938(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  68,15,40,37,64,233,0,0              ; movaps        0xe940(%rip),%xmm12        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,21,24,0,1,0                ; addps         0x10018(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  68,15,40,37,32,0,1,0                ; movaps        0x10020(%rip),%xmm12        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -28178,22 +28178,22 @@
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,26,230,0,0              ; movaps        0xe61a(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,250,252,0,0             ; movaps        0xfcfa(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,14,233,0,0              ; addps         0xe90e(%rip),%xmm9        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  68,15,40,37,22,233,0,0              ; movaps        0xe916(%rip),%xmm12        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,13,238,255,0,0             ; addps         0xffee(%rip),%xmm9        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  68,15,40,37,246,255,0,0             ; movaps        0xfff6(%rip),%xmm12        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,22,233,0,0              ; movaps        0xe916(%rip),%xmm12        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,37,246,255,0,0             ; movaps        0xfff6(%rip),%xmm12        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,26,233,0,0              ; movaps        0xe91a(%rip),%xmm13        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,45,250,255,0,0             ; movaps        0xfffa(%rip),%xmm13        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,26,233,0,0              ; mulps         0xe91a(%rip),%xmm13        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  68,15,89,45,250,255,0,0             ; mulps         0xfffa(%rip),%xmm13        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -28227,15 +28227,15 @@
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,234,231,0,0             ; mulps         0xe7ea(%rip),%xmm9        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,21,242,231,0,0             ; andps         0xe7f2(%rip),%xmm10        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,21,26,229,0,0              ; orps          0xe51a(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,13,242,231,0,0             ; addps         0xe7f2(%rip),%xmm9        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  68,15,40,37,250,231,0,0             ; movaps        0xe7fa(%rip),%xmm12        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,13,202,254,0,0             ; mulps         0xfeca(%rip),%xmm9        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,21,210,254,0,0             ; andps         0xfed2(%rip),%xmm10        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,21,250,251,0,0             ; orps          0xfbfa(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,13,210,254,0,0             ; addps         0xfed2(%rip),%xmm9        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  68,15,40,37,218,254,0,0             ; movaps        0xfeda(%rip),%xmm12        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,250,231,0,0             ; addps         0xe7fa(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  68,15,40,37,2,232,0,0               ; movaps        0xe802(%rip),%xmm12        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,21,218,254,0,0             ; addps         0xfeda(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  68,15,40,37,226,254,0,0             ; movaps        0xfee2(%rip),%xmm12        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -28243,22 +28243,22 @@
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,220,228,0,0             ; movaps        0xe4dc(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,188,251,0,0             ; movaps        0xfbbc(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,208,231,0,0             ; addps         0xe7d0(%rip),%xmm9        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  68,15,40,37,216,231,0,0             ; movaps        0xe7d8(%rip),%xmm12        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,13,176,254,0,0             ; addps         0xfeb0(%rip),%xmm9        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  68,15,40,37,184,254,0,0             ; movaps        0xfeb8(%rip),%xmm12        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,216,231,0,0             ; movaps        0xe7d8(%rip),%xmm12        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,37,184,254,0,0             ; movaps        0xfeb8(%rip),%xmm12        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,220,231,0,0             ; movaps        0xe7dc(%rip),%xmm13        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,45,188,254,0,0             ; movaps        0xfebc(%rip),%xmm13        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,220,231,0,0             ; mulps         0xe7dc(%rip),%xmm13        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  68,15,89,45,188,254,0,0             ; mulps         0xfebc(%rip),%xmm13        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -28292,15 +28292,15 @@
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,172,230,0,0             ; mulps         0xe6ac(%rip),%xmm9        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,21,180,230,0,0             ; andps         0xe6b4(%rip),%xmm10        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,21,220,227,0,0             ; orps          0xe3dc(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,13,180,230,0,0             ; addps         0xe6b4(%rip),%xmm9        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  68,15,40,37,188,230,0,0             ; movaps        0xe6bc(%rip),%xmm12        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,13,140,253,0,0             ; mulps         0xfd8c(%rip),%xmm9        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,21,148,253,0,0             ; andps         0xfd94(%rip),%xmm10        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,21,188,250,0,0             ; orps          0xfabc(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,13,148,253,0,0             ; addps         0xfd94(%rip),%xmm9        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  68,15,40,37,156,253,0,0             ; movaps        0xfd9c(%rip),%xmm12        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,188,230,0,0             ; addps         0xe6bc(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  68,15,40,37,196,230,0,0             ; movaps        0xe6c4(%rip),%xmm12        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,21,156,253,0,0             ; addps         0xfd9c(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  68,15,40,37,164,253,0,0             ; movaps        0xfda4(%rip),%xmm12        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -28308,22 +28308,22 @@
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,158,227,0,0             ; movaps        0xe39e(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,126,250,0,0             ; movaps        0xfa7e(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,146,230,0,0             ; addps         0xe692(%rip),%xmm9        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  68,15,40,37,154,230,0,0             ; movaps        0xe69a(%rip),%xmm12        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,13,114,253,0,0             ; addps         0xfd72(%rip),%xmm9        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  68,15,40,37,122,253,0,0             ; movaps        0xfd7a(%rip),%xmm12        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,154,230,0,0             ; movaps        0xe69a(%rip),%xmm12        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,37,122,253,0,0             ; movaps        0xfd7a(%rip),%xmm12        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,158,230,0,0             ; movaps        0xe69e(%rip),%xmm13        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,45,126,253,0,0             ; movaps        0xfd7e(%rip),%xmm13        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,158,230,0,0             ; mulps         0xe69e(%rip),%xmm13        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  68,15,89,45,126,253,0,0             ; mulps         0xfd7e(%rip),%xmm13        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -28357,15 +28357,15 @@
   DB  69,15,88,209                        ; addps         %xmm9,%xmm10
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,91,202                        ; cvtdq2ps      %xmm10,%xmm9
-  DB  68,15,89,13,110,229,0,0             ; mulps         0xe56e(%rip),%xmm9        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  68,15,84,21,118,229,0,0             ; andps         0xe576(%rip),%xmm10        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,86,21,158,226,0,0             ; orps          0xe29e(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
-  DB  68,15,88,13,118,229,0,0             ; addps         0xe576(%rip),%xmm9        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
-  DB  68,15,40,37,126,229,0,0             ; movaps        0xe57e(%rip),%xmm12        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,89,13,78,252,0,0              ; mulps         0xfc4e(%rip),%xmm9        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  68,15,84,21,86,252,0,0              ; andps         0xfc56(%rip),%xmm10        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,86,21,126,249,0,0             ; orps          0xf97e(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
+  DB  68,15,88,13,86,252,0,0              ; addps         0xfc56(%rip),%xmm9        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
+  DB  68,15,40,37,94,252,0,0              ; movaps        0xfc5e(%rip),%xmm12        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  69,15,89,226                        ; mulps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,88,21,126,229,0,0             ; addps         0xe57e(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
-  DB  68,15,40,37,134,229,0,0             ; movaps        0xe586(%rip),%xmm12        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,88,21,94,252,0,0              ; addps         0xfc5e(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
+  DB  68,15,40,37,102,252,0,0             ; movaps        0xfc66(%rip),%xmm12        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  69,15,94,226                        ; divps         %xmm10,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
   DB  69,15,89,203                        ; mulps         %xmm11,%xmm9
@@ -28373,22 +28373,22 @@
   DB  69,15,91,226                        ; cvtdq2ps      %xmm10,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,194,236,1                     ; cmpltps       %xmm12,%xmm13
-  DB  68,15,40,21,96,226,0,0              ; movaps        0xe260(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,64,249,0,0              ; movaps        0xf940(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,234                        ; andps         %xmm10,%xmm13
   DB  69,15,87,219                        ; xorps         %xmm11,%xmm11
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
   DB  69,15,40,233                        ; movaps        %xmm9,%xmm13
   DB  69,15,92,236                        ; subps         %xmm12,%xmm13
-  DB  68,15,88,13,84,229,0,0              ; addps         0xe554(%rip),%xmm9        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
-  DB  68,15,40,37,92,229,0,0              ; movaps        0xe55c(%rip),%xmm12        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  68,15,88,13,52,252,0,0              ; addps         0xfc34(%rip),%xmm9        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
+  DB  68,15,40,37,60,252,0,0              ; movaps        0xfc3c(%rip),%xmm12        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  69,15,89,229                        ; mulps         %xmm13,%xmm12
   DB  69,15,92,204                        ; subps         %xmm12,%xmm9
-  DB  68,15,40,37,92,229,0,0              ; movaps        0xe55c(%rip),%xmm12        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  68,15,40,37,60,252,0,0              ; movaps        0xfc3c(%rip),%xmm12        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  69,15,92,229                        ; subps         %xmm13,%xmm12
-  DB  68,15,40,45,96,229,0,0              ; movaps        0xe560(%rip),%xmm13        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,45,64,252,0,0              ; movaps        0xfc40(%rip),%xmm13        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  69,15,94,236                        ; divps         %xmm12,%xmm13
   DB  69,15,88,233                        ; addps         %xmm9,%xmm13
-  DB  68,15,89,45,96,229,0,0              ; mulps         0xe560(%rip),%xmm13        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  68,15,89,45,64,252,0,0              ; mulps         0xfc40(%rip),%xmm13        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  102,69,15,91,205                    ; cvtps2dq      %xmm13,%xmm9
   DB  243,68,15,16,96,20                  ; movss         0x14(%rax),%xmm12
   DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
@@ -28412,19 +28412,19 @@
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,91,194                           ; cvtdq2ps      %xmm2,%xmm0
-  DB  15,89,5,104,228,0,0                 ; mulps         0xe468(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  15,84,21,113,228,0,0                ; andps         0xe471(%rip),%xmm2        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  68,15,40,53,153,225,0,0             ; movaps        0xe199(%rip),%xmm14        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,89,5,72,251,0,0                  ; mulps         0xfb48(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  15,84,21,81,251,0,0                 ; andps         0xfb51(%rip),%xmm2        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  68,15,40,53,121,248,0,0             ; movaps        0xf879(%rip),%xmm14        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,86,214                        ; orps          %xmm14,%xmm2
-  DB  68,15,40,37,109,228,0,0             ; movaps        0xe46d(%rip),%xmm12        # 2c700 <_sk_srcover_rgba_8888_sse2_lowp+0xcf8>
+  DB  68,15,40,37,77,251,0,0              ; movaps        0xfb4d(%rip),%xmm12        # 2dde0 <_sk_srcover_rgba_8888_sse2_lowp+0xd2c>
   DB  65,15,88,196                        ; addps         %xmm12,%xmm0
-  DB  68,15,40,29,113,228,0,0             ; movaps        0xe471(%rip),%xmm11        # 2c710 <_sk_srcover_rgba_8888_sse2_lowp+0xd08>
+  DB  68,15,40,29,81,251,0,0              ; movaps        0xfb51(%rip),%xmm11        # 2ddf0 <_sk_srcover_rgba_8888_sse2_lowp+0xd3c>
   DB  15,40,226                           ; movaps        %xmm2,%xmm4
   DB  65,15,89,227                        ; mulps         %xmm11,%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  68,15,40,21,111,228,0,0             ; movaps        0xe46f(%rip),%xmm10        # 2c720 <_sk_srcover_rgba_8888_sse2_lowp+0xd18>
+  DB  68,15,40,21,79,251,0,0              ; movaps        0xfb4f(%rip),%xmm10        # 2de00 <_sk_srcover_rgba_8888_sse2_lowp+0xd4c>
   DB  65,15,88,210                        ; addps         %xmm10,%xmm2
-  DB  68,15,40,45,115,228,0,0             ; movaps        0xe473(%rip),%xmm13        # 2c730 <_sk_srcover_rgba_8888_sse2_lowp+0xd28>
+  DB  68,15,40,45,83,251,0,0              ; movaps        0xfb53(%rip),%xmm13        # 2de10 <_sk_srcover_rgba_8888_sse2_lowp+0xd5c>
   DB  65,15,40,229                        ; movaps        %xmm13,%xmm4
   DB  15,94,226                           ; divps         %xmm2,%xmm4
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -28436,27 +28436,27 @@
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
   DB  15,194,226,1                        ; cmpltps       %xmm2,%xmm4
-  DB  68,15,40,13,67,225,0,0              ; movaps        0xe143(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,13,35,248,0,0              ; movaps        0xf823(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  65,15,84,225                        ; andps         %xmm9,%xmm4
   DB  15,92,212                           ; subps         %xmm4,%xmm2
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
   DB  15,92,226                           ; subps         %xmm2,%xmm4
-  DB  15,40,53,63,228,0,0                 ; movaps        0xe43f(%rip),%xmm6        # 2c740 <_sk_srcover_rgba_8888_sse2_lowp+0xd38>
+  DB  15,40,53,31,251,0,0                 ; movaps        0xfb1f(%rip),%xmm6        # 2de20 <_sk_srcover_rgba_8888_sse2_lowp+0xd6c>
   DB  15,88,198                           ; addps         %xmm6,%xmm0
-  DB  15,40,61,85,228,0,0                 ; movaps        0xe455(%rip),%xmm7        # 2c760 <_sk_srcover_rgba_8888_sse2_lowp+0xd58>
+  DB  15,40,61,53,251,0,0                 ; movaps        0xfb35(%rip),%xmm7        # 2de40 <_sk_srcover_rgba_8888_sse2_lowp+0xd8c>
   DB  15,40,239                           ; movaps        %xmm7,%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,40,212                           ; movaps        %xmm4,%xmm2
-  DB  15,40,37,53,228,0,0                 ; movaps        0xe435(%rip),%xmm4        # 2c750 <_sk_srcover_rgba_8888_sse2_lowp+0xd48>
+  DB  15,40,37,21,251,0,0                 ; movaps        0xfb15(%rip),%xmm4        # 2de30 <_sk_srcover_rgba_8888_sse2_lowp+0xd7c>
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  15,92,194                           ; subps         %xmm2,%xmm0
-  DB  68,15,40,61,71,228,0,0              ; movaps        0xe447(%rip),%xmm15        # 2c770 <_sk_srcover_rgba_8888_sse2_lowp+0xd68>
+  DB  68,15,40,61,39,251,0,0              ; movaps        0xfb27(%rip),%xmm15        # 2de50 <_sk_srcover_rgba_8888_sse2_lowp+0xd9c>
   DB  65,15,40,215                        ; movaps        %xmm15,%xmm2
   DB  15,94,213                           ; divps         %xmm5,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  15,89,5,163,227,0,0                 ; mulps         0xe3a3(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  15,84,13,172,227,0,0                ; andps         0xe3ac(%rip),%xmm1        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
+  DB  15,89,5,131,250,0,0                 ; mulps         0xfa83(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  15,84,13,140,250,0,0                ; andps         0xfa8c(%rip),%xmm1        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
   DB  65,15,86,206                        ; orps          %xmm14,%xmm1
   DB  65,15,88,196                        ; addps         %xmm12,%xmm0
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
@@ -28486,9 +28486,9 @@
   DB  15,94,206                           ; divps         %xmm6,%xmm1
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  15,91,195                           ; cvtdq2ps      %xmm3,%xmm0
-  DB  15,89,5,50,227,0,0                  ; mulps         0xe332(%rip),%xmm0        # 2c6e0 <_sk_srcover_rgba_8888_sse2_lowp+0xcd8>
-  DB  15,84,29,59,227,0,0                 ; andps         0xe33b(%rip),%xmm3        # 2c6f0 <_sk_srcover_rgba_8888_sse2_lowp+0xce8>
-  DB  15,86,29,100,224,0,0                ; orps          0xe064(%rip),%xmm3        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,89,5,18,250,0,0                  ; mulps         0xfa12(%rip),%xmm0        # 2ddc0 <_sk_srcover_rgba_8888_sse2_lowp+0xd0c>
+  DB  15,84,29,27,250,0,0                 ; andps         0xfa1b(%rip),%xmm3        # 2ddd0 <_sk_srcover_rgba_8888_sse2_lowp+0xd1c>
+  DB  15,86,29,68,247,0,0                 ; orps          0xf744(%rip),%xmm3        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,88,196                        ; addps         %xmm12,%xmm0
   DB  68,15,89,219                        ; mulps         %xmm3,%xmm11
   DB  65,15,92,195                        ; subps         %xmm11,%xmm0
@@ -28510,7 +28510,7 @@
   DB  15,92,253                           ; subps         %xmm5,%xmm7
   DB  68,15,94,255                        ; divps         %xmm7,%xmm15
   DB  68,15,88,248                        ; addps         %xmm0,%xmm15
-  DB  15,40,5,113,227,0,0                 ; movaps        0xe371(%rip),%xmm0        # 2c780 <_sk_srcover_rgba_8888_sse2_lowp+0xd78>
+  DB  15,40,5,81,250,0,0                  ; movaps        0xfa51(%rip),%xmm0        # 2de60 <_sk_srcover_rgba_8888_sse2_lowp+0xdac>
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  68,15,89,248                        ; mulps         %xmm0,%xmm15
@@ -28528,29 +28528,29 @@
 
 PUBLIC _sk_lab_to_xyz_sse2
 _sk_lab_to_xyz_sse2 LABEL PROC
-  DB  15,89,5,67,227,0,0                  ; mulps         0xe343(%rip),%xmm0        # 2c790 <_sk_srcover_rgba_8888_sse2_lowp+0xd88>
-  DB  68,15,40,5,187,224,0,0              ; movaps        0xe0bb(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  15,89,5,35,250,0,0                  ; mulps         0xfa23(%rip),%xmm0        # 2de70 <_sk_srcover_rgba_8888_sse2_lowp+0xdbc>
+  DB  68,15,40,5,155,247,0,0              ; movaps        0xf79b(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
-  DB  68,15,40,13,63,227,0,0              ; movaps        0xe33f(%rip),%xmm9        # 2c7a0 <_sk_srcover_rgba_8888_sse2_lowp+0xd98>
+  DB  68,15,40,13,31,250,0,0              ; movaps        0xfa1f(%rip),%xmm9        # 2de80 <_sk_srcover_rgba_8888_sse2_lowp+0xdcc>
   DB  65,15,88,201                        ; addps         %xmm9,%xmm1
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  65,15,88,209                        ; addps         %xmm9,%xmm2
-  DB  15,88,5,60,227,0,0                  ; addps         0xe33c(%rip),%xmm0        # 2c7b0 <_sk_srcover_rgba_8888_sse2_lowp+0xda8>
-  DB  15,89,5,69,227,0,0                  ; mulps         0xe345(%rip),%xmm0        # 2c7c0 <_sk_srcover_rgba_8888_sse2_lowp+0xdb8>
-  DB  15,89,13,78,227,0,0                 ; mulps         0xe34e(%rip),%xmm1        # 2c7d0 <_sk_srcover_rgba_8888_sse2_lowp+0xdc8>
+  DB  15,88,5,28,250,0,0                  ; addps         0xfa1c(%rip),%xmm0        # 2de90 <_sk_srcover_rgba_8888_sse2_lowp+0xddc>
+  DB  15,89,5,37,250,0,0                  ; mulps         0xfa25(%rip),%xmm0        # 2dea0 <_sk_srcover_rgba_8888_sse2_lowp+0xdec>
+  DB  15,89,13,46,250,0,0                 ; mulps         0xfa2e(%rip),%xmm1        # 2deb0 <_sk_srcover_rgba_8888_sse2_lowp+0xdfc>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,21,84,227,0,0                 ; mulps         0xe354(%rip),%xmm2        # 2c7e0 <_sk_srcover_rgba_8888_sse2_lowp+0xdd8>
+  DB  15,89,21,52,250,0,0                 ; mulps         0xfa34(%rip),%xmm2        # 2dec0 <_sk_srcover_rgba_8888_sse2_lowp+0xe0c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  68,15,92,202                        ; subps         %xmm2,%xmm9
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
   DB  68,15,89,225                        ; mulps         %xmm1,%xmm12
-  DB  15,40,21,73,227,0,0                 ; movaps        0xe349(%rip),%xmm2        # 2c7f0 <_sk_srcover_rgba_8888_sse2_lowp+0xde8>
+  DB  15,40,21,41,250,0,0                 ; movaps        0xfa29(%rip),%xmm2        # 2ded0 <_sk_srcover_rgba_8888_sse2_lowp+0xe1c>
   DB  68,15,40,194                        ; movaps        %xmm2,%xmm8
   DB  69,15,194,196,1                     ; cmpltps       %xmm12,%xmm8
-  DB  68,15,40,21,72,227,0,0              ; movaps        0xe348(%rip),%xmm10        # 2c800 <_sk_srcover_rgba_8888_sse2_lowp+0xdf8>
+  DB  68,15,40,21,40,250,0,0              ; movaps        0xfa28(%rip),%xmm10        # 2dee0 <_sk_srcover_rgba_8888_sse2_lowp+0xe2c>
   DB  65,15,88,202                        ; addps         %xmm10,%xmm1
-  DB  68,15,40,29,76,227,0,0              ; movaps        0xe34c(%rip),%xmm11        # 2c810 <_sk_srcover_rgba_8888_sse2_lowp+0xe08>
+  DB  68,15,40,29,44,250,0,0              ; movaps        0xfa2c(%rip),%xmm11        # 2def0 <_sk_srcover_rgba_8888_sse2_lowp+0xe3c>
   DB  65,15,89,203                        ; mulps         %xmm11,%xmm1
   DB  69,15,84,224                        ; andps         %xmm8,%xmm12
   DB  68,15,85,193                        ; andnps        %xmm1,%xmm8
@@ -28574,8 +28574,8 @@
   DB  15,84,194                           ; andps         %xmm2,%xmm0
   DB  65,15,85,209                        ; andnps        %xmm9,%xmm2
   DB  15,86,208                           ; orps          %xmm0,%xmm2
-  DB  68,15,89,5,252,226,0,0              ; mulps         0xe2fc(%rip),%xmm8        # 2c820 <_sk_srcover_rgba_8888_sse2_lowp+0xe18>
-  DB  15,89,21,5,227,0,0                  ; mulps         0xe305(%rip),%xmm2        # 2c830 <_sk_srcover_rgba_8888_sse2_lowp+0xe28>
+  DB  68,15,89,5,220,249,0,0              ; mulps         0xf9dc(%rip),%xmm8        # 2df00 <_sk_srcover_rgba_8888_sse2_lowp+0xe4c>
+  DB  15,89,21,229,249,0,0                ; mulps         0xf9e5(%rip),%xmm2        # 2df10 <_sk_srcover_rgba_8888_sse2_lowp+0xe5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -28593,9 +28593,9 @@
   DB  102,67,15,110,4,16                  ; movd          (%r8,%r10,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  102,15,97,192                       ; punpcklwd     %xmm0,%xmm0
-  DB  102,15,219,5,127,223,0,0            ; pand          0xdf7f(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,5,95,246,0,0             ; pand          0xf65f(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,213,224,0,0                ; mulps         0xe0d5(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,181,247,0,0                ; mulps         0xf7b5(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,87,201                       ; xorpd         %xmm1,%xmm1
@@ -28636,9 +28636,9 @@
   DB  102,67,15,110,36,16                 ; movd          (%r8,%r10,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,37,235,222,0,0           ; pand          0xdeeb(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,37,203,245,0,0           ; pand          0xf5cb(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,252                           ; cvtdq2ps      %xmm4,%xmm7
-  DB  15,89,61,65,224,0,0                 ; mulps         0xe041(%rip),%xmm7        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,61,33,247,0,0                 ; mulps         0xf721(%rip),%xmm7        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  102,15,87,237                       ; xorpd         %xmm5,%xmm5
@@ -28715,7 +28715,7 @@
   DB  102,65,15,96,192                    ; punpcklbw     %xmm8,%xmm0
   DB  102,65,15,97,192                    ; punpcklwd     %xmm8,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,29,27,223,0,0                 ; mulps         0xdf1b(%rip),%xmm3        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,29,251,245,0,0                ; mulps         0xf5fb(%rip),%xmm3        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -28731,7 +28731,7 @@
   DB  77,15,175,193                       ; imul          %r9,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,210                           ; movslq        %edx,%r10
-  DB  68,15,40,5,190,221,0,0              ; movaps        0xddbe(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,158,244,0,0              ; movaps        0xf49e(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,89,195                        ; mulps         %xmm3,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
   DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
@@ -28757,7 +28757,7 @@
   DB  117,213                             ; jne           1e77f <_sk_store_a8_sse2+0x4c>
   DB  102,65,15,197,192,4                 ; pextrw        $0x4,%xmm8,%eax
   DB  67,136,68,16,2                      ; mov           %al,0x2(%r8,%r10,1)
-  DB  102,68,15,219,5,34,221,0,0          ; pand          0xdd22(%rip),%xmm8        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,68,15,219,5,2,244,0,0           ; pand          0xf402(%rip),%xmm8        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
   DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
   DB  102,68,15,126,192                   ; movd          %xmm8,%eax
@@ -28780,11 +28780,11 @@
   DB  102,67,15,110,4,16                  ; movd          (%r8,%r10,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  102,15,97,192                       ; punpcklwd     %xmm0,%xmm0
-  DB  102,15,219,5,211,220,0,0            ; pand          0xdcd3(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,5,179,243,0,0            ; pand          0xf3b3(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,41,222,0,0                  ; mulps         0xde29(%rip),%xmm0        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,5,9,245,0,0                   ; mulps         0xf509(%rip),%xmm0        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,16,220,0,0                 ; movaps        0xdc10(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,240,242,0,0                ; movaps        0xf2f0(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -28823,11 +28823,11 @@
   DB  102,67,15,110,36,16                 ; movd          (%r8,%r10,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,37,60,220,0,0            ; pand          0xdc3c(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,219,37,28,243,0,0            ; pand          0xf31c(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,146,221,0,0                ; mulps         0xdd92(%rip),%xmm4        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,37,114,244,0,0                ; mulps         0xf472(%rip),%xmm4        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,121,219,0,0                ; movaps        0xdb79(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,61,89,242,0,0                 ; movaps        0xf259(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  255,224                             ; jmpq          *%rax
@@ -28902,9 +28902,9 @@
   DB  102,65,15,96,192                    ; punpcklbw     %xmm8,%xmm0
   DB  102,65,15,97,192                    ; punpcklwd     %xmm8,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,105,220,0,0                 ; mulps         0xdc69(%rip),%xmm0        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  15,89,5,73,243,0,0                  ; mulps         0xf349(%rip),%xmm0        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,80,218,0,0                 ; movaps        0xda50(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,48,241,0,0                 ; movaps        0xf130(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
@@ -28922,19 +28922,19 @@
   DB  117,83                              ; jne           1ea56 <_sk_load_565_sse2+0x6e>
   DB  243,67,15,126,20,80                 ; movq          (%r8,%r10,2),%xmm2
   DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
-  DB  102,15,111,5,59,220,0,0             ; movdqa        0xdc3b(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,27,243,0,0             ; movdqa        0xf31b(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,61,220,0,0                  ; mulps         0xdc3d(%rip),%xmm0        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,13,69,220,0,0            ; movdqa        0xdc45(%rip),%xmm1        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,5,29,243,0,0                  ; mulps         0xf31d(%rip),%xmm0        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,13,37,243,0,0            ; movdqa        0xf325(%rip),%xmm1        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,71,220,0,0                 ; mulps         0xdc47(%rip),%xmm1        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,21,79,220,0,0            ; pand          0xdc4f(%rip),%xmm2        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,13,39,243,0,0                 ; mulps         0xf327(%rip),%xmm1        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,21,47,243,0,0            ; pand          0xf32f(%rip),%xmm2        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,85,220,0,0                 ; mulps         0xdc55(%rip),%xmm2        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,21,53,243,0,0                 ; mulps         0xf335(%rip),%xmm2        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,220,217,0,0                ; movaps        0xd9dc(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,188,240,0,0                ; movaps        0xf0bc(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,3                        ; and           $0x3,%r9b
@@ -28969,19 +28969,19 @@
   DB  117,83                              ; jne           1eb12 <_sk_load_565_dst_sse2+0x6e>
   DB  243,67,15,126,52,80                 ; movq          (%r8,%r10,2),%xmm6
   DB  102,15,97,240                       ; punpcklwd     %xmm0,%xmm6
-  DB  102,15,111,37,127,219,0,0           ; movdqa        0xdb7f(%rip),%xmm4        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,37,95,242,0,0            ; movdqa        0xf25f(%rip),%xmm4        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,129,219,0,0                ; mulps         0xdb81(%rip),%xmm4        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,45,137,219,0,0           ; movdqa        0xdb89(%rip),%xmm5        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,37,97,242,0,0                 ; mulps         0xf261(%rip),%xmm4        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,45,105,242,0,0           ; movdqa        0xf269(%rip),%xmm5        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,238                      ; pand          %xmm6,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,45,139,219,0,0                ; mulps         0xdb8b(%rip),%xmm5        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,53,147,219,0,0           ; pand          0xdb93(%rip),%xmm6        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,45,107,242,0,0                ; mulps         0xf26b(%rip),%xmm5        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,53,115,242,0,0           ; pand          0xf273(%rip),%xmm6        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,53,153,219,0,0                ; mulps         0xdb99(%rip),%xmm6        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,53,121,242,0,0                ; mulps         0xf279(%rip),%xmm6        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,32,217,0,0                 ; movaps        0xd920(%rip),%xmm7        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,61,0,240,0,0                  ; movaps        0xf000(%rip),%xmm7        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
   DB  65,128,225,3                        ; and           $0x3,%r9b
@@ -29048,19 +29048,19 @@
   DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
   DB  102,15,196,208,3                    ; pinsrw        $0x3,%eax,%xmm2
   DB  102,65,15,97,208                    ; punpcklwd     %xmm8,%xmm2
-  DB  102,15,111,5,46,218,0,0             ; movdqa        0xda2e(%rip),%xmm0        # 2c650 <_sk_srcover_rgba_8888_sse2_lowp+0xc48>
+  DB  102,15,111,5,14,241,0,0             ; movdqa        0xf10e(%rip),%xmm0        # 2dd30 <_sk_srcover_rgba_8888_sse2_lowp+0xc7c>
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,48,218,0,0                  ; mulps         0xda30(%rip),%xmm0        # 2c660 <_sk_srcover_rgba_8888_sse2_lowp+0xc58>
-  DB  102,15,111,13,56,218,0,0            ; movdqa        0xda38(%rip),%xmm1        # 2c670 <_sk_srcover_rgba_8888_sse2_lowp+0xc68>
+  DB  15,89,5,16,241,0,0                  ; mulps         0xf110(%rip),%xmm0        # 2dd40 <_sk_srcover_rgba_8888_sse2_lowp+0xc8c>
+  DB  102,15,111,13,24,241,0,0            ; movdqa        0xf118(%rip),%xmm1        # 2dd50 <_sk_srcover_rgba_8888_sse2_lowp+0xc9c>
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,58,218,0,0                 ; mulps         0xda3a(%rip),%xmm1        # 2c680 <_sk_srcover_rgba_8888_sse2_lowp+0xc78>
-  DB  102,15,219,21,66,218,0,0            ; pand          0xda42(%rip),%xmm2        # 2c690 <_sk_srcover_rgba_8888_sse2_lowp+0xc88>
+  DB  15,89,13,26,241,0,0                 ; mulps         0xf11a(%rip),%xmm1        # 2dd60 <_sk_srcover_rgba_8888_sse2_lowp+0xcac>
+  DB  102,15,219,21,34,241,0,0            ; pand          0xf122(%rip),%xmm2        # 2dd70 <_sk_srcover_rgba_8888_sse2_lowp+0xcbc>
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,72,218,0,0                 ; mulps         0xda48(%rip),%xmm2        # 2c6a0 <_sk_srcover_rgba_8888_sse2_lowp+0xc98>
+  DB  15,89,21,40,241,0,0                 ; mulps         0xf128(%rip),%xmm2        # 2dd80 <_sk_srcover_rgba_8888_sse2_lowp+0xccc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,207,215,0,0                ; movaps        0xd7cf(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,175,238,0,0                ; movaps        0xeeaf(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
 
 PUBLIC _sk_store_565_sse2
@@ -29072,12 +29072,12 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,210                           ; movslq        %edx,%r10
-  DB  68,15,40,5,207,219,0,0              ; movaps        0xdbcf(%rip),%xmm8        # 2c850 <_sk_srcover_rgba_8888_sse2_lowp+0xe48>
+  DB  68,15,40,5,175,242,0,0              ; movaps        0xf2af(%rip),%xmm8        # 2df30 <_sk_srcover_rgba_8888_sse2_lowp+0xe7c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
   DB  102,65,15,114,241,11                ; pslld         $0xb,%xmm9
-  DB  68,15,40,21,196,219,0,0             ; movaps        0xdbc4(%rip),%xmm10        # 2c860 <_sk_srcover_rgba_8888_sse2_lowp+0xe58>
+  DB  68,15,40,21,164,242,0,0             ; movaps        0xf2a4(%rip),%xmm10        # 2df40 <_sk_srcover_rgba_8888_sse2_lowp+0xe8c>
   DB  68,15,89,209                        ; mulps         %xmm1,%xmm10
   DB  102,69,15,91,210                    ; cvtps2dq      %xmm10,%xmm10
   DB  102,65,15,114,242,5                 ; pslld         $0x5,%xmm10
@@ -29124,21 +29124,21 @@
   DB  117,98                              ; jne           1ed9f <_sk_load_4444_sse2+0x7d>
   DB  243,67,15,126,28,80                 ; movq          (%r8,%r10,2),%xmm3
   DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
-  DB  102,15,111,5,33,219,0,0             ; movdqa        0xdb21(%rip),%xmm0        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,5,1,242,0,0              ; movdqa        0xf201(%rip),%xmm0        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,35,219,0,0                  ; mulps         0xdb23(%rip),%xmm0        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,13,43,219,0,0            ; movdqa        0xdb2b(%rip),%xmm1        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,5,3,242,0,0                   ; mulps         0xf203(%rip),%xmm0        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,13,11,242,0,0            ; movdqa        0xf20b(%rip),%xmm1        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,45,219,0,0                 ; mulps         0xdb2d(%rip),%xmm1        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,21,53,219,0,0            ; movdqa        0xdb35(%rip),%xmm2        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,13,13,242,0,0                 ; mulps         0xf20d(%rip),%xmm1        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,21,21,242,0,0            ; movdqa        0xf215(%rip),%xmm2        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,55,219,0,0                 ; mulps         0xdb37(%rip),%xmm2        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,29,63,219,0,0            ; pand          0xdb3f(%rip),%xmm3        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,21,23,242,0,0                 ; mulps         0xf217(%rip),%xmm2        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,29,31,242,0,0            ; pand          0xf21f(%rip),%xmm3        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,69,219,0,0                 ; mulps         0xdb45(%rip),%xmm3        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,29,37,242,0,0                 ; mulps         0xf225(%rip),%xmm3        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
@@ -29174,21 +29174,21 @@
   DB  117,98                              ; jne           1ee6a <_sk_load_4444_dst_sse2+0x7d>
   DB  243,67,15,126,60,80                 ; movq          (%r8,%r10,2),%xmm7
   DB  102,15,97,248                       ; punpcklwd     %xmm0,%xmm7
-  DB  102,15,111,37,86,218,0,0            ; movdqa        0xda56(%rip),%xmm4        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,37,54,241,0,0            ; movdqa        0xf136(%rip),%xmm4        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,231                      ; pand          %xmm7,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,37,88,218,0,0                 ; mulps         0xda58(%rip),%xmm4        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,45,96,218,0,0            ; movdqa        0xda60(%rip),%xmm5        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,37,56,241,0,0                 ; mulps         0xf138(%rip),%xmm4        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,45,64,241,0,0            ; movdqa        0xf140(%rip),%xmm5        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,239                      ; pand          %xmm7,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,45,98,218,0,0                 ; mulps         0xda62(%rip),%xmm5        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,53,106,218,0,0           ; movdqa        0xda6a(%rip),%xmm6        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,45,66,241,0,0                 ; mulps         0xf142(%rip),%xmm5        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,53,74,241,0,0            ; movdqa        0xf14a(%rip),%xmm6        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,247                      ; pand          %xmm7,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,53,108,218,0,0                ; mulps         0xda6c(%rip),%xmm6        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,61,116,218,0,0           ; pand          0xda74(%rip),%xmm7        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,53,76,241,0,0                 ; mulps         0xf14c(%rip),%xmm6        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,61,84,241,0,0            ; pand          0xf154(%rip),%xmm7        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,255                           ; cvtdq2ps      %xmm7,%xmm7
-  DB  15,89,61,122,218,0,0                ; mulps         0xda7a(%rip),%xmm7        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,61,90,241,0,0                 ; mulps         0xf15a(%rip),%xmm7        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,137,249                          ; mov           %edi,%r9d
@@ -29256,21 +29256,21 @@
   DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
   DB  102,15,196,216,3                    ; pinsrw        $0x3,%eax,%xmm3
   DB  102,65,15,97,216                    ; punpcklwd     %xmm8,%xmm3
-  DB  102,15,111,5,246,216,0,0            ; movdqa        0xd8f6(%rip),%xmm0        # 2c870 <_sk_srcover_rgba_8888_sse2_lowp+0xe68>
+  DB  102,15,111,5,214,239,0,0            ; movdqa        0xefd6(%rip),%xmm0        # 2df50 <_sk_srcover_rgba_8888_sse2_lowp+0xe9c>
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,5,248,216,0,0                 ; mulps         0xd8f8(%rip),%xmm0        # 2c880 <_sk_srcover_rgba_8888_sse2_lowp+0xe78>
-  DB  102,15,111,13,0,217,0,0             ; movdqa        0xd900(%rip),%xmm1        # 2c890 <_sk_srcover_rgba_8888_sse2_lowp+0xe88>
+  DB  15,89,5,216,239,0,0                 ; mulps         0xefd8(%rip),%xmm0        # 2df60 <_sk_srcover_rgba_8888_sse2_lowp+0xeac>
+  DB  102,15,111,13,224,239,0,0           ; movdqa        0xefe0(%rip),%xmm1        # 2df70 <_sk_srcover_rgba_8888_sse2_lowp+0xebc>
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,13,2,217,0,0                  ; mulps         0xd902(%rip),%xmm1        # 2c8a0 <_sk_srcover_rgba_8888_sse2_lowp+0xe98>
-  DB  102,15,111,21,10,217,0,0            ; movdqa        0xd90a(%rip),%xmm2        # 2c8b0 <_sk_srcover_rgba_8888_sse2_lowp+0xea8>
+  DB  15,89,13,226,239,0,0                ; mulps         0xefe2(%rip),%xmm1        # 2df80 <_sk_srcover_rgba_8888_sse2_lowp+0xecc>
+  DB  102,15,111,21,234,239,0,0           ; movdqa        0xefea(%rip),%xmm2        # 2df90 <_sk_srcover_rgba_8888_sse2_lowp+0xedc>
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,21,12,217,0,0                 ; mulps         0xd90c(%rip),%xmm2        # 2c8c0 <_sk_srcover_rgba_8888_sse2_lowp+0xeb8>
-  DB  102,15,219,29,20,217,0,0            ; pand          0xd914(%rip),%xmm3        # 2c8d0 <_sk_srcover_rgba_8888_sse2_lowp+0xec8>
+  DB  15,89,21,236,239,0,0                ; mulps         0xefec(%rip),%xmm2        # 2dfa0 <_sk_srcover_rgba_8888_sse2_lowp+0xeec>
+  DB  102,15,219,29,244,239,0,0           ; pand          0xeff4(%rip),%xmm3        # 2dfb0 <_sk_srcover_rgba_8888_sse2_lowp+0xefc>
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,29,26,217,0,0                 ; mulps         0xd91a(%rip),%xmm3        # 2c8e0 <_sk_srcover_rgba_8888_sse2_lowp+0xed8>
+  DB  15,89,29,250,239,0,0                ; mulps         0xeffa(%rip),%xmm3        # 2dfc0 <_sk_srcover_rgba_8888_sse2_lowp+0xf0c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
@@ -29283,7 +29283,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  76,99,210                           ; movslq        %edx,%r10
-  DB  68,15,40,5,8,217,0,0                ; movaps        0xd908(%rip),%xmm8        # 2c8f0 <_sk_srcover_rgba_8888_sse2_lowp+0xee8>
+  DB  68,15,40,5,232,239,0,0              ; movaps        0xefe8(%rip),%xmm8        # 2dfd0 <_sk_srcover_rgba_8888_sse2_lowp+0xf1c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -29339,11 +29339,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,98                              ; jne           1f11b <_sk_load_8888_sse2+0x7e>
   DB  243,69,15,111,12,128                ; movdqu        (%r8,%rax,4),%xmm9
-  DB  102,15,111,21,25,212,0,0            ; movdqa        0xd419(%rip),%xmm2        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,21,249,234,0,0           ; movdqa        0xeaf9(%rip),%xmm2        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,101,213,0,0              ; movaps        0xd565(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,69,236,0,0               ; movaps        0xec45(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -29388,11 +29388,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,98                              ; jne           1f1da <_sk_load_8888_dst_sse2+0x7e>
   DB  243,69,15,111,12,128                ; movdqu        (%r8,%rax,4),%xmm9
-  DB  102,15,111,53,90,211,0,0            ; movdqa        0xd35a(%rip),%xmm6        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,53,58,234,0,0            ; movdqa        0xea3a(%rip),%xmm6        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,225                   ; movdqa        %xmm9,%xmm4
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  68,15,40,5,166,212,0,0              ; movaps        0xd4a6(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,134,235,0,0              ; movaps        0xeb86(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,224                        ; mulps         %xmm8,%xmm4
   DB  102,65,15,111,233                   ; movdqa        %xmm9,%xmm5
   DB  102,15,114,213,8                    ; psrld         $0x8,%xmm5
@@ -29467,11 +29467,11 @@
   DB  102,67,15,110,12,136                ; movd          (%r8,%r9,4),%xmm1
   DB  102,15,98,200                       ; punpckldq     %xmm0,%xmm1
   DB  102,68,15,108,201                   ; punpcklqdq    %xmm1,%xmm9
-  DB  102,15,111,21,13,210,0,0            ; movdqa        0xd20d(%rip),%xmm2        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,21,237,232,0,0           ; movdqa        0xe8ed(%rip),%xmm2        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  68,15,40,5,89,211,0,0               ; movaps        0xd359(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,57,234,0,0               ; movaps        0xea39(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -29498,7 +29498,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,202,209,0,0              ; movaps        0xd1ca(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,170,232,0,0              ; movaps        0xe8aa(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -29548,11 +29548,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,98                              ; jne           1f45a <_sk_load_bgra_sse2+0x7e>
   DB  243,69,15,111,12,128                ; movdqu        (%r8,%rax,4),%xmm9
-  DB  102,15,111,5,218,208,0,0            ; movdqa        0xd0da(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,186,231,0,0            ; movdqa        0xe7ba(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
   DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
-  DB  68,15,40,5,38,210,0,0               ; movaps        0xd226(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,6,233,0,0                ; movaps        0xe906(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -29597,11 +29597,11 @@
   DB  72,133,255                          ; test          %rdi,%rdi
   DB  117,98                              ; jne           1f519 <_sk_load_bgra_dst_sse2+0x7e>
   DB  243,69,15,111,12,128                ; movdqu        (%r8,%rax,4),%xmm9
-  DB  102,15,111,37,27,208,0,0            ; movdqa        0xd01b(%rip),%xmm4        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,37,251,230,0,0           ; movdqa        0xe6fb(%rip),%xmm4        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,233                   ; movdqa        %xmm9,%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,245                           ; cvtdq2ps      %xmm5,%xmm6
-  DB  68,15,40,5,103,209,0,0              ; movaps        0xd167(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,71,232,0,0               ; movaps        0xe847(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,240                        ; mulps         %xmm8,%xmm6
   DB  102,65,15,111,233                   ; movdqa        %xmm9,%xmm5
   DB  102,15,114,213,8                    ; psrld         $0x8,%xmm5
@@ -29676,11 +29676,11 @@
   DB  102,67,15,110,12,136                ; movd          (%r8,%r9,4),%xmm1
   DB  102,15,98,200                       ; punpckldq     %xmm0,%xmm1
   DB  102,68,15,108,201                   ; punpcklqdq    %xmm1,%xmm9
-  DB  102,15,111,5,206,206,0,0            ; movdqa        0xcece(%rip),%xmm0        # 2c4e0 <_sk_srcover_rgba_8888_sse2_lowp+0xad8>
+  DB  102,15,111,5,174,229,0,0            ; movdqa        0xe5ae(%rip),%xmm0        # 2dbc0 <_sk_srcover_rgba_8888_sse2_lowp+0xb0c>
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
   DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
-  DB  68,15,40,5,26,208,0,0               ; movaps        0xd01a(%rip),%xmm8        # 2c640 <_sk_srcover_rgba_8888_sse2_lowp+0xc38>
+  DB  68,15,40,5,250,230,0,0              ; movaps        0xe6fa(%rip),%xmm8        # 2dd20 <_sk_srcover_rgba_8888_sse2_lowp+0xc6c>
   DB  65,15,89,208                        ; mulps         %xmm8,%xmm2
   DB  102,65,15,111,201                   ; movdqa        %xmm9,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -29707,7 +29707,7 @@
   DB  73,193,224,2                        ; shl           $0x2,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  72,99,194                           ; movslq        %edx,%rax
-  DB  68,15,40,5,139,206,0,0              ; movaps        0xce8b(%rip),%xmm8        # 2c510 <_sk_srcover_rgba_8888_sse2_lowp+0xb08>
+  DB  68,15,40,5,107,229,0,0              ; movaps        0xe56b(%rip),%xmm8        # 2dbf0 <_sk_srcover_rgba_8888_sse2_lowp+0xb3c>
   DB  68,15,40,202                        ; movaps        %xmm2,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  102,69,15,91,201                    ; cvtps2dq      %xmm9,%xmm9
@@ -29767,18 +29767,18 @@
   DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,97,202                    ; punpcklwd     %xmm10,%xmm1
-  DB  102,68,15,111,13,133,209,0,0        ; movdqa        0xd185(%rip),%xmm9        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,13,101,232,0,0        ; movdqa        0xe865(%rip),%xmm9        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,68,15,111,225                   ; movdqa        %xmm1,%xmm12
   DB  102,69,15,219,225                   ; pand          %xmm9,%xmm12
-  DB  102,68,15,111,29,130,209,0,0        ; movdqa        0xd182(%rip),%xmm11        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,29,98,232,0,0         ; movdqa        0xe862(%rip),%xmm11        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,203                   ; pand          %xmm11,%xmm1
-  DB  102,15,111,29,133,209,0,0           ; movdqa        0xd185(%rip),%xmm3        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,29,101,232,0,0           ; movdqa        0xe865(%rip),%xmm3        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,65,15,114,244,16                ; pslld         $0x10,%xmm12
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,102,193                      ; pcmpgtd       %xmm1,%xmm0
   DB  102,15,114,241,13                   ; pslld         $0xd,%xmm1
   DB  102,65,15,235,204                   ; por           %xmm12,%xmm1
-  DB  102,68,15,111,37,116,209,0,0        ; movdqa        0xd174(%rip),%xmm12        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,37,84,232,0,0         ; movdqa        0xe854(%rip),%xmm12        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,204                   ; paddd         %xmm12,%xmm1
   DB  102,15,223,193                      ; pandn         %xmm1,%xmm0
   DB  102,65,15,105,210                   ; punpckhwd     %xmm10,%xmm2
@@ -29850,18 +29850,18 @@
   DB  102,69,15,239,210                   ; pxor          %xmm10,%xmm10
   DB  102,15,111,238                      ; movdqa        %xmm6,%xmm5
   DB  102,65,15,97,234                    ; punpcklwd     %xmm10,%xmm5
-  DB  102,68,15,111,13,252,207,0,0        ; movdqa        0xcffc(%rip),%xmm9        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,13,220,230,0,0        ; movdqa        0xe6dc(%rip),%xmm9        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,68,15,111,229                   ; movdqa        %xmm5,%xmm12
   DB  102,69,15,219,225                   ; pand          %xmm9,%xmm12
-  DB  102,68,15,111,29,249,207,0,0        ; movdqa        0xcff9(%rip),%xmm11        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,29,217,230,0,0        ; movdqa        0xe6d9(%rip),%xmm11        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,235                   ; pand          %xmm11,%xmm5
-  DB  102,15,111,61,252,207,0,0           ; movdqa        0xcffc(%rip),%xmm7        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,61,220,230,0,0           ; movdqa        0xe6dc(%rip),%xmm7        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,65,15,114,244,16                ; pslld         $0x10,%xmm12
   DB  102,15,111,231                      ; movdqa        %xmm7,%xmm4
   DB  102,15,102,229                      ; pcmpgtd       %xmm5,%xmm4
   DB  102,15,114,245,13                   ; pslld         $0xd,%xmm5
   DB  102,65,15,235,236                   ; por           %xmm12,%xmm5
-  DB  102,68,15,111,37,235,207,0,0        ; movdqa        0xcfeb(%rip),%xmm12        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,37,203,230,0,0        ; movdqa        0xe6cb(%rip),%xmm12        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,236                   ; paddd         %xmm12,%xmm5
   DB  102,15,223,229                      ; pandn         %xmm5,%xmm4
   DB  102,65,15,105,242                   ; punpckhwd     %xmm10,%xmm6
@@ -29961,18 +29961,18 @@
   DB  102,68,15,105,201                   ; punpckhwd     %xmm1,%xmm9
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,65,15,97,200                    ; punpcklwd     %xmm8,%xmm1
-  DB  102,68,15,111,21,250,205,0,0        ; movdqa        0xcdfa(%rip),%xmm10        # 2c900 <_sk_srcover_rgba_8888_sse2_lowp+0xef8>
+  DB  102,68,15,111,21,218,228,0,0        ; movdqa        0xe4da(%rip),%xmm10        # 2dfe0 <_sk_srcover_rgba_8888_sse2_lowp+0xf2c>
   DB  102,68,15,111,225                   ; movdqa        %xmm1,%xmm12
   DB  102,69,15,219,226                   ; pand          %xmm10,%xmm12
-  DB  102,68,15,111,29,247,205,0,0        ; movdqa        0xcdf7(%rip),%xmm11        # 2c910 <_sk_srcover_rgba_8888_sse2_lowp+0xf08>
+  DB  102,68,15,111,29,215,228,0,0        ; movdqa        0xe4d7(%rip),%xmm11        # 2dff0 <_sk_srcover_rgba_8888_sse2_lowp+0xf3c>
   DB  102,65,15,219,203                   ; pand          %xmm11,%xmm1
-  DB  102,15,111,29,250,205,0,0           ; movdqa        0xcdfa(%rip),%xmm3        # 2c920 <_sk_srcover_rgba_8888_sse2_lowp+0xf18>
+  DB  102,15,111,29,218,228,0,0           ; movdqa        0xe4da(%rip),%xmm3        # 2e000 <_sk_srcover_rgba_8888_sse2_lowp+0xf4c>
   DB  102,65,15,114,244,16                ; pslld         $0x10,%xmm12
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,102,193                      ; pcmpgtd       %xmm1,%xmm0
   DB  102,15,114,241,13                   ; pslld         $0xd,%xmm1
   DB  102,65,15,235,204                   ; por           %xmm12,%xmm1
-  DB  102,68,15,111,37,233,205,0,0        ; movdqa        0xcde9(%rip),%xmm12        # 2c930 <_sk_srcover_rgba_8888_sse2_lowp+0xf28>
+  DB  102,68,15,111,37,201,228,0,0        ; movdqa        0xe4c9(%rip),%xmm12        # 2e010 <_sk_srcover_rgba_8888_sse2_lowp+0xf5c>
   DB  102,65,15,254,204                   ; paddd         %xmm12,%xmm1
   DB  102,15,223,193                      ; pandn         %xmm1,%xmm0
   DB  102,65,15,105,208                   ; punpckhwd     %xmm8,%xmm2
@@ -30014,19 +30014,19 @@
 _sk_store_f16_sse2 LABEL PROC
   DB  72,131,236,24                       ; sub           $0x18,%rsp
   DB  15,41,60,36                         ; movaps        %xmm7,(%rsp)
-  DB  102,68,15,111,29,58,205,0,0         ; movdqa        0xcd3a(%rip),%xmm11        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
+  DB  102,68,15,111,29,26,228,0,0         ; movdqa        0xe41a(%rip),%xmm11        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
   DB  102,69,15,219,195                   ; pand          %xmm11,%xmm8
-  DB  102,68,15,111,21,55,205,0,0         ; movdqa        0xcd37(%rip),%xmm10        # 2c950 <_sk_srcover_rgba_8888_sse2_lowp+0xf48>
+  DB  102,68,15,111,21,23,228,0,0         ; movdqa        0xe417(%rip),%xmm10        # 2e030 <_sk_srcover_rgba_8888_sse2_lowp+0xf7c>
   DB  102,68,15,111,240                   ; movdqa        %xmm0,%xmm14
   DB  102,69,15,219,242                   ; pand          %xmm10,%xmm14
-  DB  102,15,111,61,53,205,0,0            ; movdqa        0xcd35(%rip),%xmm7        # 2c960 <_sk_srcover_rgba_8888_sse2_lowp+0xf58>
-  DB  102,68,15,111,37,220,206,0,0        ; movdqa        0xcedc(%rip),%xmm12        # 2cb10 <_sk_srcover_rgba_8888_sse2_lowp+0x1108>
+  DB  102,15,111,61,21,228,0,0            ; movdqa        0xe415(%rip),%xmm7        # 2e040 <_sk_srcover_rgba_8888_sse2_lowp+0xf8c>
+  DB  102,68,15,111,37,188,229,0,0        ; movdqa        0xe5bc(%rip),%xmm12        # 2e1f0 <_sk_srcover_rgba_8888_sse2_lowp+0x113c>
   DB  102,68,15,111,248                   ; movdqa        %xmm0,%xmm15
   DB  102,65,15,114,247,3                 ; pslld         $0x3,%xmm15
   DB  102,69,15,219,252                   ; pand          %xmm12,%xmm15
   DB  102,69,15,254,248                   ; paddd         %xmm8,%xmm15
-  DB  102,68,15,111,45,142,201,0,0        ; movdqa        0xc98e(%rip),%xmm13        # 2c5e0 <_sk_srcover_rgba_8888_sse2_lowp+0xbd8>
+  DB  102,68,15,111,45,110,224,0,0        ; movdqa        0xe06e(%rip),%xmm13        # 2dcc0 <_sk_srcover_rgba_8888_sse2_lowp+0xc0c>
   DB  102,69,15,254,253                   ; paddd         %xmm13,%xmm15
   DB  102,65,15,114,231,16                ; psrad         $0x10,%xmm15
   DB  102,68,15,111,199                   ; movdqa        %xmm7,%xmm8
@@ -30132,7 +30132,7 @@
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,65,15,97,201                    ; punpcklwd     %xmm9,%xmm1
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  68,15,40,5,160,200,0,0              ; movaps        0xc8a0(%rip),%xmm8        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  68,15,40,5,128,223,0,0              ; movaps        0xdf80(%rip),%xmm8        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -30205,7 +30205,7 @@
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,65,15,97,200                    ; punpcklwd     %xmm8,%xmm1
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  68,15,40,13,106,199,0,0             ; movaps        0xc76a(%rip),%xmm9        # 2c6d0 <_sk_srcover_rgba_8888_sse2_lowp+0xcc8>
+  DB  68,15,40,13,74,222,0,0              ; movaps        0xde4a(%rip),%xmm9        # 2ddb0 <_sk_srcover_rgba_8888_sse2_lowp+0xcfc>
   DB  65,15,89,193                        ; mulps         %xmm9,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -30222,7 +30222,7 @@
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,129,196,0,0                ; movaps        0xc481(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,97,219,0,0                 ; movaps        0xdb61(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,20,64                 ; movd          (%r8,%rax,2),%xmm2
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
@@ -30254,7 +30254,7 @@
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  73,99,193                           ; movslq        %r9d,%rax
-  DB  68,15,40,21,74,201,0,0              ; movaps        0xc94a(%rip),%xmm10        # 2c980 <_sk_srcover_rgba_8888_sse2_lowp+0xf78>
+  DB  68,15,40,21,42,224,0,0              ; movaps        0xe02a(%rip),%xmm10        # 2e060 <_sk_srcover_rgba_8888_sse2_lowp+0xfac>
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
   DB  102,69,15,91,192                    ; cvtps2dq      %xmm8,%xmm8
@@ -30453,7 +30453,7 @@
   DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,84,13,233,192,0,0             ; andps         0xc0e9(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,84,13,201,215,0,0             ; andps         0xd7c9(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
@@ -30471,7 +30471,7 @@
   DB  243,69,15,91,209                    ; cvttps2dq     %xmm9,%xmm10
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,84,13,168,192,0,0             ; andps         0xc0a8(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,84,13,136,215,0,0             ; andps         0xd788(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  69,15,89,194                        ; mulps         %xmm10,%xmm8
@@ -30489,13 +30489,13 @@
   DB  65,15,92,194                        ; subps         %xmm10,%xmm0
   DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
-  DB  243,68,15,89,13,234,184,0,0         ; mulss         0xb8ea(%rip),%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,13,150,207,0,0         ; mulss         0xcf96(%rip),%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
   DB  243,69,15,91,217                    ; cvttps2dq     %xmm9,%xmm11
   DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
   DB  69,15,194,203,1                     ; cmpltps       %xmm11,%xmm9
-  DB  68,15,84,13,71,192,0,0              ; andps         0xc047(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,84,13,39,215,0,0              ; andps         0xd727(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,217                        ; subps         %xmm9,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -30516,13 +30516,13 @@
   DB  65,15,92,202                        ; subps         %xmm10,%xmm1
   DB  243,69,15,88,192                    ; addss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
-  DB  243,68,15,89,13,126,184,0,0         ; mulss         0xb87e(%rip),%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,13,42,207,0,0          ; mulss         0xcf2a(%rip),%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
   DB  243,69,15,91,217                    ; cvttps2dq     %xmm9,%xmm11
   DB  69,15,91,219                        ; cvtdq2ps      %xmm11,%xmm11
   DB  69,15,194,203,1                     ; cmpltps       %xmm11,%xmm9
-  DB  68,15,84,13,219,191,0,0             ; andps         0xbfdb(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,84,13,187,214,0,0             ; andps         0xd6bb(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,87,228                        ; xorps         %xmm12,%xmm12
   DB  69,15,92,217                        ; subps         %xmm9,%xmm11
   DB  69,15,89,216                        ; mulps         %xmm8,%xmm11
@@ -30537,7 +30537,7 @@
 _sk_clamp_x_1_sse2 LABEL PROC
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
   DB  68,15,95,192                        ; maxps         %xmm0,%xmm8
-  DB  68,15,93,5,171,191,0,0              ; minps         0xbfab(%rip),%xmm8        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,93,5,139,214,0,0              ; minps         0xd68b(%rip),%xmm8        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  65,15,40,192                        ; movaps        %xmm8,%xmm0
   DB  255,224                             ; jmpq          *%rax
@@ -30548,7 +30548,7 @@
   DB  69,15,91,200                        ; cvtdq2ps      %xmm8,%xmm9
   DB  68,15,40,208                        ; movaps        %xmm0,%xmm10
   DB  69,15,194,209,1                     ; cmpltps       %xmm9,%xmm10
-  DB  68,15,40,29,137,191,0,0             ; movaps        0xbf89(%rip),%xmm11        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,29,105,214,0,0             ; movaps        0xd669(%rip),%xmm11        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,211                        ; andps         %xmm11,%xmm10
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
@@ -30561,14 +30561,14 @@
 
 PUBLIC _sk_mirror_x_1_sse2
 _sk_mirror_x_1_sse2 LABEL PROC
-  DB  68,15,40,13,193,191,0,0             ; movaps        0xbfc1(%rip),%xmm9        # 2c490 <_sk_srcover_rgba_8888_sse2_lowp+0xa88>
+  DB  68,15,40,13,161,214,0,0             ; movaps        0xd6a1(%rip),%xmm9        # 2db70 <_sk_srcover_rgba_8888_sse2_lowp+0xabc>
   DB  65,15,88,193                        ; addps         %xmm9,%xmm0
-  DB  68,15,40,21,69,191,0,0              ; movaps        0xbf45(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,21,37,214,0,0              ; movaps        0xd625(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
   DB  243,69,15,91,194                    ; cvttps2dq     %xmm10,%xmm8
   DB  69,15,91,216                        ; cvtdq2ps      %xmm8,%xmm11
   DB  69,15,194,211,1                     ; cmpltps       %xmm11,%xmm10
-  DB  68,15,40,37,59,191,0,0              ; movaps        0xbf3b(%rip),%xmm12        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,37,27,214,0,0              ; movaps        0xd61b(%rip),%xmm12        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,212                        ; andps         %xmm12,%xmm10
   DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
   DB  69,15,92,218                        ; subps         %xmm10,%xmm11
@@ -30587,10 +30587,10 @@
 PUBLIC _sk_luminance_to_alpha_sse2
 _sk_luminance_to_alpha_sse2 LABEL PROC
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
-  DB  15,89,5,93,196,0,0                  ; mulps         0xc45d(%rip),%xmm0        # 2c990 <_sk_srcover_rgba_8888_sse2_lowp+0xf88>
-  DB  15,89,13,102,196,0,0                ; mulps         0xc466(%rip),%xmm1        # 2c9a0 <_sk_srcover_rgba_8888_sse2_lowp+0xf98>
+  DB  15,89,5,61,219,0,0                  ; mulps         0xdb3d(%rip),%xmm0        # 2e070 <_sk_srcover_rgba_8888_sse2_lowp+0xfbc>
+  DB  15,89,13,70,219,0,0                 ; mulps         0xdb46(%rip),%xmm1        # 2e080 <_sk_srcover_rgba_8888_sse2_lowp+0xfcc>
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,29,108,196,0,0                ; mulps         0xc46c(%rip),%xmm3        # 2c9b0 <_sk_srcover_rgba_8888_sse2_lowp+0xfa8>
+  DB  15,89,29,76,219,0,0                 ; mulps         0xdb4c(%rip),%xmm3        # 2e090 <_sk_srcover_rgba_8888_sse2_lowp+0xfdc>
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -30984,15 +30984,15 @@
 
 PUBLIC _sk_gauss_a_to_rgba_sse2
 _sk_gauss_a_to_rgba_sse2 LABEL PROC
-  DB  15,40,5,52,190,0,0                  ; movaps        0xbe34(%rip),%xmm0        # 2c9c0 <_sk_srcover_rgba_8888_sse2_lowp+0xfb8>
+  DB  15,40,5,20,213,0,0                  ; movaps        0xd514(%rip),%xmm0        # 2e0a0 <_sk_srcover_rgba_8888_sse2_lowp+0xfec>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,58,190,0,0                  ; addps         0xbe3a(%rip),%xmm0        # 2c9d0 <_sk_srcover_rgba_8888_sse2_lowp+0xfc8>
+  DB  15,88,5,26,213,0,0                  ; addps         0xd51a(%rip),%xmm0        # 2e0b0 <_sk_srcover_rgba_8888_sse2_lowp+0xffc>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,64,190,0,0                  ; addps         0xbe40(%rip),%xmm0        # 2c9e0 <_sk_srcover_rgba_8888_sse2_lowp+0xfd8>
+  DB  15,88,5,32,213,0,0                  ; addps         0xd520(%rip),%xmm0        # 2e0c0 <_sk_srcover_rgba_8888_sse2_lowp+0x100c>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,70,190,0,0                  ; addps         0xbe46(%rip),%xmm0        # 2c9f0 <_sk_srcover_rgba_8888_sse2_lowp+0xfe8>
+  DB  15,88,5,38,213,0,0                  ; addps         0xd526(%rip),%xmm0        # 2e0d0 <_sk_srcover_rgba_8888_sse2_lowp+0x101c>
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,5,76,190,0,0                  ; addps         0xbe4c(%rip),%xmm0        # 2ca00 <_sk_srcover_rgba_8888_sse2_lowp+0xff8>
+  DB  15,88,5,44,213,0,0                  ; addps         0xd52c(%rip),%xmm0        # 2e0e0 <_sk_srcover_rgba_8888_sse2_lowp+0x102c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
@@ -31152,29 +31152,29 @@
   DB  69,15,94,220                        ; divps         %xmm12,%xmm11
   DB  69,15,40,227                        ; movaps        %xmm11,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
-  DB  68,15,40,45,220,187,0,0             ; movaps        0xbbdc(%rip),%xmm13        # 2ca10 <_sk_srcover_rgba_8888_sse2_lowp+0x1008>
+  DB  68,15,40,45,188,210,0,0             ; movaps        0xd2bc(%rip),%xmm13        # 2e0f0 <_sk_srcover_rgba_8888_sse2_lowp+0x103c>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,224,187,0,0             ; addps         0xbbe0(%rip),%xmm13        # 2ca20 <_sk_srcover_rgba_8888_sse2_lowp+0x1018>
+  DB  68,15,88,45,192,210,0,0             ; addps         0xd2c0(%rip),%xmm13        # 2e100 <_sk_srcover_rgba_8888_sse2_lowp+0x104c>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,228,187,0,0             ; addps         0xbbe4(%rip),%xmm13        # 2ca30 <_sk_srcover_rgba_8888_sse2_lowp+0x1028>
+  DB  68,15,88,45,196,210,0,0             ; addps         0xd2c4(%rip),%xmm13        # 2e110 <_sk_srcover_rgba_8888_sse2_lowp+0x105c>
   DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
-  DB  68,15,88,45,232,187,0,0             ; addps         0xbbe8(%rip),%xmm13        # 2ca40 <_sk_srcover_rgba_8888_sse2_lowp+0x1038>
+  DB  68,15,88,45,200,210,0,0             ; addps         0xd2c8(%rip),%xmm13        # 2e120 <_sk_srcover_rgba_8888_sse2_lowp+0x106c>
   DB  69,15,89,235                        ; mulps         %xmm11,%xmm13
   DB  69,15,194,202,1                     ; cmpltps       %xmm10,%xmm9
-  DB  68,15,40,21,231,187,0,0             ; movaps        0xbbe7(%rip),%xmm10        # 2ca50 <_sk_srcover_rgba_8888_sse2_lowp+0x1048>
+  DB  68,15,40,21,199,210,0,0             ; movaps        0xd2c7(%rip),%xmm10        # 2e130 <_sk_srcover_rgba_8888_sse2_lowp+0x107c>
   DB  69,15,92,213                        ; subps         %xmm13,%xmm10
   DB  69,15,84,209                        ; andps         %xmm9,%xmm10
   DB  69,15,85,205                        ; andnps        %xmm13,%xmm9
   DB  69,15,86,202                        ; orps          %xmm10,%xmm9
   DB  68,15,194,192,1                     ; cmpltps       %xmm0,%xmm8
-  DB  68,15,40,21,154,181,0,0             ; movaps        0xb59a(%rip),%xmm10        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,21,122,204,0,0             ; movaps        0xcc7a(%rip),%xmm10        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,92,209                        ; subps         %xmm9,%xmm10
   DB  69,15,84,208                        ; andps         %xmm8,%xmm10
   DB  69,15,85,193                        ; andnps        %xmm9,%xmm8
   DB  69,15,86,194                        ; orps          %xmm10,%xmm8
   DB  68,15,40,201                        ; movaps        %xmm1,%xmm9
   DB  68,15,194,200,1                     ; cmpltps       %xmm0,%xmm9
-  DB  68,15,40,21,137,181,0,0             ; movaps        0xb589(%rip),%xmm10        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,21,105,204,0,0             ; movaps        0xcc69(%rip),%xmm10        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,208                        ; subps         %xmm8,%xmm10
   DB  69,15,84,209                        ; andps         %xmm9,%xmm10
   DB  69,15,85,200                        ; andnps        %xmm8,%xmm9
@@ -31205,7 +31205,7 @@
   DB  243,69,15,89,203                    ; mulss         %xmm11,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,80,187,0,0              ; mulps         0xbb50(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,48,210,0,0              ; mulps         0xd230(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
@@ -31213,7 +31213,7 @@
   DB  243,69,15,89,219                    ; mulss         %xmm11,%xmm11
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,92,227                        ; subps         %xmm11,%xmm12
-  DB  68,15,89,21,59,187,0,0              ; mulps         0xbb3b(%rip),%xmm10        # 2ca70 <_sk_srcover_rgba_8888_sse2_lowp+0x1068>
+  DB  68,15,89,21,27,210,0,0              ; mulps         0xd21b(%rip),%xmm10        # 2e150 <_sk_srcover_rgba_8888_sse2_lowp+0x109c>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
@@ -31222,8 +31222,8 @@
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,92,193                        ; subps         %xmm9,%xmm0
-  DB  68,15,87,13,227,185,0,0             ; xorps         0xb9e3(%rip),%xmm9        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
-  DB  68,15,89,5,187,180,0,0              ; mulps         0xb4bb(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,87,13,195,208,0,0             ; xorps         0xd0c3(%rip),%xmm9        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
+  DB  68,15,89,5,155,203,0,0              ; mulps         0xcb9b(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -31242,7 +31242,7 @@
   DB  243,69,15,89,203                    ; mulss         %xmm11,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,178,186,0,0             ; mulps         0xbab2(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,146,209,0,0             ; mulps         0xd192(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,225                        ; movaps        %xmm1,%xmm12
   DB  69,15,89,228                        ; mulps         %xmm12,%xmm12
@@ -31250,7 +31250,7 @@
   DB  243,69,15,89,219                    ; mulss         %xmm11,%xmm11
   DB  69,15,198,219,0                     ; shufps        $0x0,%xmm11,%xmm11
   DB  69,15,92,227                        ; subps         %xmm11,%xmm12
-  DB  68,15,89,21,157,186,0,0             ; mulps         0xba9d(%rip),%xmm10        # 2ca70 <_sk_srcover_rgba_8888_sse2_lowp+0x1068>
+  DB  68,15,89,21,125,209,0,0             ; mulps         0xd17d(%rip),%xmm10        # 2e150 <_sk_srcover_rgba_8888_sse2_lowp+0x109c>
   DB  69,15,89,212                        ; mulps         %xmm12,%xmm10
   DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
@@ -31259,8 +31259,8 @@
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,40,194                        ; movaps        %xmm10,%xmm0
   DB  65,15,92,193                        ; subps         %xmm9,%xmm0
-  DB  68,15,87,13,69,185,0,0              ; xorps         0xb945(%rip),%xmm9        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
-  DB  68,15,89,5,29,180,0,0               ; mulps         0xb41d(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,87,13,37,208,0,0              ; xorps         0xd025(%rip),%xmm9        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
+  DB  68,15,89,5,253,202,0,0              ; mulps         0xcafd(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
@@ -31276,7 +31276,7 @@
   DB  243,69,15,89,200                    ; mulss         %xmm8,%xmm9
   DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
   DB  68,15,88,200                        ; addps         %xmm0,%xmm9
-  DB  68,15,89,13,37,186,0,0              ; mulps         0xba25(%rip),%xmm9        # 2ca60 <_sk_srcover_rgba_8888_sse2_lowp+0x1058>
+  DB  68,15,89,13,5,209,0,0               ; mulps         0xd105(%rip),%xmm9        # 2e140 <_sk_srcover_rgba_8888_sse2_lowp+0x108c>
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  68,15,40,209                        ; movaps        %xmm1,%xmm10
   DB  69,15,89,210                        ; mulps         %xmm10,%xmm10
@@ -31284,7 +31284,7 @@
   DB  243,69,15,89,192                    ; mulss         %xmm8,%xmm8
   DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
   DB  65,15,92,192                        ; subps         %xmm8,%xmm0
-  DB  15,87,5,225,184,0,0                 ; xorps         0xb8e1(%rip),%xmm0        # 2c940 <_sk_srcover_rgba_8888_sse2_lowp+0xf38>
+  DB  15,87,5,193,207,0,0                 ; xorps         0xcfc1(%rip),%xmm0        # 2e020 <_sk_srcover_rgba_8888_sse2_lowp+0xf6c>
   DB  65,15,94,193                        ; divps         %xmm9,%xmm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31324,7 +31324,7 @@
 PUBLIC _sk_save_xy_sse2
 _sk_save_xy_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,72,179,0,0               ; movaps        0xb348(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,40,202,0,0               ; movaps        0xca28(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,17,0                             ; movups        %xmm0,(%rax)
   DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
   DB  69,15,88,200                        ; addps         %xmm8,%xmm9
@@ -31332,7 +31332,7 @@
   DB  69,15,91,210                        ; cvtdq2ps      %xmm10,%xmm10
   DB  69,15,40,217                        ; movaps        %xmm9,%xmm11
   DB  69,15,194,218,1                     ; cmpltps       %xmm10,%xmm11
-  DB  68,15,40,37,51,179,0,0              ; movaps        0xb333(%rip),%xmm12        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  68,15,40,37,19,202,0,0              ; movaps        0xca13(%rip),%xmm12        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,84,220                        ; andps         %xmm12,%xmm11
   DB  69,15,92,211                        ; subps         %xmm11,%xmm10
   DB  69,15,92,202                        ; subps         %xmm10,%xmm9
@@ -31375,8 +31375,8 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,227,184,0,0                 ; addps         0xb8e3(%rip),%xmm0        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,139,178,0,0             ; movaps        0xb28b(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,195,207,0,0                 ; addps         0xcfc3(%rip),%xmm0        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,107,201,0,0             ; movaps        0xc96b(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,0,1,0,0                ; movups        %xmm9,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31387,7 +31387,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,87,178,0,0                  ; addps         0xb257(%rip),%xmm0        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,5,55,201,0,0                  ; addps         0xc937(%rip),%xmm0        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31397,8 +31397,8 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,150,184,0,0                ; addps         0xb896(%rip),%xmm1        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,62,178,0,0              ; movaps        0xb23e(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,118,207,0,0                ; addps         0xcf76(%rip),%xmm1        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,30,201,0,0              ; movaps        0xc91e(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  68,15,17,136,64,1,0,0               ; movups        %xmm9,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31409,7 +31409,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,9,178,0,0                  ; addps         0xb209(%rip),%xmm1        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,13,233,200,0,0                ; addps         0xc8e9(%rip),%xmm1        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31419,13 +31419,13 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,89,184,0,0                  ; addps         0xb859(%rip),%xmm0        # 2ca90 <_sk_srcover_rgba_8888_sse2_lowp+0x1088>
-  DB  68,15,40,13,241,177,0,0             ; movaps        0xb1f1(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,57,207,0,0                  ; addps         0xcf39(%rip),%xmm0        # 2e170 <_sk_srcover_rgba_8888_sse2_lowp+0x10bc>
+  DB  68,15,40,13,209,200,0,0             ; movaps        0xc8d1(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,77,184,0,0              ; mulps         0xb84d(%rip),%xmm9        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,13,213,179,0,0             ; addps         0xb3d5(%rip),%xmm9        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,13,45,207,0,0              ; mulps         0xcf2d(%rip),%xmm9        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,13,181,202,0,0             ; addps         0xcab5(%rip),%xmm9        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,0,1,0,0                ; movups        %xmm9,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31436,16 +31436,16 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,1,184,0,0                   ; addps         0xb801(%rip),%xmm0        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,169,177,0,0             ; movaps        0xb1a9(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,5,225,206,0,0                 ; addps         0xcee1(%rip),%xmm0        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,137,200,0,0             ; movaps        0xc889(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,29,184,0,0               ; movaps        0xb81d(%rip),%xmm8        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,5,253,206,0,0              ; movaps        0xcefd(%rip),%xmm8        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,33,184,0,0               ; addps         0xb821(%rip),%xmm8        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,5,1,207,0,0                ; addps         0xcf01(%rip),%xmm8        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,117,177,0,0              ; addps         0xb175(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,88,5,85,200,0,0               ; addps         0xc855(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,25,184,0,0               ; addps         0xb819(%rip),%xmm8        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,5,249,206,0,0              ; addps         0xcef9(%rip),%xmm8        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31453,17 +31453,17 @@
 PUBLIC _sk_bicubic_p1x_sse2
 _sk_bicubic_p1x_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,83,177,0,0               ; movaps        0xb153(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,51,200,0,0               ; movaps        0xc833(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,136,128,0,0,0              ; movups        0x80(%rax),%xmm9
   DB  65,15,88,192                        ; addps         %xmm8,%xmm0
-  DB  68,15,40,21,204,183,0,0             ; movaps        0xb7cc(%rip),%xmm10        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,21,172,206,0,0             ; movaps        0xceac(%rip),%xmm10        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,208,183,0,0             ; addps         0xb7d0(%rip),%xmm10        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,21,176,206,0,0             ; addps         0xceb0(%rip),%xmm10        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,204,183,0,0             ; addps         0xb7cc(%rip),%xmm10        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,21,172,206,0,0             ; addps         0xceac(%rip),%xmm10        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,144,0,1,0,0                ; movups        %xmm10,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31473,11 +31473,11 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,0                             ; movups        (%rax),%xmm0
   DB  68,15,16,128,128,0,0,0              ; movups        0x80(%rax),%xmm8
-  DB  15,88,5,156,183,0,0                 ; addps         0xb79c(%rip),%xmm0        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  15,88,5,124,206,0,0                 ; addps         0xce7c(%rip),%xmm0        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,108,183,0,0              ; mulps         0xb76c(%rip),%xmm8        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,5,244,178,0,0              ; addps         0xb2f4(%rip),%xmm8        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,5,76,206,0,0               ; mulps         0xce4c(%rip),%xmm8        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,5,212,201,0,0              ; addps         0xc9d4(%rip),%xmm8        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,0,1,0,0                ; movups        %xmm8,0x100(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31488,13 +31488,13 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,47,183,0,0                 ; addps         0xb72f(%rip),%xmm1        # 2ca90 <_sk_srcover_rgba_8888_sse2_lowp+0x1088>
-  DB  68,15,40,13,199,176,0,0             ; movaps        0xb0c7(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,15,206,0,0                 ; addps         0xce0f(%rip),%xmm1        # 2e170 <_sk_srcover_rgba_8888_sse2_lowp+0x10bc>
+  DB  68,15,40,13,167,199,0,0             ; movaps        0xc7a7(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
   DB  69,15,40,193                        ; movaps        %xmm9,%xmm8
   DB  69,15,89,192                        ; mulps         %xmm8,%xmm8
-  DB  68,15,89,13,35,183,0,0              ; mulps         0xb723(%rip),%xmm9        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,13,171,178,0,0             ; addps         0xb2ab(%rip),%xmm9        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,13,3,206,0,0               ; mulps         0xce03(%rip),%xmm9        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,13,139,201,0,0             ; addps         0xc98b(%rip),%xmm9        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,200                        ; mulps         %xmm8,%xmm9
   DB  68,15,17,136,64,1,0,0               ; movups        %xmm9,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31505,16 +31505,16 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,214,182,0,0                ; addps         0xb6d6(%rip),%xmm1        # 2ca80 <_sk_srcover_rgba_8888_sse2_lowp+0x1078>
-  DB  68,15,40,13,126,176,0,0             ; movaps        0xb07e(%rip),%xmm9        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,88,13,182,205,0,0                ; addps         0xcdb6(%rip),%xmm1        # 2e160 <_sk_srcover_rgba_8888_sse2_lowp+0x10ac>
+  DB  68,15,40,13,94,199,0,0              ; movaps        0xc75e(%rip),%xmm9        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  69,15,92,200                        ; subps         %xmm8,%xmm9
-  DB  68,15,40,5,242,182,0,0              ; movaps        0xb6f2(%rip),%xmm8        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,5,210,205,0,0              ; movaps        0xcdd2(%rip),%xmm8        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,246,182,0,0              ; addps         0xb6f6(%rip),%xmm8        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,5,214,205,0,0              ; addps         0xcdd6(%rip),%xmm8        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,74,176,0,0               ; addps         0xb04a(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,88,5,42,199,0,0               ; addps         0xc72a(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
-  DB  68,15,88,5,238,182,0,0              ; addps         0xb6ee(%rip),%xmm8        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,5,206,205,0,0              ; addps         0xcdce(%rip),%xmm8        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31522,17 +31522,17 @@
 PUBLIC _sk_bicubic_p1y_sse2
 _sk_bicubic_p1y_sse2 LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  68,15,40,5,40,176,0,0               ; movaps        0xb028(%rip),%xmm8        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  68,15,40,5,8,199,0,0                ; movaps        0xc708(%rip),%xmm8        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,136,192,0,0,0              ; movups        0xc0(%rax),%xmm9
   DB  65,15,88,200                        ; addps         %xmm8,%xmm1
-  DB  68,15,40,21,160,182,0,0             ; movaps        0xb6a0(%rip),%xmm10        # 2cab0 <_sk_srcover_rgba_8888_sse2_lowp+0x10a8>
+  DB  68,15,40,21,128,205,0,0             ; movaps        0xcd80(%rip),%xmm10        # 2e190 <_sk_srcover_rgba_8888_sse2_lowp+0x10dc>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,164,182,0,0             ; addps         0xb6a4(%rip),%xmm10        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  68,15,88,21,132,205,0,0             ; addps         0xcd84(%rip),%xmm10        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
   DB  69,15,88,208                        ; addps         %xmm8,%xmm10
   DB  69,15,89,209                        ; mulps         %xmm9,%xmm10
-  DB  68,15,88,21,160,182,0,0             ; addps         0xb6a0(%rip),%xmm10        # 2cad0 <_sk_srcover_rgba_8888_sse2_lowp+0x10c8>
+  DB  68,15,88,21,128,205,0,0             ; addps         0xcd80(%rip),%xmm10        # 2e1b0 <_sk_srcover_rgba_8888_sse2_lowp+0x10fc>
   DB  68,15,17,144,64,1,0,0               ; movups        %xmm10,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -31542,11 +31542,11 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,16,72,64                         ; movups        0x40(%rax),%xmm1
   DB  68,15,16,128,192,0,0,0              ; movups        0xc0(%rax),%xmm8
-  DB  15,88,13,111,182,0,0                ; addps         0xb66f(%rip),%xmm1        # 2cac0 <_sk_srcover_rgba_8888_sse2_lowp+0x10b8>
+  DB  15,88,13,79,205,0,0                 ; addps         0xcd4f(%rip),%xmm1        # 2e1a0 <_sk_srcover_rgba_8888_sse2_lowp+0x10ec>
   DB  69,15,40,200                        ; movaps        %xmm8,%xmm9
   DB  69,15,89,201                        ; mulps         %xmm9,%xmm9
-  DB  68,15,89,5,63,182,0,0               ; mulps         0xb63f(%rip),%xmm8        # 2caa0 <_sk_srcover_rgba_8888_sse2_lowp+0x1098>
-  DB  68,15,88,5,199,177,0,0              ; addps         0xb1c7(%rip),%xmm8        # 2c630 <_sk_srcover_rgba_8888_sse2_lowp+0xc28>
+  DB  68,15,89,5,31,205,0,0               ; mulps         0xcd1f(%rip),%xmm8        # 2e180 <_sk_srcover_rgba_8888_sse2_lowp+0x10cc>
+  DB  68,15,88,5,167,200,0,0              ; addps         0xc8a7(%rip),%xmm8        # 2dd10 <_sk_srcover_rgba_8888_sse2_lowp+0xc5c>
   DB  69,15,89,193                        ; mulps         %xmm9,%xmm8
   DB  68,15,17,128,64,1,0,0               ; movups        %xmm8,0x140(%rax)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -31675,7 +31675,7 @@
   DB  102,65,15,254,205                   ; paddd         %xmm13,%xmm1
   DB  102,68,15,127,108,36,80             ; movdqa        %xmm13,0x50(%rsp)
   DB  102,15,254,217                      ; paddd         %xmm1,%xmm3
-  DB  102,68,15,111,5,156,180,0,0         ; movdqa        0xb49c(%rip),%xmm8        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,68,15,111,5,124,203,0,0         ; movdqa        0xcb7c(%rip),%xmm8        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,112,195,245                  ; pshufd        $0xf5,%xmm3,%xmm0
   DB  102,65,15,244,216                   ; pmuludq       %xmm8,%xmm3
   DB  102,65,15,244,192                   ; pmuludq       %xmm8,%xmm0
@@ -31714,7 +31714,7 @@
   DB  243,66,15,16,44,139                 ; movss         (%rbx,%r9,4),%xmm5
   DB  15,20,236                           ; unpcklps      %xmm4,%xmm5
   DB  102,15,20,221                       ; unpcklpd      %xmm5,%xmm3
-  DB  102,68,15,111,37,83,173,0,0         ; movdqa        0xad53(%rip),%xmm12        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,37,51,196,0,0         ; movdqa        0xc433(%rip),%xmm12        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,196                   ; paddd         %xmm12,%xmm0
   DB  102,15,112,224,78                   ; pshufd        $0x4e,%xmm0,%xmm4
   DB  102,73,15,126,224                   ; movq          %xmm4,%r8
@@ -31730,7 +31730,7 @@
   DB  243,66,15,16,36,139                 ; movss         (%rbx,%r9,4),%xmm4
   DB  15,20,224                           ; unpcklps      %xmm0,%xmm4
   DB  102,15,20,236                       ; unpcklpd      %xmm4,%xmm5
-  DB  15,40,37,136,179,0,0                ; movaps        0xb388(%rip),%xmm4        # 2cae0 <_sk_srcover_rgba_8888_sse2_lowp+0x10d8>
+  DB  15,40,37,104,202,0,0                ; movaps        0xca68(%rip),%xmm4        # 2e1c0 <_sk_srcover_rgba_8888_sse2_lowp+0x110c>
   DB  68,15,88,220                        ; addps         %xmm4,%xmm11
   DB  68,15,41,92,36,64                   ; movaps        %xmm11,0x40(%rsp)
   DB  68,15,40,223                        ; movaps        %xmm7,%xmm11
@@ -31766,7 +31766,7 @@
   DB  15,20,249                           ; unpcklps      %xmm1,%xmm7
   DB  102,15,20,252                       ; unpcklpd      %xmm4,%xmm7
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
-  DB  102,15,250,13,253,178,0,0           ; psubd         0xb2fd(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,221,201,0,0           ; psubd         0xc9dd(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,225,78                   ; pshufd        $0x4e,%xmm1,%xmm4
   DB  102,73,15,126,224                   ; movq          %xmm4,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -31856,7 +31856,7 @@
   DB  243,15,16,44,131                    ; movss         (%rbx,%rax,4),%xmm5
   DB  15,20,233                           ; unpcklps      %xmm1,%xmm5
   DB  102,15,20,221                       ; unpcklpd      %xmm5,%xmm3
-  DB  102,68,15,111,45,197,170,0,0        ; movdqa        0xaac5(%rip),%xmm13        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,45,165,193,0,0        ; movdqa        0xc1a5(%rip),%xmm13        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,197                   ; paddd         %xmm13,%xmm0
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
   DB  102,73,15,126,200                   ; movq          %xmm1,%r8
@@ -31953,7 +31953,7 @@
   DB  102,15,111,206                      ; movdqa        %xmm6,%xmm1
   DB  102,65,15,254,202                   ; paddd         %xmm10,%xmm1
   DB  102,15,112,209,245                  ; pshufd        $0xf5,%xmm1,%xmm2
-  DB  102,15,111,29,154,175,0,0           ; movdqa        0xaf9a(%rip),%xmm3        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,29,122,198,0,0           ; movdqa        0xc67a(%rip),%xmm3        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,203                      ; pmuludq       %xmm3,%xmm1
   DB  102,15,244,211                      ; pmuludq       %xmm3,%xmm2
   DB  102,15,111,251                      ; movdqa        %xmm3,%xmm7
@@ -31991,7 +31991,7 @@
   DB  243,15,16,44,131                    ; movss         (%rbx,%rax,4),%xmm5
   DB  15,20,236                           ; unpcklps      %xmm4,%xmm5
   DB  102,15,20,213                       ; unpcklpd      %xmm5,%xmm2
-  DB  102,15,111,5,90,168,0,0             ; movdqa        0xa85a(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,111,5,58,191,0,0             ; movdqa        0xbf3a(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,254,216                      ; paddd         %xmm0,%xmm3
   DB  102,15,112,227,78                   ; pshufd        $0x4e,%xmm3,%xmm4
   DB  102,73,15,126,224                   ; movq          %xmm4,%r8
@@ -32072,7 +32072,7 @@
   DB  102,15,111,214                      ; movdqa        %xmm6,%xmm2
   DB  102,65,15,254,212                   ; paddd         %xmm12,%xmm2
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,13,120,173,0,0           ; movdqa        0xad78(%rip),%xmm1        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,13,88,196,0,0            ; movdqa        0xc458(%rip),%xmm1        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
   DB  102,15,244,193                      ; pmuludq       %xmm1,%xmm0
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
@@ -32094,7 +32094,7 @@
   DB  68,15,20,209                        ; unpcklps      %xmm1,%xmm10
   DB  102,68,15,20,210                    ; unpcklpd      %xmm2,%xmm10
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  102,15,250,13,29,173,0,0            ; psubd         0xad1d(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,253,195,0,0           ; psubd         0xc3fd(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
   DB  102,73,15,126,208                   ; movq          %xmm2,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -32109,7 +32109,7 @@
   DB  243,15,16,20,131                    ; movss         (%rbx,%rax,4),%xmm2
   DB  15,20,209                           ; unpcklps      %xmm1,%xmm2
   DB  102,15,20,226                       ; unpcklpd      %xmm2,%xmm4
-  DB  102,15,254,5,56,166,0,0             ; paddd         0xa638(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,5,24,189,0,0             ; paddd         0xbd18(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
   DB  102,73,15,126,200                   ; movq          %xmm1,%r8
   DB  102,72,15,126,192                   ; movq          %xmm0,%rax
@@ -32146,7 +32146,7 @@
   DB  15,20,199                           ; unpcklps      %xmm7,%xmm0
   DB  102,15,20,193                       ; unpcklpd      %xmm1,%xmm0
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
-  DB  102,15,250,13,39,172,0,0            ; psubd         0xac27(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,7,195,0,0             ; psubd         0xc307(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,249,78                   ; pshufd        $0x4e,%xmm1,%xmm7
   DB  102,73,15,126,248                   ; movq          %xmm7,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -32161,7 +32161,7 @@
   DB  243,15,16,52,131                    ; movss         (%rbx,%rax,4),%xmm6
   DB  15,20,247                           ; unpcklps      %xmm7,%xmm6
   DB  102,15,20,206                       ; unpcklpd      %xmm6,%xmm1
-  DB  102,15,254,21,66,165,0,0            ; paddd         0xa542(%rip),%xmm2        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,21,34,188,0,0            ; paddd         0xbc22(%rip),%xmm2        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,242,78                   ; pshufd        $0x4e,%xmm2,%xmm6
   DB  102,73,15,126,240                   ; movq          %xmm6,%r8
   DB  102,72,15,126,208                   ; movq          %xmm2,%rax
@@ -32305,7 +32305,7 @@
   DB  102,65,15,254,210                   ; paddd         %xmm10,%xmm2
   DB  102,15,111,220                      ; movdqa        %xmm4,%xmm3
   DB  102,15,254,218                      ; paddd         %xmm2,%xmm3
-  DB  102,15,111,37,75,169,0,0            ; movdqa        0xa94b(%rip),%xmm4        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,37,43,192,0,0            ; movdqa        0xc02b(%rip),%xmm4        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,112,195,245                  ; pshufd        $0xf5,%xmm3,%xmm0
   DB  102,15,244,220                      ; pmuludq       %xmm4,%xmm3
   DB  102,15,244,196                      ; pmuludq       %xmm4,%xmm0
@@ -32344,7 +32344,7 @@
   DB  243,66,15,16,52,139                 ; movss         (%rbx,%r9,4),%xmm6
   DB  15,20,244                           ; unpcklps      %xmm4,%xmm6
   DB  102,15,20,238                       ; unpcklpd      %xmm6,%xmm5
-  DB  102,15,254,5,5,162,0,0              ; paddd         0xa205(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,5,229,184,0,0            ; paddd         0xb8e5(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,224,78                   ; pshufd        $0x4e,%xmm0,%xmm4
   DB  102,73,15,126,224                   ; movq          %xmm4,%r8
   DB  102,73,15,126,193                   ; movq          %xmm0,%r9
@@ -32359,7 +32359,7 @@
   DB  243,66,15,16,52,139                 ; movss         (%rbx,%r9,4),%xmm6
   DB  15,20,244                           ; unpcklps      %xmm4,%xmm6
   DB  102,15,20,222                       ; unpcklpd      %xmm6,%xmm3
-  DB  15,40,53,63,168,0,0                 ; movaps        0xa83f(%rip),%xmm6        # 2cae0 <_sk_srcover_rgba_8888_sse2_lowp+0x10d8>
+  DB  15,40,53,31,191,0,0                 ; movaps        0xbf1f(%rip),%xmm6        # 2e1c0 <_sk_srcover_rgba_8888_sse2_lowp+0x110c>
   DB  68,15,88,198                        ; addps         %xmm6,%xmm8
   DB  68,15,41,68,36,48                   ; movaps        %xmm8,0x30(%rsp)
   DB  68,15,88,254                        ; addps         %xmm6,%xmm15
@@ -32377,7 +32377,7 @@
   DB  102,65,15,254,215                   ; paddd         %xmm15,%xmm2
   DB  102,68,15,127,124,36,96             ; movdqa        %xmm15,0x60(%rsp)
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,13,242,167,0,0           ; movdqa        0xa7f2(%rip),%xmm1        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,13,210,190,0,0           ; movdqa        0xbed2(%rip),%xmm1        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
   DB  102,15,244,193                      ; pmuludq       %xmm1,%xmm0
   DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
@@ -32398,7 +32398,7 @@
   DB  15,20,240                           ; unpcklps      %xmm0,%xmm6
   DB  102,15,20,242                       ; unpcklpd      %xmm2,%xmm6
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
-  DB  102,15,250,5,157,167,0,0            ; psubd         0xa79d(%rip),%xmm0        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,5,125,190,0,0            ; psubd         0xbe7d(%rip),%xmm0        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,208,78                   ; pshufd        $0x4e,%xmm0,%xmm2
   DB  102,73,15,126,208                   ; movq          %xmm2,%r8
   DB  102,72,15,126,192                   ; movq          %xmm0,%rax
@@ -32413,7 +32413,7 @@
   DB  243,68,15,16,28,131                 ; movss         (%rbx,%rax,4),%xmm11
   DB  68,15,20,216                        ; unpcklps      %xmm0,%xmm11
   DB  102,65,15,20,211                    ; unpcklpd      %xmm11,%xmm2
-  DB  102,15,254,13,181,160,0,0           ; paddd         0xa0b5(%rip),%xmm1        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,13,149,183,0,0           ; paddd         0xb795(%rip),%xmm1        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,193,78                   ; pshufd        $0x4e,%xmm1,%xmm0
   DB  102,73,15,126,192                   ; movq          %xmm0,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -32451,7 +32451,7 @@
   DB  102,65,15,111,194                   ; movdqa        %xmm10,%xmm0
   DB  102,15,254,68,36,32                 ; paddd         0x20(%rsp),%xmm0
   DB  102,15,112,216,245                  ; pshufd        $0xf5,%xmm0,%xmm3
-  DB  102,15,111,37,153,166,0,0           ; movdqa        0xa699(%rip),%xmm4        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,37,121,189,0,0           ; movdqa        0xbd79(%rip),%xmm4        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,196                      ; pmuludq       %xmm4,%xmm0
   DB  102,15,244,220                      ; pmuludq       %xmm4,%xmm3
   DB  102,68,15,111,220                   ; movdqa        %xmm4,%xmm11
@@ -32489,7 +32489,7 @@
   DB  243,15,16,60,131                    ; movss         (%rbx,%rax,4),%xmm7
   DB  15,20,253                           ; unpcklps      %xmm5,%xmm7
   DB  102,15,20,199                       ; unpcklpd      %xmm7,%xmm0
-  DB  102,68,15,111,5,85,159,0,0          ; movdqa        0x9f55(%rip),%xmm8        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,5,53,182,0,0          ; movdqa        0xb635(%rip),%xmm8        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,224                   ; paddd         %xmm8,%xmm4
   DB  102,15,112,236,78                   ; pshufd        $0x4e,%xmm4,%xmm5
   DB  102,73,15,126,232                   ; movq          %xmm5,%r8
@@ -32635,7 +32635,7 @@
   DB  243,15,16,36,131                    ; movss         (%rbx,%rax,4),%xmm4
   DB  15,20,226                           ; unpcklps      %xmm2,%xmm4
   DB  102,15,20,252                       ; unpcklpd      %xmm4,%xmm7
-  DB  102,68,15,111,61,164,156,0,0        ; movdqa        0x9ca4(%rip),%xmm15        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,61,132,179,0,0        ; movdqa        0xb384(%rip),%xmm15        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,207                   ; paddd         %xmm15,%xmm1
   DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
   DB  102,73,15,126,208                   ; movq          %xmm2,%r8
@@ -32718,7 +32718,7 @@
   DB  102,65,15,111,217                   ; movdqa        %xmm9,%xmm3
   DB  102,65,15,254,216                   ; paddd         %xmm8,%xmm3
   DB  102,15,112,211,245                  ; pshufd        $0xf5,%xmm3,%xmm2
-  DB  102,15,111,45,189,161,0,0           ; movdqa        0xa1bd(%rip),%xmm5        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,45,157,184,0,0           ; movdqa        0xb89d(%rip),%xmm5        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,221                      ; pmuludq       %xmm5,%xmm3
   DB  102,15,244,213                      ; pmuludq       %xmm5,%xmm2
   DB  102,15,112,234,232                  ; pshufd        $0xe8,%xmm2,%xmm5
@@ -32739,7 +32739,7 @@
   DB  68,15,20,198                        ; unpcklps      %xmm6,%xmm8
   DB  102,68,15,20,197                    ; unpcklpd      %xmm5,%xmm8
   DB  102,15,111,234                      ; movdqa        %xmm2,%xmm5
-  DB  102,15,250,45,102,161,0,0           ; psubd         0xa166(%rip),%xmm5        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,45,70,184,0,0            ; psubd         0xb846(%rip),%xmm5        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,245,78                   ; pshufd        $0x4e,%xmm5,%xmm6
   DB  102,73,15,126,240                   ; movq          %xmm6,%r8
   DB  102,72,15,126,232                   ; movq          %xmm5,%rax
@@ -32771,7 +32771,7 @@
   DB  102,15,20,214                       ; unpcklpd      %xmm6,%xmm2
   DB  102,69,15,254,205                   ; paddd         %xmm13,%xmm9
   DB  102,65,15,112,233,245               ; pshufd        $0xf5,%xmm9,%xmm5
-  DB  102,15,111,29,196,160,0,0           ; movdqa        0xa0c4(%rip),%xmm3        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,29,164,183,0,0           ; movdqa        0xb7a4(%rip),%xmm3        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,68,15,244,203                   ; pmuludq       %xmm3,%xmm9
   DB  102,15,244,235                      ; pmuludq       %xmm3,%xmm5
   DB  102,15,112,237,232                  ; pshufd        $0xe8,%xmm5,%xmm5
@@ -32792,7 +32792,7 @@
   DB  69,15,20,233                        ; unpcklps      %xmm9,%xmm13
   DB  102,68,15,20,237                    ; unpcklpd      %xmm5,%xmm13
   DB  102,65,15,111,239                   ; movdqa        %xmm15,%xmm5
-  DB  102,15,250,45,100,160,0,0           ; psubd         0xa064(%rip),%xmm5        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,45,68,183,0,0            ; psubd         0xb744(%rip),%xmm5        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,221,78                   ; pshufd        $0x4e,%xmm5,%xmm3
   DB  102,73,15,126,216                   ; movq          %xmm3,%r8
   DB  102,72,15,126,232                   ; movq          %xmm5,%rax
@@ -32807,7 +32807,7 @@
   DB  243,15,16,44,131                    ; movss         (%rbx,%rax,4),%xmm5
   DB  15,20,235                           ; unpcklps      %xmm3,%xmm5
   DB  102,15,20,245                       ; unpcklpd      %xmm5,%xmm6
-  DB  102,68,15,111,13,126,153,0,0        ; movdqa        0x997e(%rip),%xmm9        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,13,94,176,0,0         ; movdqa        0xb05e(%rip),%xmm9        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,69,15,254,249                   ; paddd         %xmm9,%xmm15
   DB  102,65,15,112,223,78                ; pshufd        $0x4e,%xmm15,%xmm3
   DB  102,73,15,126,216                   ; movq          %xmm3,%r8
@@ -32869,7 +32869,7 @@
   DB  102,65,15,111,195                   ; movdqa        %xmm11,%xmm0
   DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
   DB  102,15,112,200,245                  ; pshufd        $0xf5,%xmm0,%xmm1
-  DB  102,15,111,21,248,158,0,0           ; movdqa        0x9ef8(%rip),%xmm2        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,21,216,181,0,0           ; movdqa        0xb5d8(%rip),%xmm2        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
   DB  102,15,244,202                      ; pmuludq       %xmm2,%xmm1
   DB  102,68,15,111,234                   ; movdqa        %xmm2,%xmm13
@@ -33062,7 +33062,7 @@
   DB  68,15,20,238                        ; unpcklps      %xmm6,%xmm13
   DB  102,68,15,20,234                    ; unpcklpd      %xmm2,%xmm13
   DB  102,15,111,212                      ; movdqa        %xmm4,%xmm2
-  DB  102,15,250,21,145,155,0,0           ; psubd         0x9b91(%rip),%xmm2        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,21,113,178,0,0           ; psubd         0xb271(%rip),%xmm2        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,242,78                   ; pshufd        $0x4e,%xmm2,%xmm6
   DB  102,73,15,126,240                   ; movq          %xmm6,%r8
   DB  102,72,15,126,208                   ; movq          %xmm2,%rax
@@ -33122,7 +33122,7 @@
   DB  102,68,15,111,116,36,32             ; movdqa        0x20(%rsp),%xmm14
   DB  102,65,15,254,198                   ; paddd         %xmm14,%xmm0
   DB  102,15,112,200,245                  ; pshufd        $0xf5,%xmm0,%xmm1
-  DB  102,68,15,111,5,110,154,0,0         ; movdqa        0x9a6e(%rip),%xmm8        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,68,15,111,5,78,177,0,0          ; movdqa        0xb14e(%rip),%xmm8        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,65,15,244,192                   ; pmuludq       %xmm8,%xmm0
   DB  102,65,15,244,200                   ; pmuludq       %xmm8,%xmm1
   DB  102,65,15,111,240                   ; movdqa        %xmm8,%xmm6
@@ -33160,7 +33160,7 @@
   DB  243,15,16,44,131                    ; movss         (%rbx,%rax,4),%xmm5
   DB  15,20,235                           ; unpcklps      %xmm3,%xmm5
   DB  102,15,20,205                       ; unpcklpd      %xmm5,%xmm1
-  DB  102,68,15,111,5,44,147,0,0          ; movdqa        0x932c(%rip),%xmm8        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,68,15,111,5,12,170,0,0          ; movdqa        0xaa0c(%rip),%xmm8        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,65,15,254,208                   ; paddd         %xmm8,%xmm2
   DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
   DB  102,73,15,126,216                   ; movq          %xmm3,%r8
@@ -33243,7 +33243,7 @@
   DB  102,15,254,208                      ; paddd         %xmm0,%xmm2
   DB  102,15,111,240                      ; movdqa        %xmm0,%xmm6
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,37,70,152,0,0            ; movdqa        0x9846(%rip),%xmm4        # 2caf0 <_sk_srcover_rgba_8888_sse2_lowp+0x10e8>
+  DB  102,15,111,37,38,175,0,0            ; movdqa        0xaf26(%rip),%xmm4        # 2e1d0 <_sk_srcover_rgba_8888_sse2_lowp+0x111c>
   DB  102,15,244,212                      ; pmuludq       %xmm4,%xmm2
   DB  102,15,244,196                      ; pmuludq       %xmm4,%xmm0
   DB  102,15,112,200,232                  ; pshufd        $0xe8,%xmm0,%xmm1
@@ -33264,7 +33264,7 @@
   DB  68,15,20,193                        ; unpcklps      %xmm1,%xmm8
   DB  102,68,15,20,194                    ; unpcklpd      %xmm2,%xmm8
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  102,15,250,13,239,151,0,0           ; psubd         0x97ef(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,207,174,0,0           ; psubd         0xaecf(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
   DB  102,73,15,126,208                   ; movq          %xmm2,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -33279,7 +33279,7 @@
   DB  243,15,16,20,131                    ; movss         (%rbx,%rax,4),%xmm2
   DB  15,20,209                           ; unpcklps      %xmm1,%xmm2
   DB  102,68,15,20,242                    ; unpcklpd      %xmm2,%xmm14
-  DB  102,15,254,5,8,145,0,0              ; paddd         0x9108(%rip),%xmm0        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,5,232,167,0,0            ; paddd         0xa7e8(%rip),%xmm0        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
   DB  102,73,15,126,200                   ; movq          %xmm1,%r8
   DB  102,72,15,126,192                   ; movq          %xmm0,%rax
@@ -33317,7 +33317,7 @@
   DB  15,20,198                           ; unpcklps      %xmm6,%xmm0
   DB  102,15,20,193                       ; unpcklpd      %xmm1,%xmm0
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
-  DB  102,15,250,13,247,150,0,0           ; psubd         0x96f7(%rip),%xmm1        # 2cb00 <_sk_srcover_rgba_8888_sse2_lowp+0x10f8>
+  DB  102,15,250,13,215,173,0,0           ; psubd         0xadd7(%rip),%xmm1        # 2e1e0 <_sk_srcover_rgba_8888_sse2_lowp+0x112c>
   DB  102,15,112,241,78                   ; pshufd        $0x4e,%xmm1,%xmm6
   DB  102,73,15,126,240                   ; movq          %xmm6,%r8
   DB  102,72,15,126,200                   ; movq          %xmm1,%rax
@@ -33332,7 +33332,7 @@
   DB  243,15,16,36,131                    ; movss         (%rbx,%rax,4),%xmm4
   DB  15,20,230                           ; unpcklps      %xmm6,%xmm4
   DB  102,15,20,204                       ; unpcklpd      %xmm4,%xmm1
-  DB  102,15,254,21,18,144,0,0            ; paddd         0x9012(%rip),%xmm2        # 2c460 <_sk_srcover_rgba_8888_sse2_lowp+0xa58>
+  DB  102,15,254,21,242,166,0,0           ; paddd         0xa6f2(%rip),%xmm2        # 2db40 <_sk_srcover_rgba_8888_sse2_lowp+0xa8c>
   DB  102,15,112,226,78                   ; pshufd        $0x4e,%xmm2,%xmm4
   DB  102,73,15,126,224                   ; movq          %xmm4,%r8
   DB  102,72,15,126,208                   ; movq          %xmm2,%rax
@@ -33391,7 +33391,7 @@
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  65,15,88,215                        ; addps         %xmm15,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,248,142,0,0                ; movaps        0x8ef8(%rip),%xmm3        # 2c430 <_sk_srcover_rgba_8888_sse2_lowp+0xa28>
+  DB  15,40,29,216,165,0,0                ; movaps        0xa5d8(%rip),%xmm3        # 2db10 <_sk_srcover_rgba_8888_sse2_lowp+0xa5c>
   DB  15,40,164,36,224,0,0,0              ; movaps        0xe0(%rsp),%xmm4
   DB  15,40,172,36,240,0,0,0              ; movaps        0xf0(%rsp),%xmm5
   DB  15,40,180,36,0,1,0,0                ; movaps        0x100(%rsp),%xmm6
@@ -33513,12 +33513,38 @@
   DB  197,249,110,209                     ; vmovd         %ecx,%xmm2
   DB  196,226,125,88,210                  ; vpbroadcastd  %xmm2,%ymm2
   DB  197,252,91,210                      ; vcvtdq2ps     %ymm2,%ymm2
-  DB  196,226,125,24,29,154,133,0,0       ; vbroadcastss  0x859a(%rip),%ymm3        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  196,226,125,24,29,70,156,0,0        ; vbroadcastss  0x9c46(%rip),%ymm3        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  197,236,88,211                      ; vaddps        %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,252,40,218                      ; vmovaps       %ymm2,%ymm3
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_translate_hsw_lowp
+_sk_matrix_translate_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
+  DB  196,193,124,88,192                  ; vaddps        %ymm8,%ymm0,%ymm0
+  DB  196,193,116,88,200                  ; vaddps        %ymm8,%ymm1,%ymm1
+  DB  196,98,125,24,64,4                  ; vbroadcastss  0x4(%rax),%ymm8
+  DB  196,193,108,88,208                  ; vaddps        %ymm8,%ymm2,%ymm2
+  DB  196,193,100,88,216                  ; vaddps        %ymm8,%ymm3,%ymm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_matrix_scale_translate_hsw_lowp
+_sk_matrix_scale_translate_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,24,0                     ; vbroadcastss  (%rax),%ymm8
+  DB  196,98,125,24,72,8                  ; vbroadcastss  0x8(%rax),%ymm9
+  DB  196,194,61,168,193                  ; vfmadd213ps   %ymm9,%ymm8,%ymm0
+  DB  196,194,61,168,201                  ; vfmadd213ps   %ymm9,%ymm8,%ymm1
+  DB  196,98,125,24,64,4                  ; vbroadcastss  0x4(%rax),%ymm8
+  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
+  DB  196,194,61,168,209                  ; vfmadd213ps   %ymm9,%ymm8,%ymm2
+  DB  196,194,61,168,217                  ; vfmadd213ps   %ymm9,%ymm8,%ymm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_matrix_2x3_hsw_lowp
 _sk_matrix_2x3_hsw_lowp LABEL PROC
   DB  197,124,40,192                      ; vmovaps       %ymm0,%ymm8
@@ -33543,6 +33569,41 @@
   DB  197,124,41,194                      ; vmovaps       %ymm8,%ymm2
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_perspective_hsw_lowp
+_sk_matrix_perspective_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,24,8                     ; vbroadcastss  (%rax),%ymm9
+  DB  196,98,125,24,64,4                  ; vbroadcastss  0x4(%rax),%ymm8
+  DB  196,98,125,24,80,8                  ; vbroadcastss  0x8(%rax),%ymm10
+  DB  196,65,124,40,216                   ; vmovaps       %ymm8,%ymm11
+  DB  196,66,109,168,218                  ; vfmadd213ps   %ymm10,%ymm2,%ymm11
+  DB  196,66,101,168,194                  ; vfmadd213ps   %ymm10,%ymm3,%ymm8
+  DB  196,66,117,184,193                  ; vfmadd231ps   %ymm9,%ymm1,%ymm8
+  DB  196,66,125,184,217                  ; vfmadd231ps   %ymm9,%ymm0,%ymm11
+  DB  196,98,125,24,72,12                 ; vbroadcastss  0xc(%rax),%ymm9
+  DB  196,98,125,24,80,16                 ; vbroadcastss  0x10(%rax),%ymm10
+  DB  196,98,125,24,96,20                 ; vbroadcastss  0x14(%rax),%ymm12
+  DB  196,65,124,40,234                   ; vmovaps       %ymm10,%ymm13
+  DB  196,66,109,168,236                  ; vfmadd213ps   %ymm12,%ymm2,%ymm13
+  DB  196,66,101,168,212                  ; vfmadd213ps   %ymm12,%ymm3,%ymm10
+  DB  196,66,117,184,209                  ; vfmadd231ps   %ymm9,%ymm1,%ymm10
+  DB  196,66,125,184,233                  ; vfmadd231ps   %ymm9,%ymm0,%ymm13
+  DB  196,98,125,24,72,24                 ; vbroadcastss  0x18(%rax),%ymm9
+  DB  196,98,125,24,96,28                 ; vbroadcastss  0x1c(%rax),%ymm12
+  DB  196,98,125,24,112,32                ; vbroadcastss  0x20(%rax),%ymm14
+  DB  196,194,29,168,214                  ; vfmadd213ps   %ymm14,%ymm12,%ymm2
+  DB  196,66,101,168,230                  ; vfmadd213ps   %ymm14,%ymm3,%ymm12
+  DB  196,194,53,168,204                  ; vfmadd213ps   %ymm12,%ymm9,%ymm1
+  DB  196,194,125,184,209                 ; vfmadd231ps   %ymm9,%ymm0,%ymm2
+  DB  197,252,83,210                      ; vrcpps        %ymm2,%ymm2
+  DB  197,252,83,217                      ; vrcpps        %ymm1,%ymm3
+  DB  197,164,89,194                      ; vmulps        %ymm2,%ymm11,%ymm0
+  DB  197,188,89,203                      ; vmulps        %ymm3,%ymm8,%ymm1
+  DB  197,148,89,210                      ; vmulps        %ymm2,%ymm13,%ymm2
+  DB  197,172,89,219                      ; vmulps        %ymm3,%ymm10,%ymm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_uniform_color_hsw_lowp
 _sk_uniform_color_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -33556,7 +33617,7 @@
 PUBLIC _sk_black_color_hsw_lowp
 _sk_black_color_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,29,208,138,0,0      ; vpbroadcastw  0x8ad0(%rip),%ymm3        # 2c280 <_sk_srcover_rgba_8888_sse2_lowp+0x878>
+  DB  196,226,125,121,29,129,160,0,0      ; vpbroadcastw  0xa081(%rip),%ymm3        # 2d920 <_sk_srcover_rgba_8888_sse2_lowp+0x86c>
   DB  197,252,87,192                      ; vxorps        %ymm0,%ymm0,%ymm0
   DB  197,244,87,201                      ; vxorps        %ymm1,%ymm1,%ymm1
   DB  197,236,87,210                      ; vxorps        %ymm2,%ymm2,%ymm2
@@ -33565,7 +33626,7 @@
 PUBLIC _sk_white_color_hsw_lowp
 _sk_white_color_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,5,185,138,0,0       ; vpbroadcastw  0x8ab9(%rip),%ymm0        # 2c282 <_sk_srcover_rgba_8888_sse2_lowp+0x87a>
+  DB  196,226,125,121,5,106,160,0,0       ; vpbroadcastw  0xa06a(%rip),%ymm0        # 2d922 <_sk_srcover_rgba_8888_sse2_lowp+0x86e>
   DB  197,253,111,200                     ; vmovdqa       %ymm0,%ymm1
   DB  197,253,111,208                     ; vmovdqa       %ymm0,%ymm2
   DB  197,253,111,216                     ; vmovdqa       %ymm0,%ymm3
@@ -33574,8 +33635,8 @@
 PUBLIC _sk_set_rgb_hsw_lowp
 _sk_set_rgb_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  197,122,16,5,211,132,0,0            ; vmovss        0x84d3(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
-  DB  197,250,16,21,251,132,0,0           ; vmovss        0x84fb(%rip),%xmm2        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  197,122,16,5,144,154,0,0            ; vmovss        0x9a90(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,250,16,21,184,154,0,0           ; vmovss        0x9ab8(%rip),%xmm2        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  197,250,16,0                        ; vmovss        (%rax),%xmm0
   DB  196,226,57,153,194                  ; vfmadd132ss   %xmm2,%xmm8,%xmm0
   DB  197,122,44,192                      ; vcvttss2si    %xmm0,%r8d
@@ -33596,7 +33657,7 @@
 PUBLIC _sk_premul_hsw_lowp
 _sk_premul_hsw_lowp LABEL PROC
   DB  197,229,213,192                     ; vpmullw       %ymm0,%ymm3,%ymm0
-  DB  196,98,125,121,5,72,138,0,0         ; vpbroadcastw  0x8a48(%rip),%ymm8        # 2c284 <_sk_srcover_rgba_8888_sse2_lowp+0x87c>
+  DB  196,98,125,121,5,249,159,0,0        ; vpbroadcastw  0x9ff9(%rip),%ymm8        # 2d924 <_sk_srcover_rgba_8888_sse2_lowp+0x870>
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,229,213,201                     ; vpmullw       %ymm1,%ymm3,%ymm1
@@ -33636,7 +33697,7 @@
 
 PUBLIC _sk_invert_hsw_lowp
 _sk_invert_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,223,137,0,0        ; vpbroadcastw  0x89df(%rip),%ymm8        # 2c286 <_sk_srcover_rgba_8888_sse2_lowp+0x87e>
+  DB  196,98,125,121,5,144,159,0,0        ; vpbroadcastw  0x9f90(%rip),%ymm8        # 2d926 <_sk_srcover_rgba_8888_sse2_lowp+0x872>
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
   DB  197,189,249,201                     ; vpsubw        %ymm1,%ymm8,%ymm1
   DB  197,189,249,210                     ; vpsubw        %ymm2,%ymm8,%ymm2
@@ -33656,7 +33717,7 @@
 PUBLIC _sk_srcatop_hsw_lowp
 _sk_srcatop_hsw_lowp LABEL PROC
   DB  197,197,213,192                     ; vpmullw       %ymm0,%ymm7,%ymm0
-  DB  196,98,125,121,5,172,137,0,0        ; vpbroadcastw  0x89ac(%rip),%ymm8        # 2c288 <_sk_srcover_rgba_8888_sse2_lowp+0x880>
+  DB  196,98,125,121,5,93,159,0,0         ; vpbroadcastw  0x9f5d(%rip),%ymm8        # 2d928 <_sk_srcover_rgba_8888_sse2_lowp+0x874>
   DB  197,189,249,219                     ; vpsubw        %ymm3,%ymm8,%ymm3
   DB  197,101,213,204                     ; vpmullw       %ymm4,%ymm3,%ymm9
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
@@ -33681,7 +33742,7 @@
 PUBLIC _sk_dstatop_hsw_lowp
 _sk_dstatop_hsw_lowp LABEL PROC
   DB  197,93,213,195                      ; vpmullw       %ymm3,%ymm4,%ymm8
-  DB  196,98,125,121,13,76,137,0,0        ; vpbroadcastw  0x894c(%rip),%ymm9        # 2c28a <_sk_srcover_rgba_8888_sse2_lowp+0x882>
+  DB  196,98,125,121,13,253,158,0,0       ; vpbroadcastw  0x9efd(%rip),%ymm9        # 2d92a <_sk_srcover_rgba_8888_sse2_lowp+0x876>
   DB  197,53,249,215                      ; vpsubw        %ymm7,%ymm9,%ymm10
   DB  197,173,213,192                     ; vpmullw       %ymm0,%ymm10,%ymm0
   DB  196,65,61,253,193                   ; vpaddw        %ymm9,%ymm8,%ymm8
@@ -33706,7 +33767,7 @@
 PUBLIC _sk_srcin_hsw_lowp
 _sk_srcin_hsw_lowp LABEL PROC
   DB  197,197,213,192                     ; vpmullw       %ymm0,%ymm7,%ymm0
-  DB  196,98,125,121,5,233,136,0,0        ; vpbroadcastw  0x88e9(%rip),%ymm8        # 2c28c <_sk_srcover_rgba_8888_sse2_lowp+0x884>
+  DB  196,98,125,121,5,154,158,0,0        ; vpbroadcastw  0x9e9a(%rip),%ymm8        # 2d92c <_sk_srcover_rgba_8888_sse2_lowp+0x878>
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,197,213,201                     ; vpmullw       %ymm1,%ymm7,%ymm1
@@ -33724,7 +33785,7 @@
 PUBLIC _sk_dstin_hsw_lowp
 _sk_dstin_hsw_lowp LABEL PROC
   DB  197,221,213,195                     ; vpmullw       %ymm3,%ymm4,%ymm0
-  DB  196,98,125,121,5,166,136,0,0        ; vpbroadcastw  0x88a6(%rip),%ymm8        # 2c28e <_sk_srcover_rgba_8888_sse2_lowp+0x886>
+  DB  196,98,125,121,5,87,158,0,0         ; vpbroadcastw  0x9e57(%rip),%ymm8        # 2d92e <_sk_srcover_rgba_8888_sse2_lowp+0x87a>
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,213,213,203                     ; vpmullw       %ymm3,%ymm5,%ymm1
@@ -33741,7 +33802,7 @@
 
 PUBLIC _sk_srcout_hsw_lowp
 _sk_srcout_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,103,136,0,0        ; vpbroadcastw  0x8867(%rip),%ymm8        # 2c290 <_sk_srcover_rgba_8888_sse2_lowp+0x888>
+  DB  196,98,125,121,5,24,158,0,0         ; vpbroadcastw  0x9e18(%rip),%ymm8        # 2d930 <_sk_srcover_rgba_8888_sse2_lowp+0x87c>
   DB  197,61,249,207                      ; vpsubw        %ymm7,%ymm8,%ymm9
   DB  197,181,213,192                     ; vpmullw       %ymm0,%ymm9,%ymm0
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
@@ -33760,7 +33821,7 @@
 
 PUBLIC _sk_dstout_hsw_lowp
 _sk_dstout_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,32,136,0,0         ; vpbroadcastw  0x8820(%rip),%ymm8        # 2c292 <_sk_srcover_rgba_8888_sse2_lowp+0x88a>
+  DB  196,98,125,121,5,209,157,0,0        ; vpbroadcastw  0x9dd1(%rip),%ymm8        # 2d932 <_sk_srcover_rgba_8888_sse2_lowp+0x87e>
   DB  197,189,249,219                     ; vpsubw        %ymm3,%ymm8,%ymm3
   DB  197,229,213,196                     ; vpmullw       %ymm4,%ymm3,%ymm0
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
@@ -33779,7 +33840,7 @@
 
 PUBLIC _sk_srcover_hsw_lowp
 _sk_srcover_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,217,135,0,0        ; vpbroadcastw  0x87d9(%rip),%ymm8        # 2c294 <_sk_srcover_rgba_8888_sse2_lowp+0x88c>
+  DB  196,98,125,121,5,138,157,0,0        ; vpbroadcastw  0x9d8a(%rip),%ymm8        # 2d934 <_sk_srcover_rgba_8888_sse2_lowp+0x880>
   DB  197,61,249,203                      ; vpsubw        %ymm3,%ymm8,%ymm9
   DB  197,53,213,212                      ; vpmullw       %ymm4,%ymm9,%ymm10
   DB  196,65,45,253,208                   ; vpaddw        %ymm8,%ymm10,%ymm10
@@ -33802,7 +33863,7 @@
 
 PUBLIC _sk_dstover_hsw_lowp
 _sk_dstover_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,126,135,0,0        ; vpbroadcastw  0x877e(%rip),%ymm8        # 2c296 <_sk_srcover_rgba_8888_sse2_lowp+0x88e>
+  DB  196,98,125,121,5,47,157,0,0         ; vpbroadcastw  0x9d2f(%rip),%ymm8        # 2d936 <_sk_srcover_rgba_8888_sse2_lowp+0x882>
   DB  197,61,249,207                      ; vpsubw        %ymm7,%ymm8,%ymm9
   DB  197,181,213,192                     ; vpmullw       %ymm0,%ymm9,%ymm0
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
@@ -33826,7 +33887,7 @@
 PUBLIC _sk_modulate_hsw_lowp
 _sk_modulate_hsw_lowp LABEL PROC
   DB  197,221,213,192                     ; vpmullw       %ymm0,%ymm4,%ymm0
-  DB  196,98,125,121,5,35,135,0,0         ; vpbroadcastw  0x8723(%rip),%ymm8        # 2c298 <_sk_srcover_rgba_8888_sse2_lowp+0x890>
+  DB  196,98,125,121,5,212,156,0,0        ; vpbroadcastw  0x9cd4(%rip),%ymm8        # 2d938 <_sk_srcover_rgba_8888_sse2_lowp+0x884>
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,213,213,201                     ; vpmullw       %ymm1,%ymm5,%ymm1
@@ -33843,7 +33904,7 @@
 
 PUBLIC _sk_multiply_hsw_lowp
 _sk_multiply_hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,228,134,0,0        ; vpbroadcastw  0x86e4(%rip),%ymm8        # 2c29a <_sk_srcover_rgba_8888_sse2_lowp+0x892>
+  DB  196,98,125,121,5,149,156,0,0        ; vpbroadcastw  0x9c95(%rip),%ymm8        # 2d93a <_sk_srcover_rgba_8888_sse2_lowp+0x886>
   DB  197,61,249,203                      ; vpsubw        %ymm3,%ymm8,%ymm9
   DB  197,53,213,212                      ; vpmullw       %ymm4,%ymm9,%ymm10
   DB  196,65,93,253,216                   ; vpaddw        %ymm8,%ymm4,%ymm11
@@ -33877,7 +33938,7 @@
 PUBLIC _sk_plus__hsw_lowp
 _sk_plus__hsw_lowp LABEL PROC
   DB  197,221,253,192                     ; vpaddw        %ymm0,%ymm4,%ymm0
-  DB  196,98,125,121,5,90,134,0,0         ; vpbroadcastw  0x865a(%rip),%ymm8        # 2c29c <_sk_srcover_rgba_8888_sse2_lowp+0x894>
+  DB  196,98,125,121,5,11,156,0,0         ; vpbroadcastw  0x9c0b(%rip),%ymm8        # 2d93c <_sk_srcover_rgba_8888_sse2_lowp+0x888>
   DB  196,194,125,58,192                  ; vpminuw       %ymm8,%ymm0,%ymm0
   DB  197,213,253,201                     ; vpaddw        %ymm1,%ymm5,%ymm1
   DB  196,194,117,58,200                  ; vpminuw       %ymm8,%ymm1,%ymm1
@@ -33892,7 +33953,7 @@
 _sk_screen_hsw_lowp LABEL PROC
   DB  197,93,253,192                      ; vpaddw        %ymm0,%ymm4,%ymm8
   DB  197,221,213,192                     ; vpmullw       %ymm0,%ymm4,%ymm0
-  DB  196,98,125,121,13,39,134,0,0        ; vpbroadcastw  0x8627(%rip),%ymm9        # 2c29e <_sk_srcover_rgba_8888_sse2_lowp+0x896>
+  DB  196,98,125,121,13,216,155,0,0       ; vpbroadcastw  0x9bd8(%rip),%ymm9        # 2d93e <_sk_srcover_rgba_8888_sse2_lowp+0x88a>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
@@ -33916,7 +33977,7 @@
 
 PUBLIC _sk_xor__hsw_lowp
 _sk_xor__hsw_lowp LABEL PROC
-  DB  196,98,125,121,5,204,133,0,0        ; vpbroadcastw  0x85cc(%rip),%ymm8        # 2c2a0 <_sk_srcover_rgba_8888_sse2_lowp+0x898>
+  DB  196,98,125,121,5,125,155,0,0        ; vpbroadcastw  0x9b7d(%rip),%ymm8        # 2d940 <_sk_srcover_rgba_8888_sse2_lowp+0x88c>
   DB  197,61,249,207                      ; vpsubw        %ymm7,%ymm8,%ymm9
   DB  197,181,213,192                     ; vpmullw       %ymm0,%ymm9,%ymm0
   DB  197,61,249,211                      ; vpsubw        %ymm3,%ymm8,%ymm10
@@ -33948,7 +34009,7 @@
   DB  197,197,213,192                     ; vpmullw       %ymm0,%ymm7,%ymm0
   DB  197,93,213,203                      ; vpmullw       %ymm3,%ymm4,%ymm9
   DB  196,194,125,62,193                  ; vpmaxuw       %ymm9,%ymm0,%ymm0
-  DB  196,98,125,121,13,80,133,0,0        ; vpbroadcastw  0x8550(%rip),%ymm9        # 2c2a2 <_sk_srcover_rgba_8888_sse2_lowp+0x89a>
+  DB  196,98,125,121,13,1,155,0,0         ; vpbroadcastw  0x9b01(%rip),%ymm9        # 2d942 <_sk_srcover_rgba_8888_sse2_lowp+0x88e>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
@@ -33980,7 +34041,7 @@
   DB  197,197,213,192                     ; vpmullw       %ymm0,%ymm7,%ymm0
   DB  197,93,213,203                      ; vpmullw       %ymm3,%ymm4,%ymm9
   DB  196,194,125,58,193                  ; vpminuw       %ymm9,%ymm0,%ymm0
-  DB  196,98,125,121,13,209,132,0,0       ; vpbroadcastw  0x84d1(%rip),%ymm9        # 2c2a4 <_sk_srcover_rgba_8888_sse2_lowp+0x89c>
+  DB  196,98,125,121,13,130,154,0,0       ; vpbroadcastw  0x9a82(%rip),%ymm9        # 2d944 <_sk_srcover_rgba_8888_sse2_lowp+0x890>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
@@ -34012,10 +34073,10 @@
   DB  197,197,213,192                     ; vpmullw       %ymm0,%ymm7,%ymm0
   DB  197,93,213,203                      ; vpmullw       %ymm3,%ymm4,%ymm9
   DB  196,194,125,58,193                  ; vpminuw       %ymm9,%ymm0,%ymm0
-  DB  196,98,125,121,13,82,132,0,0        ; vpbroadcastw  0x8452(%rip),%ymm9        # 2c2a6 <_sk_srcover_rgba_8888_sse2_lowp+0x89e>
+  DB  196,98,125,121,13,3,154,0,0         ; vpbroadcastw  0x9a03(%rip),%ymm9        # 2d946 <_sk_srcover_rgba_8888_sse2_lowp+0x892>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,7                   ; vpsrlw        $0x7,%ymm0,%ymm0
-  DB  196,98,125,121,21,65,132,0,0        ; vpbroadcastw  0x8441(%rip),%ymm10        # 2c2a8 <_sk_srcover_rgba_8888_sse2_lowp+0x8a0>
+  DB  196,98,125,121,21,242,153,0,0       ; vpbroadcastw  0x99f2(%rip),%ymm10        # 2d948 <_sk_srcover_rgba_8888_sse2_lowp+0x894>
   DB  196,193,125,219,194                 ; vpand         %ymm10,%ymm0,%ymm0
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
   DB  197,85,253,193                      ; vpaddw        %ymm1,%ymm5,%ymm8
@@ -34046,10 +34107,10 @@
 _sk_exclusion_hsw_lowp LABEL PROC
   DB  197,93,253,192                      ; vpaddw        %ymm0,%ymm4,%ymm8
   DB  197,221,213,192                     ; vpmullw       %ymm0,%ymm4,%ymm0
-  DB  196,98,125,121,13,198,131,0,0       ; vpbroadcastw  0x83c6(%rip),%ymm9        # 2c2aa <_sk_srcover_rgba_8888_sse2_lowp+0x8a2>
+  DB  196,98,125,121,13,119,153,0,0       ; vpbroadcastw  0x9977(%rip),%ymm9        # 2d94a <_sk_srcover_rgba_8888_sse2_lowp+0x896>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,7                   ; vpsrlw        $0x7,%ymm0,%ymm0
-  DB  196,98,125,121,21,181,131,0,0       ; vpbroadcastw  0x83b5(%rip),%ymm10        # 2c2ac <_sk_srcover_rgba_8888_sse2_lowp+0x8a4>
+  DB  196,98,125,121,21,102,153,0,0       ; vpbroadcastw  0x9966(%rip),%ymm10        # 2d94c <_sk_srcover_rgba_8888_sse2_lowp+0x898>
   DB  196,193,125,219,194                 ; vpand         %ymm10,%ymm0,%ymm0
   DB  197,189,249,192                     ; vpsubw        %ymm0,%ymm8,%ymm0
   DB  197,85,253,193                      ; vpaddw        %ymm1,%ymm5,%ymm8
@@ -34075,7 +34136,7 @@
 PUBLIC _sk_hardlight_hsw_lowp
 _sk_hardlight_hsw_lowp LABEL PROC
   DB  197,125,253,192                     ; vpaddw        %ymm0,%ymm0,%ymm8
-  DB  196,98,125,121,13,82,131,0,0        ; vpbroadcastw  0x8352(%rip),%ymm9        # 2c2b0 <_sk_srcover_rgba_8888_sse2_lowp+0x8a8>
+  DB  196,98,125,121,13,3,153,0,0         ; vpbroadcastw  0x9903(%rip),%ymm9        # 2d950 <_sk_srcover_rgba_8888_sse2_lowp+0x89c>
   DB  196,65,101,239,209                  ; vpxor         %ymm9,%ymm3,%ymm10
   DB  196,65,61,239,217                   ; vpxor         %ymm9,%ymm8,%ymm11
   DB  196,65,37,101,218                   ; vpcmpgtw      %ymm10,%ymm11,%ymm11
@@ -34103,7 +34164,7 @@
   DB  197,101,249,210                     ; vpsubw        %ymm2,%ymm3,%ymm10
   DB  197,69,249,246                      ; vpsubw        %ymm6,%ymm7,%ymm14
   DB  196,65,45,213,214                   ; vpmullw       %ymm14,%ymm10,%ymm10
-  DB  196,98,125,121,53,201,130,0,0       ; vpbroadcastw  0x82c9(%rip),%ymm14        # 2c2ae <_sk_srcover_rgba_8888_sse2_lowp+0x8a6>
+  DB  196,98,125,121,53,122,152,0,0       ; vpbroadcastw  0x987a(%rip),%ymm14        # 2d94e <_sk_srcover_rgba_8888_sse2_lowp+0x89a>
   DB  196,65,45,253,210                   ; vpaddw        %ymm10,%ymm10,%ymm10
   DB  196,65,61,249,194                   ; vpsubw        %ymm10,%ymm8,%ymm8
   DB  197,13,249,215                      ; vpsubw        %ymm7,%ymm14,%ymm10
@@ -34138,7 +34199,7 @@
 PUBLIC _sk_overlay_hsw_lowp
 _sk_overlay_hsw_lowp LABEL PROC
   DB  197,93,253,196                      ; vpaddw        %ymm4,%ymm4,%ymm8
-  DB  196,98,125,121,13,62,130,0,0        ; vpbroadcastw  0x823e(%rip),%ymm9        # 2c2b4 <_sk_srcover_rgba_8888_sse2_lowp+0x8ac>
+  DB  196,98,125,121,13,239,151,0,0       ; vpbroadcastw  0x97ef(%rip),%ymm9        # 2d954 <_sk_srcover_rgba_8888_sse2_lowp+0x8a0>
   DB  196,65,69,239,209                   ; vpxor         %ymm9,%ymm7,%ymm10
   DB  196,65,61,239,193                   ; vpxor         %ymm9,%ymm8,%ymm8
   DB  196,65,61,101,218                   ; vpcmpgtw      %ymm10,%ymm8,%ymm11
@@ -34162,7 +34223,7 @@
   DB  196,65,21,253,237                   ; vpaddw        %ymm13,%ymm13,%ymm13
   DB  196,65,61,249,237                   ; vpsubw        %ymm13,%ymm8,%ymm13
   DB  196,67,13,76,229,192                ; vpblendvb     %ymm12,%ymm13,%ymm14,%ymm12
-  DB  196,98,125,121,45,198,129,0,0       ; vpbroadcastw  0x81c6(%rip),%ymm13        # 2c2b2 <_sk_srcover_rgba_8888_sse2_lowp+0x8aa>
+  DB  196,98,125,121,45,119,151,0,0       ; vpbroadcastw  0x9777(%rip),%ymm13        # 2d952 <_sk_srcover_rgba_8888_sse2_lowp+0x89e>
   DB  197,77,253,246                      ; vpaddw        %ymm6,%ymm6,%ymm14
   DB  196,65,13,239,201                   ; vpxor         %ymm9,%ymm14,%ymm9
   DB  196,65,53,101,202                   ; vpcmpgtw      %ymm10,%ymm9,%ymm9
@@ -34212,10 +34273,10 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,55                              ; ja            241e2 <_sk_load_8888_hsw_lowp+0x52>
+  DB  119,55                              ; ja            242d1 <_sk_load_8888_hsw_lowp+0x52>
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 2435c <_sk_load_8888_hsw_lowp+0x1cc>
+  DB  76,141,13,160,1,0,0                 ; lea           0x1a0(%rip),%r9        # 24448 <_sk_load_8888_hsw_lowp+0x1c9>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34225,10 +34286,10 @@
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,12,144              ; vmovd         (%r8,%rdx,4),%xmm1
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
-  DB  233,39,1,0,0                        ; jmpq          24309 <_sk_load_8888_hsw_lowp+0x179>
+  DB  233,39,1,0,0                        ; jmpq          243f8 <_sk_load_8888_hsw_lowp+0x179>
   DB  196,193,126,111,12,144              ; vmovdqu       (%r8,%rdx,4),%ymm1
   DB  196,193,126,111,68,144,32           ; vmovdqu       0x20(%r8,%rdx,4),%ymm0
-  DB  233,21,1,0,0                        ; jmpq          24309 <_sk_load_8888_hsw_lowp+0x179>
+  DB  233,21,1,0,0                        ; jmpq          243f8 <_sk_load_8888_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,8            ; vmovd         0x8(%r8,%rdx,4),%xmm0
   DB  196,226,121,89,192                  ; vpbroadcastq  %xmm0,%xmm0
   DB  197,229,239,219                     ; vpxor         %ymm3,%ymm3,%ymm3
@@ -34237,7 +34298,7 @@
   DB  197,249,112,201,232                 ; vpshufd       $0xe8,%xmm1,%xmm1
   DB  196,227,125,2,201,3                 ; vpblendd      $0x3,%ymm1,%ymm0,%ymm1
   DB  197,253,111,195                     ; vmovdqa       %ymm3,%ymm0
-  DB  233,229,0,0,0                       ; jmpq          24309 <_sk_load_8888_hsw_lowp+0x179>
+  DB  233,229,0,0,0                       ; jmpq          243f8 <_sk_load_8888_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,24           ; vmovd         0x18(%r8,%rdx,4),%xmm0
   DB  196,226,125,89,192                  ; vpbroadcastq  %xmm0,%ymm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34253,7 +34314,7 @@
   DB  196,193,122,111,12,144              ; vmovdqu       (%r8,%rdx,4),%xmm1
   DB  196,227,117,2,200,240               ; vpblendd      $0xf0,%ymm0,%ymm1,%ymm1
   DB  197,125,127,192                     ; vmovdqa       %ymm8,%ymm0
-  DB  233,138,0,0,0                       ; jmpq          24309 <_sk_load_8888_hsw_lowp+0x179>
+  DB  233,138,0,0,0                       ; jmpq          243f8 <_sk_load_8888_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,40           ; vmovd         0x28(%r8,%rdx,4),%xmm0
   DB  196,226,121,89,192                  ; vpbroadcastq  %xmm0,%xmm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34263,7 +34324,7 @@
   DB  196,193,121,110,76,144,32           ; vmovd         0x20(%r8,%rdx,4),%xmm1
   DB  196,227,125,2,193,1                 ; vpblendd      $0x1,%ymm1,%ymm0,%ymm0
   DB  196,193,126,111,12,144              ; vmovdqu       (%r8,%rdx,4),%ymm1
-  DB  235,81                              ; jmp           24309 <_sk_load_8888_hsw_lowp+0x179>
+  DB  235,81                              ; jmp           243f8 <_sk_load_8888_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,56           ; vmovd         0x38(%r8,%rdx,4),%xmm0
   DB  196,226,125,89,192                  ; vpbroadcastq  %xmm0,%ymm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34283,7 +34344,7 @@
   DB  196,227,101,14,200,170              ; vpblendw      $0xaa,%ymm0,%ymm3,%ymm1
   DB  196,227,109,14,192,170              ; vpblendw      $0xaa,%ymm0,%ymm2,%ymm0
   DB  196,226,125,43,201                  ; vpackusdw     %ymm1,%ymm0,%ymm1
-  DB  197,125,111,5,142,127,0,0           ; vmovdqa       0x7f8e(%rip),%ymm8        # 2c2c0 <_sk_srcover_rgba_8888_sse2_lowp+0x8b8>
+  DB  197,125,111,5,63,149,0,0            ; vmovdqa       0x953f(%rip),%ymm8        # 2d960 <_sk_srcover_rgba_8888_sse2_lowp+0x8ac>
   DB  196,193,117,219,192                 ; vpand         %ymm8,%ymm1,%ymm0
   DB  197,245,113,209,8                   ; vpsrlw        $0x8,%ymm1,%ymm1
   DB  197,229,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm3
@@ -34293,39 +34354,41 @@
   DB  197,229,113,211,8                   ; vpsrlw        $0x8,%ymm3,%ymm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
-  DB  15,31,0                             ; nopl          (%rax)
-  DB  119,254                             ; ja            2435c <_sk_load_8888_hsw_lowp+0x1cc>
+  DB  122,254                             ; jp            24448 <_sk_load_8888_hsw_lowp+0x1c9>
   DB  255                                 ; (bad)
-  DB  255,174,254,255,255,152             ; ljmp          *-0x67000002(%rsi)
+  DB  255,177,254,255,255,155             ; pushq         -0x64000002(%rcx)
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,14                              ; decl          (%rsi)
+  DB  255,17                              ; callq         *(%rcx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,246                             ; push          %rsi
+  DB  255                                 ; (bad)
+  DB  249                                 ; stc
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  222,254                             ; fdivrp        %st,%st(6)
-  DB  255                                 ; (bad)
-  DB  255,200                             ; dec           %eax
+  DB  255,225                             ; jmpq          *%rcx
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,84,255,255                      ; callq         *-0x1(%rdi,%rdi,8)
-  DB  255,71,255                          ; incl          -0x1(%rdi)
+  DB  255,203                             ; dec           %ebx
+  DB  254                                 ; (bad)
+  DB  255                                 ; (bad)
+  DB  255,87,255                          ; callq         *-0x1(%rdi)
+  DB  255                                 ; (bad)
+  DB  255,74,255                          ; decl          -0x1(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  57,255                              ; cmp           %edi,%edi
+  DB  60,255                              ; cmp           $0xff,%al
   DB  255                                 ; (bad)
-  DB  255,35                              ; jmpq          *(%rbx)
+  DB  255,38                              ; jmpq          *(%rsi)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,154,255,255,255,134             ; lcall         *-0x79000001(%rdx)
+  DB  255,157,255,255,255,137             ; lcall         *-0x76000001(%rbp)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,114,255                         ; pushq         -0x1(%rdx)
+  DB  255,117,255                         ; pushq         -0x1(%rbp)
   DB  255                                 ; (bad)
-  DB  255,92,255,255                      ; lcall         *-0x1(%rdi,%rdi,8)
+  DB  255,95,255                          ; lcall         *-0x1(%rdi)
+  DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
 PUBLIC _sk_load_8888_dst_hsw_lowp
@@ -34339,10 +34402,10 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,55                              ; ja            243ea <_sk_load_8888_dst_hsw_lowp+0x52>
+  DB  119,55                              ; ja            244d6 <_sk_load_8888_dst_hsw_lowp+0x52>
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 24564 <_sk_load_8888_dst_hsw_lowp+0x1cc>
+  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 24650 <_sk_load_8888_dst_hsw_lowp+0x1cc>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34352,10 +34415,10 @@
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,44,144              ; vmovd         (%r8,%rdx,4),%xmm5
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
-  DB  233,39,1,0,0                        ; jmpq          24511 <_sk_load_8888_dst_hsw_lowp+0x179>
+  DB  233,39,1,0,0                        ; jmpq          245fd <_sk_load_8888_dst_hsw_lowp+0x179>
   DB  196,193,126,111,44,144              ; vmovdqu       (%r8,%rdx,4),%ymm5
   DB  196,193,126,111,100,144,32          ; vmovdqu       0x20(%r8,%rdx,4),%ymm4
-  DB  233,21,1,0,0                        ; jmpq          24511 <_sk_load_8888_dst_hsw_lowp+0x179>
+  DB  233,21,1,0,0                        ; jmpq          245fd <_sk_load_8888_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,8           ; vmovd         0x8(%r8,%rdx,4),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,197,239,255                     ; vpxor         %ymm7,%ymm7,%ymm7
@@ -34364,7 +34427,7 @@
   DB  197,249,112,237,232                 ; vpshufd       $0xe8,%xmm5,%xmm5
   DB  196,227,93,2,237,3                  ; vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   DB  197,253,111,231                     ; vmovdqa       %ymm7,%ymm4
-  DB  233,229,0,0,0                       ; jmpq          24511 <_sk_load_8888_dst_hsw_lowp+0x179>
+  DB  233,229,0,0,0                       ; jmpq          245fd <_sk_load_8888_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,24          ; vmovd         0x18(%r8,%rdx,4),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34380,7 +34443,7 @@
   DB  196,193,122,111,44,144              ; vmovdqu       (%r8,%rdx,4),%xmm5
   DB  196,227,85,2,236,240                ; vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   DB  197,125,127,196                     ; vmovdqa       %ymm8,%ymm4
-  DB  233,138,0,0,0                       ; jmpq          24511 <_sk_load_8888_dst_hsw_lowp+0x179>
+  DB  233,138,0,0,0                       ; jmpq          245fd <_sk_load_8888_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,40          ; vmovd         0x28(%r8,%rdx,4),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34390,7 +34453,7 @@
   DB  196,193,121,110,108,144,32          ; vmovd         0x20(%r8,%rdx,4),%xmm5
   DB  196,227,93,2,229,1                  ; vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   DB  196,193,126,111,44,144              ; vmovdqu       (%r8,%rdx,4),%ymm5
-  DB  235,81                              ; jmp           24511 <_sk_load_8888_dst_hsw_lowp+0x179>
+  DB  235,81                              ; jmp           245fd <_sk_load_8888_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,56          ; vmovd         0x38(%r8,%rdx,4),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34410,7 +34473,7 @@
   DB  196,227,69,14,236,170               ; vpblendw      $0xaa,%ymm4,%ymm7,%ymm5
   DB  196,227,77,14,228,170               ; vpblendw      $0xaa,%ymm4,%ymm6,%ymm4
   DB  196,226,93,43,237                   ; vpackusdw     %ymm5,%ymm4,%ymm5
-  DB  197,125,111,5,166,125,0,0           ; vmovdqa       0x7da6(%rip),%ymm8        # 2c2e0 <_sk_srcover_rgba_8888_sse2_lowp+0x8d8>
+  DB  197,125,111,5,90,147,0,0            ; vmovdqa       0x935a(%rip),%ymm8        # 2d980 <_sk_srcover_rgba_8888_sse2_lowp+0x8cc>
   DB  196,193,85,219,224                  ; vpand         %ymm8,%ymm5,%ymm4
   DB  197,213,113,213,8                   ; vpsrlw        $0x8,%ymm5,%ymm5
   DB  197,197,114,215,16                  ; vpsrld        $0x10,%ymm7,%ymm7
@@ -34421,7 +34484,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  119,254                             ; ja            24564 <_sk_load_8888_dst_hsw_lowp+0x1cc>
+  DB  119,254                             ; ja            24650 <_sk_load_8888_dst_hsw_lowp+0x1cc>
   DB  255                                 ; (bad)
   DB  255,174,254,255,255,152             ; ljmp          *-0x67000002(%rsi)
   DB  254                                 ; (bad)
@@ -34480,20 +34543,20 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,30                              ; ja            24621 <_sk_store_8888_hsw_lowp+0x81>
+  DB  119,30                              ; ja            2470d <_sk_store_8888_hsw_lowp+0x81>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 246d0 <_sk_store_8888_hsw_lowp+0x130>
+  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 247bc <_sk_store_8888_hsw_lowp+0x130>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  196,65,121,126,4,144                ; vmovd         %xmm8,(%r8,%rdx,4)
-  DB  233,171,0,0,0                       ; jmpq          246cc <_sk_store_8888_hsw_lowp+0x12c>
+  DB  233,171,0,0,0                       ; jmpq          247b8 <_sk_store_8888_hsw_lowp+0x12c>
   DB  196,65,126,127,4,144                ; vmovdqu       %ymm8,(%r8,%rdx,4)
   DB  196,65,126,127,76,144,32            ; vmovdqu       %ymm9,0x20(%r8,%rdx,4)
-  DB  233,153,0,0,0                       ; jmpq          246cc <_sk_store_8888_hsw_lowp+0x12c>
+  DB  233,153,0,0,0                       ; jmpq          247b8 <_sk_store_8888_hsw_lowp+0x12c>
   DB  196,67,121,22,68,144,8,2            ; vpextrd       $0x2,%xmm8,0x8(%r8,%rdx,4)
   DB  196,65,121,214,4,144                ; vmovq         %xmm8,(%r8,%rdx,4)
-  DB  233,134,0,0,0                       ; jmpq          246cc <_sk_store_8888_hsw_lowp+0x12c>
+  DB  233,134,0,0,0                       ; jmpq          247b8 <_sk_store_8888_hsw_lowp+0x12c>
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,22,76,144,24,2           ; vpextrd       $0x2,%xmm9,0x18(%r8,%rdx,4)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
@@ -34501,12 +34564,12 @@
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,65,121,126,76,144,16            ; vmovd         %xmm9,0x10(%r8,%rdx,4)
   DB  196,65,122,127,4,144                ; vmovdqu       %xmm8,(%r8,%rdx,4)
-  DB  235,85                              ; jmp           246cc <_sk_store_8888_hsw_lowp+0x12c>
+  DB  235,85                              ; jmp           247b8 <_sk_store_8888_hsw_lowp+0x12c>
   DB  196,67,121,22,76,144,40,2           ; vpextrd       $0x2,%xmm9,0x28(%r8,%rdx,4)
   DB  196,67,121,22,76,144,36,1           ; vpextrd       $0x1,%xmm9,0x24(%r8,%rdx,4)
   DB  196,65,121,126,76,144,32            ; vmovd         %xmm9,0x20(%r8,%rdx,4)
   DB  196,65,126,127,4,144                ; vmovdqu       %ymm8,(%r8,%rdx,4)
-  DB  235,54                              ; jmp           246cc <_sk_store_8888_hsw_lowp+0x12c>
+  DB  235,54                              ; jmp           247b8 <_sk_store_8888_hsw_lowp+0x12c>
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
   DB  196,67,121,22,84,144,56,2           ; vpextrd       $0x2,%xmm10,0x38(%r8,%rdx,4)
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
@@ -34561,10 +34624,10 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,55                              ; ja            2475e <_sk_load_bgra_hsw_lowp+0x52>
+  DB  119,55                              ; ja            2484a <_sk_load_bgra_hsw_lowp+0x52>
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 248d8 <_sk_load_bgra_hsw_lowp+0x1cc>
+  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 249c4 <_sk_load_bgra_hsw_lowp+0x1cc>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34574,10 +34637,10 @@
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,12,144              ; vmovd         (%r8,%rdx,4),%xmm1
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
-  DB  233,39,1,0,0                        ; jmpq          24885 <_sk_load_bgra_hsw_lowp+0x179>
+  DB  233,39,1,0,0                        ; jmpq          24971 <_sk_load_bgra_hsw_lowp+0x179>
   DB  196,193,126,111,12,144              ; vmovdqu       (%r8,%rdx,4),%ymm1
   DB  196,193,126,111,68,144,32           ; vmovdqu       0x20(%r8,%rdx,4),%ymm0
-  DB  233,21,1,0,0                        ; jmpq          24885 <_sk_load_bgra_hsw_lowp+0x179>
+  DB  233,21,1,0,0                        ; jmpq          24971 <_sk_load_bgra_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,8            ; vmovd         0x8(%r8,%rdx,4),%xmm0
   DB  196,226,121,89,192                  ; vpbroadcastq  %xmm0,%xmm0
   DB  197,229,239,219                     ; vpxor         %ymm3,%ymm3,%ymm3
@@ -34586,7 +34649,7 @@
   DB  197,249,112,201,232                 ; vpshufd       $0xe8,%xmm1,%xmm1
   DB  196,227,125,2,201,3                 ; vpblendd      $0x3,%ymm1,%ymm0,%ymm1
   DB  197,253,111,195                     ; vmovdqa       %ymm3,%ymm0
-  DB  233,229,0,0,0                       ; jmpq          24885 <_sk_load_bgra_hsw_lowp+0x179>
+  DB  233,229,0,0,0                       ; jmpq          24971 <_sk_load_bgra_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,24           ; vmovd         0x18(%r8,%rdx,4),%xmm0
   DB  196,226,125,89,192                  ; vpbroadcastq  %xmm0,%ymm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34602,7 +34665,7 @@
   DB  196,193,122,111,12,144              ; vmovdqu       (%r8,%rdx,4),%xmm1
   DB  196,227,117,2,200,240               ; vpblendd      $0xf0,%ymm0,%ymm1,%ymm1
   DB  197,125,127,192                     ; vmovdqa       %ymm8,%ymm0
-  DB  233,138,0,0,0                       ; jmpq          24885 <_sk_load_bgra_hsw_lowp+0x179>
+  DB  233,138,0,0,0                       ; jmpq          24971 <_sk_load_bgra_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,40           ; vmovd         0x28(%r8,%rdx,4),%xmm0
   DB  196,226,121,89,192                  ; vpbroadcastq  %xmm0,%xmm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34612,7 +34675,7 @@
   DB  196,193,121,110,76,144,32           ; vmovd         0x20(%r8,%rdx,4),%xmm1
   DB  196,227,125,2,193,1                 ; vpblendd      $0x1,%ymm1,%ymm0,%ymm0
   DB  196,193,126,111,12,144              ; vmovdqu       (%r8,%rdx,4),%ymm1
-  DB  235,81                              ; jmp           24885 <_sk_load_bgra_hsw_lowp+0x179>
+  DB  235,81                              ; jmp           24971 <_sk_load_bgra_hsw_lowp+0x179>
   DB  196,193,121,110,68,144,56           ; vmovd         0x38(%r8,%rdx,4),%xmm0
   DB  196,226,125,89,192                  ; vpbroadcastq  %xmm0,%ymm0
   DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
@@ -34632,7 +34695,7 @@
   DB  196,227,125,14,209,170              ; vpblendw      $0xaa,%ymm1,%ymm0,%ymm2
   DB  196,227,101,14,201,170              ; vpblendw      $0xaa,%ymm1,%ymm3,%ymm1
   DB  196,226,117,43,202                  ; vpackusdw     %ymm2,%ymm1,%ymm1
-  DB  197,125,111,5,82,122,0,0            ; vmovdqa       0x7a52(%rip),%ymm8        # 2c300 <_sk_srcover_rgba_8888_sse2_lowp+0x8f8>
+  DB  197,125,111,5,6,144,0,0             ; vmovdqa       0x9006(%rip),%ymm8        # 2d9a0 <_sk_srcover_rgba_8888_sse2_lowp+0x8ec>
   DB  196,193,117,219,208                 ; vpand         %ymm8,%ymm1,%ymm2
   DB  197,245,113,209,8                   ; vpsrlw        $0x8,%ymm1,%ymm1
   DB  197,253,114,208,16                  ; vpsrld        $0x10,%ymm0,%ymm0
@@ -34643,7 +34706,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  119,254                             ; ja            248d8 <_sk_load_bgra_hsw_lowp+0x1cc>
+  DB  119,254                             ; ja            249c4 <_sk_load_bgra_hsw_lowp+0x1cc>
   DB  255                                 ; (bad)
   DB  255,174,254,255,255,152             ; ljmp          *-0x67000002(%rsi)
   DB  254                                 ; (bad)
@@ -34688,10 +34751,10 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,55                              ; ja            24966 <_sk_load_bgra_dst_hsw_lowp+0x52>
+  DB  119,55                              ; ja            24a52 <_sk_load_bgra_dst_hsw_lowp+0x52>
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 24ae0 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
+  DB  76,141,13,163,1,0,0                 ; lea           0x1a3(%rip),%r9        # 24bcc <_sk_load_bgra_dst_hsw_lowp+0x1cc>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34701,10 +34764,10 @@
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,44,144              ; vmovd         (%r8,%rdx,4),%xmm5
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
-  DB  233,39,1,0,0                        ; jmpq          24a8d <_sk_load_bgra_dst_hsw_lowp+0x179>
+  DB  233,39,1,0,0                        ; jmpq          24b79 <_sk_load_bgra_dst_hsw_lowp+0x179>
   DB  196,193,126,111,44,144              ; vmovdqu       (%r8,%rdx,4),%ymm5
   DB  196,193,126,111,100,144,32          ; vmovdqu       0x20(%r8,%rdx,4),%ymm4
-  DB  233,21,1,0,0                        ; jmpq          24a8d <_sk_load_bgra_dst_hsw_lowp+0x179>
+  DB  233,21,1,0,0                        ; jmpq          24b79 <_sk_load_bgra_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,8           ; vmovd         0x8(%r8,%rdx,4),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,197,239,255                     ; vpxor         %ymm7,%ymm7,%ymm7
@@ -34713,7 +34776,7 @@
   DB  197,249,112,237,232                 ; vpshufd       $0xe8,%xmm5,%xmm5
   DB  196,227,93,2,237,3                  ; vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   DB  197,253,111,231                     ; vmovdqa       %ymm7,%ymm4
-  DB  233,229,0,0,0                       ; jmpq          24a8d <_sk_load_bgra_dst_hsw_lowp+0x179>
+  DB  233,229,0,0,0                       ; jmpq          24b79 <_sk_load_bgra_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,24          ; vmovd         0x18(%r8,%rdx,4),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34729,7 +34792,7 @@
   DB  196,193,122,111,44,144              ; vmovdqu       (%r8,%rdx,4),%xmm5
   DB  196,227,85,2,236,240                ; vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   DB  197,125,127,196                     ; vmovdqa       %ymm8,%ymm4
-  DB  233,138,0,0,0                       ; jmpq          24a8d <_sk_load_bgra_dst_hsw_lowp+0x179>
+  DB  233,138,0,0,0                       ; jmpq          24b79 <_sk_load_bgra_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,40          ; vmovd         0x28(%r8,%rdx,4),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34739,7 +34802,7 @@
   DB  196,193,121,110,108,144,32          ; vmovd         0x20(%r8,%rdx,4),%xmm5
   DB  196,227,93,2,229,1                  ; vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   DB  196,193,126,111,44,144              ; vmovdqu       (%r8,%rdx,4),%ymm5
-  DB  235,81                              ; jmp           24a8d <_sk_load_bgra_dst_hsw_lowp+0x179>
+  DB  235,81                              ; jmp           24b79 <_sk_load_bgra_dst_hsw_lowp+0x179>
   DB  196,193,121,110,100,144,56          ; vmovd         0x38(%r8,%rdx,4),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -34759,7 +34822,7 @@
   DB  196,227,93,14,245,170               ; vpblendw      $0xaa,%ymm5,%ymm4,%ymm6
   DB  196,227,69,14,237,170               ; vpblendw      $0xaa,%ymm5,%ymm7,%ymm5
   DB  196,226,85,43,238                   ; vpackusdw     %ymm6,%ymm5,%ymm5
-  DB  197,125,111,5,106,120,0,0           ; vmovdqa       0x786a(%rip),%ymm8        # 2c320 <_sk_srcover_rgba_8888_sse2_lowp+0x918>
+  DB  197,125,111,5,30,142,0,0            ; vmovdqa       0x8e1e(%rip),%ymm8        # 2d9c0 <_sk_srcover_rgba_8888_sse2_lowp+0x90c>
   DB  196,193,85,219,240                  ; vpand         %ymm8,%ymm5,%ymm6
   DB  197,213,113,213,8                   ; vpsrlw        $0x8,%ymm5,%ymm5
   DB  197,221,114,212,16                  ; vpsrld        $0x10,%ymm4,%ymm4
@@ -34770,7 +34833,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  119,254                             ; ja            24ae0 <_sk_load_bgra_dst_hsw_lowp+0x1cc>
+  DB  119,254                             ; ja            24bcc <_sk_load_bgra_dst_hsw_lowp+0x1cc>
   DB  255                                 ; (bad)
   DB  255,174,254,255,255,152             ; ljmp          *-0x67000002(%rsi)
   DB  254                                 ; (bad)
@@ -34829,20 +34892,20 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,30                              ; ja            24b9d <_sk_store_bgra_hsw_lowp+0x81>
+  DB  119,30                              ; ja            24c89 <_sk_store_bgra_hsw_lowp+0x81>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 24c4c <_sk_store_bgra_hsw_lowp+0x130>
+  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 24d38 <_sk_store_bgra_hsw_lowp+0x130>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  196,65,121,126,4,144                ; vmovd         %xmm8,(%r8,%rdx,4)
-  DB  233,171,0,0,0                       ; jmpq          24c48 <_sk_store_bgra_hsw_lowp+0x12c>
+  DB  233,171,0,0,0                       ; jmpq          24d34 <_sk_store_bgra_hsw_lowp+0x12c>
   DB  196,65,126,127,4,144                ; vmovdqu       %ymm8,(%r8,%rdx,4)
   DB  196,65,126,127,76,144,32            ; vmovdqu       %ymm9,0x20(%r8,%rdx,4)
-  DB  233,153,0,0,0                       ; jmpq          24c48 <_sk_store_bgra_hsw_lowp+0x12c>
+  DB  233,153,0,0,0                       ; jmpq          24d34 <_sk_store_bgra_hsw_lowp+0x12c>
   DB  196,67,121,22,68,144,8,2            ; vpextrd       $0x2,%xmm8,0x8(%r8,%rdx,4)
   DB  196,65,121,214,4,144                ; vmovq         %xmm8,(%r8,%rdx,4)
-  DB  233,134,0,0,0                       ; jmpq          24c48 <_sk_store_bgra_hsw_lowp+0x12c>
+  DB  233,134,0,0,0                       ; jmpq          24d34 <_sk_store_bgra_hsw_lowp+0x12c>
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,22,76,144,24,2           ; vpextrd       $0x2,%xmm9,0x18(%r8,%rdx,4)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
@@ -34850,12 +34913,12 @@
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,65,121,126,76,144,16            ; vmovd         %xmm9,0x10(%r8,%rdx,4)
   DB  196,65,122,127,4,144                ; vmovdqu       %xmm8,(%r8,%rdx,4)
-  DB  235,85                              ; jmp           24c48 <_sk_store_bgra_hsw_lowp+0x12c>
+  DB  235,85                              ; jmp           24d34 <_sk_store_bgra_hsw_lowp+0x12c>
   DB  196,67,121,22,76,144,40,2           ; vpextrd       $0x2,%xmm9,0x28(%r8,%rdx,4)
   DB  196,67,121,22,76,144,36,1           ; vpextrd       $0x1,%xmm9,0x24(%r8,%rdx,4)
   DB  196,65,121,126,76,144,32            ; vmovd         %xmm9,0x20(%r8,%rdx,4)
   DB  196,65,126,127,4,144                ; vmovdqu       %ymm8,(%r8,%rdx,4)
-  DB  235,54                              ; jmp           24c48 <_sk_store_bgra_hsw_lowp+0x12c>
+  DB  235,54                              ; jmp           24d34 <_sk_store_bgra_hsw_lowp+0x12c>
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
   DB  196,67,121,22,84,144,56,2           ; vpextrd       $0x2,%xmm10,0x38(%r8,%rdx,4)
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
@@ -34975,7 +35038,7 @@
   DB  196,195,109,14,192,170              ; vpblendw      $0xaa,%ymm8,%ymm2,%ymm0
   DB  196,195,101,14,200,170              ; vpblendw      $0xaa,%ymm8,%ymm3,%ymm1
   DB  196,226,117,43,200                  ; vpackusdw     %ymm0,%ymm1,%ymm1
-  DB  197,125,111,5,54,117,0,0            ; vmovdqa       0x7536(%rip),%ymm8        # 2c340 <_sk_srcover_rgba_8888_sse2_lowp+0x938>
+  DB  197,125,111,5,234,138,0,0           ; vmovdqa       0x8aea(%rip),%ymm8        # 2d9e0 <_sk_srcover_rgba_8888_sse2_lowp+0x92c>
   DB  196,193,117,219,192                 ; vpand         %ymm8,%ymm1,%ymm0
   DB  197,245,113,209,8                   ; vpsrlw        $0x8,%ymm1,%ymm1
   DB  197,237,114,210,16                  ; vpsrld        $0x10,%ymm2,%ymm2
@@ -34986,6 +35049,93 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_bgra_hsw_lowp
+_sk_gather_bgra_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,88,64,12                 ; vpbroadcastd  0xc(%rax),%ymm8
+  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
+  DB  196,65,61,254,209                   ; vpaddd        %ymm9,%ymm8,%ymm10
+  DB  196,65,61,239,192                   ; vpxor         %ymm8,%ymm8,%ymm8
+  DB  196,193,124,95,192                  ; vmaxps        %ymm8,%ymm0,%ymm0
+  DB  196,193,116,95,200                  ; vmaxps        %ymm8,%ymm1,%ymm1
+  DB  196,193,116,93,202                  ; vminps        %ymm10,%ymm1,%ymm1
+  DB  196,193,124,93,194                  ; vminps        %ymm10,%ymm0,%ymm0
+  DB  196,98,125,88,80,16                 ; vpbroadcastd  0x10(%rax),%ymm10
+  DB  196,65,45,254,201                   ; vpaddd        %ymm9,%ymm10,%ymm9
+  DB  196,193,100,95,216                  ; vmaxps        %ymm8,%ymm3,%ymm3
+  DB  196,193,108,95,208                  ; vmaxps        %ymm8,%ymm2,%ymm2
+  DB  196,193,108,93,209                  ; vminps        %ymm9,%ymm2,%ymm2
+  DB  196,193,100,93,217                  ; vminps        %ymm9,%ymm3,%ymm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  197,254,91,219                      ; vcvttps2dq    %ymm3,%ymm3
+  DB  197,254,91,210                      ; vcvttps2dq    %ymm2,%ymm2
+  DB  196,98,125,88,72,8                  ; vpbroadcastd  0x8(%rax),%ymm9
+  DB  196,226,53,64,210                   ; vpmulld       %ymm2,%ymm9,%ymm2
+  DB  196,226,53,64,219                   ; vpmulld       %ymm3,%ymm9,%ymm3
+  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
+  DB  197,109,254,208                     ; vpaddd        %ymm0,%ymm2,%ymm10
+  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
+  DB  197,229,254,201                     ; vpaddd        %ymm1,%ymm3,%ymm1
+  DB  196,67,125,57,209,1                 ; vextracti128  $0x1,%ymm10,%xmm9
+  DB  196,227,125,57,203,1                ; vextracti128  $0x1,%ymm1,%xmm3
+  DB  196,225,249,126,216                 ; vmovq         %xmm3,%rax
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  196,129,121,110,20,136              ; vmovd         (%r8,%r9,4),%xmm2
+  DB  196,195,249,22,217,1                ; vpextrq       $0x1,%xmm3,%r9
+  DB  196,195,105,34,20,128,1             ; vpinsrd       $0x1,(%r8,%rax,4),%xmm2,%xmm2
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  196,195,105,34,20,128,2             ; vpinsrd       $0x2,(%r8,%rax,4),%xmm2,%xmm2
+  DB  196,225,249,126,200                 ; vmovq         %xmm1,%rax
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  196,131,105,34,20,136,3             ; vpinsrd       $0x3,(%r8,%r9,4),%xmm2,%xmm2
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  196,129,121,110,28,136              ; vmovd         (%r8,%r9,4),%xmm3
+  DB  196,195,249,22,201,1                ; vpextrq       $0x1,%xmm1,%r9
+  DB  196,195,97,34,12,128,1              ; vpinsrd       $0x1,(%r8,%rax,4),%xmm3,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  196,195,113,34,12,128,2             ; vpinsrd       $0x2,(%r8,%rax,4),%xmm1,%xmm1
+  DB  196,97,249,126,200                  ; vmovq         %xmm9,%rax
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  196,131,113,34,12,136,3             ; vpinsrd       $0x3,(%r8,%r9,4),%xmm1,%xmm1
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  196,129,121,110,28,136              ; vmovd         (%r8,%r9,4),%xmm3
+  DB  196,67,249,22,201,1                 ; vpextrq       $0x1,%xmm9,%r9
+  DB  196,195,97,34,28,128,1              ; vpinsrd       $0x1,(%r8,%rax,4),%xmm3,%xmm3
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  196,195,97,34,28,128,2              ; vpinsrd       $0x2,(%r8,%rax,4),%xmm3,%xmm3
+  DB  196,97,249,126,208                  ; vmovq         %xmm10,%rax
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  196,131,97,34,28,136,3              ; vpinsrd       $0x3,(%r8,%r9,4),%xmm3,%xmm3
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  196,129,121,110,4,136               ; vmovd         (%r8,%r9,4),%xmm0
+  DB  196,67,249,22,209,1                 ; vpextrq       $0x1,%xmm10,%r9
+  DB  196,195,121,34,4,128,1              ; vpinsrd       $0x1,(%r8,%rax,4),%xmm0,%xmm0
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  196,195,121,34,4,128,2              ; vpinsrd       $0x2,(%r8,%rax,4),%xmm0,%xmm0
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  196,131,121,34,4,136,3              ; vpinsrd       $0x3,(%r8,%r9,4),%xmm0,%xmm0
+  DB  196,227,117,56,210,1                ; vinserti128   $0x1,%xmm2,%ymm1,%ymm2
+  DB  196,227,125,56,195,1                ; vinserti128   $0x1,%xmm3,%ymm0,%ymm0
+  DB  196,227,125,56,217,1                ; vinserti128   $0x1,%xmm1,%ymm0,%ymm3
+  DB  196,227,125,70,194,49               ; vperm2i128    $0x31,%ymm2,%ymm0,%ymm0
+  DB  196,195,125,14,200,170              ; vpblendw      $0xaa,%ymm8,%ymm0,%ymm1
+  DB  196,195,101,14,208,170              ; vpblendw      $0xaa,%ymm8,%ymm3,%ymm2
+  DB  196,226,109,43,201                  ; vpackusdw     %ymm1,%ymm2,%ymm1
+  DB  197,125,111,5,97,137,0,0            ; vmovdqa       0x8961(%rip),%ymm8        # 2da00 <_sk_srcover_rgba_8888_sse2_lowp+0x94c>
+  DB  196,193,117,219,208                 ; vpand         %ymm8,%ymm1,%ymm2
+  DB  197,245,113,209,8                   ; vpsrlw        $0x8,%ymm1,%ymm1
+  DB  197,253,114,208,16                  ; vpsrld        $0x10,%ymm0,%ymm0
+  DB  197,229,114,211,16                  ; vpsrld        $0x10,%ymm3,%ymm3
+  DB  196,226,101,43,216                  ; vpackusdw     %ymm0,%ymm3,%ymm3
+  DB  196,193,101,219,192                 ; vpand         %ymm8,%ymm3,%ymm0
+  DB  197,229,113,211,8                   ; vpsrlw        $0x8,%ymm3,%ymm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_565_hsw_lowp
 _sk_load_565_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -34997,24 +35147,24 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            24e70 <_sk_load_565_hsw_lowp+0x3f>
+  DB  119,37                              ; ja            25105 <_sk_load_565_hsw_lowp+0x3f>
   DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,131,1,0,0                 ; lea           0x183(%rip),%r9        # 24fdc <_sk_load_565_hsw_lowp+0x1ab>
+  DB  76,141,13,130,1,0,0                 ; lea           0x182(%rip),%r9        # 25270 <_sk_load_565_hsw_lowp+0x1aa>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  197,249,110,192                     ; vmovd         %eax,%xmm0
-  DB  233,8,1,0,0                         ; jmpq          24f78 <_sk_load_565_hsw_lowp+0x147>
+  DB  233,8,1,0,0                         ; jmpq          2520d <_sk_load_565_hsw_lowp+0x147>
   DB  196,193,126,111,4,80                ; vmovdqu       (%r8,%rdx,2),%ymm0
-  DB  233,253,0,0,0                       ; jmpq          24f78 <_sk_load_565_hsw_lowp+0x147>
+  DB  233,253,0,0,0                       ; jmpq          2520d <_sk_load_565_hsw_lowp+0x147>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,193,121,196,76,80,4,2           ; vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm0,%xmm1
   DB  196,227,117,56,192,1                ; vinserti128   $0x1,%xmm0,%ymm1,%ymm0
   DB  196,193,121,110,12,80               ; vmovd         (%r8,%rdx,2),%xmm1
   DB  196,227,125,2,193,1                 ; vpblendd      $0x1,%ymm1,%ymm0,%ymm0
-  DB  233,218,0,0,0                       ; jmpq          24f78 <_sk_load_565_hsw_lowp+0x147>
+  DB  233,218,0,0,0                       ; jmpq          2520d <_sk_load_565_hsw_lowp+0x147>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,193,121,196,76,80,12,6          ; vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm0,%xmm1
   DB  196,227,117,56,192,1                ; vinserti128   $0x1,%xmm0,%ymm1,%ymm0
@@ -35024,7 +35174,7 @@
   DB  196,227,125,2,193,15                ; vpblendd      $0xf,%ymm1,%ymm0,%ymm0
   DB  196,193,122,126,12,80               ; vmovq         (%r8,%rdx,2),%xmm1
   DB  196,227,125,2,193,3                 ; vpblendd      $0x3,%ymm1,%ymm0,%ymm0
-  DB  233,155,0,0,0                       ; jmpq          24f78 <_sk_load_565_hsw_lowp+0x147>
+  DB  233,155,0,0,0                       ; jmpq          2520d <_sk_load_565_hsw_lowp+0x147>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,193,121,196,76,80,20,2          ; vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm0,%xmm1
   DB  196,227,125,56,193,1                ; vinserti128   $0x1,%xmm1,%ymm0,%ymm0
@@ -35036,7 +35186,7 @@
   DB  196,227,125,56,193,1                ; vinserti128   $0x1,%xmm1,%ymm0,%ymm0
   DB  196,193,122,111,12,80               ; vmovdqu       (%r8,%rdx,2),%xmm1
   DB  196,227,117,2,192,240               ; vpblendd      $0xf0,%ymm0,%ymm1,%ymm0
-  DB  235,83                              ; jmp           24f78 <_sk_load_565_hsw_lowp+0x147>
+  DB  235,83                              ; jmp           2520d <_sk_load_565_hsw_lowp+0x147>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,193,121,196,76,80,28,6          ; vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm0,%xmm1
   DB  196,227,125,56,193,1                ; vinserti128   $0x1,%xmm1,%ymm0,%ymm0
@@ -35050,13 +35200,13 @@
   DB  196,193,122,111,20,80               ; vmovdqu       (%r8,%rdx,2),%xmm2
   DB  196,227,109,56,201,1                ; vinserti128   $0x1,%xmm1,%ymm2,%ymm1
   DB  196,227,117,2,192,192               ; vpblendd      $0xc0,%ymm0,%ymm1,%ymm0
-  DB  196,226,125,121,13,223,115,0,0      ; vpbroadcastw  0x73df(%rip),%ymm1        # 2c360 <_sk_srcover_rgba_8888_sse2_lowp+0x958>
+  DB  196,226,125,121,13,10,136,0,0       ; vpbroadcastw  0x880a(%rip),%ymm1        # 2da20 <_sk_srcover_rgba_8888_sse2_lowp+0x96c>
   DB  197,237,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm2
   DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
   DB  197,237,113,208,5                   ; vpsrlw        $0x5,%ymm0,%ymm2
-  DB  196,226,125,121,29,202,115,0,0      ; vpbroadcastw  0x73ca(%rip),%ymm3        # 2c362 <_sk_srcover_rgba_8888_sse2_lowp+0x95a>
+  DB  196,226,125,121,29,245,135,0,0      ; vpbroadcastw  0x87f5(%rip),%ymm3        # 2da22 <_sk_srcover_rgba_8888_sse2_lowp+0x96e>
   DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
-  DB  196,226,125,121,29,191,115,0,0      ; vpbroadcastw  0x73bf(%rip),%ymm3        # 2c364 <_sk_srcover_rgba_8888_sse2_lowp+0x95c>
+  DB  196,226,125,121,29,234,135,0,0      ; vpbroadcastw  0x87ea(%rip),%ymm3        # 2da24 <_sk_srcover_rgba_8888_sse2_lowp+0x970>
   DB  197,253,219,219                     ; vpand         %ymm3,%ymm0,%ymm3
   DB  197,253,113,208,13                  ; vpsrlw        $0xd,%ymm0,%ymm0
   DB  197,245,235,192                     ; vpor          %ymm0,%ymm1,%ymm0
@@ -35067,44 +35217,40 @@
   DB  197,229,113,211,2                   ; vpsrlw        $0x2,%ymm3,%ymm3
   DB  197,237,235,211                     ; vpor          %ymm3,%ymm2,%ymm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,29,141,115,0,0      ; vpbroadcastw  0x738d(%rip),%ymm3        # 2c366 <_sk_srcover_rgba_8888_sse2_lowp+0x95e>
+  DB  196,226,125,121,29,184,135,0,0      ; vpbroadcastw  0x87b8(%rip),%ymm3        # 2da26 <_sk_srcover_rgba_8888_sse2_lowp+0x972>
   DB  255,224                             ; jmpq          *%rax
-  DB  144                                 ; nop
-  DB  134,254                             ; xchg          %bh,%dh
+  DB  135,254                             ; xchg          %edi,%esi
   DB  255                                 ; (bad)
-  DB  255,177,254,255,255,159             ; pushq         -0x60000002(%rcx)
+  DB  255,178,254,255,255,160             ; pushq         -0x5f000002(%rdx)
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,240                             ; push          %rax
+  DB  255,241                             ; push          %rcx
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,226                             ; jmpq          *%rdx
+  DB  255,227                             ; jmpq          *%rbx
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,212                             ; callq         *%rsp
+  DB  255,213                             ; callq         *%rbp
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,194                             ; inc           %edx
+  DB  255,195                             ; inc           %ebx
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  59,255                              ; cmp           %edi,%edi
+  DB  60,255                              ; cmp           $0xff,%al
   DB  255                                 ; (bad)
-  DB  255,39                              ; jmpq          *(%rdi)
+  DB  255,40                              ; ljmp          *(%rax)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,19                              ; callq         *(%rbx)
+  DB  255,20,255                          ; callq         *(%rdi,%rdi,8)
+  DB  255                                 ; (bad)
+  DB  255,2                               ; incl          (%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,1                               ; incl          (%rcx)
+  DB  255,132,255,255,255,112,255         ; incl          -0x8f0001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  255,131,255,255,255,111             ; incl          0x6fffffff(%rbx)
-  DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  255,91,255                          ; lcall         *-0x1(%rbx)
-  DB  255                                 ; (bad)
-  DB  255,73,255                          ; decl          -0x1(%rcx)
+  DB  255,92,255,255                      ; lcall         *-0x1(%rdi,%rdi,8)
+  DB  255,74,255                          ; decl          -0x1(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
@@ -35119,24 +35265,24 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            25057 <_sk_load_565_dst_hsw_lowp+0x3f>
+  DB  119,37                              ; ja            252eb <_sk_load_565_dst_hsw_lowp+0x3f>
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,132,1,0,0                 ; lea           0x184(%rip),%r9        # 251c4 <_sk_load_565_dst_hsw_lowp+0x1ac>
+  DB  76,141,13,132,1,0,0                 ; lea           0x184(%rip),%r9        # 25458 <_sk_load_565_dst_hsw_lowp+0x1ac>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  197,249,110,224                     ; vmovd         %eax,%xmm4
-  DB  233,8,1,0,0                         ; jmpq          2515f <_sk_load_565_dst_hsw_lowp+0x147>
+  DB  233,8,1,0,0                         ; jmpq          253f3 <_sk_load_565_dst_hsw_lowp+0x147>
   DB  196,193,126,111,36,80               ; vmovdqu       (%r8,%rdx,2),%ymm4
-  DB  233,253,0,0,0                       ; jmpq          2515f <_sk_load_565_dst_hsw_lowp+0x147>
+  DB  233,253,0,0,0                       ; jmpq          253f3 <_sk_load_565_dst_hsw_lowp+0x147>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,193,89,196,108,80,4,2           ; vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm4,%xmm5
   DB  196,227,85,56,228,1                 ; vinserti128   $0x1,%xmm4,%ymm5,%ymm4
   DB  196,193,121,110,44,80               ; vmovd         (%r8,%rdx,2),%xmm5
   DB  196,227,93,2,229,1                  ; vpblendd      $0x1,%ymm5,%ymm4,%ymm4
-  DB  233,218,0,0,0                       ; jmpq          2515f <_sk_load_565_dst_hsw_lowp+0x147>
+  DB  233,218,0,0,0                       ; jmpq          253f3 <_sk_load_565_dst_hsw_lowp+0x147>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,193,89,196,108,80,12,6          ; vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm4,%xmm5
   DB  196,227,85,56,228,1                 ; vinserti128   $0x1,%xmm4,%ymm5,%ymm4
@@ -35146,7 +35292,7 @@
   DB  196,227,93,2,229,15                 ; vpblendd      $0xf,%ymm5,%ymm4,%ymm4
   DB  196,193,122,126,44,80               ; vmovq         (%r8,%rdx,2),%xmm5
   DB  196,227,93,2,229,3                  ; vpblendd      $0x3,%ymm5,%ymm4,%ymm4
-  DB  233,155,0,0,0                       ; jmpq          2515f <_sk_load_565_dst_hsw_lowp+0x147>
+  DB  233,155,0,0,0                       ; jmpq          253f3 <_sk_load_565_dst_hsw_lowp+0x147>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,193,89,196,108,80,20,2          ; vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm4,%xmm5
   DB  196,227,93,56,229,1                 ; vinserti128   $0x1,%xmm5,%ymm4,%ymm4
@@ -35158,7 +35304,7 @@
   DB  196,227,93,56,229,1                 ; vinserti128   $0x1,%xmm5,%ymm4,%ymm4
   DB  196,193,122,111,44,80               ; vmovdqu       (%r8,%rdx,2),%xmm5
   DB  196,227,85,2,228,240                ; vpblendd      $0xf0,%ymm4,%ymm5,%ymm4
-  DB  235,83                              ; jmp           2515f <_sk_load_565_dst_hsw_lowp+0x147>
+  DB  235,83                              ; jmp           253f3 <_sk_load_565_dst_hsw_lowp+0x147>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,193,89,196,108,80,28,6          ; vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm4,%xmm5
   DB  196,227,93,56,229,1                 ; vinserti128   $0x1,%xmm5,%ymm4,%ymm4
@@ -35172,13 +35318,13 @@
   DB  196,193,122,111,52,80               ; vmovdqu       (%r8,%rdx,2),%xmm6
   DB  196,227,77,56,237,1                 ; vinserti128   $0x1,%xmm5,%ymm6,%ymm5
   DB  196,227,85,2,228,192                ; vpblendd      $0xc0,%ymm4,%ymm5,%ymm4
-  DB  196,226,125,121,45,0,114,0,0        ; vpbroadcastw  0x7200(%rip),%ymm5        # 2c368 <_sk_srcover_rgba_8888_sse2_lowp+0x960>
+  DB  196,226,125,121,45,44,134,0,0       ; vpbroadcastw  0x862c(%rip),%ymm5        # 2da28 <_sk_srcover_rgba_8888_sse2_lowp+0x974>
   DB  197,205,113,212,8                   ; vpsrlw        $0x8,%ymm4,%ymm6
   DB  197,205,219,237                     ; vpand         %ymm5,%ymm6,%ymm5
   DB  197,205,113,212,5                   ; vpsrlw        $0x5,%ymm4,%ymm6
-  DB  196,226,125,121,61,235,113,0,0      ; vpbroadcastw  0x71eb(%rip),%ymm7        # 2c36a <_sk_srcover_rgba_8888_sse2_lowp+0x962>
+  DB  196,226,125,121,61,23,134,0,0       ; vpbroadcastw  0x8617(%rip),%ymm7        # 2da2a <_sk_srcover_rgba_8888_sse2_lowp+0x976>
   DB  197,205,219,247                     ; vpand         %ymm7,%ymm6,%ymm6
-  DB  196,226,125,121,61,224,113,0,0      ; vpbroadcastw  0x71e0(%rip),%ymm7        # 2c36c <_sk_srcover_rgba_8888_sse2_lowp+0x964>
+  DB  196,226,125,121,61,12,134,0,0       ; vpbroadcastw  0x860c(%rip),%ymm7        # 2da2c <_sk_srcover_rgba_8888_sse2_lowp+0x978>
   DB  197,221,219,255                     ; vpand         %ymm7,%ymm4,%ymm7
   DB  197,221,113,212,13                  ; vpsrlw        $0xd,%ymm4,%ymm4
   DB  197,213,235,228                     ; vpor          %ymm4,%ymm5,%ymm4
@@ -35189,7 +35335,7 @@
   DB  197,197,113,215,2                   ; vpsrlw        $0x2,%ymm7,%ymm7
   DB  197,205,235,247                     ; vpor          %ymm7,%ymm6,%ymm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,61,174,113,0,0      ; vpbroadcastw  0x71ae(%rip),%ymm7        # 2c36e <_sk_srcover_rgba_8888_sse2_lowp+0x966>
+  DB  196,226,125,121,61,218,133,0,0      ; vpbroadcastw  0x85da(%rip),%ymm7        # 2da2e <_sk_srcover_rgba_8888_sse2_lowp+0x97a>
   DB  255,224                             ; jmpq          *%rax
   DB  102,144                             ; xchg          %ax,%ax
   DB  133,254                             ; test          %edi,%esi
@@ -35238,10 +35384,10 @@
   DB  76,15,175,193                       ; imul          %rcx,%r8
   DB  77,1,192                            ; add           %r8,%r8
   DB  76,3,0                              ; add           (%rax),%r8
-  DB  196,98,125,121,5,87,113,0,0         ; vpbroadcastw  0x7157(%rip),%ymm8        # 2c370 <_sk_srcover_rgba_8888_sse2_lowp+0x968>
+  DB  196,98,125,121,5,131,133,0,0        ; vpbroadcastw  0x8583(%rip),%ymm8        # 2da30 <_sk_srcover_rgba_8888_sse2_lowp+0x97c>
   DB  197,181,113,240,8                   ; vpsllw        $0x8,%ymm0,%ymm9
   DB  196,65,53,219,192                   ; vpand         %ymm8,%ymm9,%ymm8
-  DB  196,98,125,121,13,70,113,0,0        ; vpbroadcastw  0x7146(%rip),%ymm9        # 2c372 <_sk_srcover_rgba_8888_sse2_lowp+0x96a>
+  DB  196,98,125,121,13,114,133,0,0       ; vpbroadcastw  0x8572(%rip),%ymm9        # 2da32 <_sk_srcover_rgba_8888_sse2_lowp+0x97e>
   DB  197,173,113,241,3                   ; vpsllw        $0x3,%ymm1,%ymm10
   DB  196,65,45,219,201                   ; vpand         %ymm9,%ymm10,%ymm9
   DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
@@ -35251,31 +35397,31 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,31                              ; ja            2526e <_sk_store_565_hsw_lowp+0x6e>
+  DB  119,31                              ; ja            25502 <_sk_store_565_hsw_lowp+0x6e>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 2531c <_sk_store_565_hsw_lowp+0x11c>
+  DB  76,141,13,195,0,0,0                 ; lea           0xc3(%rip),%r9        # 255b0 <_sk_store_565_hsw_lowp+0x11c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  196,67,121,21,4,80,0                ; vpextrw       $0x0,%xmm8,(%r8,%rdx,2)
-  DB  233,167,0,0,0                       ; jmpq          25315 <_sk_store_565_hsw_lowp+0x115>
+  DB  233,167,0,0,0                       ; jmpq          255a9 <_sk_store_565_hsw_lowp+0x115>
   DB  196,65,126,127,4,80                 ; vmovdqu       %ymm8,(%r8,%rdx,2)
-  DB  233,156,0,0,0                       ; jmpq          25315 <_sk_store_565_hsw_lowp+0x115>
+  DB  233,156,0,0,0                       ; jmpq          255a9 <_sk_store_565_hsw_lowp+0x115>
   DB  196,67,121,21,68,80,4,2             ; vpextrw       $0x2,%xmm8,0x4(%r8,%rdx,2)
   DB  196,65,121,126,4,80                 ; vmovd         %xmm8,(%r8,%rdx,2)
-  DB  233,137,0,0,0                       ; jmpq          25315 <_sk_store_565_hsw_lowp+0x115>
+  DB  233,137,0,0,0                       ; jmpq          255a9 <_sk_store_565_hsw_lowp+0x115>
   DB  196,67,121,21,68,80,12,6            ; vpextrw       $0x6,%xmm8,0xc(%r8,%rdx,2)
   DB  196,67,121,21,68,80,10,5            ; vpextrw       $0x5,%xmm8,0xa(%r8,%rdx,2)
   DB  196,67,121,21,68,80,8,4             ; vpextrw       $0x4,%xmm8,0x8(%r8,%rdx,2)
   DB  196,65,121,214,4,80                 ; vmovq         %xmm8,(%r8,%rdx,2)
-  DB  235,105                             ; jmp           25315 <_sk_store_565_hsw_lowp+0x115>
+  DB  235,105                             ; jmp           255a9 <_sk_store_565_hsw_lowp+0x115>
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,21,76,80,20,2            ; vpextrw       $0x2,%xmm9,0x14(%r8,%rdx,2)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,21,76,80,18,1            ; vpextrw       $0x1,%xmm9,0x12(%r8,%rdx,2)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,21,76,80,16,0            ; vpextrw       $0x0,%xmm9,0x10(%r8,%rdx,2)
-  DB  235,55                              ; jmp           2530f <_sk_store_565_hsw_lowp+0x10f>
+  DB  235,55                              ; jmp           255a3 <_sk_store_565_hsw_lowp+0x10f>
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,21,76,80,28,6            ; vpextrw       $0x6,%xmm9,0x1c(%r8,%rdx,2)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
@@ -35298,7 +35444,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  120,255                             ; js            25331 <_sk_store_565_hsw_lowp+0x131>
+  DB  120,255                             ; js            255c5 <_sk_store_565_hsw_lowp+0x131>
   DB  255                                 ; (bad)
   DB  255,112,255                         ; pushq         -0x1(%rax)
   DB  255                                 ; (bad)
@@ -35322,6 +35468,112 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_565_hsw_lowp
+_sk_gather_565_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,88,64,12                 ; vpbroadcastd  0xc(%rax),%ymm8
+  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
+  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
+  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
+  DB  196,193,124,95,194                  ; vmaxps        %ymm10,%ymm0,%ymm0
+  DB  196,193,116,95,202                  ; vmaxps        %ymm10,%ymm1,%ymm1
+  DB  196,193,116,93,200                  ; vminps        %ymm8,%ymm1,%ymm1
+  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
+  DB  196,98,125,88,64,16                 ; vpbroadcastd  0x10(%rax),%ymm8
+  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
+  DB  196,193,100,95,218                  ; vmaxps        %ymm10,%ymm3,%ymm3
+  DB  196,193,108,95,210                  ; vmaxps        %ymm10,%ymm2,%ymm2
+  DB  196,193,108,93,208                  ; vminps        %ymm8,%ymm2,%ymm2
+  DB  196,193,100,93,216                  ; vminps        %ymm8,%ymm3,%ymm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  197,254,91,219                      ; vcvttps2dq    %ymm3,%ymm3
+  DB  197,254,91,210                      ; vcvttps2dq    %ymm2,%ymm2
+  DB  196,98,125,88,64,8                  ; vpbroadcastd  0x8(%rax),%ymm8
+  DB  196,226,61,64,210                   ; vpmulld       %ymm2,%ymm8,%ymm2
+  DB  196,226,61,64,219                   ; vpmulld       %ymm3,%ymm8,%ymm3
+  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
+  DB  197,237,254,192                     ; vpaddd        %ymm0,%ymm2,%ymm0
+  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
+  DB  197,229,254,201                     ; vpaddd        %ymm1,%ymm3,%ymm1
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,183,4,72                      ; movzwl        (%r8,%r9,2),%eax
+  DB  197,233,196,208,1                   ; vpinsrw       $0x1,%eax,%xmm2,%xmm2
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,201,1                ; vextracti128  $0x1,%ymm1,%xmm1
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,233,196,208,2                   ; vpinsrw       $0x2,%eax,%xmm2,%xmm2
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  67,15,183,4,80                      ; movzwl        (%r8,%r10,2),%eax
+  DB  197,233,196,208,3                   ; vpinsrw       $0x3,%eax,%xmm2,%xmm2
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,233,196,208,4                   ; vpinsrw       $0x4,%eax,%xmm2,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,183,4,72                      ; movzwl        (%r8,%r9,2),%eax
+  DB  197,233,196,200,5                   ; vpinsrw       $0x5,%eax,%xmm2,%xmm1
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,241,196,200,6                   ; vpinsrw       $0x6,%eax,%xmm1,%xmm1
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  67,15,183,4,80                      ; movzwl        (%r8,%r10,2),%eax
+  DB  197,241,196,200,7                   ; vpinsrw       $0x7,%eax,%xmm1,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,183,4,72                      ; movzwl        (%r8,%r9,2),%eax
+  DB  197,233,196,208,1                   ; vpinsrw       $0x1,%eax,%xmm2,%xmm2
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,233,196,208,2                   ; vpinsrw       $0x2,%eax,%xmm2,%xmm2
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  67,15,183,4,80                      ; movzwl        (%r8,%r10,2),%eax
+  DB  197,233,196,208,3                   ; vpinsrw       $0x3,%eax,%xmm2,%xmm2
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,233,196,208,4                   ; vpinsrw       $0x4,%eax,%xmm2,%xmm2
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,183,4,72                      ; movzwl        (%r8,%r9,2),%eax
+  DB  197,233,196,192,5                   ; vpinsrw       $0x5,%eax,%xmm2,%xmm0
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  197,249,196,192,6                   ; vpinsrw       $0x6,%eax,%xmm0,%xmm0
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  67,15,183,4,80                      ; movzwl        (%r8,%r10,2),%eax
+  DB  197,249,196,192,7                   ; vpinsrw       $0x7,%eax,%xmm0,%xmm0
+  DB  196,227,125,56,193,1                ; vinserti128   $0x1,%xmm1,%ymm0,%ymm0
+  DB  196,226,125,121,13,182,130,0,0      ; vpbroadcastw  0x82b6(%rip),%ymm1        # 2da34 <_sk_srcover_rgba_8888_sse2_lowp+0x980>
+  DB  197,237,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm2
+  DB  197,237,219,201                     ; vpand         %ymm1,%ymm2,%ymm1
+  DB  197,237,113,208,5                   ; vpsrlw        $0x5,%ymm0,%ymm2
+  DB  196,226,125,121,29,161,130,0,0      ; vpbroadcastw  0x82a1(%rip),%ymm3        # 2da36 <_sk_srcover_rgba_8888_sse2_lowp+0x982>
+  DB  197,237,219,211                     ; vpand         %ymm3,%ymm2,%ymm2
+  DB  196,226,125,121,29,150,130,0,0      ; vpbroadcastw  0x8296(%rip),%ymm3        # 2da38 <_sk_srcover_rgba_8888_sse2_lowp+0x984>
+  DB  197,253,219,219                     ; vpand         %ymm3,%ymm0,%ymm3
+  DB  197,253,113,208,13                  ; vpsrlw        $0xd,%ymm0,%ymm0
+  DB  197,245,235,192                     ; vpor          %ymm0,%ymm1,%ymm0
+  DB  197,245,113,242,2                   ; vpsllw        $0x2,%ymm2,%ymm1
+  DB  197,237,113,210,4                   ; vpsrlw        $0x4,%ymm2,%ymm2
+  DB  197,245,235,202                     ; vpor          %ymm2,%ymm1,%ymm1
+  DB  197,237,113,243,3                   ; vpsllw        $0x3,%ymm3,%ymm2
+  DB  197,229,113,211,2                   ; vpsrlw        $0x2,%ymm3,%ymm3
+  DB  197,237,235,211                     ; vpor          %ymm3,%ymm2,%ymm2
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,226,125,121,29,100,130,0,0      ; vpbroadcastw  0x8264(%rip),%ymm3        # 2da3a <_sk_srcover_rgba_8888_sse2_lowp+0x986>
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_a8_hsw_lowp
 _sk_load_a8_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -35332,38 +35584,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            25394 <_sk_load_a8_hsw_lowp+0x3c>
+  DB  119,37                              ; ja            25814 <_sk_load_a8_hsw_lowp+0x3c>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,219,0,0,0                 ; lea           0xdb(%rip),%r9        # 25458 <_sk_load_a8_hsw_lowp+0x100>
+  DB  76,141,13,219,0,0,0                 ; lea           0xdb(%rip),%r9        # 258d8 <_sk_load_a8_hsw_lowp+0x100>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,249,110,192                     ; vmovd         %eax,%xmm0
-  DB  233,175,0,0,0                       ; jmpq          25443 <_sk_load_a8_hsw_lowp+0xeb>
+  DB  233,175,0,0,0                       ; jmpq          258c3 <_sk_load_a8_hsw_lowp+0xeb>
   DB  196,193,122,111,4,16                ; vmovdqu       (%r8,%rdx,1),%xmm0
-  DB  233,164,0,0,0                       ; jmpq          25443 <_sk_load_a8_hsw_lowp+0xeb>
+  DB  233,164,0,0,0                       ; jmpq          258c3 <_sk_load_a8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,2,2            ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm0,%xmm0
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,249,110,200                     ; vmovd         %eax,%xmm1
   DB  196,227,121,14,193,1                ; vpblendw      $0x1,%xmm1,%xmm0,%xmm0
-  DB  233,132,0,0,0                       ; jmpq          25443 <_sk_load_a8_hsw_lowp+0xeb>
+  DB  233,132,0,0,0                       ; jmpq          258c3 <_sk_load_a8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,6,6            ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,5,5            ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,4,4            ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,193,121,110,12,16               ; vmovd         (%r8,%rdx,1),%xmm1
   DB  196,227,121,2,193,1                 ; vpblendd      $0x1,%xmm1,%xmm0,%xmm0
-  DB  235,90                              ; jmp           25443 <_sk_load_a8_hsw_lowp+0xeb>
+  DB  235,90                              ; jmp           258c3 <_sk_load_a8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,10,10          ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,9,9            ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,8,8            ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,193,122,126,12,16               ; vmovq         (%r8,%rdx,1),%xmm1
   DB  196,227,113,2,192,12                ; vpblendd      $0xc,%xmm0,%xmm1,%xmm0
-  DB  235,48                              ; jmp           25443 <_sk_load_a8_hsw_lowp+0xeb>
+  DB  235,48                              ; jmp           258c3 <_sk_load_a8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,14,14          ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,13,13          ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm0,%xmm0
@@ -35421,38 +35673,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            254d0 <_sk_load_a8_dst_hsw_lowp+0x3c>
+  DB  119,37                              ; ja            25950 <_sk_load_a8_dst_hsw_lowp+0x3c>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,219,0,0,0                 ; lea           0xdb(%rip),%r9        # 25594 <_sk_load_a8_dst_hsw_lowp+0x100>
+  DB  76,141,13,219,0,0,0                 ; lea           0xdb(%rip),%r9        # 25a14 <_sk_load_a8_dst_hsw_lowp+0x100>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,249,110,224                     ; vmovd         %eax,%xmm4
-  DB  233,175,0,0,0                       ; jmpq          2557f <_sk_load_a8_dst_hsw_lowp+0xeb>
+  DB  233,175,0,0,0                       ; jmpq          259ff <_sk_load_a8_dst_hsw_lowp+0xeb>
   DB  196,193,122,111,36,16               ; vmovdqu       (%r8,%rdx,1),%xmm4
-  DB  233,164,0,0,0                       ; jmpq          2557f <_sk_load_a8_dst_hsw_lowp+0xeb>
+  DB  233,164,0,0,0                       ; jmpq          259ff <_sk_load_a8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,2,2            ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm4,%xmm4
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,249,110,232                     ; vmovd         %eax,%xmm5
   DB  196,227,89,14,229,1                 ; vpblendw      $0x1,%xmm5,%xmm4,%xmm4
-  DB  233,132,0,0,0                       ; jmpq          2557f <_sk_load_a8_dst_hsw_lowp+0xeb>
+  DB  233,132,0,0,0                       ; jmpq          259ff <_sk_load_a8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,6,6            ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,5,5            ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,4,4            ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,193,121,110,44,16               ; vmovd         (%r8,%rdx,1),%xmm5
   DB  196,227,89,2,229,1                  ; vpblendd      $0x1,%xmm5,%xmm4,%xmm4
-  DB  235,90                              ; jmp           2557f <_sk_load_a8_dst_hsw_lowp+0xeb>
+  DB  235,90                              ; jmp           259ff <_sk_load_a8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,10,10          ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,9,9            ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,8,8            ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,193,122,126,44,16               ; vmovq         (%r8,%rdx,1),%xmm5
   DB  196,227,81,2,228,12                 ; vpblendd      $0xc,%xmm4,%xmm5,%xmm4
-  DB  235,48                              ; jmp           2557f <_sk_load_a8_dst_hsw_lowp+0xeb>
+  DB  235,48                              ; jmp           259ff <_sk_load_a8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,14,14          ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,13,13          ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm4,%xmm4
@@ -35507,7 +35759,7 @@
   DB  76,15,175,193                       ; imul          %rcx,%r8
   DB  76,3,0                              ; add           (%rax),%r8
   DB  196,195,125,57,216,1                ; vextracti128  $0x1,%ymm3,%xmm8
-  DB  197,121,111,13,53,117,0,0           ; vmovdqa       0x7535(%rip),%xmm9        # 2cb20 <_sk_srcover_rgba_8888_sse2_lowp+0x1118>
+  DB  197,121,111,13,149,135,0,0          ; vmovdqa       0x8795(%rip),%xmm9        # 2e200 <_sk_srcover_rgba_8888_sse2_lowp+0x114c>
   DB  196,66,57,0,193                     ; vpshufb       %xmm9,%xmm8,%xmm8
   DB  196,66,97,0,201                     ; vpshufb       %xmm9,%xmm3,%xmm9
   DB  196,65,49,108,192                   ; vpunpcklqdq   %xmm8,%xmm9,%xmm8
@@ -35515,28 +35767,28 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,28                              ; ja            25620 <_sk_store_a8_hsw_lowp+0x50>
+  DB  119,28                              ; ja            25aa0 <_sk_store_a8_hsw_lowp+0x50>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,146,0,0,0                 ; lea           0x92(%rip),%r9        # 256a0 <_sk_store_a8_hsw_lowp+0xd0>
+  DB  76,141,13,146,0,0,0                 ; lea           0x92(%rip),%r9        # 25b20 <_sk_store_a8_hsw_lowp+0xd0>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  196,67,121,20,4,16,0                ; vpextrb       $0x0,%xmm8,(%r8,%rdx,1)
-  DB  235,121                             ; jmp           25699 <_sk_store_a8_hsw_lowp+0xc9>
+  DB  235,121                             ; jmp           25b19 <_sk_store_a8_hsw_lowp+0xc9>
   DB  196,65,122,127,4,16                 ; vmovdqu       %xmm8,(%r8,%rdx,1)
-  DB  235,113                             ; jmp           25699 <_sk_store_a8_hsw_lowp+0xc9>
+  DB  235,113                             ; jmp           25b19 <_sk_store_a8_hsw_lowp+0xc9>
   DB  196,67,121,20,68,16,2,2             ; vpextrb       $0x2,%xmm8,0x2(%r8,%rdx,1)
   DB  196,67,121,21,4,16,0                ; vpextrw       $0x0,%xmm8,(%r8,%rdx,1)
-  DB  235,96                              ; jmp           25699 <_sk_store_a8_hsw_lowp+0xc9>
+  DB  235,96                              ; jmp           25b19 <_sk_store_a8_hsw_lowp+0xc9>
   DB  196,67,121,20,68,16,6,6             ; vpextrb       $0x6,%xmm8,0x6(%r8,%rdx,1)
   DB  196,67,121,20,68,16,5,5             ; vpextrb       $0x5,%xmm8,0x5(%r8,%rdx,1)
   DB  196,67,121,20,68,16,4,4             ; vpextrb       $0x4,%xmm8,0x4(%r8,%rdx,1)
   DB  196,65,121,126,4,16                 ; vmovd         %xmm8,(%r8,%rdx,1)
-  DB  235,64                              ; jmp           25699 <_sk_store_a8_hsw_lowp+0xc9>
+  DB  235,64                              ; jmp           25b19 <_sk_store_a8_hsw_lowp+0xc9>
   DB  196,67,121,20,68,16,10,10           ; vpextrb       $0xa,%xmm8,0xa(%r8,%rdx,1)
   DB  196,67,121,20,68,16,9,9             ; vpextrb       $0x9,%xmm8,0x9(%r8,%rdx,1)
   DB  196,67,121,20,68,16,8,8             ; vpextrb       $0x8,%xmm8,0x8(%r8,%rdx,1)
-  DB  235,32                              ; jmp           25693 <_sk_store_a8_hsw_lowp+0xc3>
+  DB  235,32                              ; jmp           25b13 <_sk_store_a8_hsw_lowp+0xc3>
   DB  196,67,121,20,68,16,14,14           ; vpextrb       $0xe,%xmm8,0xe(%r8,%rdx,1)
   DB  196,67,121,20,68,16,13,13           ; vpextrb       $0xd,%xmm8,0xd(%r8,%rdx,1)
   DB  196,67,121,20,68,16,12,12           ; vpextrb       $0xc,%xmm8,0xc(%r8,%rdx,1)
@@ -35545,7 +35797,7 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  119,255                             ; ja            256a1 <_sk_store_a8_hsw_lowp+0xd1>
+  DB  119,255                             ; ja            25b21 <_sk_store_a8_hsw_lowp+0xd1>
   DB  255                                 ; (bad)
   DB  255,144,255,255,255,136             ; callq         *-0x77000001(%rax)
   DB  255                                 ; (bad)
@@ -35580,6 +35832,98 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_a8_hsw_lowp
+_sk_gather_a8_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,88,64,12                 ; vpbroadcastd  0xc(%rax),%ymm8
+  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
+  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
+  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
+  DB  196,193,124,95,194                  ; vmaxps        %ymm10,%ymm0,%ymm0
+  DB  196,193,116,95,202                  ; vmaxps        %ymm10,%ymm1,%ymm1
+  DB  196,65,116,93,216                   ; vminps        %ymm8,%ymm1,%ymm11
+  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
+  DB  196,226,125,88,72,16                ; vpbroadcastd  0x10(%rax),%ymm1
+  DB  196,193,117,254,201                 ; vpaddd        %ymm9,%ymm1,%ymm1
+  DB  196,193,100,95,218                  ; vmaxps        %ymm10,%ymm3,%ymm3
+  DB  196,193,108,95,210                  ; vmaxps        %ymm10,%ymm2,%ymm2
+  DB  197,236,93,209                      ; vminps        %ymm1,%ymm2,%ymm2
+  DB  197,228,93,201                      ; vminps        %ymm1,%ymm3,%ymm1
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
+  DB  197,254,91,210                      ; vcvttps2dq    %ymm2,%ymm2
+  DB  196,226,125,88,88,8                 ; vpbroadcastd  0x8(%rax),%ymm3
+  DB  196,226,101,64,210                  ; vpmulld       %ymm2,%ymm3,%ymm2
+  DB  196,226,101,64,217                  ; vpmulld       %ymm1,%ymm3,%ymm3
+  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
+  DB  197,237,254,200                     ; vpaddd        %ymm0,%ymm2,%ymm1
+  DB  196,193,126,91,195                  ; vcvttps2dq    %ymm11,%ymm0
+  DB  197,229,254,192                     ; vpaddd        %ymm0,%ymm3,%ymm0
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,105,32,208,1                ; vpinsrb       $0x1,%eax,%xmm2,%xmm2
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,201,1                ; vextracti128  $0x1,%ymm1,%xmm1
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,105,32,208,2                ; vpinsrb       $0x2,%eax,%xmm2,%xmm2
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,105,32,208,3                ; vpinsrb       $0x3,%eax,%xmm2,%xmm2
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,105,32,208,4                ; vpinsrb       $0x4,%eax,%xmm2,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,105,32,200,5                ; vpinsrb       $0x5,%eax,%xmm2,%xmm1
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,6                ; vpinsrb       $0x6,%eax,%xmm1,%xmm1
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,113,32,200,7                ; vpinsrb       $0x7,%eax,%xmm1,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,8                ; vpinsrb       $0x8,%eax,%xmm1,%xmm1
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,113,32,200,9                ; vpinsrb       $0x9,%eax,%xmm1,%xmm1
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,10               ; vpinsrb       $0xa,%eax,%xmm1,%xmm1
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,113,32,200,11               ; vpinsrb       $0xb,%eax,%xmm1,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,12               ; vpinsrb       $0xc,%eax,%xmm1,%xmm1
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,113,32,192,13               ; vpinsrb       $0xd,%eax,%xmm1,%xmm0
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,121,32,192,14               ; vpinsrb       $0xe,%eax,%xmm0,%xmm0
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,121,32,192,15               ; vpinsrb       $0xf,%eax,%xmm0,%xmm0
+  DB  196,226,125,48,216                  ; vpmovzxbw     %xmm0,%ymm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  197,253,239,192                     ; vpxor         %ymm0,%ymm0,%ymm0
+  DB  197,245,239,201                     ; vpxor         %ymm1,%ymm1,%ymm1
+  DB  197,237,239,210                     ; vpxor         %ymm2,%ymm2,%ymm2
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_g8_hsw_lowp
 _sk_load_g8_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -35590,38 +35934,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            25718 <_sk_load_g8_hsw_lowp+0x3c>
+  DB  119,37                              ; ja            25d3f <_sk_load_g8_hsw_lowp+0x3c>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,227,0,0,0                 ; lea           0xe3(%rip),%r9        # 257e4 <_sk_load_g8_hsw_lowp+0x108>
+  DB  76,141,13,224,0,0,0                 ; lea           0xe0(%rip),%r9        # 25e08 <_sk_load_g8_hsw_lowp+0x105>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,249,110,192                     ; vmovd         %eax,%xmm0
-  DB  233,175,0,0,0                       ; jmpq          257c7 <_sk_load_g8_hsw_lowp+0xeb>
+  DB  233,175,0,0,0                       ; jmpq          25dee <_sk_load_g8_hsw_lowp+0xeb>
   DB  196,193,122,111,4,16                ; vmovdqu       (%r8,%rdx,1),%xmm0
-  DB  233,164,0,0,0                       ; jmpq          257c7 <_sk_load_g8_hsw_lowp+0xeb>
+  DB  233,164,0,0,0                       ; jmpq          25dee <_sk_load_g8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,2,2            ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm0,%xmm0
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,249,110,200                     ; vmovd         %eax,%xmm1
   DB  196,227,121,14,193,1                ; vpblendw      $0x1,%xmm1,%xmm0,%xmm0
-  DB  233,132,0,0,0                       ; jmpq          257c7 <_sk_load_g8_hsw_lowp+0xeb>
+  DB  233,132,0,0,0                       ; jmpq          25dee <_sk_load_g8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,6,6            ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,5,5            ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,4,4            ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,193,121,110,12,16               ; vmovd         (%r8,%rdx,1),%xmm1
   DB  196,227,121,2,193,1                 ; vpblendd      $0x1,%xmm1,%xmm0,%xmm0
-  DB  235,90                              ; jmp           257c7 <_sk_load_g8_hsw_lowp+0xeb>
+  DB  235,90                              ; jmp           25dee <_sk_load_g8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,10,10          ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,9,9            ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,8,8            ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,193,122,126,12,16               ; vmovq         (%r8,%rdx,1),%xmm1
   DB  196,227,113,2,192,12                ; vpblendd      $0xc,%xmm0,%xmm1,%xmm0
-  DB  235,48                              ; jmp           257c7 <_sk_load_g8_hsw_lowp+0xeb>
+  DB  235,48                              ; jmp           25dee <_sk_load_g8_hsw_lowp+0xeb>
   DB  197,249,239,192                     ; vpxor         %xmm0,%xmm0,%xmm0
   DB  196,195,121,32,68,16,14,14          ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm0,%xmm0
   DB  196,195,121,32,68,16,13,13          ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm0,%xmm0
@@ -35631,43 +35975,42 @@
   DB  196,227,113,2,192,8                 ; vpblendd      $0x8,%xmm0,%xmm1,%xmm0
   DB  196,226,125,48,192                  ; vpmovzxbw     %xmm0,%ymm0
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,29,157,107,0,0      ; vpbroadcastw  0x6b9d(%rip),%ymm3        # 2c374 <_sk_srcover_rgba_8888_sse2_lowp+0x96c>
+  DB  196,226,125,121,29,62,124,0,0       ; vpbroadcastw  0x7c3e(%rip),%ymm3        # 2da3c <_sk_srcover_rgba_8888_sse2_lowp+0x988>
   DB  197,253,111,200                     ; vmovdqa       %ymm0,%ymm1
   DB  197,253,111,208                     ; vmovdqa       %ymm0,%ymm2
   DB  255,224                             ; jmpq          *%rax
-  DB  15,31,0                             ; nopl          (%rax)
-  DB  38,255                              ; es            (bad)
+  DB  41,255                              ; sub           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,75,255                          ; decl          -0x1(%rbx)
+  DB  255,78,255                          ; decl          -0x1(%rsi)
+  DB  255                                 ; (bad)
+  DB  255,66,255                          ; incl          -0x1(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  63                                  ; (bad)
+  DB  126,255                             ; jle           25e15 <_sk_load_g8_hsw_lowp+0x112>
+  DB  255                                 ; (bad)
+  DB  255,118,255                         ; pushq         -0x1(%rsi)
+  DB  255                                 ; (bad)
+  DB  255,110,255                         ; ljmp          *-0x1(%rsi)
+  DB  255                                 ; (bad)
+  DB  255,98,255                          ; jmpq          *-0x1(%rdx)
+  DB  255                                 ; (bad)
+  DB  255,168,255,255,255,160             ; ljmp          *-0x5f000001(%rax)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           257f1 <_sk_load_g8_hsw_lowp+0x115>
-  DB  255                                 ; (bad)
-  DB  255,115,255                         ; pushq         -0x1(%rbx)
-  DB  255                                 ; (bad)
-  DB  255,107,255                         ; ljmp          *-0x1(%rbx)
-  DB  255                                 ; (bad)
-  DB  255,95,255                          ; lcall         *-0x1(%rdi)
-  DB  255                                 ; (bad)
-  DB  255,165,255,255,255,157             ; jmpq          *-0x62000001(%rbp)
+  DB  255,152,255,255,255,140             ; lcall         *-0x73000001(%rax)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,149,255,255,255,137             ; callq         *-0x76000001(%rbp)
+  DB  255,210                             ; callq         *%rdx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,207                             ; dec           %edi
+  DB  255,202                             ; dec           %edx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,199                             ; inc           %edi
+  DB  255,194                             ; inc           %edx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  191,255,255,255,179                 ; mov           $0xb3ffffff,%edi
-  DB  255                                 ; (bad)
+  DB  255                                 ; .byte         0xff
+  DB  182,255                             ; mov           $0xff,%dh
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
@@ -35681,38 +36024,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,37                              ; ja            2585c <_sk_load_g8_dst_hsw_lowp+0x3c>
+  DB  119,37                              ; ja            25e80 <_sk_load_g8_dst_hsw_lowp+0x3c>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,227,0,0,0                 ; lea           0xe3(%rip),%r9        # 25928 <_sk_load_g8_dst_hsw_lowp+0x108>
+  DB  76,141,13,227,0,0,0                 ; lea           0xe3(%rip),%r9        # 25f4c <_sk_load_g8_dst_hsw_lowp+0x108>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,249,110,224                     ; vmovd         %eax,%xmm4
-  DB  233,175,0,0,0                       ; jmpq          2590b <_sk_load_g8_dst_hsw_lowp+0xeb>
+  DB  233,175,0,0,0                       ; jmpq          25f2f <_sk_load_g8_dst_hsw_lowp+0xeb>
   DB  196,193,122,111,36,16               ; vmovdqu       (%r8,%rdx,1),%xmm4
-  DB  233,164,0,0,0                       ; jmpq          2590b <_sk_load_g8_dst_hsw_lowp+0xeb>
+  DB  233,164,0,0,0                       ; jmpq          25f2f <_sk_load_g8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,2,2            ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm4,%xmm4
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,249,110,232                     ; vmovd         %eax,%xmm5
   DB  196,227,89,14,229,1                 ; vpblendw      $0x1,%xmm5,%xmm4,%xmm4
-  DB  233,132,0,0,0                       ; jmpq          2590b <_sk_load_g8_dst_hsw_lowp+0xeb>
+  DB  233,132,0,0,0                       ; jmpq          25f2f <_sk_load_g8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,6,6            ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,5,5            ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,4,4            ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,193,121,110,44,16               ; vmovd         (%r8,%rdx,1),%xmm5
   DB  196,227,89,2,229,1                  ; vpblendd      $0x1,%xmm5,%xmm4,%xmm4
-  DB  235,90                              ; jmp           2590b <_sk_load_g8_dst_hsw_lowp+0xeb>
+  DB  235,90                              ; jmp           25f2f <_sk_load_g8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,10,10          ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,9,9            ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,8,8            ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,193,122,126,44,16               ; vmovq         (%r8,%rdx,1),%xmm5
   DB  196,227,81,2,228,12                 ; vpblendd      $0xc,%xmm4,%xmm5,%xmm4
-  DB  235,48                              ; jmp           2590b <_sk_load_g8_dst_hsw_lowp+0xeb>
+  DB  235,48                              ; jmp           25f2f <_sk_load_g8_dst_hsw_lowp+0xeb>
   DB  197,217,239,228                     ; vpxor         %xmm4,%xmm4,%xmm4
   DB  196,195,89,32,100,16,14,14          ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm4,%xmm4
   DB  196,195,89,32,100,16,13,13          ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm4,%xmm4
@@ -35722,7 +36065,7 @@
   DB  196,227,81,2,228,8                  ; vpblendd      $0x8,%xmm4,%xmm5,%xmm4
   DB  196,226,125,48,228                  ; vpmovzxbw     %xmm4,%ymm4
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  196,226,125,121,61,91,106,0,0       ; vpbroadcastw  0x6a5b(%rip),%ymm7        # 2c376 <_sk_srcover_rgba_8888_sse2_lowp+0x96e>
+  DB  196,226,125,121,61,255,122,0,0      ; vpbroadcastw  0x7aff(%rip),%ymm7        # 2da3e <_sk_srcover_rgba_8888_sse2_lowp+0x98a>
   DB  197,253,111,236                     ; vmovdqa       %ymm4,%ymm5
   DB  197,253,111,244                     ; vmovdqa       %ymm4,%ymm6
   DB  255,224                             ; jmpq          *%rax
@@ -35736,7 +36079,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           25935 <_sk_load_g8_dst_hsw_lowp+0x115>
+  DB  123,255                             ; jnp           25f59 <_sk_load_g8_dst_hsw_lowp+0x115>
   DB  255                                 ; (bad)
   DB  255,115,255                         ; pushq         -0x1(%rbx)
   DB  255                                 ; (bad)
@@ -35764,12 +36107,12 @@
 
 PUBLIC _sk_luminance_to_alpha_hsw_lowp
 _sk_luminance_to_alpha_hsw_lowp LABEL PROC
-  DB  196,226,125,121,29,11,106,0,0       ; vpbroadcastw  0x6a0b(%rip),%ymm3        # 2c378 <_sk_srcover_rgba_8888_sse2_lowp+0x970>
+  DB  196,226,125,121,29,175,122,0,0      ; vpbroadcastw  0x7aaf(%rip),%ymm3        # 2da40 <_sk_srcover_rgba_8888_sse2_lowp+0x98c>
   DB  197,253,213,195                     ; vpmullw       %ymm3,%ymm0,%ymm0
-  DB  196,226,125,121,29,0,106,0,0        ; vpbroadcastw  0x6a00(%rip),%ymm3        # 2c37a <_sk_srcover_rgba_8888_sse2_lowp+0x972>
+  DB  196,226,125,121,29,164,122,0,0      ; vpbroadcastw  0x7aa4(%rip),%ymm3        # 2da42 <_sk_srcover_rgba_8888_sse2_lowp+0x98e>
   DB  197,245,213,203                     ; vpmullw       %ymm3,%ymm1,%ymm1
   DB  197,245,253,192                     ; vpaddw        %ymm0,%ymm1,%ymm0
-  DB  196,226,125,121,13,241,105,0,0      ; vpbroadcastw  0x69f1(%rip),%ymm1        # 2c37c <_sk_srcover_rgba_8888_sse2_lowp+0x974>
+  DB  196,226,125,121,13,149,122,0,0      ; vpbroadcastw  0x7a95(%rip),%ymm1        # 2da44 <_sk_srcover_rgba_8888_sse2_lowp+0x990>
   DB  197,237,213,201                     ; vpmullw       %ymm1,%ymm2,%ymm1
   DB  197,253,253,193                     ; vpaddw        %ymm1,%ymm0,%ymm0
   DB  197,229,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm3
@@ -35779,17 +36122,109 @@
   DB  197,237,239,210                     ; vpxor         %ymm2,%ymm2,%ymm2
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_g8_hsw_lowp
+_sk_gather_g8_hsw_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,98,125,88,64,12                 ; vpbroadcastd  0xc(%rax),%ymm8
+  DB  196,65,53,118,201                   ; vpcmpeqd      %ymm9,%ymm9,%ymm9
+  DB  196,65,61,254,193                   ; vpaddd        %ymm9,%ymm8,%ymm8
+  DB  196,65,44,87,210                    ; vxorps        %ymm10,%ymm10,%ymm10
+  DB  196,193,124,95,194                  ; vmaxps        %ymm10,%ymm0,%ymm0
+  DB  196,193,116,95,202                  ; vmaxps        %ymm10,%ymm1,%ymm1
+  DB  196,65,116,93,216                   ; vminps        %ymm8,%ymm1,%ymm11
+  DB  196,193,124,93,192                  ; vminps        %ymm8,%ymm0,%ymm0
+  DB  196,226,125,88,72,16                ; vpbroadcastd  0x10(%rax),%ymm1
+  DB  196,193,117,254,201                 ; vpaddd        %ymm9,%ymm1,%ymm1
+  DB  196,193,100,95,218                  ; vmaxps        %ymm10,%ymm3,%ymm3
+  DB  196,193,108,95,210                  ; vmaxps        %ymm10,%ymm2,%ymm2
+  DB  197,236,93,209                      ; vminps        %ymm1,%ymm2,%ymm2
+  DB  197,228,93,201                      ; vminps        %ymm1,%ymm3,%ymm1
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  197,254,91,201                      ; vcvttps2dq    %ymm1,%ymm1
+  DB  197,254,91,210                      ; vcvttps2dq    %ymm2,%ymm2
+  DB  196,226,125,88,88,8                 ; vpbroadcastd  0x8(%rax),%ymm3
+  DB  196,226,101,64,210                  ; vpmulld       %ymm2,%ymm3,%ymm2
+  DB  196,226,101,64,217                  ; vpmulld       %ymm1,%ymm3,%ymm3
+  DB  197,254,91,192                      ; vcvttps2dq    %ymm0,%ymm0
+  DB  197,237,254,200                     ; vpaddd        %ymm0,%ymm2,%ymm1
+  DB  196,193,126,91,195                  ; vcvttps2dq    %ymm11,%ymm0
+  DB  197,229,254,192                     ; vpaddd        %ymm0,%ymm3,%ymm0
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  197,249,110,208                     ; vmovd         %eax,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,105,32,208,1                ; vpinsrb       $0x1,%eax,%xmm2,%xmm2
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,201,1                ; vextracti128  $0x1,%ymm1,%xmm1
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,105,32,208,2                ; vpinsrb       $0x2,%eax,%xmm2,%xmm2
+  DB  196,193,249,126,201                 ; vmovq         %xmm1,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,105,32,208,3                ; vpinsrb       $0x3,%eax,%xmm2,%xmm2
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,105,32,208,4                ; vpinsrb       $0x4,%eax,%xmm2,%xmm2
+  DB  196,195,249,22,202,1                ; vpextrq       $0x1,%xmm1,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,105,32,200,5                ; vpinsrb       $0x5,%eax,%xmm2,%xmm1
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,6                ; vpinsrb       $0x6,%eax,%xmm1,%xmm1
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,113,32,200,7                ; vpinsrb       $0x7,%eax,%xmm1,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,8                ; vpinsrb       $0x8,%eax,%xmm1,%xmm1
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,113,32,200,9                ; vpinsrb       $0x9,%eax,%xmm1,%xmm1
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  196,227,125,57,192,1                ; vextracti128  $0x1,%ymm0,%xmm0
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,10               ; vpinsrb       $0xa,%eax,%xmm1,%xmm1
+  DB  196,193,249,126,193                 ; vmovq         %xmm0,%r9
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,113,32,200,11               ; vpinsrb       $0xb,%eax,%xmm1,%xmm1
+  DB  68,137,200                          ; mov           %r9d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,113,32,200,12               ; vpinsrb       $0xc,%eax,%xmm1,%xmm1
+  DB  196,195,249,22,194,1                ; vpextrq       $0x1,%xmm0,%r10
+  DB  73,193,233,32                       ; shr           $0x20,%r9
+  DB  67,15,182,4,8                       ; movzbl        (%r8,%r9,1),%eax
+  DB  196,227,113,32,192,13               ; vpinsrb       $0xd,%eax,%xmm1,%xmm0
+  DB  68,137,208                          ; mov           %r10d,%eax
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  196,227,121,32,192,14               ; vpinsrb       $0xe,%eax,%xmm0,%xmm0
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  67,15,182,4,16                      ; movzbl        (%r8,%r10,1),%eax
+  DB  196,227,121,32,192,15               ; vpinsrb       $0xf,%eax,%xmm0,%xmm0
+  DB  196,226,125,48,192                  ; vpmovzxbw     %xmm0,%ymm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  196,226,125,121,29,216,120,0,0      ; vpbroadcastw  0x78d8(%rip),%ymm3        # 2da46 <_sk_srcover_rgba_8888_sse2_lowp+0x992>
+  DB  197,253,111,200                     ; vmovdqa       %ymm0,%ymm1
+  DB  197,253,111,208                     ; vmovdqa       %ymm0,%ymm2
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_scale_1_float_hsw_lowp
 _sk_scale_1_float_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,122,16,0                        ; vmovss        (%rax),%xmm8
-  DB  197,122,16,13,46,99,0,0             ; vmovss        0x632e(%rip),%xmm9        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  196,98,57,169,13,245,98,0,0         ; vfmadd213ss   0x62f5(%rip),%xmm8,%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,122,16,13,10,114,0,0            ; vmovss        0x720a(%rip),%xmm9        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,57,169,13,209,113,0,0        ; vfmadd213ss   0x71d1(%rip),%xmm8,%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,122,44,193                  ; vcvttss2si    %xmm9,%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
   DB  196,66,125,121,192                  ; vpbroadcastw  %xmm8,%ymm8
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
-  DB  196,98,125,121,13,164,105,0,0       ; vpbroadcastw  0x69a4(%rip),%ymm9        # 2c37e <_sk_srcover_rgba_8888_sse2_lowp+0x976>
+  DB  196,98,125,121,13,158,120,0,0       ; vpbroadcastw  0x789e(%rip),%ymm9        # 2da48 <_sk_srcover_rgba_8888_sse2_lowp+0x994>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,189,213,201                     ; vpmullw       %ymm1,%ymm8,%ymm1
@@ -35808,12 +36243,12 @@
 _sk_lerp_1_float_hsw_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  197,122,16,0                        ; vmovss        (%rax),%xmm8
-  DB  197,122,16,13,196,98,0,0            ; vmovss        0x62c4(%rip),%xmm9        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  196,98,57,169,13,139,98,0,0         ; vfmadd213ss   0x628b(%rip),%xmm8,%xmm9        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  197,122,16,13,160,113,0,0           ; vmovss        0x71a0(%rip),%xmm9        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  196,98,57,169,13,103,113,0,0        ; vfmadd213ss   0x7167(%rip),%xmm8,%xmm9        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  196,193,122,44,193                  ; vcvttss2si    %xmm9,%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
   DB  196,66,125,121,192                  ; vpbroadcastw  %xmm8,%ymm8
-  DB  196,98,125,121,13,64,105,0,0        ; vpbroadcastw  0x6940(%rip),%ymm9        # 2c380 <_sk_srcover_rgba_8888_sse2_lowp+0x978>
+  DB  196,98,125,121,13,58,120,0,0        ; vpbroadcastw  0x783a(%rip),%ymm9        # 2da4a <_sk_srcover_rgba_8888_sse2_lowp+0x996>
   DB  196,65,53,249,208                   ; vpsubw        %ymm8,%ymm9,%ymm10
   DB  197,45,213,220                      ; vpmullw       %ymm4,%ymm10,%ymm11
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
@@ -35848,38 +36283,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,38                              ; ja            25ade <_sk_scale_u8_hsw_lowp+0x3d>
+  DB  119,38                              ; ja            262ae <_sk_scale_u8_hsw_lowp+0x3d>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,21,1,0,0                  ; lea           0x115(%rip),%r9        # 25bdc <_sk_scale_u8_hsw_lowp+0x13b>
+  DB  76,141,13,21,1,0,0                  ; lea           0x115(%rip),%r9        # 263ac <_sk_scale_u8_hsw_lowp+0x13b>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
-  DB  233,179,0,0,0                       ; jmpq          25b91 <_sk_scale_u8_hsw_lowp+0xf0>
+  DB  233,179,0,0,0                       ; jmpq          26361 <_sk_scale_u8_hsw_lowp+0xf0>
   DB  196,65,122,111,4,16                 ; vmovdqu       (%r8,%rdx,1),%xmm8
-  DB  233,168,0,0,0                       ; jmpq          25b91 <_sk_scale_u8_hsw_lowp+0xf0>
+  DB  233,168,0,0,0                       ; jmpq          26361 <_sk_scale_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,2,2              ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm8,%xmm8
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,121,110,200                     ; vmovd         %eax,%xmm9
   DB  196,67,57,14,193,1                  ; vpblendw      $0x1,%xmm9,%xmm8,%xmm8
-  DB  233,135,0,0,0                       ; jmpq          25b91 <_sk_scale_u8_hsw_lowp+0xf0>
+  DB  233,135,0,0,0                       ; jmpq          26361 <_sk_scale_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,6,6              ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,5,5              ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,4,4              ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,65,121,110,12,16                ; vmovd         (%r8,%rdx,1),%xmm9
   DB  196,67,57,2,193,1                   ; vpblendd      $0x1,%xmm9,%xmm8,%xmm8
-  DB  235,92                              ; jmp           25b91 <_sk_scale_u8_hsw_lowp+0xf0>
+  DB  235,92                              ; jmp           26361 <_sk_scale_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,10,10            ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,9,9              ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,8,8              ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,65,122,126,12,16                ; vmovq         (%r8,%rdx,1),%xmm9
   DB  196,67,49,2,192,12                  ; vpblendd      $0xc,%xmm8,%xmm9,%xmm8
-  DB  235,49                              ; jmp           25b91 <_sk_scale_u8_hsw_lowp+0xf0>
+  DB  235,49                              ; jmp           26361 <_sk_scale_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,14,14            ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,13,13            ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm8,%xmm8
@@ -35889,7 +36324,7 @@
   DB  196,67,49,2,192,8                   ; vpblendd      $0x8,%xmm8,%xmm9,%xmm8
   DB  196,66,125,48,192                   ; vpmovzxbw     %xmm8,%ymm8
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
-  DB  196,98,125,121,13,223,103,0,0       ; vpbroadcastw  0x67df(%rip),%ymm9        # 2c382 <_sk_srcover_rgba_8888_sse2_lowp+0x97a>
+  DB  196,98,125,121,13,217,118,0,0       ; vpbroadcastw  0x76d9(%rip),%ymm9        # 2da4c <_sk_srcover_rgba_8888_sse2_lowp+0x998>
   DB  196,193,125,253,193                 ; vpaddw        %ymm9,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,189,213,201                     ; vpmullw       %ymm1,%ymm8,%ymm1
@@ -35910,7 +36345,7 @@
   DB  255,26                              ; lcall         *(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,75               ; decl          0x4bffffff(%rip)        # 4c025be8 <_sk_srcover_rgba_8888_sse2_lowp+0x4bffa1e0>
+  DB  255,13,255,255,255,75               ; decl          0x4bffffff(%rip)        # 4c0263b8 <_sk_srcover_rgba_8888_sse2_lowp+0x4bff9304>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,67,255                          ; incl          -0x1(%rbx)
@@ -35947,38 +36382,38 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,38                              ; ja            25c55 <_sk_lerp_u8_hsw_lowp+0x3d>
+  DB  119,38                              ; ja            26425 <_sk_lerp_u8_hsw_lowp+0x3d>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,58,1,0,0                  ; lea           0x13a(%rip),%r9        # 25d78 <_sk_lerp_u8_hsw_lowp+0x160>
+  DB  76,141,13,58,1,0,0                  ; lea           0x13a(%rip),%r9        # 26548 <_sk_lerp_u8_hsw_lowp+0x160>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
-  DB  233,179,0,0,0                       ; jmpq          25d08 <_sk_lerp_u8_hsw_lowp+0xf0>
+  DB  233,179,0,0,0                       ; jmpq          264d8 <_sk_lerp_u8_hsw_lowp+0xf0>
   DB  196,65,122,111,4,16                 ; vmovdqu       (%r8,%rdx,1),%xmm8
-  DB  233,168,0,0,0                       ; jmpq          25d08 <_sk_lerp_u8_hsw_lowp+0xf0>
+  DB  233,168,0,0,0                       ; jmpq          264d8 <_sk_lerp_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,2,2              ; vpinsrb       $0x2,0x2(%r8,%rdx,1),%xmm8,%xmm8
   DB  65,15,183,4,16                      ; movzwl        (%r8,%rdx,1),%eax
   DB  197,121,110,200                     ; vmovd         %eax,%xmm9
   DB  196,67,57,14,193,1                  ; vpblendw      $0x1,%xmm9,%xmm8,%xmm8
-  DB  233,135,0,0,0                       ; jmpq          25d08 <_sk_lerp_u8_hsw_lowp+0xf0>
+  DB  233,135,0,0,0                       ; jmpq          264d8 <_sk_lerp_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,6,6              ; vpinsrb       $0x6,0x6(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,5,5              ; vpinsrb       $0x5,0x5(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,4,4              ; vpinsrb       $0x4,0x4(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,65,121,110,12,16                ; vmovd         (%r8,%rdx,1),%xmm9
   DB  196,67,57,2,193,1                   ; vpblendd      $0x1,%xmm9,%xmm8,%xmm8
-  DB  235,92                              ; jmp           25d08 <_sk_lerp_u8_hsw_lowp+0xf0>
+  DB  235,92                              ; jmp           264d8 <_sk_lerp_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,10,10            ; vpinsrb       $0xa,0xa(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,9,9              ; vpinsrb       $0x9,0x9(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,8,8              ; vpinsrb       $0x8,0x8(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,65,122,126,12,16                ; vmovq         (%r8,%rdx,1),%xmm9
   DB  196,67,49,2,192,12                  ; vpblendd      $0xc,%xmm8,%xmm9,%xmm8
-  DB  235,49                              ; jmp           25d08 <_sk_lerp_u8_hsw_lowp+0xf0>
+  DB  235,49                              ; jmp           264d8 <_sk_lerp_u8_hsw_lowp+0xf0>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,67,57,32,68,16,14,14            ; vpinsrb       $0xe,0xe(%r8,%rdx,1),%xmm8,%xmm8
   DB  196,67,57,32,68,16,13,13            ; vpinsrb       $0xd,0xd(%r8,%rdx,1),%xmm8,%xmm8
@@ -35987,7 +36422,7 @@
   DB  196,67,49,34,76,16,8,2              ; vpinsrd       $0x2,0x8(%r8,%rdx,1),%xmm9,%xmm9
   DB  196,67,49,2,192,8                   ; vpblendd      $0x8,%xmm8,%xmm9,%xmm8
   DB  196,66,125,48,192                   ; vpmovzxbw     %xmm8,%ymm8
-  DB  196,98,125,121,13,110,102,0,0       ; vpbroadcastw  0x666e(%rip),%ymm9        # 2c384 <_sk_srcover_rgba_8888_sse2_lowp+0x97c>
+  DB  196,98,125,121,13,104,117,0,0       ; vpbroadcastw  0x7568(%rip),%ymm9        # 2da4e <_sk_srcover_rgba_8888_sse2_lowp+0x99a>
   DB  196,65,61,239,209                   ; vpxor         %ymm9,%ymm8,%ymm10
   DB  197,45,213,220                      ; vpmullw       %ymm4,%ymm10,%ymm11
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
@@ -36019,7 +36454,7 @@
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  232,254,255,255,38                  ; callq         27025d83 <_sk_srcover_rgba_8888_sse2_lowp+0x26ffa37b>
+  DB  232,254,255,255,38                  ; callq         27026553 <_sk_srcover_rgba_8888_sse2_lowp+0x26ff949f>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,30                              ; lcall         *(%rsi)
@@ -36040,7 +36475,7 @@
   DB  255,52,255                          ; pushq         (%rdi,%rdi,8)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  124,255                             ; jl            25da5 <_sk_lerp_u8_hsw_lowp+0x18d>
+  DB  124,255                             ; jl            26575 <_sk_lerp_u8_hsw_lowp+0x18d>
   DB  255                                 ; (bad)
   DB  255,116,255,255                     ; pushq         -0x1(%rdi,%rdi,8)
   DB  255,108,255,255                     ; ljmp          *-0x1(%rdi,%rdi,8)
@@ -36059,24 +36494,24 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,38                              ; ja            25df4 <_sk_scale_565_hsw_lowp+0x40>
+  DB  119,38                              ; ja            265c4 <_sk_scale_565_hsw_lowp+0x40>
   DB  196,65,61,239,192                   ; vpxor         %ymm8,%ymm8,%ymm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,255,1,0,0                 ; lea           0x1ff(%rip),%r9        # 25fdc <_sk_scale_565_hsw_lowp+0x228>
+  DB  76,141,13,255,1,0,0                 ; lea           0x1ff(%rip),%r9        # 267ac <_sk_scale_565_hsw_lowp+0x228>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
-  DB  233,12,1,0,0                        ; jmpq          25f00 <_sk_scale_565_hsw_lowp+0x14c>
+  DB  233,12,1,0,0                        ; jmpq          266d0 <_sk_scale_565_hsw_lowp+0x14c>
   DB  196,65,126,111,4,80                 ; vmovdqu       (%r8,%rdx,2),%ymm8
-  DB  233,1,1,0,0                         ; jmpq          25f00 <_sk_scale_565_hsw_lowp+0x14c>
+  DB  233,1,1,0,0                         ; jmpq          266d0 <_sk_scale_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,4,2             ; vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,53,56,192,1                  ; vinserti128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,121,110,12,80                ; vmovd         (%r8,%rdx,2),%xmm9
   DB  196,67,61,2,193,1                   ; vpblendd      $0x1,%ymm9,%ymm8,%ymm8
-  DB  233,221,0,0,0                       ; jmpq          25f00 <_sk_scale_565_hsw_lowp+0x14c>
+  DB  233,221,0,0,0                       ; jmpq          266d0 <_sk_scale_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,12,6            ; vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,53,56,192,1                  ; vinserti128   $0x1,%xmm8,%ymm9,%ymm8
@@ -36086,7 +36521,7 @@
   DB  196,67,61,2,193,15                  ; vpblendd      $0xf,%ymm9,%ymm8,%ymm8
   DB  196,65,122,126,12,80                ; vmovq         (%r8,%rdx,2),%xmm9
   DB  196,67,61,2,193,3                   ; vpblendd      $0x3,%ymm9,%ymm8,%ymm8
-  DB  233,157,0,0,0                       ; jmpq          25f00 <_sk_scale_565_hsw_lowp+0x14c>
+  DB  233,157,0,0,0                       ; jmpq          266d0 <_sk_scale_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,20,2            ; vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -36098,7 +36533,7 @@
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
   DB  196,65,122,111,12,80                ; vmovdqu       (%r8,%rdx,2),%xmm9
   DB  196,67,53,2,192,240                 ; vpblendd      $0xf0,%ymm8,%ymm9,%ymm8
-  DB  235,84                              ; jmp           25f00 <_sk_scale_565_hsw_lowp+0x14c>
+  DB  235,84                              ; jmp           266d0 <_sk_scale_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,28,6            ; vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -36112,13 +36547,13 @@
   DB  196,65,122,111,20,80                ; vmovdqu       (%r8,%rdx,2),%xmm10
   DB  196,67,45,56,201,1                  ; vinserti128   $0x1,%xmm9,%ymm10,%ymm9
   DB  196,67,53,2,192,192                 ; vpblendd      $0xc0,%ymm8,%ymm9,%ymm8
-  DB  196,98,125,121,13,125,100,0,0       ; vpbroadcastw  0x647d(%rip),%ymm9        # 2c386 <_sk_srcover_rgba_8888_sse2_lowp+0x97e>
+  DB  196,98,125,121,13,119,115,0,0       ; vpbroadcastw  0x7377(%rip),%ymm9        # 2da50 <_sk_srcover_rgba_8888_sse2_lowp+0x99c>
   DB  196,193,45,113,208,8                ; vpsrlw        $0x8,%ymm8,%ymm10
   DB  196,65,45,219,201                   ; vpand         %ymm9,%ymm10,%ymm9
   DB  196,193,45,113,208,5                ; vpsrlw        $0x5,%ymm8,%ymm10
-  DB  196,98,125,121,29,101,100,0,0       ; vpbroadcastw  0x6465(%rip),%ymm11        # 2c388 <_sk_srcover_rgba_8888_sse2_lowp+0x980>
+  DB  196,98,125,121,29,95,115,0,0        ; vpbroadcastw  0x735f(%rip),%ymm11        # 2da52 <_sk_srcover_rgba_8888_sse2_lowp+0x99e>
   DB  196,65,45,219,211                   ; vpand         %ymm11,%ymm10,%ymm10
-  DB  196,98,125,121,29,89,100,0,0        ; vpbroadcastw  0x6459(%rip),%ymm11        # 2c38a <_sk_srcover_rgba_8888_sse2_lowp+0x982>
+  DB  196,98,125,121,29,83,115,0,0        ; vpbroadcastw  0x7353(%rip),%ymm11        # 2da54 <_sk_srcover_rgba_8888_sse2_lowp+0x9a0>
   DB  196,65,61,219,219                   ; vpand         %ymm11,%ymm8,%ymm11
   DB  196,193,61,113,208,13               ; vpsrlw        $0xd,%ymm8,%ymm8
   DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
@@ -36128,7 +36563,7 @@
   DB  196,193,45,113,243,3                ; vpsllw        $0x3,%ymm11,%ymm10
   DB  196,193,37,113,211,2                ; vpsrlw        $0x2,%ymm11,%ymm11
   DB  196,65,45,235,211                   ; vpor          %ymm11,%ymm10,%ymm10
-  DB  196,98,125,121,29,32,100,0,0        ; vpbroadcastw  0x6420(%rip),%ymm11        # 2c38c <_sk_srcover_rgba_8888_sse2_lowp+0x984>
+  DB  196,98,125,121,29,26,115,0,0        ; vpbroadcastw  0x731a(%rip),%ymm11        # 2da56 <_sk_srcover_rgba_8888_sse2_lowp+0x9a2>
   DB  196,65,101,239,227                  ; vpxor         %ymm11,%ymm3,%ymm12
   DB  196,65,69,239,219                   ; vpxor         %ymm11,%ymm7,%ymm11
   DB  196,65,37,101,220                   ; vpcmpgtw      %ymm12,%ymm11,%ymm11
@@ -36138,7 +36573,7 @@
   DB  196,66,21,62,232                    ; vpmaxuw       %ymm8,%ymm13,%ymm13
   DB  196,67,21,76,220,176                ; vpblendvb     %ymm11,%ymm12,%ymm13,%ymm11
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
-  DB  196,98,125,121,5,236,99,0,0         ; vpbroadcastw  0x63ec(%rip),%ymm8        # 2c38e <_sk_srcover_rgba_8888_sse2_lowp+0x986>
+  DB  196,98,125,121,5,230,114,0,0        ; vpbroadcastw  0x72e6(%rip),%ymm8        # 2da58 <_sk_srcover_rgba_8888_sse2_lowp+0x9a4>
   DB  196,193,125,253,192                 ; vpaddw        %ymm8,%ymm0,%ymm0
   DB  197,253,113,208,8                   ; vpsrlw        $0x8,%ymm0,%ymm0
   DB  197,181,213,201                     ; vpmullw       %ymm1,%ymm9,%ymm1
@@ -36200,24 +36635,24 @@
   DB  36,15                               ; and           $0xf,%al
   DB  254,200                             ; dec           %al
   DB  60,14                               ; cmp           $0xe,%al
-  DB  119,38                              ; ja            26058 <_sk_lerp_565_hsw_lowp+0x40>
+  DB  119,38                              ; ja            26828 <_sk_lerp_565_hsw_lowp+0x40>
   DB  196,65,61,239,192                   ; vpxor         %ymm8,%ymm8,%ymm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,51,2,0,0                  ; lea           0x233(%rip),%r9        # 26274 <_sk_lerp_565_hsw_lowp+0x25c>
+  DB  76,141,13,51,2,0,0                  ; lea           0x233(%rip),%r9        # 26a44 <_sk_lerp_565_hsw_lowp+0x25c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  197,121,110,192                     ; vmovd         %eax,%xmm8
-  DB  233,12,1,0,0                        ; jmpq          26164 <_sk_lerp_565_hsw_lowp+0x14c>
+  DB  233,12,1,0,0                        ; jmpq          26934 <_sk_lerp_565_hsw_lowp+0x14c>
   DB  196,65,126,111,4,80                 ; vmovdqu       (%r8,%rdx,2),%ymm8
-  DB  233,1,1,0,0                         ; jmpq          26164 <_sk_lerp_565_hsw_lowp+0x14c>
+  DB  233,1,1,0,0                         ; jmpq          26934 <_sk_lerp_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,4,2             ; vpinsrw       $0x2,0x4(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,53,56,192,1                  ; vinserti128   $0x1,%xmm8,%ymm9,%ymm8
   DB  196,65,121,110,12,80                ; vmovd         (%r8,%rdx,2),%xmm9
   DB  196,67,61,2,193,1                   ; vpblendd      $0x1,%ymm9,%ymm8,%ymm8
-  DB  233,221,0,0,0                       ; jmpq          26164 <_sk_lerp_565_hsw_lowp+0x14c>
+  DB  233,221,0,0,0                       ; jmpq          26934 <_sk_lerp_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,12,6            ; vpinsrw       $0x6,0xc(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,53,56,192,1                  ; vinserti128   $0x1,%xmm8,%ymm9,%ymm8
@@ -36227,7 +36662,7 @@
   DB  196,67,61,2,193,15                  ; vpblendd      $0xf,%ymm9,%ymm8,%ymm8
   DB  196,65,122,126,12,80                ; vmovq         (%r8,%rdx,2),%xmm9
   DB  196,67,61,2,193,3                   ; vpblendd      $0x3,%ymm9,%ymm8,%ymm8
-  DB  233,157,0,0,0                       ; jmpq          26164 <_sk_lerp_565_hsw_lowp+0x14c>
+  DB  233,157,0,0,0                       ; jmpq          26934 <_sk_lerp_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,20,2            ; vpinsrw       $0x2,0x14(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -36239,7 +36674,7 @@
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
   DB  196,65,122,111,12,80                ; vmovdqu       (%r8,%rdx,2),%xmm9
   DB  196,67,53,2,192,240                 ; vpblendd      $0xf0,%ymm8,%ymm9,%ymm8
-  DB  235,84                              ; jmp           26164 <_sk_lerp_565_hsw_lowp+0x14c>
+  DB  235,84                              ; jmp           26934 <_sk_lerp_565_hsw_lowp+0x14c>
   DB  196,65,57,239,192                   ; vpxor         %xmm8,%xmm8,%xmm8
   DB  196,65,57,196,76,80,28,6            ; vpinsrw       $0x6,0x1c(%r8,%rdx,2),%xmm8,%xmm9
   DB  196,67,61,56,193,1                  ; vinserti128   $0x1,%xmm9,%ymm8,%ymm8
@@ -36253,13 +36688,13 @@
   DB  196,65,122,111,20,80                ; vmovdqu       (%r8,%rdx,2),%xmm10
   DB  196,67,45,56,201,1                  ; vinserti128   $0x1,%xmm9,%ymm10,%ymm9
   DB  196,67,53,2,192,192                 ; vpblendd      $0xc0,%ymm8,%ymm9,%ymm8
-  DB  196,98,125,121,13,35,98,0,0         ; vpbroadcastw  0x6223(%rip),%ymm9        # 2c390 <_sk_srcover_rgba_8888_sse2_lowp+0x988>
+  DB  196,98,125,121,13,29,113,0,0        ; vpbroadcastw  0x711d(%rip),%ymm9        # 2da5a <_sk_srcover_rgba_8888_sse2_lowp+0x9a6>
   DB  196,193,45,113,208,8                ; vpsrlw        $0x8,%ymm8,%ymm10
   DB  196,65,45,219,201                   ; vpand         %ymm9,%ymm10,%ymm9
   DB  196,193,45,113,208,5                ; vpsrlw        $0x5,%ymm8,%ymm10
-  DB  196,98,125,121,29,11,98,0,0         ; vpbroadcastw  0x620b(%rip),%ymm11        # 2c392 <_sk_srcover_rgba_8888_sse2_lowp+0x98a>
+  DB  196,98,125,121,29,5,113,0,0         ; vpbroadcastw  0x7105(%rip),%ymm11        # 2da5c <_sk_srcover_rgba_8888_sse2_lowp+0x9a8>
   DB  196,65,45,219,211                   ; vpand         %ymm11,%ymm10,%ymm10
-  DB  196,98,125,121,29,255,97,0,0        ; vpbroadcastw  0x61ff(%rip),%ymm11        # 2c394 <_sk_srcover_rgba_8888_sse2_lowp+0x98c>
+  DB  196,98,125,121,29,249,112,0,0       ; vpbroadcastw  0x70f9(%rip),%ymm11        # 2da5e <_sk_srcover_rgba_8888_sse2_lowp+0x9aa>
   DB  196,65,61,219,219                   ; vpand         %ymm11,%ymm8,%ymm11
   DB  196,193,61,113,208,13               ; vpsrlw        $0xd,%ymm8,%ymm8
   DB  196,65,53,235,192                   ; vpor          %ymm8,%ymm9,%ymm8
@@ -36269,7 +36704,7 @@
   DB  196,193,45,113,243,3                ; vpsllw        $0x3,%ymm11,%ymm10
   DB  196,193,37,113,211,2                ; vpsrlw        $0x2,%ymm11,%ymm11
   DB  196,65,45,235,211                   ; vpor          %ymm11,%ymm10,%ymm10
-  DB  196,98,125,121,29,198,97,0,0        ; vpbroadcastw  0x61c6(%rip),%ymm11        # 2c396 <_sk_srcover_rgba_8888_sse2_lowp+0x98e>
+  DB  196,98,125,121,29,192,112,0,0       ; vpbroadcastw  0x70c0(%rip),%ymm11        # 2da60 <_sk_srcover_rgba_8888_sse2_lowp+0x9ac>
   DB  196,65,101,239,227                  ; vpxor         %ymm11,%ymm3,%ymm12
   DB  196,65,69,239,219                   ; vpxor         %ymm11,%ymm7,%ymm11
   DB  196,65,37,101,220                   ; vpcmpgtw      %ymm12,%ymm11,%ymm11
@@ -36278,7 +36713,7 @@
   DB  196,66,53,62,234                    ; vpmaxuw       %ymm10,%ymm9,%ymm13
   DB  196,66,21,62,232                    ; vpmaxuw       %ymm8,%ymm13,%ymm13
   DB  196,67,21,76,220,176                ; vpblendvb     %ymm11,%ymm12,%ymm13,%ymm11
-  DB  196,98,125,121,37,150,97,0,0        ; vpbroadcastw  0x6196(%rip),%ymm12        # 2c398 <_sk_srcover_rgba_8888_sse2_lowp+0x990>
+  DB  196,98,125,121,37,144,112,0,0       ; vpbroadcastw  0x7090(%rip),%ymm12        # 2da62 <_sk_srcover_rgba_8888_sse2_lowp+0x9ae>
   DB  196,65,61,239,236                   ; vpxor         %ymm12,%ymm8,%ymm13
   DB  197,21,213,236                      ; vpmullw       %ymm4,%ymm13,%ymm13
   DB  197,189,213,192                     ; vpmullw       %ymm0,%ymm8,%ymm0
@@ -36357,9 +36792,9 @@
   DB  65,254,202                          ; dec           %r10b
   DB  69,15,182,194                       ; movzbl        %r10b,%r8d
   DB  65,128,248,14                       ; cmp           $0xe,%r8b
-  DB  119,51                              ; ja            2630c <_sk_srcover_rgba_8888_hsw_lowp+0x5c>
+  DB  119,51                              ; ja            26adc <_sk_srcover_rgba_8888_hsw_lowp+0x5c>
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
-  DB  76,141,29,236,2,0,0                 ; lea           0x2ec(%rip),%r11        # 265d0 <_sk_srcover_rgba_8888_hsw_lowp+0x320>
+  DB  76,141,29,236,2,0,0                 ; lea           0x2ec(%rip),%r11        # 26da0 <_sk_srcover_rgba_8888_hsw_lowp+0x320>
   DB  75,99,4,131                         ; movslq        (%r11,%r8,4),%rax
   DB  76,1,216                            ; add           %r11,%rax
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -36369,10 +36804,10 @@
   DB  255,224                             ; jmpq          *%rax
   DB  196,193,121,110,41                  ; vmovd         (%r9),%xmm5
   DB  197,221,239,228                     ; vpxor         %ymm4,%ymm4,%ymm4
-  DB  233,22,1,0,0                        ; jmpq          26422 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  DB  233,22,1,0,0                        ; jmpq          26bf2 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   DB  196,193,126,111,41                  ; vmovdqu       (%r9),%ymm5
   DB  196,193,126,111,97,32               ; vmovdqu       0x20(%r9),%ymm4
-  DB  233,6,1,0,0                         ; jmpq          26422 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  DB  233,6,1,0,0                         ; jmpq          26bf2 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   DB  196,193,121,110,97,8                ; vmovd         0x8(%r9),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,197,239,255                     ; vpxor         %ymm7,%ymm7,%ymm7
@@ -36381,7 +36816,7 @@
   DB  197,249,112,237,232                 ; vpshufd       $0xe8,%xmm5,%xmm5
   DB  196,227,93,2,237,3                  ; vpblendd      $0x3,%ymm5,%ymm4,%ymm5
   DB  197,253,111,231                     ; vmovdqa       %ymm7,%ymm4
-  DB  233,216,0,0,0                       ; jmpq          26422 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  DB  233,216,0,0,0                       ; jmpq          26bf2 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   DB  196,193,121,110,97,24               ; vmovd         0x18(%r9),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -36397,7 +36832,7 @@
   DB  196,193,122,111,41                  ; vmovdqu       (%r9),%xmm5
   DB  196,227,85,2,236,240                ; vpblendd      $0xf0,%ymm4,%ymm5,%ymm5
   DB  197,125,127,196                     ; vmovdqa       %ymm8,%ymm4
-  DB  233,129,0,0,0                       ; jmpq          26422 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  DB  233,129,0,0,0                       ; jmpq          26bf2 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   DB  196,193,121,110,97,40               ; vmovd         0x28(%r9),%xmm4
   DB  196,226,121,89,228                  ; vpbroadcastq  %xmm4,%xmm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -36407,7 +36842,7 @@
   DB  196,193,121,110,105,32              ; vmovd         0x20(%r9),%xmm5
   DB  196,227,93,2,229,1                  ; vpblendd      $0x1,%ymm5,%ymm4,%ymm4
   DB  196,193,126,111,41                  ; vmovdqu       (%r9),%ymm5
-  DB  235,76                              ; jmp           26422 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
+  DB  235,76                              ; jmp           26bf2 <_sk_srcover_rgba_8888_hsw_lowp+0x172>
   DB  196,193,121,110,97,56               ; vmovd         0x38(%r9),%xmm4
   DB  196,226,125,89,228                  ; vpbroadcastq  %xmm4,%ymm4
   DB  197,213,239,237                     ; vpxor         %ymm5,%ymm5,%ymm5
@@ -36427,9 +36862,9 @@
   DB  196,227,69,14,236,170               ; vpblendw      $0xaa,%ymm4,%ymm7,%ymm5
   DB  196,227,77,14,228,170               ; vpblendw      $0xaa,%ymm4,%ymm6,%ymm4
   DB  196,226,93,43,237                   ; vpackusdw     %ymm5,%ymm4,%ymm5
-  DB  197,125,111,5,85,95,0,0             ; vmovdqa       0x5f55(%rip),%ymm8        # 2c3a0 <_sk_srcover_rgba_8888_sse2_lowp+0x998>
+  DB  197,125,111,5,101,110,0,0           ; vmovdqa       0x6e65(%rip),%ymm8        # 2da80 <_sk_srcover_rgba_8888_sse2_lowp+0x9cc>
   DB  196,193,85,219,224                  ; vpand         %ymm8,%ymm5,%ymm4
-  DB  196,98,125,121,13,103,95,0,0        ; vpbroadcastw  0x5f67(%rip),%ymm9        # 2c3c0 <_sk_srcover_rgba_8888_sse2_lowp+0x9b8>
+  DB  196,98,125,121,13,119,110,0,0       ; vpbroadcastw  0x6e77(%rip),%ymm9        # 2daa0 <_sk_srcover_rgba_8888_sse2_lowp+0x9ec>
   DB  197,213,113,213,8                   ; vpsrlw        $0x8,%ymm5,%ymm5
   DB  197,197,114,215,16                  ; vpsrld        $0x10,%ymm7,%ymm7
   DB  197,205,114,214,16                  ; vpsrld        $0x10,%ymm6,%ymm6
@@ -36468,19 +36903,19 @@
   DB  196,193,53,114,243,16               ; vpslld        $0x10,%ymm11,%ymm9
   DB  196,65,53,235,202                   ; vpor          %ymm10,%ymm9,%ymm9
   DB  65,128,250,14                       ; cmp           $0xe,%r10b
-  DB  119,26                              ; ja            26533 <_sk_srcover_rgba_8888_hsw_lowp+0x283>
-  DB  76,141,21,236,0,0,0                 ; lea           0xec(%rip),%r10        # 2660c <_sk_srcover_rgba_8888_hsw_lowp+0x35c>
+  DB  119,26                              ; ja            26d03 <_sk_srcover_rgba_8888_hsw_lowp+0x283>
+  DB  76,141,21,236,0,0,0                 ; lea           0xec(%rip),%r10        # 26ddc <_sk_srcover_rgba_8888_hsw_lowp+0x35c>
   DB  75,99,4,130                         ; movslq        (%r10,%r8,4),%rax
   DB  76,1,208                            ; add           %r10,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  196,65,121,126,1                    ; vmovd         %xmm8,(%r9)
-  DB  233,151,0,0,0                       ; jmpq          265ca <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  DB  233,151,0,0,0                       ; jmpq          26d9a <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   DB  196,65,126,127,1                    ; vmovdqu       %ymm8,(%r9)
   DB  196,65,126,127,73,32                ; vmovdqu       %ymm9,0x20(%r9)
-  DB  233,135,0,0,0                       ; jmpq          265ca <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  DB  233,135,0,0,0                       ; jmpq          26d9a <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   DB  196,67,121,22,65,8,2                ; vpextrd       $0x2,%xmm8,0x8(%r9)
   DB  196,65,121,214,1                    ; vmovq         %xmm8,(%r9)
-  DB  235,121                             ; jmp           265ca <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  DB  235,121                             ; jmp           26d9a <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,67,121,22,73,24,2               ; vpextrd       $0x2,%xmm9,0x18(%r9)
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
@@ -36488,12 +36923,12 @@
   DB  196,67,125,57,193,1                 ; vextracti128  $0x1,%ymm8,%xmm9
   DB  196,65,121,126,73,16                ; vmovd         %xmm9,0x10(%r9)
   DB  196,65,122,127,1                    ; vmovdqu       %xmm8,(%r9)
-  DB  235,76                              ; jmp           265ca <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  DB  235,76                              ; jmp           26d9a <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   DB  196,67,121,22,73,40,2               ; vpextrd       $0x2,%xmm9,0x28(%r9)
   DB  196,67,121,22,73,36,1               ; vpextrd       $0x1,%xmm9,0x24(%r9)
   DB  196,65,121,126,73,32                ; vmovd         %xmm9,0x20(%r9)
   DB  196,65,126,127,1                    ; vmovdqu       %ymm8,(%r9)
-  DB  235,49                              ; jmp           265ca <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
+  DB  235,49                              ; jmp           26d9a <_sk_srcover_rgba_8888_hsw_lowp+0x31a>
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
   DB  196,67,121,22,81,56,2               ; vpextrd       $0x2,%xmm10,0x38(%r9)
   DB  196,67,125,57,202,1                 ; vextracti128  $0x1,%ymm9,%xmm10
@@ -36541,7 +36976,7 @@
   DB  255,6                               ; incl          (%rsi)
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,62               ; lcall         *0x3effffff(%rip)        # 3f026610 <_sk_srcover_rgba_8888_sse2_lowp+0x3effac08>
+  DB  255,29,255,255,255,62               ; lcall         *0x3effffff(%rip)        # 3f026de0 <_sk_srcover_rgba_8888_sse2_lowp+0x3eff9d2c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,55                              ; pushq         (%rdi)
@@ -36559,7 +36994,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  121,255                             ; jns           26631 <_sk_srcover_rgba_8888_hsw_lowp+0x381>
+  DB  121,255                             ; jns           26e01 <_sk_srcover_rgba_8888_hsw_lowp+0x381>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -36602,13 +37037,13 @@
   DB  73,137,246                          ; mov           %rsi,%r14
   DB  76,137,141,8,255,255,255            ; mov           %r9,-0xf8(%rbp)
   DB  76,57,203                           ; cmp           %r9,%rbx
-  DB  15,131,153,0,0,0                    ; jae           26762 <_sk_start_pipeline_sse41_lowp+0x11a>
+  DB  15,131,153,0,0,0                    ; jae           26f32 <_sk_start_pipeline_sse41_lowp+0x11a>
   DB  72,139,133,24,255,255,255           ; mov           -0xe8(%rbp),%rax
   DB  72,141,64,8                         ; lea           0x8(%rax),%rax
   DB  72,137,133,16,255,255,255           ; mov           %rax,-0xf0(%rbp)
   DB  76,57,189,16,255,255,255            ; cmp           %r15,-0xf0(%rbp)
   DB  72,139,149,24,255,255,255           ; mov           -0xe8(%rbp),%rdx
-  DB  119,62                              ; ja            26729 <_sk_start_pipeline_sse41_lowp+0xe1>
+  DB  119,62                              ; ja            26ef9 <_sk_start_pipeline_sse41_lowp+0xe1>
   DB  76,139,165,24,255,255,255           ; mov           -0xe8(%rbp),%r12
   DB  49,255                              ; xor           %edi,%edi
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -36627,10 +37062,10 @@
   DB  73,131,196,16                       ; add           $0x10,%r12
   DB  77,57,252                           ; cmp           %r15,%r12
   DB  73,137,212                          ; mov           %rdx,%r12
-  DB  118,201                             ; jbe           266f2 <_sk_start_pipeline_sse41_lowp+0xaa>
+  DB  118,201                             ; jbe           26ec2 <_sk_start_pipeline_sse41_lowp+0xaa>
   DB  76,137,255                          ; mov           %r15,%rdi
   DB  72,41,215                           ; sub           %rdx,%rdi
-  DB  116,33                              ; je            26752 <_sk_start_pipeline_sse41_lowp+0x10a>
+  DB  116,33                              ; je            26f22 <_sk_start_pipeline_sse41_lowp+0x10a>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -36644,7 +37079,7 @@
   DB  65,255,213                          ; callq         *%r13
   DB  72,255,195                          ; inc           %rbx
   DB  72,59,157,8,255,255,255             ; cmp           -0xf8(%rbp),%rbx
-  DB  15,133,121,255,255,255              ; jne           266db <_sk_start_pipeline_sse41_lowp+0x93>
+  DB  15,133,121,255,255,255              ; jne           26eab <_sk_start_pipeline_sse41_lowp+0x93>
   DB  15,40,181,32,255,255,255            ; movaps        -0xe0(%rbp),%xmm6
   DB  15,40,189,48,255,255,255            ; movaps        -0xd0(%rbp),%xmm7
   DB  68,15,40,133,64,255,255,255         ; movaps        -0xc0(%rbp),%xmm8
@@ -36683,11 +37118,47 @@
   DB  102,15,110,209                      ; movd          %ecx,%xmm2
   DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,88,21,57,92,0,0                  ; addps         0x5c39(%rip),%xmm2        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,21,73,107,0,0                 ; addps         0x6b49(%rip),%xmm2        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_translate_sse41_lowp
+_sk_matrix_translate_sse41_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
+  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
+  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_matrix_scale_translate_sse41_lowp
+_sk_matrix_scale_translate_sse41_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
+  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
+  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
+  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
+  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
+  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
+  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
+  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_matrix_2x3_sse41_lowp
 _sk_matrix_2x3_sse41_lowp LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
@@ -36727,6 +37198,68 @@
   DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_perspective_sse41_lowp
+_sk_matrix_perspective_sse41_lowp LABEL PROC
+  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,15,16,8                         ; movss         (%rax),%xmm1
+  DB  243,68,15,16,80,4                   ; movss         0x4(%rax),%xmm10
+  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  243,68,15,16,72,8                   ; movss         0x8(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
+  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
+  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
+  DB  69,15,88,209                        ; addps         %xmm9,%xmm10
+  DB  69,15,88,217                        ; addps         %xmm9,%xmm11
+  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
+  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
+  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
+  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
+  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
+  DB  243,68,15,16,96,12                  ; movss         0xc(%rax),%xmm12
+  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
+  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  243,68,15,16,104,20                 ; movss         0x14(%rax),%xmm13
+  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
+  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
+  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
+  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
+  DB  69,15,88,213                        ; addps         %xmm13,%xmm10
+  DB  69,15,88,221                        ; addps         %xmm13,%xmm11
+  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
+  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
+  DB  69,15,88,213                        ; addps         %xmm13,%xmm10
+  DB  69,15,89,224                        ; mulps         %xmm8,%xmm12
+  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
+  DB  243,68,15,16,96,24                  ; movss         0x18(%rax),%xmm12
+  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
+  DB  243,68,15,16,104,28                 ; movss         0x1c(%rax),%xmm13
+  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
+  DB  243,68,15,16,112,32                 ; movss         0x20(%rax),%xmm14
+  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
+  DB  65,15,89,221                        ; mulps         %xmm13,%xmm3
+  DB  68,15,89,234                        ; mulps         %xmm2,%xmm13
+  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
+  DB  65,15,88,222                        ; addps         %xmm14,%xmm3
+  DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
+  DB  65,15,88,197                        ; addps         %xmm13,%xmm0
+  DB  69,15,89,224                        ; mulps         %xmm8,%xmm12
+  DB  68,15,88,227                        ; addps         %xmm3,%xmm12
+  DB  15,83,192                           ; rcpps         %xmm0,%xmm0
+  DB  65,15,83,212                        ; rcpps         %xmm12,%xmm2
+  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
+  DB  15,89,202                           ; mulps         %xmm2,%xmm1
+  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
+  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
+  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
+  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_uniform_color_sse41_lowp
 _sk_uniform_color_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -36750,7 +37283,7 @@
 PUBLIC _sk_black_color_sse41_lowp
 _sk_black_color_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,246,90,0,0                 ; movaps        0x5af6(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,142,104,0,0                ; movaps        0x688e(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -36759,7 +37292,7 @@
 PUBLIC _sk_white_color_sse41_lowp
 _sk_white_color_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,5,226,90,0,0                  ; movaps        0x5ae2(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,5,122,104,0,0                 ; movaps        0x687a(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -36768,10 +37301,10 @@
 PUBLIC _sk_set_rgb_sse41_lowp
 _sk_set_rgb_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  243,15,16,21,225,83,0,0             ; movss         0x53e1(%rip),%xmm2        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,15,16,21,69,97,0,0              ; movss         0x6145(%rip),%xmm2        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  243,15,16,0                         ; movss         (%rax),%xmm0
   DB  243,15,89,194                       ; mulss         %xmm2,%xmm0
-  DB  243,68,15,16,5,160,83,0,0           ; movss         0x53a0(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,16,5,4,97,0,0             ; movss         0x6104(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,88,192                    ; addss         %xmm8,%xmm0
   DB  243,68,15,44,192                    ; cvttss2si     %xmm0,%r8d
   DB  102,65,15,110,192                   ; movd          %r8d,%xmm0
@@ -36796,7 +37329,7 @@
 PUBLIC _sk_premul_sse41_lowp
 _sk_premul_sse41_lowp LABEL PROC
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,68,15,111,5,84,90,0,0           ; movdqa        0x5a54(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,236,103,0,0         ; movdqa        0x67ec(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
@@ -36836,7 +37369,7 @@
 
 PUBLIC _sk_invert_sse41_lowp
 _sk_invert_sse41_lowp LABEL PROC
-  DB  102,68,15,111,5,242,89,0,0          ; movdqa        0x59f2(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,138,103,0,0         ; movdqa        0x678a(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,200                   ; psubw         %xmm0,%xmm9
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -36864,7 +37397,7 @@
 _sk_srcatop_sse41_lowp LABEL PROC
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,15,111,29,150,89,0,0            ; movdqa        0x5996(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,29,46,103,0,0            ; movdqa        0x672e(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,111,203                   ; movdqa        %xmm3,%xmm9
   DB  102,69,15,249,200                   ; psubw         %xmm8,%xmm9
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
@@ -36894,7 +37427,7 @@
 _sk_dstatop_sse41_lowp LABEL PROC
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
   DB  102,68,15,213,195                   ; pmullw        %xmm3,%xmm8
-  DB  102,68,15,111,13,22,89,0,0          ; movdqa        0x5916(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,174,102,0,0        ; movdqa        0x66ae(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,253,193                   ; paddw         %xmm9,%xmm8
   DB  102,68,15,111,213                   ; movdqa        %xmm5,%xmm10
   DB  102,68,15,213,211                   ; pmullw        %xmm3,%xmm10
@@ -36921,7 +37454,7 @@
 PUBLIC _sk_srcin_sse41_lowp
 _sk_srcin_sse41_lowp LABEL PROC
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,68,15,111,5,161,88,0,0          ; movdqa        0x58a1(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,57,102,0,0          ; movdqa        0x6639(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,207                      ; pmullw        %xmm7,%xmm1
@@ -36940,7 +37473,7 @@
 _sk_dstin_sse41_lowp LABEL PROC
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,68,15,111,5,88,88,0,0           ; movdqa        0x5858(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,240,101,0,0         ; movdqa        0x65f0(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
@@ -36959,7 +37492,7 @@
 
 PUBLIC _sk_srcout_sse41_lowp
 _sk_srcout_sse41_lowp LABEL PROC
-  DB  102,68,15,111,5,15,88,0,0           ; movdqa        0x580f(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,167,101,0,0         ; movdqa        0x65a7(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,207                   ; psubw         %xmm7,%xmm9
   DB  102,65,15,213,193                   ; pmullw        %xmm9,%xmm0
@@ -36980,7 +37513,7 @@
 PUBLIC _sk_dstout_sse41_lowp
 _sk_dstout_sse41_lowp LABEL PROC
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,68,15,111,5,184,87,0,0          ; movdqa        0x57b8(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,80,101,0,0          ; movdqa        0x6550(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  102,15,249,216                      ; psubw         %xmm0,%xmm3
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
@@ -37003,7 +37536,7 @@
 
 PUBLIC _sk_srcover_sse41_lowp
 _sk_srcover_sse41_lowp LABEL PROC
-  DB  102,68,15,111,13,94,87,0,0          ; movdqa        0x575e(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,246,100,0,0        ; movdqa        0x64f6(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,68,15,249,195                   ; psubw         %xmm3,%xmm8
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -37030,7 +37563,7 @@
 
 PUBLIC _sk_dstover_sse41_lowp
 _sk_dstover_sse41_lowp LABEL PROC
-  DB  102,68,15,111,5,228,86,0,0          ; movdqa        0x56e4(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,124,100,0,0         ; movdqa        0x647c(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,207                   ; psubw         %xmm7,%xmm9
   DB  102,65,15,213,193                   ; pmullw        %xmm9,%xmm0
@@ -37055,7 +37588,7 @@
 PUBLIC _sk_modulate_sse41_lowp
 _sk_modulate_sse41_lowp LABEL PROC
   DB  102,15,213,196                      ; pmullw        %xmm4,%xmm0
-  DB  102,68,15,111,5,125,86,0,0          ; movdqa        0x567d(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,21,100,0,0          ; movdqa        0x6415(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,205                      ; pmullw        %xmm5,%xmm1
@@ -37072,7 +37605,7 @@
 
 PUBLIC _sk_multiply_sse41_lowp
 _sk_multiply_sse41_lowp LABEL PROC
-  DB  102,68,15,111,13,60,86,0,0          ; movdqa        0x563c(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,212,99,0,0         ; movdqa        0x63d4(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,68,15,249,195                   ; psubw         %xmm3,%xmm8
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -37113,7 +37646,7 @@
 PUBLIC _sk_plus__sse41_lowp
 _sk_plus__sse41_lowp LABEL PROC
   DB  102,15,253,196                      ; paddw         %xmm4,%xmm0
-  DB  102,68,15,111,5,129,85,0,0          ; movdqa        0x5581(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,25,99,0,0           ; movdqa        0x6319(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,56,58,192                 ; pminuw        %xmm8,%xmm0
   DB  102,15,253,205                      ; paddw         %xmm5,%xmm1
   DB  102,65,15,56,58,200                 ; pminuw        %xmm8,%xmm1
@@ -37131,7 +37664,7 @@
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,204                      ; pmullw        %xmm4,%xmm1
-  DB  102,68,15,111,21,59,85,0,0          ; movdqa        0x553b(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,21,211,98,0,0         ; movdqa        0x62d3(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,202                   ; paddw         %xmm10,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -37161,7 +37694,7 @@
 PUBLIC _sk_xor__sse41_lowp
 _sk_xor__sse41_lowp LABEL PROC
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
-  DB  102,68,15,111,13,185,84,0,0         ; movdqa        0x54b9(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,81,98,0,0          ; movdqa        0x6251(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,209                   ; movdqa        %xmm9,%xmm10
   DB  102,68,15,249,215                   ; psubw         %xmm7,%xmm10
   DB  102,65,15,213,194                   ; pmullw        %xmm10,%xmm0
@@ -37202,7 +37735,7 @@
   DB  102,68,15,213,207                   ; pmullw        %xmm7,%xmm9
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
   DB  102,65,15,56,62,201                 ; pmaxuw        %xmm9,%xmm1
-  DB  102,68,15,111,13,0,84,0,0           ; movdqa        0x5400(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,152,97,0,0         ; movdqa        0x6198(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,201                   ; paddw         %xmm9,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -37244,7 +37777,7 @@
   DB  102,68,15,213,207                   ; pmullw        %xmm7,%xmm9
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
   DB  102,65,15,56,58,201                 ; pminuw        %xmm9,%xmm1
-  DB  102,68,15,111,13,67,83,0,0          ; movdqa        0x5343(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,219,96,0,0         ; movdqa        0x60db(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,201                   ; paddw         %xmm9,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -37286,10 +37819,10 @@
   DB  102,68,15,213,207                   ; pmullw        %xmm7,%xmm9
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
   DB  102,65,15,56,58,201                 ; pminuw        %xmm9,%xmm1
-  DB  102,68,15,111,13,134,82,0,0         ; movdqa        0x5286(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,30,96,0,0          ; movdqa        0x601e(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,201                   ; paddw         %xmm9,%xmm1
   DB  102,15,113,209,7                    ; psrlw         $0x7,%xmm1
-  DB  102,68,15,111,21,211,89,0,0         ; movdqa        0x59d3(%rip),%xmm10        # 2cb30 <_sk_srcover_rgba_8888_sse2_lowp+0x1128>
+  DB  102,68,15,111,21,107,103,0,0        ; movdqa        0x676b(%rip),%xmm10        # 2e210 <_sk_srcover_rgba_8888_sse2_lowp+0x115c>
   DB  102,65,15,219,202                   ; pand          %xmm10,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
   DB  102,68,15,111,221                   ; movdqa        %xmm5,%xmm11
@@ -37329,10 +37862,10 @@
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,204                      ; pmullw        %xmm4,%xmm1
-  DB  102,68,15,111,13,194,81,0,0         ; movdqa        0x51c2(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,90,95,0,0          ; movdqa        0x5f5a(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,201                   ; paddw         %xmm9,%xmm1
   DB  102,15,113,209,7                    ; psrlw         $0x7,%xmm1
-  DB  102,68,15,111,21,15,89,0,0          ; movdqa        0x590f(%rip),%xmm10        # 2cb30 <_sk_srcover_rgba_8888_sse2_lowp+0x1128>
+  DB  102,68,15,111,21,167,102,0,0        ; movdqa        0x66a7(%rip),%xmm10        # 2e210 <_sk_srcover_rgba_8888_sse2_lowp+0x115c>
   DB  102,65,15,219,202                   ; pand          %xmm10,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
@@ -37367,14 +37900,14 @@
   DB  102,15,111,245                      ; movdqa        %xmm5,%xmm6
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
-  DB  102,68,15,111,29,26,81,0,0          ; movdqa        0x511a(%rip),%xmm11        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,29,178,94,0,0         ; movdqa        0x5eb2(%rip),%xmm11        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,249,220                   ; psubw         %xmm12,%xmm11
   DB  102,65,15,111,227                   ; movdqa        %xmm11,%xmm4
   DB  102,65,15,213,224                   ; pmullw        %xmm8,%xmm4
   DB  102,68,15,111,251                   ; movdqa        %xmm3,%xmm15
   DB  102,69,15,249,248                   ; psubw         %xmm8,%xmm15
   DB  102,69,15,253,192                   ; paddw         %xmm8,%xmm8
-  DB  102,68,15,111,13,99,88,0,0          ; movdqa        0x5863(%rip),%xmm9        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,13,251,101,0,0        ; movdqa        0x65fb(%rip),%xmm9        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,68,15,111,243                   ; movdqa        %xmm3,%xmm14
   DB  102,69,15,239,241                   ; pxor          %xmm9,%xmm14
   DB  102,65,15,111,196                   ; movdqa        %xmm12,%xmm0
@@ -37421,7 +37954,7 @@
   DB  102,68,15,111,232                   ; movdqa        %xmm0,%xmm13
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,65,15,56,16,210                 ; pblendvb      %xmm0,%xmm10,%xmm2
-  DB  102,68,15,111,13,12,80,0,0          ; movdqa        0x500c(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,164,93,0,0         ; movdqa        0x5da4(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,249,195                      ; psubw         %xmm3,%xmm0
   DB  102,15,111,248                      ; movdqa        %xmm0,%xmm7
@@ -37460,7 +37993,7 @@
   DB  72,131,236,24                       ; sub           $0x18,%rsp
   DB  102,68,15,111,231                   ; movdqa        %xmm7,%xmm12
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
-  DB  102,68,15,111,29,103,79,0,0         ; movdqa        0x4f67(%rip),%xmm11        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,29,255,92,0,0         ; movdqa        0x5cff(%rip),%xmm11        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,249,220                   ; psubw         %xmm12,%xmm11
   DB  102,65,15,111,195                   ; movdqa        %xmm11,%xmm0
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
@@ -37472,7 +38005,7 @@
   DB  102,68,15,249,252                   ; psubw         %xmm4,%xmm15
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,253,192                      ; paddw         %xmm0,%xmm0
-  DB  102,68,15,111,13,153,86,0,0         ; movdqa        0x5699(%rip),%xmm9        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,13,49,100,0,0         ; movdqa        0x6431(%rip),%xmm9        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,69,15,111,212                   ; movdqa        %xmm12,%xmm10
   DB  102,68,15,213,211                   ; pmullw        %xmm3,%xmm10
   DB  102,69,15,213,254                   ; pmullw        %xmm14,%xmm15
@@ -37518,7 +38051,7 @@
   DB  102,15,253,210                      ; paddw         %xmm2,%xmm2
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,65,15,56,16,210                 ; pblendvb      %xmm0,%xmm10,%xmm2
-  DB  102,68,15,111,13,75,78,0,0          ; movdqa        0x4e4b(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,227,91,0,0         ; movdqa        0x5be3(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,249,195                      ; psubw         %xmm3,%xmm0
   DB  102,15,111,248                      ; movdqa        %xmm0,%xmm7
@@ -37560,49 +38093,49 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            2764a <_sk_load_8888_sse41_lowp+0x44>
+  DB  119,41                              ; ja            27f92 <_sk_load_8888_sse41_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,240,0,0,0                 ; lea           0xf0(%rip),%r9        # 27720 <_sk_load_8888_sse41_lowp+0x11a>
+  DB  76,141,13,240,0,0,0                 ; lea           0xf0(%rip),%r9        # 28068 <_sk_load_8888_sse41_lowp+0x11a>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,20,144                ; movd          (%r8,%rdx,4),%xmm2
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,89                              ; jmp           276a3 <_sk_load_8888_sse41_lowp+0x9d>
+  DB  235,89                              ; jmp           27feb <_sk_load_8888_sse41_lowp+0x9d>
   DB  243,65,15,111,20,144                ; movdqu        (%r8,%rdx,4),%xmm2
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,74                              ; jmp           276a3 <_sk_load_8888_sse41_lowp+0x9d>
+  DB  235,74                              ; jmp           27feb <_sk_load_8888_sse41_lowp+0x9d>
   DB  102,65,15,110,68,144,8              ; movd          0x8(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  243,65,15,126,20,144                ; movq          (%r8,%rdx,4),%xmm2
   DB  102,65,15,58,14,208,240             ; pblendw       $0xf0,%xmm8,%xmm2
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
-  DB  235,37                              ; jmp           276a3 <_sk_load_8888_sse41_lowp+0x9d>
+  DB  235,37                              ; jmp           27feb <_sk_load_8888_sse41_lowp+0x9d>
   DB  102,65,15,110,68,144,24             ; movd          0x18(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,69,15,58,34,68,144,20,1         ; pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   DB  102,69,15,58,34,68,144,16,0         ; pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   DB  243,65,15,111,20,144                ; movdqu        (%r8,%rdx,4),%xmm2
-  DB  102,15,111,5,165,84,0,0             ; movdqa        0x54a5(%rip),%xmm0        # 2cb50 <_sk_srcover_rgba_8888_sse2_lowp+0x1148>
+  DB  102,15,111,5,61,98,0,0              ; movdqa        0x623d(%rip),%xmm0        # 2e230 <_sk_srcover_rgba_8888_sse2_lowp+0x117c>
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,15,56,0,200                     ; pshufb        %xmm0,%xmm1
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  102,15,56,0,216                     ; pshufb        %xmm0,%xmm3
   DB  102,15,108,203                      ; punpcklqdq    %xmm3,%xmm1
-  DB  102,68,15,111,13,5,77,0,0           ; movdqa        0x4d05(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,157,90,0,0         ; movdqa        0x5a9d(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
-  DB  102,68,15,111,21,126,84,0,0         ; movdqa        0x547e(%rip),%xmm10        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,21,22,98,0,0          ; movdqa        0x6216(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
   DB  102,65,15,56,0,218                  ; pshufb        %xmm10,%xmm3
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
   DB  102,65,15,108,219                   ; punpcklqdq    %xmm11,%xmm3
-  DB  102,68,15,111,21,107,84,0,0         ; movdqa        0x546b(%rip),%xmm10        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,21,3,98,0,0           ; movdqa        0x6203(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,210                  ; pshufb        %xmm10,%xmm2
   DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
   DB  102,65,15,108,208                   ; punpcklqdq    %xmm8,%xmm2
@@ -37617,7 +38150,7 @@
   DB  57,255                              ; cmp           %edi,%edi
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  125,255                             ; jge           2772d <_sk_load_8888_sse41_lowp+0x127>
+  DB  125,255                             ; jge           28075 <_sk_load_8888_sse41_lowp+0x127>
   DB  255                                 ; (bad)
   DB  255,116,255,255                     ; pushq         -0x1(%rdi,%rdi,8)
   DB  255,107,255                         ; ljmp          *-0x1(%rbx)
@@ -37637,49 +38170,49 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            27780 <_sk_load_8888_dst_sse41_lowp+0x44>
+  DB  119,41                              ; ja            280c8 <_sk_load_8888_dst_sse41_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 27858 <_sk_load_8888_dst_sse41_lowp+0x11c>
+  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 281a0 <_sk_load_8888_dst_sse41_lowp+0x11c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,52,144                ; movd          (%r8,%rdx,4),%xmm6
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,89                              ; jmp           277d9 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  DB  235,89                              ; jmp           28121 <_sk_load_8888_dst_sse41_lowp+0x9d>
   DB  243,65,15,111,52,144                ; movdqu        (%r8,%rdx,4),%xmm6
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,74                              ; jmp           277d9 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  DB  235,74                              ; jmp           28121 <_sk_load_8888_dst_sse41_lowp+0x9d>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  243,65,15,126,52,144                ; movq          (%r8,%rdx,4),%xmm6
   DB  102,65,15,58,14,240,240             ; pblendw       $0xf0,%xmm8,%xmm6
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
-  DB  235,37                              ; jmp           277d9 <_sk_load_8888_dst_sse41_lowp+0x9d>
+  DB  235,37                              ; jmp           28121 <_sk_load_8888_dst_sse41_lowp+0x9d>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,69,15,58,34,68,144,20,1         ; pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   DB  102,69,15,58,34,68,144,16,0         ; pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   DB  243,65,15,111,52,144                ; movdqu        (%r8,%rdx,4),%xmm6
-  DB  102,15,111,37,111,83,0,0            ; movdqa        0x536f(%rip),%xmm4        # 2cb50 <_sk_srcover_rgba_8888_sse2_lowp+0x1148>
+  DB  102,15,111,37,7,97,0,0              ; movdqa        0x6107(%rip),%xmm4        # 2e230 <_sk_srcover_rgba_8888_sse2_lowp+0x117c>
   DB  102,15,111,238                      ; movdqa        %xmm6,%xmm5
   DB  102,15,56,0,236                     ; pshufb        %xmm4,%xmm5
   DB  102,65,15,111,248                   ; movdqa        %xmm8,%xmm7
   DB  102,15,56,0,252                     ; pshufb        %xmm4,%xmm7
   DB  102,15,108,239                      ; punpcklqdq    %xmm7,%xmm5
-  DB  102,68,15,111,13,207,75,0,0         ; movdqa        0x4bcf(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,103,89,0,0         ; movdqa        0x5967(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,229                      ; movdqa        %xmm5,%xmm4
   DB  102,65,15,219,225                   ; pand          %xmm9,%xmm4
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
-  DB  102,68,15,111,21,72,83,0,0          ; movdqa        0x5348(%rip),%xmm10        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,21,224,96,0,0         ; movdqa        0x60e0(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,254                      ; movdqa        %xmm6,%xmm7
   DB  102,65,15,56,0,250                  ; pshufb        %xmm10,%xmm7
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
   DB  102,65,15,108,251                   ; punpcklqdq    %xmm11,%xmm7
-  DB  102,68,15,111,21,53,83,0,0          ; movdqa        0x5335(%rip),%xmm10        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,21,205,96,0,0         ; movdqa        0x60cd(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,242                  ; pshufb        %xmm10,%xmm6
   DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
   DB  102,65,15,108,240                   ; punpcklqdq    %xmm8,%xmm6
@@ -37695,7 +38228,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           27865 <_sk_load_8888_dst_sse41_lowp+0x129>
+  DB  123,255                             ; jnp           281ad <_sk_load_8888_dst_sse41_lowp+0x129>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -37730,20 +38263,20 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,27                              ; ja            278fa <_sk_store_8888_sse41_lowp+0x86>
+  DB  119,27                              ; ja            28242 <_sk_store_8888_sse41_lowp+0x86>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,87,0,0,0                  ; lea           0x57(%rip),%r9        # 27940 <_sk_store_8888_sse41_lowp+0xcc>
+  DB  76,141,13,87,0,0,0                  ; lea           0x57(%rip),%r9        # 28288 <_sk_store_8888_sse41_lowp+0xcc>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,126,4,144                 ; movd          %xmm8,(%r8,%rdx,4)
-  DB  235,63                              ; jmp           27939 <_sk_store_8888_sse41_lowp+0xc5>
+  DB  235,63                              ; jmp           28281 <_sk_store_8888_sse41_lowp+0xc5>
   DB  243,69,15,127,4,144                 ; movdqu        %xmm8,(%r8,%rdx,4)
   DB  243,69,15,127,76,144,16             ; movdqu        %xmm9,0x10(%r8,%rdx,4)
-  DB  235,48                              ; jmp           27939 <_sk_store_8888_sse41_lowp+0xc5>
+  DB  235,48                              ; jmp           28281 <_sk_store_8888_sse41_lowp+0xc5>
   DB  102,69,15,58,22,68,144,8,2          ; pextrd        $0x2,%xmm8,0x8(%r8,%rdx,4)
   DB  102,69,15,214,4,144                 ; movq          %xmm8,(%r8,%rdx,4)
-  DB  235,31                              ; jmp           27939 <_sk_store_8888_sse41_lowp+0xc5>
+  DB  235,31                              ; jmp           28281 <_sk_store_8888_sse41_lowp+0xc5>
   DB  102,69,15,58,22,76,144,24,2         ; pextrd        $0x2,%xmm9,0x18(%r8,%rdx,4)
   DB  102,69,15,58,22,76,144,20,1         ; pextrd        $0x1,%xmm9,0x14(%r8,%rdx,4)
   DB  102,69,15,126,76,144,16             ; movd          %xmm9,0x10(%r8,%rdx,4)
@@ -37785,49 +38318,49 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            279a0 <_sk_load_bgra_sse41_lowp+0x44>
+  DB  119,41                              ; ja            282e8 <_sk_load_bgra_sse41_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 27a78 <_sk_load_bgra_sse41_lowp+0x11c>
+  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 283c0 <_sk_load_bgra_sse41_lowp+0x11c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,4,144                 ; movd          (%r8,%rdx,4),%xmm0
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,89                              ; jmp           279f9 <_sk_load_bgra_sse41_lowp+0x9d>
+  DB  235,89                              ; jmp           28341 <_sk_load_bgra_sse41_lowp+0x9d>
   DB  243,65,15,111,4,144                 ; movdqu        (%r8,%rdx,4),%xmm0
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,74                              ; jmp           279f9 <_sk_load_bgra_sse41_lowp+0x9d>
+  DB  235,74                              ; jmp           28341 <_sk_load_bgra_sse41_lowp+0x9d>
   DB  102,65,15,110,68,144,8              ; movd          0x8(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  243,65,15,126,4,144                 ; movq          (%r8,%rdx,4),%xmm0
   DB  102,65,15,58,14,192,240             ; pblendw       $0xf0,%xmm8,%xmm0
   DB  102,68,15,111,193                   ; movdqa        %xmm1,%xmm8
-  DB  235,37                              ; jmp           279f9 <_sk_load_bgra_sse41_lowp+0x9d>
+  DB  235,37                              ; jmp           28341 <_sk_load_bgra_sse41_lowp+0x9d>
   DB  102,65,15,110,68,144,24             ; movd          0x18(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,69,15,58,34,68,144,20,1         ; pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   DB  102,69,15,58,34,68,144,16,0         ; pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   DB  243,65,15,111,4,144                 ; movdqu        (%r8,%rdx,4),%xmm0
-  DB  102,15,111,21,79,81,0,0             ; movdqa        0x514f(%rip),%xmm2        # 2cb50 <_sk_srcover_rgba_8888_sse2_lowp+0x1148>
+  DB  102,15,111,21,231,94,0,0            ; movdqa        0x5ee7(%rip),%xmm2        # 2e230 <_sk_srcover_rgba_8888_sse2_lowp+0x117c>
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,56,0,202                     ; pshufb        %xmm2,%xmm1
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  102,15,56,0,218                     ; pshufb        %xmm2,%xmm3
   DB  102,15,108,203                      ; punpcklqdq    %xmm3,%xmm1
-  DB  102,68,15,111,13,175,73,0,0         ; movdqa        0x49af(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,71,87,0,0          ; movdqa        0x5747(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
-  DB  102,68,15,111,21,40,81,0,0          ; movdqa        0x5128(%rip),%xmm10        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,21,192,94,0,0         ; movdqa        0x5ec0(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
   DB  102,65,15,56,0,218                  ; pshufb        %xmm10,%xmm3
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
   DB  102,65,15,108,219                   ; punpcklqdq    %xmm11,%xmm3
-  DB  102,68,15,111,21,21,81,0,0          ; movdqa        0x5115(%rip),%xmm10        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,21,173,94,0,0         ; movdqa        0x5ead(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,194                  ; pshufb        %xmm10,%xmm0
   DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
   DB  102,65,15,108,192                   ; punpcklqdq    %xmm8,%xmm0
@@ -37843,7 +38376,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           27a85 <_sk_load_bgra_sse41_lowp+0x129>
+  DB  123,255                             ; jnp           283cd <_sk_load_bgra_sse41_lowp+0x129>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -37863,49 +38396,49 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            27ad8 <_sk_load_bgra_dst_sse41_lowp+0x44>
+  DB  119,41                              ; ja            28420 <_sk_load_bgra_dst_sse41_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 27bb0 <_sk_load_bgra_dst_sse41_lowp+0x11c>
+  DB  76,141,13,242,0,0,0                 ; lea           0xf2(%rip),%r9        # 284f8 <_sk_load_bgra_dst_sse41_lowp+0x11c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,36,144                ; movd          (%r8,%rdx,4),%xmm4
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,89                              ; jmp           27b31 <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  DB  235,89                              ; jmp           28479 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   DB  243,65,15,111,36,144                ; movdqu        (%r8,%rdx,4),%xmm4
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,74                              ; jmp           27b31 <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  DB  235,74                              ; jmp           28479 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  243,65,15,126,36,144                ; movq          (%r8,%rdx,4),%xmm4
   DB  102,65,15,58,14,224,240             ; pblendw       $0xf0,%xmm8,%xmm4
   DB  102,68,15,111,197                   ; movdqa        %xmm5,%xmm8
-  DB  235,37                              ; jmp           27b31 <_sk_load_bgra_dst_sse41_lowp+0x9d>
+  DB  235,37                              ; jmp           28479 <_sk_load_bgra_dst_sse41_lowp+0x9d>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,69,15,58,34,68,144,20,1         ; pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   DB  102,69,15,58,34,68,144,16,0         ; pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   DB  243,65,15,111,36,144                ; movdqu        (%r8,%rdx,4),%xmm4
-  DB  102,15,111,53,23,80,0,0             ; movdqa        0x5017(%rip),%xmm6        # 2cb50 <_sk_srcover_rgba_8888_sse2_lowp+0x1148>
+  DB  102,15,111,53,175,93,0,0            ; movdqa        0x5daf(%rip),%xmm6        # 2e230 <_sk_srcover_rgba_8888_sse2_lowp+0x117c>
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,56,0,238                     ; pshufb        %xmm6,%xmm5
   DB  102,65,15,111,248                   ; movdqa        %xmm8,%xmm7
   DB  102,15,56,0,254                     ; pshufb        %xmm6,%xmm7
   DB  102,15,108,239                      ; punpcklqdq    %xmm7,%xmm5
-  DB  102,68,15,111,13,119,72,0,0         ; movdqa        0x4877(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,15,86,0,0          ; movdqa        0x560f(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,245                      ; movdqa        %xmm5,%xmm6
   DB  102,65,15,219,241                   ; pand          %xmm9,%xmm6
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
-  DB  102,68,15,111,21,240,79,0,0         ; movdqa        0x4ff0(%rip),%xmm10        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,21,136,93,0,0         ; movdqa        0x5d88(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,252                      ; movdqa        %xmm4,%xmm7
   DB  102,65,15,56,0,250                  ; pshufb        %xmm10,%xmm7
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
   DB  102,65,15,108,251                   ; punpcklqdq    %xmm11,%xmm7
-  DB  102,68,15,111,21,221,79,0,0         ; movdqa        0x4fdd(%rip),%xmm10        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,21,117,93,0,0         ; movdqa        0x5d75(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,226                  ; pshufb        %xmm10,%xmm4
   DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
   DB  102,65,15,108,224                   ; punpcklqdq    %xmm8,%xmm4
@@ -37921,7 +38454,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           27bbd <_sk_load_bgra_dst_sse41_lowp+0x129>
+  DB  123,255                             ; jnp           28505 <_sk_load_bgra_dst_sse41_lowp+0x129>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -37956,20 +38489,20 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,27                              ; ja            27c52 <_sk_store_bgra_sse41_lowp+0x86>
+  DB  119,27                              ; ja            2859a <_sk_store_bgra_sse41_lowp+0x86>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,87,0,0,0                  ; lea           0x57(%rip),%r9        # 27c98 <_sk_store_bgra_sse41_lowp+0xcc>
+  DB  76,141,13,87,0,0,0                  ; lea           0x57(%rip),%r9        # 285e0 <_sk_store_bgra_sse41_lowp+0xcc>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,126,4,144                 ; movd          %xmm8,(%r8,%rdx,4)
-  DB  235,63                              ; jmp           27c91 <_sk_store_bgra_sse41_lowp+0xc5>
+  DB  235,63                              ; jmp           285d9 <_sk_store_bgra_sse41_lowp+0xc5>
   DB  243,69,15,127,4,144                 ; movdqu        %xmm8,(%r8,%rdx,4)
   DB  243,69,15,127,76,144,16             ; movdqu        %xmm9,0x10(%r8,%rdx,4)
-  DB  235,48                              ; jmp           27c91 <_sk_store_bgra_sse41_lowp+0xc5>
+  DB  235,48                              ; jmp           285d9 <_sk_store_bgra_sse41_lowp+0xc5>
   DB  102,69,15,58,22,68,144,8,2          ; pextrd        $0x2,%xmm8,0x8(%r8,%rdx,4)
   DB  102,69,15,214,4,144                 ; movq          %xmm8,(%r8,%rdx,4)
-  DB  235,31                              ; jmp           27c91 <_sk_store_bgra_sse41_lowp+0xc5>
+  DB  235,31                              ; jmp           285d9 <_sk_store_bgra_sse41_lowp+0xc5>
   DB  102,69,15,58,22,76,144,24,2         ; pextrd        $0x2,%xmm9,0x18(%r8,%rdx,4)
   DB  102,69,15,58,22,76,144,20,1         ; pextrd        $0x1,%xmm9,0x14(%r8,%rdx,4)
   DB  102,69,15,126,76,144,16             ; movd          %xmm9,0x10(%r8,%rdx,4)
@@ -38070,17 +38603,17 @@
   DB  102,65,15,196,202,5                 ; pinsrw        $0x5,%r10d,%xmm1
   DB  102,65,15,196,201,6                 ; pinsrw        $0x6,%r9d,%xmm1
   DB  102,15,196,200,7                    ; pinsrw        $0x7,%eax,%xmm1
-  DB  102,68,15,111,13,224,69,0,0         ; movdqa        0x45e0(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,120,83,0,0         ; movdqa        0x5378(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
-  DB  102,68,15,111,21,89,77,0,0          ; movdqa        0x4d59(%rip),%xmm10        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,21,241,90,0,0         ; movdqa        0x5af1(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
   DB  102,65,15,56,0,218                  ; pshufb        %xmm10,%xmm3
   DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
   DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
   DB  102,65,15,108,219                   ; punpcklqdq    %xmm11,%xmm3
-  DB  102,68,15,111,21,70,77,0,0          ; movdqa        0x4d46(%rip),%xmm10        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,21,222,90,0,0         ; movdqa        0x5ade(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,210                  ; pshufb        %xmm10,%xmm2
   DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
   DB  102,65,15,108,208                   ; punpcklqdq    %xmm8,%xmm2
@@ -38093,6 +38626,99 @@
   DB  93                                  ; pop           %rbp
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_bgra_sse41_lowp
+_sk_gather_bgra_sse41_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  243,68,15,91,195                    ; cvttps2dq     %xmm3,%xmm8
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,15,56,64,211                    ; pmulld        %xmm3,%xmm2
+  DB  102,68,15,56,64,195                 ; pmulld        %xmm3,%xmm8
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,200                   ; paddd         %xmm8,%xmm1
+  DB  102,72,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%rax
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
+  DB  69,137,211                          ; mov           %r10d,%r11d
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  102,72,15,58,22,195,1               ; pextrq        $0x1,%xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  67,139,44,184                       ; mov           (%r8,%r15,4),%ebp
+  DB  71,139,52,176                       ; mov           (%r8,%r14,4),%r14d
+  DB  65,139,28,152                       ; mov           (%r8,%rbx,4),%ebx
+  DB  71,139,28,152                       ; mov           (%r8,%r11,4),%r11d
+  DB  71,139,20,144                       ; mov           (%r8,%r10,4),%r10d
+  DB  71,139,12,136                       ; mov           (%r8,%r9,4),%r9d
+  DB  65,139,4,128                        ; mov           (%r8,%rax,4),%eax
+  DB  102,69,15,110,195                   ; movd          %r11d,%xmm8
+  DB  102,69,15,58,34,194,1               ; pinsrd        $0x1,%r10d,%xmm8
+  DB  102,69,15,58,34,193,2               ; pinsrd        $0x2,%r9d,%xmm8
+  DB  102,68,15,58,34,192,3               ; pinsrd        $0x3,%eax,%xmm8
+  DB  102,67,15,110,12,160                ; movd          (%r8,%r12,4),%xmm1
+  DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
+  DB  102,15,58,34,197,1                  ; pinsrd        $0x1,%ebp,%xmm0
+  DB  102,65,15,58,34,198,2               ; pinsrd        $0x2,%r14d,%xmm0
+  DB  102,15,58,34,195,3                  ; pinsrd        $0x3,%ebx,%xmm0
+  DB  102,15,196,205,1                    ; pinsrw        $0x1,%ebp,%xmm1
+  DB  102,65,15,196,206,2                 ; pinsrw        $0x2,%r14d,%xmm1
+  DB  102,15,196,203,3                    ; pinsrw        $0x3,%ebx,%xmm1
+  DB  102,65,15,196,203,4                 ; pinsrw        $0x4,%r11d,%xmm1
+  DB  102,65,15,196,202,5                 ; pinsrw        $0x5,%r10d,%xmm1
+  DB  102,65,15,196,201,6                 ; pinsrw        $0x6,%r9d,%xmm1
+  DB  102,15,196,200,7                    ; pinsrw        $0x7,%eax,%xmm1
+  DB  102,68,15,111,13,224,81,0,0         ; movdqa        0x51e0(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
+  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
+  DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
+  DB  102,68,15,111,21,89,89,0,0          ; movdqa        0x5959(%rip),%xmm10        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
+  DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
+  DB  102,65,15,56,0,218                  ; pshufb        %xmm10,%xmm3
+  DB  102,69,15,111,216                   ; movdqa        %xmm8,%xmm11
+  DB  102,69,15,56,0,218                  ; pshufb        %xmm10,%xmm11
+  DB  102,65,15,108,219                   ; punpcklqdq    %xmm11,%xmm3
+  DB  102,68,15,111,21,70,89,0,0          ; movdqa        0x5946(%rip),%xmm10        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
+  DB  102,65,15,56,0,194                  ; pshufb        %xmm10,%xmm0
+  DB  102,69,15,56,0,194                  ; pshufb        %xmm10,%xmm8
+  DB  102,65,15,108,192                   ; punpcklqdq    %xmm8,%xmm0
+  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_565_sse41_lowp
 _sk_load_565_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -38104,23 +38730,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            27e88 <_sk_load_565_sse41_lowp+0x3c>
+  DB  119,34                              ; ja            28968 <_sk_load_565_sse41_lowp+0x3c>
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,196,0,0,0                 ; lea           0xc4(%rip),%r9        # 27f38 <_sk_load_565_sse41_lowp+0xec>
+  DB  76,141,13,196,0,0,0                 ; lea           0xc4(%rip),%r9        # 28a18 <_sk_load_565_sse41_lowp+0xec>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,192                      ; movd          %eax,%xmm0
-  DB  235,74                              ; jmp           27ed2 <_sk_load_565_sse41_lowp+0x86>
+  DB  235,74                              ; jmp           289b2 <_sk_load_565_sse41_lowp+0x86>
   DB  243,65,15,111,4,80                  ; movdqu        (%r8,%rdx,2),%xmm0
-  DB  235,66                              ; jmp           27ed2 <_sk_load_565_sse41_lowp+0x86>
+  DB  235,66                              ; jmp           289b2 <_sk_load_565_sse41_lowp+0x86>
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  102,65,15,196,76,80,4,2             ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm1
   DB  102,65,15,110,4,80                  ; movd          (%r8,%rdx,2),%xmm0
   DB  102,15,58,14,193,252                ; pblendw       $0xfc,%xmm1,%xmm0
-  DB  235,40                              ; jmp           27ed2 <_sk_load_565_sse41_lowp+0x86>
+  DB  235,40                              ; jmp           289b2 <_sk_load_565_sse41_lowp+0x86>
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  102,65,15,196,76,80,12,6            ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm1
   DB  102,65,15,196,76,80,10,5            ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm1
@@ -38129,11 +38755,11 @@
   DB  102,15,58,14,193,240                ; pblendw       $0xf0,%xmm1,%xmm0
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
   DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
-  DB  102,15,219,29,157,76,0,0            ; pand          0x4c9d(%rip),%xmm3        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,29,157,88,0,0            ; pand          0x589d(%rip),%xmm3        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
-  DB  102,15,219,13,156,76,0,0            ; pand          0x4c9c(%rip),%xmm1        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,15,111,21,164,76,0,0            ; movdqa        0x4ca4(%rip),%xmm2        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,15,219,13,156,88,0,0            ; pand          0x589c(%rip),%xmm1        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,21,164,88,0,0            ; movdqa        0x58a4(%rip),%xmm2        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,15,219,208                      ; pand          %xmm0,%xmm2
   DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
   DB  102,15,235,195                      ; por           %xmm3,%xmm0
@@ -38146,7 +38772,7 @@
   DB  102,15,113,210,2                    ; psrlw         $0x2,%xmm2
   DB  102,15,235,211                      ; por           %xmm3,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,154,68,0,0                 ; movaps        0x449a(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,154,80,0,0                 ; movaps        0x509a(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  255,224                             ; jmpq          *%rax
   DB  69,255                              ; rex.RB        (bad)
   DB  255                                 ; (bad)
@@ -38157,7 +38783,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  126,255                             ; jle           27f4d <_sk_load_565_sse41_lowp+0x101>
+  DB  126,255                             ; jle           28a2d <_sk_load_565_sse41_lowp+0x101>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -38174,23 +38800,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            27f90 <_sk_load_565_dst_sse41_lowp+0x3c>
+  DB  119,34                              ; ja            28a70 <_sk_load_565_dst_sse41_lowp+0x3c>
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,196,0,0,0                 ; lea           0xc4(%rip),%r9        # 28040 <_sk_load_565_dst_sse41_lowp+0xec>
+  DB  76,141,13,196,0,0,0                 ; lea           0xc4(%rip),%r9        # 28b20 <_sk_load_565_dst_sse41_lowp+0xec>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,224                      ; movd          %eax,%xmm4
-  DB  235,74                              ; jmp           27fda <_sk_load_565_dst_sse41_lowp+0x86>
+  DB  235,74                              ; jmp           28aba <_sk_load_565_dst_sse41_lowp+0x86>
   DB  243,65,15,111,36,80                 ; movdqu        (%r8,%rdx,2),%xmm4
-  DB  235,66                              ; jmp           27fda <_sk_load_565_dst_sse41_lowp+0x86>
+  DB  235,66                              ; jmp           28aba <_sk_load_565_dst_sse41_lowp+0x86>
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  102,65,15,196,108,80,4,2            ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm5
   DB  102,65,15,110,36,80                 ; movd          (%r8,%rdx,2),%xmm4
   DB  102,15,58,14,229,252                ; pblendw       $0xfc,%xmm5,%xmm4
-  DB  235,40                              ; jmp           27fda <_sk_load_565_dst_sse41_lowp+0x86>
+  DB  235,40                              ; jmp           28aba <_sk_load_565_dst_sse41_lowp+0x86>
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  102,65,15,196,108,80,12,6           ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm5
   DB  102,65,15,196,108,80,10,5           ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm5
@@ -38199,11 +38825,11 @@
   DB  102,15,58,14,229,240                ; pblendw       $0xf0,%xmm5,%xmm4
   DB  102,15,111,252                      ; movdqa        %xmm4,%xmm7
   DB  102,15,113,215,8                    ; psrlw         $0x8,%xmm7
-  DB  102,15,219,61,149,75,0,0            ; pand          0x4b95(%rip),%xmm7        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,61,149,87,0,0            ; pand          0x5795(%rip),%xmm7        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,113,213,5                    ; psrlw         $0x5,%xmm5
-  DB  102,15,219,45,148,75,0,0            ; pand          0x4b94(%rip),%xmm5        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,15,111,53,156,75,0,0            ; movdqa        0x4b9c(%rip),%xmm6        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,15,219,45,148,87,0,0            ; pand          0x5794(%rip),%xmm5        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,53,156,87,0,0            ; movdqa        0x579c(%rip),%xmm6        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,15,219,244                      ; pand          %xmm4,%xmm6
   DB  102,15,113,212,13                   ; psrlw         $0xd,%xmm4
   DB  102,15,235,231                      ; por           %xmm7,%xmm4
@@ -38216,7 +38842,7 @@
   DB  102,15,113,214,2                    ; psrlw         $0x2,%xmm6
   DB  102,15,235,247                      ; por           %xmm7,%xmm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,146,67,0,0                 ; movaps        0x4392(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,61,146,79,0,0                 ; movaps        0x4f92(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  255,224                             ; jmpq          *%rax
   DB  69,255                              ; rex.RB        (bad)
   DB  255                                 ; (bad)
@@ -38227,7 +38853,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  126,255                             ; jle           28055 <_sk_load_565_dst_sse41_lowp+0x101>
+  DB  126,255                             ; jle           28b35 <_sk_load_565_dst_sse41_lowp+0x101>
   DB  255                                 ; (bad)
   DB  255,114,255                         ; pushq         -0x1(%rdx)
   DB  255                                 ; (bad)
@@ -38242,10 +38868,10 @@
   DB  76,3,0                              ; add           (%rax),%r8
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
   DB  102,65,15,113,240,8                 ; psllw         $0x8,%xmm8
-  DB  102,68,15,219,5,48,75,0,0           ; pand          0x4b30(%rip),%xmm8        # 2cbb0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a8>
+  DB  102,68,15,219,5,48,87,0,0           ; pand          0x5730(%rip),%xmm8        # 2e290 <_sk_srcover_rgba_8888_sse2_lowp+0x11dc>
   DB  102,68,15,111,201                   ; movdqa        %xmm1,%xmm9
   DB  102,65,15,113,241,3                 ; psllw         $0x3,%xmm9
-  DB  102,68,15,219,13,44,75,0,0          ; pand          0x4b2c(%rip),%xmm9        # 2cbc0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b8>
+  DB  102,68,15,219,13,44,87,0,0          ; pand          0x572c(%rip),%xmm9        # 2e2a0 <_sk_srcover_rgba_8888_sse2_lowp+0x11ec>
   DB  102,69,15,235,200                   ; por           %xmm8,%xmm9
   DB  102,68,15,111,194                   ; movdqa        %xmm2,%xmm8
   DB  102,65,15,113,208,3                 ; psrlw         $0x3,%xmm8
@@ -38254,19 +38880,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,29                              ; ja            280d0 <_sk_store_565_sse41_lowp+0x74>
+  DB  119,29                              ; ja            28bb0 <_sk_store_565_sse41_lowp+0x74>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,83,0,0,0                  ; lea           0x53(%rip),%r9        # 28110 <_sk_store_565_sse41_lowp+0xb4>
+  DB  76,141,13,83,0,0,0                  ; lea           0x53(%rip),%r9        # 28bf0 <_sk_store_565_sse41_lowp+0xb4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,58,21,4,80,0              ; pextrw        $0x0,%xmm8,(%r8,%rdx,2)
-  DB  235,58                              ; jmp           2810a <_sk_store_565_sse41_lowp+0xae>
+  DB  235,58                              ; jmp           28bea <_sk_store_565_sse41_lowp+0xae>
   DB  243,69,15,127,4,80                  ; movdqu        %xmm8,(%r8,%rdx,2)
-  DB  235,50                              ; jmp           2810a <_sk_store_565_sse41_lowp+0xae>
+  DB  235,50                              ; jmp           28bea <_sk_store_565_sse41_lowp+0xae>
   DB  102,69,15,58,21,68,80,4,2           ; pextrw        $0x2,%xmm8,0x4(%r8,%rdx,2)
   DB  102,69,15,126,4,80                  ; movd          %xmm8,(%r8,%rdx,2)
-  DB  235,33                              ; jmp           2810a <_sk_store_565_sse41_lowp+0xae>
+  DB  235,33                              ; jmp           28bea <_sk_store_565_sse41_lowp+0xae>
   DB  102,69,15,58,21,68,80,12,6          ; pextrw        $0x6,%xmm8,0xc(%r8,%rdx,2)
   DB  102,69,15,58,21,68,80,10,5          ; pextrw        $0x5,%xmm8,0xa(%r8,%rdx,2)
   DB  102,69,15,58,21,68,80,8,4           ; pextrw        $0x4,%xmm8,0x8(%r8,%rdx,2)
@@ -38286,7 +38912,7 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  235,255                             ; jmp           28121 <_sk_store_565_sse41_lowp+0xc5>
+  DB  235,255                             ; jmp           28c01 <_sk_store_565_sse41_lowp+0xc5>
   DB  255                                 ; (bad)
   DB  255,226                             ; jmpq          *%rdx
   DB  255                                 ; (bad)
@@ -38296,6 +38922,96 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_565_sse41_lowp
+_sk_gather_565_sse41_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  243,68,15,91,195                    ; cvttps2dq     %xmm3,%xmm8
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,15,56,64,211                    ; pmulld        %xmm3,%xmm2
+  DB  102,68,15,56,64,195                 ; pmulld        %xmm3,%xmm8
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,200                   ; paddd         %xmm8,%xmm1
+  DB  102,72,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%rax
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
+  DB  69,137,211                          ; mov           %r10d,%r11d
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  102,72,15,58,22,195,1               ; pextrq        $0x1,%xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  71,15,183,60,120                    ; movzwl        (%r8,%r15,2),%r15d
+  DB  67,15,183,44,96                     ; movzwl        (%r8,%r12,2),%ebp
+  DB  102,15,110,197                      ; movd          %ebp,%xmm0
+  DB  102,65,15,196,199,1                 ; pinsrw        $0x1,%r15d,%xmm0
+  DB  67,15,183,44,112                    ; movzwl        (%r8,%r14,2),%ebp
+  DB  102,15,196,197,2                    ; pinsrw        $0x2,%ebp,%xmm0
+  DB  65,15,183,28,88                     ; movzwl        (%r8,%rbx,2),%ebx
+  DB  102,15,196,195,3                    ; pinsrw        $0x3,%ebx,%xmm0
+  DB  67,15,183,44,88                     ; movzwl        (%r8,%r11,2),%ebp
+  DB  102,15,196,197,4                    ; pinsrw        $0x4,%ebp,%xmm0
+  DB  67,15,183,44,80                     ; movzwl        (%r8,%r10,2),%ebp
+  DB  102,15,196,197,5                    ; pinsrw        $0x5,%ebp,%xmm0
+  DB  67,15,183,44,72                     ; movzwl        (%r8,%r9,2),%ebp
+  DB  102,15,196,197,6                    ; pinsrw        $0x6,%ebp,%xmm0
+  DB  65,15,183,4,64                      ; movzwl        (%r8,%rax,2),%eax
+  DB  102,15,196,192,7                    ; pinsrw        $0x7,%eax,%xmm0
+  DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
+  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
+  DB  102,15,219,29,58,85,0,0             ; pand          0x553a(%rip),%xmm3        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
+  DB  102,15,219,13,57,85,0,0             ; pand          0x5539(%rip),%xmm1        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,21,65,85,0,0             ; movdqa        0x5541(%rip),%xmm2        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
+  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
+  DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
+  DB  102,15,235,195                      ; por           %xmm3,%xmm0
+  DB  102,15,111,217                      ; movdqa        %xmm1,%xmm3
+  DB  102,15,113,243,2                    ; psllw         $0x2,%xmm3
+  DB  102,15,113,209,4                    ; psrlw         $0x4,%xmm1
+  DB  102,15,235,203                      ; por           %xmm3,%xmm1
+  DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
+  DB  102,15,113,243,3                    ; psllw         $0x3,%xmm3
+  DB  102,15,113,210,2                    ; psrlw         $0x2,%xmm2
+  DB  102,15,235,211                      ; por           %xmm3,%xmm2
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  15,40,29,55,77,0,0                  ; movaps        0x4d37(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_a8_sse41_lowp
 _sk_load_a8_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -38306,18 +39022,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            28165 <_sk_load_a8_sse41_lowp+0x39>
+  DB  119,34                              ; ja            28dbc <_sk_load_a8_sse41_lowp+0x39>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,143,0,0,0                 ; lea           0x8f(%rip),%r9        # 281e0 <_sk_load_a8_sse41_lowp+0xb4>
+  DB  76,141,13,144,0,0,0                 ; lea           0x90(%rip),%r9        # 28e38 <_sk_load_a8_sse41_lowp+0xb5>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,216                      ; movd          %eax,%xmm3
-  DB  235,100                             ; jmp           281c9 <_sk_load_a8_sse41_lowp+0x9d>
+  DB  235,100                             ; jmp           28e20 <_sk_load_a8_sse41_lowp+0x9d>
   DB  102,65,15,56,48,28,16               ; pmovzxbw      (%r8,%rdx,1),%xmm3
-  DB  235,91                              ; jmp           281c9 <_sk_load_a8_sse41_lowp+0x9d>
+  DB  235,91                              ; jmp           28e20 <_sk_load_a8_sse41_lowp+0x9d>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,192,2                    ; pinsrw        $0x2,%eax,%xmm0
@@ -38325,7 +39041,7 @@
   DB  102,15,110,200                      ; movd          %eax,%xmm1
   DB  102,15,56,48,217                    ; pmovzxbw      %xmm1,%xmm3
   DB  102,15,58,14,216,252                ; pblendw       $0xfc,%xmm0,%xmm3
-  DB  235,54                              ; jmp           281c9 <_sk_load_a8_sse41_lowp+0x9d>
+  DB  235,54                              ; jmp           28e20 <_sk_load_a8_sse41_lowp+0x9d>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,192,6                    ; pinsrw        $0x6,%eax,%xmm0
@@ -38336,28 +39052,28 @@
   DB  102,65,15,110,12,16                 ; movd          (%r8,%rdx,1),%xmm1
   DB  102,15,56,48,217                    ; pmovzxbw      %xmm1,%xmm3
   DB  102,15,58,14,216,240                ; pblendw       $0xf0,%xmm0,%xmm3
-  DB  102,15,219,29,255,65,0,0            ; pand          0x41ff(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,29,136,76,0,0            ; pand          0x4c88(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
   DB  255,224                             ; jmpq          *%rax
-  DB  122,255                             ; jp            281e1 <_sk_load_a8_sse41_lowp+0xb5>
+  DB  144                                 ; nop
+  DB  121,255                             ; jns           28e39 <_sk_load_a8_sse41_lowp+0xb6>
   DB  255                                 ; (bad)
-  DB  255,157,255,255,255,142             ; lcall         *-0x71000001(%rbp)
+  DB  255,156,255,255,255,141,255         ; lcall         *-0x720001(%rdi,%rdi,8)
+  DB  255                                 ; (bad)
+  DB  255,215                             ; callq         *%rdi
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255                                 ; (bad)
-  DB  216,255                             ; fdivr         %st(7),%st
-  DB  255                                 ; (bad)
-  DB  255,205                             ; dec           %ebp
+  DB  255,204                             ; dec           %esp
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,194                             ; inc           %edx
+  DB  255,193                             ; inc           %ecx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
-  DB  179,255                             ; mov           $0xff,%bl
+  DB  178,255                             ; mov           $0xff,%dl
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
@@ -38371,18 +39087,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            28235 <_sk_load_a8_dst_sse41_lowp+0x39>
+  DB  119,34                              ; ja            28e8d <_sk_load_a8_dst_sse41_lowp+0x39>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,143,0,0,0                 ; lea           0x8f(%rip),%r9        # 282b0 <_sk_load_a8_dst_sse41_lowp+0xb4>
+  DB  76,141,13,143,0,0,0                 ; lea           0x8f(%rip),%r9        # 28f08 <_sk_load_a8_dst_sse41_lowp+0xb4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,248                      ; movd          %eax,%xmm7
-  DB  235,100                             ; jmp           28299 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  DB  235,100                             ; jmp           28ef1 <_sk_load_a8_dst_sse41_lowp+0x9d>
   DB  102,65,15,56,48,60,16               ; pmovzxbw      (%r8,%rdx,1),%xmm7
-  DB  235,91                              ; jmp           28299 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  DB  235,91                              ; jmp           28ef1 <_sk_load_a8_dst_sse41_lowp+0x9d>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,224,2                    ; pinsrw        $0x2,%eax,%xmm4
@@ -38390,7 +39106,7 @@
   DB  102,15,110,232                      ; movd          %eax,%xmm5
   DB  102,15,56,48,253                    ; pmovzxbw      %xmm5,%xmm7
   DB  102,15,58,14,252,252                ; pblendw       $0xfc,%xmm4,%xmm7
-  DB  235,54                              ; jmp           28299 <_sk_load_a8_dst_sse41_lowp+0x9d>
+  DB  235,54                              ; jmp           28ef1 <_sk_load_a8_dst_sse41_lowp+0x9d>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,224,6                    ; pinsrw        $0x6,%eax,%xmm4
@@ -38401,13 +39117,13 @@
   DB  102,65,15,110,44,16                 ; movd          (%r8,%rdx,1),%xmm5
   DB  102,15,56,48,253                    ; pmovzxbw      %xmm5,%xmm7
   DB  102,15,58,14,252,240                ; pblendw       $0xf0,%xmm4,%xmm7
-  DB  102,15,219,61,47,65,0,0             ; pand          0x412f(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,61,183,75,0,0            ; pand          0x4bb7(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  15,87,246                           ; xorps         %xmm6,%xmm6
   DB  255,224                             ; jmpq          *%rax
-  DB  122,255                             ; jp            282b1 <_sk_load_a8_dst_sse41_lowp+0xb5>
+  DB  122,255                             ; jp            28f09 <_sk_load_a8_dst_sse41_lowp+0xb5>
   DB  255                                 ; (bad)
   DB  255,157,255,255,255,142             ; lcall         *-0x71000001(%rbp)
   DB  255                                 ; (bad)
@@ -38436,28 +39152,28 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,29                              ; ja            28300 <_sk_store_a8_sse41_lowp+0x34>
+  DB  119,29                              ; ja            28f58 <_sk_store_a8_sse41_lowp+0x34>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,131,0,0,0                 ; lea           0x83(%rip),%r9        # 28370 <_sk_store_a8_sse41_lowp+0xa4>
+  DB  76,141,13,131,0,0,0                 ; lea           0x83(%rip),%r9        # 28fc8 <_sk_store_a8_sse41_lowp+0xa4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,58,20,28,16,0             ; pextrb        $0x0,%xmm3,(%r8,%rdx,1)
-  DB  235,105                             ; jmp           28369 <_sk_store_a8_sse41_lowp+0x9d>
+  DB  235,105                             ; jmp           28fc1 <_sk_store_a8_sse41_lowp+0x9d>
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
-  DB  102,68,15,56,0,5,17,72,0,0          ; pshufb        0x4811(%rip),%xmm8        # 2cb20 <_sk_srcover_rgba_8888_sse2_lowp+0x1118>
+  DB  102,68,15,56,0,5,153,82,0,0         ; pshufb        0x5299(%rip),%xmm8        # 2e200 <_sk_srcover_rgba_8888_sse2_lowp+0x114c>
   DB  102,69,15,214,4,16                  ; movq          %xmm8,(%r8,%rdx,1)
-  DB  235,82                              ; jmp           28369 <_sk_store_a8_sse41_lowp+0x9d>
+  DB  235,82                              ; jmp           28fc1 <_sk_store_a8_sse41_lowp+0x9d>
   DB  102,65,15,58,20,92,16,2,4           ; pextrb        $0x4,%xmm3,0x2(%r8,%rdx,1)
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
-  DB  102,68,15,56,0,5,177,64,0,0         ; pshufb        0x40b1(%rip),%xmm8        # 2c3e0 <_sk_srcover_rgba_8888_sse2_lowp+0x9d8>
+  DB  102,68,15,56,0,5,57,75,0,0          ; pshufb        0x4b39(%rip),%xmm8        # 2dac0 <_sk_srcover_rgba_8888_sse2_lowp+0xa0c>
   DB  102,69,15,58,21,4,16,0              ; pextrw        $0x0,%xmm8,(%r8,%rdx,1)
-  DB  235,48                              ; jmp           28369 <_sk_store_a8_sse41_lowp+0x9d>
+  DB  235,48                              ; jmp           28fc1 <_sk_store_a8_sse41_lowp+0x9d>
   DB  102,65,15,58,20,92,16,6,12          ; pextrb        $0xc,%xmm3,0x6(%r8,%rdx,1)
   DB  102,65,15,58,20,92,16,5,10          ; pextrb        $0xa,%xmm3,0x5(%r8,%rdx,1)
   DB  102,65,15,58,20,92,16,4,8           ; pextrb        $0x8,%xmm3,0x4(%r8,%rdx,1)
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
-  DB  102,68,15,56,0,5,141,64,0,0         ; pshufb        0x408d(%rip),%xmm8        # 2c3f0 <_sk_srcover_rgba_8888_sse2_lowp+0x9e8>
+  DB  102,68,15,56,0,5,21,75,0,0          ; pshufb        0x4b15(%rip),%xmm8        # 2dad0 <_sk_srcover_rgba_8888_sse2_lowp+0xa1c>
   DB  102,69,15,126,4,16                  ; movd          %xmm8,(%r8,%rdx,1)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
@@ -38481,6 +39197,81 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_a8_sse41_lowp
+_sk_gather_a8_sse41_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  243,68,15,91,195                    ; cvttps2dq     %xmm3,%xmm8
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,15,56,64,211                    ; pmulld        %xmm3,%xmm2
+  DB  102,68,15,56,64,195                 ; pmulld        %xmm3,%xmm8
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,200                   ; paddd         %xmm8,%xmm1
+  DB  102,72,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%rax
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
+  DB  69,137,211                          ; mov           %r10d,%r11d
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  102,72,15,58,22,195,1               ; pextrq        $0x1,%xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  71,15,182,60,56                     ; movzbl        (%r8,%r15,1),%r15d
+  DB  67,15,182,44,32                     ; movzbl        (%r8,%r12,1),%ebp
+  DB  102,15,110,197                      ; movd          %ebp,%xmm0
+  DB  102,65,15,58,32,199,1               ; pinsrb        $0x1,%r15d,%xmm0
+  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
+  DB  102,15,58,32,197,2                  ; pinsrb        $0x2,%ebp,%xmm0
+  DB  65,15,182,28,24                     ; movzbl        (%r8,%rbx,1),%ebx
+  DB  102,15,58,32,195,3                  ; pinsrb        $0x3,%ebx,%xmm0
+  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
+  DB  102,15,58,32,197,4                  ; pinsrb        $0x4,%ebp,%xmm0
+  DB  67,15,182,44,16                     ; movzbl        (%r8,%r10,1),%ebp
+  DB  102,15,58,32,197,5                  ; pinsrb        $0x5,%ebp,%xmm0
+  DB  67,15,182,44,8                      ; movzbl        (%r8,%r9,1),%ebp
+  DB  102,15,58,32,197,6                  ; pinsrb        $0x6,%ebp,%xmm0
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  102,15,58,32,192,7                  ; pinsrb        $0x7,%eax,%xmm0
+  DB  102,15,56,48,216                    ; pmovzxbw      %xmm0,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
+  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
+  DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_g8_sse41_lowp
 _sk_load_g8_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -38491,18 +39282,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            283c5 <_sk_load_g8_sse41_lowp+0x39>
+  DB  119,34                              ; ja            2914a <_sk_load_g8_sse41_lowp+0x39>
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 28444 <_sk_load_g8_sse41_lowp+0xb8>
+  DB  76,141,13,150,0,0,0                 ; lea           0x96(%rip),%r9        # 291cc <_sk_load_g8_sse41_lowp+0xbb>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,192                      ; movd          %eax,%xmm0
-  DB  235,100                             ; jmp           28429 <_sk_load_g8_sse41_lowp+0x9d>
+  DB  235,100                             ; jmp           291ae <_sk_load_g8_sse41_lowp+0x9d>
   DB  102,65,15,56,48,4,16                ; pmovzxbw      (%r8,%rdx,1),%xmm0
-  DB  235,91                              ; jmp           28429 <_sk_load_g8_sse41_lowp+0x9d>
+  DB  235,91                              ; jmp           291ae <_sk_load_g8_sse41_lowp+0x9d>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  102,15,196,200,2                    ; pinsrw        $0x2,%eax,%xmm1
@@ -38510,7 +39301,7 @@
   DB  102,15,110,192                      ; movd          %eax,%xmm0
   DB  102,15,56,48,192                    ; pmovzxbw      %xmm0,%xmm0
   DB  102,15,58,14,193,252                ; pblendw       $0xfc,%xmm1,%xmm0
-  DB  235,54                              ; jmp           28429 <_sk_load_g8_sse41_lowp+0x9d>
+  DB  235,54                              ; jmp           291ae <_sk_load_g8_sse41_lowp+0x9d>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
   DB  102,15,196,200,6                    ; pinsrw        $0x6,%eax,%xmm1
@@ -38521,25 +39312,26 @@
   DB  102,65,15,110,4,16                  ; movd          (%r8,%rdx,1),%xmm0
   DB  102,15,56,48,192                    ; pmovzxbw      %xmm0,%xmm0
   DB  102,15,58,14,193,240                ; pblendw       $0xf0,%xmm1,%xmm0
-  DB  102,15,219,5,159,63,0,0             ; pand          0x3f9f(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,5,250,72,0,0             ; pand          0x48fa(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,150,63,0,0                 ; movaps        0x3f96(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,241,72,0,0                 ; movaps        0x48f1(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
-  DB  118,255                             ; jbe           28445 <_sk_load_g8_sse41_lowp+0xb9>
+  DB  15,31,0                             ; nopl          (%rax)
+  DB  115,255                             ; jae           291cd <_sk_load_g8_sse41_lowp+0xbc>
   DB  255                                 ; (bad)
-  DB  255,153,255,255,255,138             ; lcall         *-0x75000001(%rcx)
+  DB  255,150,255,255,255,135             ; callq         *-0x78000001(%rsi)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,212                             ; callq         *%rsp
+  DB  255,209                             ; callq         *%rcx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,201                             ; dec           %ecx
+  DB  255,198                             ; inc           %esi
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  190,255,255,255,175                 ; mov           $0xafffffff,%esi
+  DB  187,255,255,255,172                 ; mov           $0xacffffff,%ebx
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -38554,18 +39346,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            28499 <_sk_load_g8_dst_sse41_lowp+0x39>
+  DB  119,34                              ; ja            29221 <_sk_load_g8_dst_sse41_lowp+0x39>
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 28518 <_sk_load_g8_dst_sse41_lowp+0xb8>
+  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 292a0 <_sk_load_g8_dst_sse41_lowp+0xb8>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,224                      ; movd          %eax,%xmm4
-  DB  235,100                             ; jmp           284fd <_sk_load_g8_dst_sse41_lowp+0x9d>
+  DB  235,100                             ; jmp           29285 <_sk_load_g8_dst_sse41_lowp+0x9d>
   DB  102,65,15,56,48,36,16               ; pmovzxbw      (%r8,%rdx,1),%xmm4
-  DB  235,91                              ; jmp           284fd <_sk_load_g8_dst_sse41_lowp+0x9d>
+  DB  235,91                              ; jmp           29285 <_sk_load_g8_dst_sse41_lowp+0x9d>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  102,15,196,232,2                    ; pinsrw        $0x2,%eax,%xmm5
@@ -38573,7 +39365,7 @@
   DB  102,15,110,224                      ; movd          %eax,%xmm4
   DB  102,15,56,48,228                    ; pmovzxbw      %xmm4,%xmm4
   DB  102,15,58,14,229,252                ; pblendw       $0xfc,%xmm5,%xmm4
-  DB  235,54                              ; jmp           284fd <_sk_load_g8_dst_sse41_lowp+0x9d>
+  DB  235,54                              ; jmp           29285 <_sk_load_g8_dst_sse41_lowp+0x9d>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,237                      ; pxor          %xmm5,%xmm5
   DB  102,15,196,232,6                    ; pinsrw        $0x6,%eax,%xmm5
@@ -38584,13 +39376,13 @@
   DB  102,65,15,110,36,16                 ; movd          (%r8,%rdx,1),%xmm4
   DB  102,15,56,48,228                    ; pmovzxbw      %xmm4,%xmm4
   DB  102,15,58,14,229,240                ; pblendw       $0xf0,%xmm5,%xmm4
-  DB  102,15,219,37,203,62,0,0            ; pand          0x3ecb(%rip),%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,37,35,72,0,0             ; pand          0x4823(%rip),%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,194,62,0,0                 ; movaps        0x3ec2(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,61,26,72,0,0                  ; movaps        0x481a(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  255,224                             ; jmpq          *%rax
-  DB  118,255                             ; jbe           28519 <_sk_load_g8_dst_sse41_lowp+0xb9>
+  DB  118,255                             ; jbe           292a1 <_sk_load_g8_dst_sse41_lowp+0xb9>
   DB  255                                 ; (bad)
   DB  255,153,255,255,255,138             ; lcall         *-0x75000001(%rcx)
   DB  255                                 ; (bad)
@@ -38610,10 +39402,10 @@
 PUBLIC _sk_luminance_to_alpha_sse41_lowp
 _sk_luminance_to_alpha_sse41_lowp LABEL PROC
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
-  DB  102,15,213,5,144,70,0,0             ; pmullw        0x4690(%rip),%xmm0        # 2cbd0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c8>
-  DB  102,15,213,13,152,70,0,0            ; pmullw        0x4698(%rip),%xmm1        # 2cbe0 <_sk_srcover_rgba_8888_sse2_lowp+0x11d8>
+  DB  102,15,213,5,232,79,0,0             ; pmullw        0x4fe8(%rip),%xmm0        # 2e2b0 <_sk_srcover_rgba_8888_sse2_lowp+0x11fc>
+  DB  102,15,213,13,240,79,0,0            ; pmullw        0x4ff0(%rip),%xmm1        # 2e2c0 <_sk_srcover_rgba_8888_sse2_lowp+0x120c>
   DB  102,15,253,200                      ; paddw         %xmm0,%xmm1
-  DB  102,15,213,29,156,70,0,0            ; pmullw        0x469c(%rip),%xmm3        # 2cbf0 <_sk_srcover_rgba_8888_sse2_lowp+0x11e8>
+  DB  102,15,213,29,244,79,0,0            ; pmullw        0x4ff4(%rip),%xmm3        # 2e2d0 <_sk_srcover_rgba_8888_sse2_lowp+0x121c>
   DB  102,15,253,217                      ; paddw         %xmm1,%xmm3
   DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -38622,18 +39414,93 @@
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_g8_sse41_lowp
+_sk_gather_g8_sse41_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  76,139,0                            ; mov           (%rax),%r8
+  DB  243,68,15,91,195                    ; cvttps2dq     %xmm3,%xmm8
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,15,56,64,211                    ; pmulld        %xmm3,%xmm2
+  DB  102,68,15,56,64,195                 ; pmulld        %xmm3,%xmm8
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,200                   ; paddd         %xmm8,%xmm1
+  DB  102,72,15,58,22,200,1               ; pextrq        $0x1,%xmm1,%rax
+  DB  65,137,193                          ; mov           %eax,%r9d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
+  DB  69,137,211                          ; mov           %r10d,%r11d
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  102,72,15,58,22,195,1               ; pextrq        $0x1,%xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  71,15,182,60,56                     ; movzbl        (%r8,%r15,1),%r15d
+  DB  67,15,182,44,32                     ; movzbl        (%r8,%r12,1),%ebp
+  DB  102,15,110,197                      ; movd          %ebp,%xmm0
+  DB  102,65,15,58,32,199,1               ; pinsrb        $0x1,%r15d,%xmm0
+  DB  67,15,182,44,48                     ; movzbl        (%r8,%r14,1),%ebp
+  DB  102,15,58,32,197,2                  ; pinsrb        $0x2,%ebp,%xmm0
+  DB  65,15,182,28,24                     ; movzbl        (%r8,%rbx,1),%ebx
+  DB  102,15,58,32,195,3                  ; pinsrb        $0x3,%ebx,%xmm0
+  DB  67,15,182,44,24                     ; movzbl        (%r8,%r11,1),%ebp
+  DB  102,15,58,32,197,4                  ; pinsrb        $0x4,%ebp,%xmm0
+  DB  67,15,182,44,16                     ; movzbl        (%r8,%r10,1),%ebp
+  DB  102,15,58,32,197,5                  ; pinsrb        $0x5,%ebp,%xmm0
+  DB  67,15,182,44,8                      ; movzbl        (%r8,%r9,1),%ebp
+  DB  102,15,58,32,197,6                  ; pinsrb        $0x6,%ebp,%xmm0
+  DB  65,15,182,4,0                       ; movzbl        (%r8,%rax,1),%eax
+  DB  102,15,58,32,192,7                  ; pinsrb        $0x7,%eax,%xmm0
+  DB  102,15,56,48,192                    ; pmovzxbw      %xmm0,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  15,40,29,157,70,0,0                 ; movaps        0x469d(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_scale_1_float_sse41_lowp
 _sk_scale_1_float_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
-  DB  243,68,15,89,5,103,55,0,0           ; mulss         0x3767(%rip),%xmm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  243,68,15,88,5,46,55,0,0            ; addss         0x372e(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,5,91,63,0,0            ; mulss         0x3f5b(%rip),%xmm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,68,15,88,5,34,63,0,0            ; addss         0x3f22(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,44,192                    ; cvttss2si     %xmm8,%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  242,69,15,112,192,0                 ; pshuflw       $0x0,%xmm8,%xmm8
   DB  102,69,15,112,192,80                ; pshufd        $0x50,%xmm8,%xmm8
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
-  DB  102,68,15,111,13,38,62,0,0          ; movdqa        0x3e26(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,78,70,0,0          ; movdqa        0x464e(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -38652,13 +39519,13 @@
 _sk_lerp_1_float_sse41_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
-  DB  243,68,15,89,5,239,54,0,0           ; mulss         0x36ef(%rip),%xmm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  243,68,15,88,5,182,54,0,0           ; addss         0x36b6(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,5,227,62,0,0           ; mulss         0x3ee3(%rip),%xmm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,68,15,88,5,170,62,0,0           ; addss         0x3eaa(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,44,192                    ; cvttss2si     %xmm8,%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  242,69,15,112,192,0                 ; pshuflw       $0x0,%xmm8,%xmm8
   DB  102,69,15,112,192,80                ; pshufd        $0x50,%xmm8,%xmm8
-  DB  102,68,15,111,13,179,61,0,0         ; movdqa        0x3db3(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,219,69,0,0         ; movdqa        0x45db(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -38696,18 +39563,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,36                              ; ja            286d4 <_sk_scale_u8_sse41_lowp+0x3b>
+  DB  119,36                              ; ja            2958c <_sk_scale_u8_sse41_lowp+0x3b>
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,221,0,0,0                 ; lea           0xdd(%rip),%r9        # 2879c <_sk_scale_u8_sse41_lowp+0x103>
+  DB  76,141,13,221,0,0,0                 ; lea           0xdd(%rip),%r9        # 29654 <_sk_scale_u8_sse41_lowp+0x103>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
-  DB  235,111                             ; jmp           28743 <_sk_scale_u8_sse41_lowp+0xaa>
+  DB  235,111                             ; jmp           295fb <_sk_scale_u8_sse41_lowp+0xaa>
   DB  102,69,15,56,48,4,16                ; pmovzxbw      (%r8,%rdx,1),%xmm8
-  DB  235,102                             ; jmp           28743 <_sk_scale_u8_sse41_lowp+0xaa>
+  DB  235,102                             ; jmp           295fb <_sk_scale_u8_sse41_lowp+0xaa>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,68,15,196,200,2                 ; pinsrw        $0x2,%eax,%xmm9
@@ -38715,7 +39582,7 @@
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  102,69,15,56,48,192                 ; pmovzxbw      %xmm8,%xmm8
   DB  102,69,15,58,14,193,252             ; pblendw       $0xfc,%xmm9,%xmm8
-  DB  235,60                              ; jmp           28743 <_sk_scale_u8_sse41_lowp+0xaa>
+  DB  235,60                              ; jmp           295fb <_sk_scale_u8_sse41_lowp+0xaa>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,68,15,196,200,6                 ; pinsrw        $0x6,%eax,%xmm9
@@ -38726,9 +39593,9 @@
   DB  102,69,15,110,4,16                  ; movd          (%r8,%rdx,1),%xmm8
   DB  102,69,15,56,48,192                 ; pmovzxbw      %xmm8,%xmm8
   DB  102,69,15,58,14,193,240             ; pblendw       $0xf0,%xmm9,%xmm8
-  DB  102,68,15,219,5,132,60,0,0          ; pand          0x3c84(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,219,5,172,68,0,0          ; pand          0x44ac(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
-  DB  102,68,15,111,13,118,60,0,0         ; movdqa        0x3c76(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,158,68,0,0         ; movdqa        0x449e(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -38753,7 +39620,7 @@
   DB  255,148,255,255,255,136,255         ; callq         *-0x770001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  124,255                             ; jl            287b1 <_sk_scale_u8_sse41_lowp+0x118>
+  DB  124,255                             ; jl            29669 <_sk_scale_u8_sse41_lowp+0x118>
   DB  255                                 ; (bad)
   DB  255,107,255                         ; ljmp          *-0x1(%rbx)
   DB  255                                 ; (bad)
@@ -38769,18 +39636,18 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,36                              ; ja            287f3 <_sk_lerp_u8_sse41_lowp+0x3b>
+  DB  119,36                              ; ja            296ab <_sk_lerp_u8_sse41_lowp+0x3b>
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,30,1,0,0                  ; lea           0x11e(%rip),%r9        # 288fc <_sk_lerp_u8_sse41_lowp+0x144>
+  DB  76,141,13,30,1,0,0                  ; lea           0x11e(%rip),%r9        # 297b4 <_sk_lerp_u8_sse41_lowp+0x144>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
-  DB  235,111                             ; jmp           28862 <_sk_lerp_u8_sse41_lowp+0xaa>
+  DB  235,111                             ; jmp           2971a <_sk_lerp_u8_sse41_lowp+0xaa>
   DB  102,69,15,56,48,4,16                ; pmovzxbw      (%r8,%rdx,1),%xmm8
-  DB  235,102                             ; jmp           28862 <_sk_lerp_u8_sse41_lowp+0xaa>
+  DB  235,102                             ; jmp           2971a <_sk_lerp_u8_sse41_lowp+0xaa>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,68,15,196,200,2                 ; pinsrw        $0x2,%eax,%xmm9
@@ -38788,7 +39655,7 @@
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  102,69,15,56,48,192                 ; pmovzxbw      %xmm8,%xmm8
   DB  102,69,15,58,14,193,252             ; pblendw       $0xfc,%xmm9,%xmm8
-  DB  235,60                              ; jmp           28862 <_sk_lerp_u8_sse41_lowp+0xaa>
+  DB  235,60                              ; jmp           2971a <_sk_lerp_u8_sse41_lowp+0xaa>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,69,15,239,201                   ; pxor          %xmm9,%xmm9
   DB  102,68,15,196,200,6                 ; pinsrw        $0x6,%eax,%xmm9
@@ -38799,8 +39666,8 @@
   DB  102,69,15,110,4,16                  ; movd          (%r8,%rdx,1),%xmm8
   DB  102,69,15,56,48,192                 ; pmovzxbw      %xmm8,%xmm8
   DB  102,69,15,58,14,193,240             ; pblendw       $0xf0,%xmm9,%xmm8
-  DB  102,68,15,219,5,101,59,0,0          ; pand          0x3b65(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
-  DB  102,68,15,111,21,92,59,0,0          ; movdqa        0x3b5c(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,219,5,141,67,0,0          ; pand          0x438d(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,68,15,111,21,132,67,0,0         ; movdqa        0x4384(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,69,15,239,202                   ; pxor          %xmm10,%xmm9
   DB  102,69,15,111,217                   ; movdqa        %xmm9,%xmm11
@@ -38830,7 +39697,7 @@
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  255,224                             ; jmpq          *%rax
   DB  144                                 ; nop
-  DB  235,254                             ; jmp           288fc <_sk_lerp_u8_sse41_lowp+0x144>
+  DB  235,254                             ; jmp           297b4 <_sk_lerp_u8_sse41_lowp+0x144>
   DB  255                                 ; (bad)
   DB  255,17                              ; callq         *(%rcx)
   DB  255                                 ; (bad)
@@ -38862,23 +39729,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,35                              ; ja            2895a <_sk_scale_565_sse41_lowp+0x42>
+  DB  119,35                              ; ja            29812 <_sk_scale_565_sse41_lowp+0x42>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,111,1,0,0                 ; lea           0x16f(%rip),%r9        # 28ab4 <_sk_scale_565_sse41_lowp+0x19c>
+  DB  76,141,13,111,1,0,0                 ; lea           0x16f(%rip),%r9        # 2996c <_sk_scale_565_sse41_lowp+0x19c>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,68,15,110,200                   ; movd          %eax,%xmm9
-  DB  235,76                              ; jmp           289a6 <_sk_scale_565_sse41_lowp+0x8e>
+  DB  235,76                              ; jmp           2985e <_sk_scale_565_sse41_lowp+0x8e>
   DB  243,69,15,111,12,80                 ; movdqu        (%r8,%rdx,2),%xmm9
-  DB  235,68                              ; jmp           289a6 <_sk_scale_565_sse41_lowp+0x8e>
+  DB  235,68                              ; jmp           2985e <_sk_scale_565_sse41_lowp+0x8e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,4,2             ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   DB  102,69,15,110,12,80                 ; movd          (%r8,%rdx,2),%xmm9
   DB  102,68,15,58,14,200,252             ; pblendw       $0xfc,%xmm0,%xmm9
-  DB  235,41                              ; jmp           289a6 <_sk_scale_565_sse41_lowp+0x8e>
+  DB  235,41                              ; jmp           2985e <_sk_scale_565_sse41_lowp+0x8e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,12,6            ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   DB  102,65,15,196,68,80,10,5            ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -38887,11 +39754,11 @@
   DB  102,68,15,58,14,200,240             ; pblendw       $0xf0,%xmm0,%xmm9
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
-  DB  102,15,219,5,200,65,0,0             ; pand          0x41c8(%rip),%xmm0        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,5,240,73,0,0             ; pand          0x49f0(%rip),%xmm0        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,69,15,111,209                   ; movdqa        %xmm9,%xmm10
   DB  102,65,15,113,210,5                 ; psrlw         $0x5,%xmm10
-  DB  102,68,15,219,21,196,65,0,0         ; pand          0x41c4(%rip),%xmm10        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,68,15,111,29,203,65,0,0         ; movdqa        0x41cb(%rip),%xmm11        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,68,15,219,21,236,73,0,0         ; pand          0x49ec(%rip),%xmm10        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,68,15,111,29,243,73,0,0         ; movdqa        0x49f3(%rip),%xmm11        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,69,15,219,217                   ; pand          %xmm9,%xmm11
   DB  102,65,15,113,209,13                ; psrlw         $0xd,%xmm9
   DB  102,68,15,235,200                   ; por           %xmm0,%xmm9
@@ -38903,7 +39770,7 @@
   DB  102,15,113,240,3                    ; psllw         $0x3,%xmm0
   DB  102,65,15,113,211,2                 ; psrlw         $0x2,%xmm11
   DB  102,68,15,235,216                   ; por           %xmm0,%xmm11
-  DB  102,15,111,5,41,65,0,0              ; movdqa        0x4129(%rip),%xmm0        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,15,111,5,81,73,0,0              ; movdqa        0x4951(%rip),%xmm0        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,68,15,111,227                   ; movdqa        %xmm3,%xmm12
   DB  102,68,15,239,224                   ; pxor          %xmm0,%xmm12
   DB  102,15,239,199                      ; pxor          %xmm7,%xmm0
@@ -38916,7 +39783,7 @@
   DB  102,69,15,56,62,225                 ; pmaxuw        %xmm9,%xmm12
   DB  102,69,15,56,16,229                 ; pblendvb      %xmm0,%xmm13,%xmm12
   DB  102,69,15,213,200                   ; pmullw        %xmm8,%xmm9
-  DB  102,15,111,5,113,57,0,0             ; movdqa        0x3971(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,5,153,65,0,0             ; movdqa        0x4199(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,253,200                   ; paddw         %xmm0,%xmm9
   DB  102,65,15,113,209,8                 ; psrlw         $0x8,%xmm9
   DB  102,68,15,213,209                   ; pmullw        %xmm1,%xmm10
@@ -38968,23 +39835,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,35                              ; ja            28b12 <_sk_lerp_565_sse41_lowp+0x42>
+  DB  119,35                              ; ja            299ca <_sk_lerp_565_sse41_lowp+0x42>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,187,1,0,0                 ; lea           0x1bb(%rip),%r9        # 28cb8 <_sk_lerp_565_sse41_lowp+0x1e8>
+  DB  76,141,13,187,1,0,0                 ; lea           0x1bb(%rip),%r9        # 29b70 <_sk_lerp_565_sse41_lowp+0x1e8>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,68,15,110,200                   ; movd          %eax,%xmm9
-  DB  235,76                              ; jmp           28b5e <_sk_lerp_565_sse41_lowp+0x8e>
+  DB  235,76                              ; jmp           29a16 <_sk_lerp_565_sse41_lowp+0x8e>
   DB  243,69,15,111,12,80                 ; movdqu        (%r8,%rdx,2),%xmm9
-  DB  235,68                              ; jmp           28b5e <_sk_lerp_565_sse41_lowp+0x8e>
+  DB  235,68                              ; jmp           29a16 <_sk_lerp_565_sse41_lowp+0x8e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,4,2             ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   DB  102,69,15,110,12,80                 ; movd          (%r8,%rdx,2),%xmm9
   DB  102,68,15,58,14,200,252             ; pblendw       $0xfc,%xmm0,%xmm9
-  DB  235,41                              ; jmp           28b5e <_sk_lerp_565_sse41_lowp+0x8e>
+  DB  235,41                              ; jmp           29a16 <_sk_lerp_565_sse41_lowp+0x8e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,12,6            ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   DB  102,65,15,196,68,80,10,5            ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -38993,11 +39860,11 @@
   DB  102,68,15,58,14,200,240             ; pblendw       $0xf0,%xmm0,%xmm9
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
-  DB  102,15,219,5,16,64,0,0              ; pand          0x4010(%rip),%xmm0        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,5,56,72,0,0              ; pand          0x4838(%rip),%xmm0        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,69,15,111,209                   ; movdqa        %xmm9,%xmm10
   DB  102,65,15,113,210,5                 ; psrlw         $0x5,%xmm10
-  DB  102,68,15,219,21,12,64,0,0          ; pand          0x400c(%rip),%xmm10        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,68,15,111,29,19,64,0,0          ; movdqa        0x4013(%rip),%xmm11        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,68,15,219,21,52,72,0,0          ; pand          0x4834(%rip),%xmm10        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,68,15,111,29,59,72,0,0          ; movdqa        0x483b(%rip),%xmm11        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,69,15,219,217                   ; pand          %xmm9,%xmm11
   DB  102,65,15,113,209,13                ; psrlw         $0xd,%xmm9
   DB  102,68,15,235,200                   ; por           %xmm0,%xmm9
@@ -39009,7 +39876,7 @@
   DB  102,15,113,240,3                    ; psllw         $0x3,%xmm0
   DB  102,65,15,113,211,2                 ; psrlw         $0x2,%xmm11
   DB  102,68,15,235,216                   ; por           %xmm0,%xmm11
-  DB  102,15,111,5,113,63,0,0             ; movdqa        0x3f71(%rip),%xmm0        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,15,111,5,153,71,0,0             ; movdqa        0x4799(%rip),%xmm0        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,68,15,111,227                   ; movdqa        %xmm3,%xmm12
   DB  102,68,15,239,224                   ; pxor          %xmm0,%xmm12
   DB  102,15,239,199                      ; pxor          %xmm7,%xmm0
@@ -39021,7 +39888,7 @@
   DB  102,69,15,56,62,227                 ; pmaxuw        %xmm11,%xmm12
   DB  102,69,15,56,62,225                 ; pmaxuw        %xmm9,%xmm12
   DB  102,69,15,56,16,229                 ; pblendvb      %xmm0,%xmm13,%xmm12
-  DB  102,68,15,111,45,189,55,0,0         ; movdqa        0x37bd(%rip),%xmm13        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,45,229,63,0,0         ; movdqa        0x3fe5(%rip),%xmm13        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,111,193                   ; movdqa        %xmm9,%xmm0
   DB  102,65,15,239,197                   ; pxor          %xmm13,%xmm0
   DB  102,15,213,196                      ; pmullw        %xmm4,%xmm0
@@ -39083,48 +39950,48 @@
   DB  65,254,202                          ; dec           %r10b
   DB  69,15,182,202                       ; movzbl        %r10b,%r9d
   DB  65,128,249,6                        ; cmp           $0x6,%r9b
-  DB  119,38                              ; ja            28d1f <_sk_srcover_rgba_8888_sse41_lowp+0x4b>
+  DB  119,38                              ; ja            29bd7 <_sk_srcover_rgba_8888_sse41_lowp+0x4b>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  76,141,29,23,2,0,0                  ; lea           0x217(%rip),%r11        # 28f1c <_sk_srcover_rgba_8888_sse41_lowp+0x248>
+  DB  76,141,29,23,2,0,0                  ; lea           0x217(%rip),%r11        # 29dd4 <_sk_srcover_rgba_8888_sse41_lowp+0x248>
   DB  75,99,4,139                         ; movslq        (%r11,%r9,4),%rax
   DB  76,1,216                            ; add           %r11,%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,110,52,144                ; movd          (%r8,%rdx,4),%xmm6
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,89                              ; jmp           28d78 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  DB  235,89                              ; jmp           29c30 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   DB  243,65,15,111,52,144                ; movdqu        (%r8,%rdx,4),%xmm6
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,74                              ; jmp           28d78 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  DB  235,74                              ; jmp           29c30 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  243,65,15,126,52,144                ; movq          (%r8,%rdx,4),%xmm6
   DB  102,65,15,58,14,240,240             ; pblendw       $0xf0,%xmm8,%xmm6
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
-  DB  235,37                              ; jmp           28d78 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
+  DB  235,37                              ; jmp           29c30 <_sk_srcover_rgba_8888_sse41_lowp+0xa4>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,69,15,58,34,68,144,20,1         ; pinsrd        $0x1,0x14(%r8,%rdx,4),%xmm8
   DB  102,69,15,58,34,68,144,16,0         ; pinsrd        $0x0,0x10(%r8,%rdx,4),%xmm8
   DB  243,65,15,111,52,144                ; movdqu        (%r8,%rdx,4),%xmm6
-  DB  102,15,111,37,208,61,0,0            ; movdqa        0x3dd0(%rip),%xmm4        # 2cb50 <_sk_srcover_rgba_8888_sse2_lowp+0x1148>
+  DB  102,15,111,37,248,69,0,0            ; movdqa        0x45f8(%rip),%xmm4        # 2e230 <_sk_srcover_rgba_8888_sse2_lowp+0x117c>
   DB  102,15,111,238                      ; movdqa        %xmm6,%xmm5
   DB  102,15,56,0,236                     ; pshufb        %xmm4,%xmm5
   DB  102,65,15,111,248                   ; movdqa        %xmm8,%xmm7
   DB  102,15,56,0,252                     ; pshufb        %xmm4,%xmm7
   DB  102,15,108,239                      ; punpcklqdq    %xmm7,%xmm5
-  DB  102,68,15,111,37,48,54,0,0          ; movdqa        0x3630(%rip),%xmm12        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,37,88,62,0,0          ; movdqa        0x3e58(%rip),%xmm12        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,229                      ; movdqa        %xmm5,%xmm4
   DB  102,65,15,219,228                   ; pand          %xmm12,%xmm4
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
-  DB  102,68,15,111,13,169,61,0,0         ; movdqa        0x3da9(%rip),%xmm9        # 2cb60 <_sk_srcover_rgba_8888_sse2_lowp+0x1158>
+  DB  102,68,15,111,13,209,69,0,0         ; movdqa        0x45d1(%rip),%xmm9        # 2e240 <_sk_srcover_rgba_8888_sse2_lowp+0x118c>
   DB  102,15,111,254                      ; movdqa        %xmm6,%xmm7
   DB  102,65,15,56,0,249                  ; pshufb        %xmm9,%xmm7
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
   DB  102,69,15,56,0,209                  ; pshufb        %xmm9,%xmm10
   DB  102,65,15,108,250                   ; punpcklqdq    %xmm10,%xmm7
-  DB  102,68,15,111,13,150,61,0,0         ; movdqa        0x3d96(%rip),%xmm9        # 2cb70 <_sk_srcover_rgba_8888_sse2_lowp+0x1168>
+  DB  102,68,15,111,13,190,69,0,0         ; movdqa        0x45be(%rip),%xmm9        # 2e250 <_sk_srcover_rgba_8888_sse2_lowp+0x119c>
   DB  102,65,15,56,0,241                  ; pshufb        %xmm9,%xmm6
   DB  102,69,15,56,0,193                  ; pshufb        %xmm9,%xmm8
   DB  102,65,15,108,240                   ; punpcklqdq    %xmm8,%xmm6
@@ -39166,19 +40033,19 @@
   DB  102,15,97,195                       ; punpcklwd     %xmm3,%xmm0
   DB  102,65,15,235,196                   ; por           %xmm12,%xmm0
   DB  65,128,250,6                        ; cmp           $0x6,%r10b
-  DB  119,24                              ; ja            28ec2 <_sk_srcover_rgba_8888_sse41_lowp+0x1ee>
-  DB  76,141,21,135,0,0,0                 ; lea           0x87(%rip),%r10        # 28f38 <_sk_srcover_rgba_8888_sse41_lowp+0x264>
+  DB  119,24                              ; ja            29d7a <_sk_srcover_rgba_8888_sse41_lowp+0x1ee>
+  DB  76,141,21,135,0,0,0                 ; lea           0x87(%rip),%r10        # 29df0 <_sk_srcover_rgba_8888_sse41_lowp+0x264>
   DB  75,99,4,138                         ; movslq        (%r10,%r9,4),%rax
   DB  76,1,208                            ; add           %r10,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,65,15,126,4,144                 ; movd          %xmm0,(%r8,%rdx,4)
-  DB  235,63                              ; jmp           28f01 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  DB  235,63                              ; jmp           29db9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   DB  243,65,15,127,4,144                 ; movdqu        %xmm0,(%r8,%rdx,4)
   DB  243,65,15,127,76,144,16             ; movdqu        %xmm1,0x10(%r8,%rdx,4)
-  DB  235,48                              ; jmp           28f01 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  DB  235,48                              ; jmp           29db9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   DB  102,65,15,58,22,68,144,8,2          ; pextrd        $0x2,%xmm0,0x8(%r8,%rdx,4)
   DB  102,65,15,214,4,144                 ; movq          %xmm0,(%r8,%rdx,4)
-  DB  235,31                              ; jmp           28f01 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
+  DB  235,31                              ; jmp           29db9 <_sk_srcover_rgba_8888_sse41_lowp+0x22d>
   DB  102,65,15,58,22,76,144,24,2         ; pextrd        $0x2,%xmm1,0x18(%r8,%rdx,4)
   DB  102,65,15,58,22,76,144,20,1         ; pextrd        $0x1,%xmm1,0x14(%r8,%rdx,4)
   DB  102,65,15,126,76,144,16             ; movd          %xmm1,0x10(%r8,%rdx,4)
@@ -39253,13 +40120,13 @@
   DB  73,137,246                          ; mov           %rsi,%r14
   DB  76,137,141,8,255,255,255            ; mov           %r9,-0xf8(%rbp)
   DB  76,57,203                           ; cmp           %r9,%rbx
-  DB  15,131,153,0,0,0                    ; jae           2906e <_sk_start_pipeline_sse2_lowp+0x11a>
+  DB  15,131,153,0,0,0                    ; jae           29f26 <_sk_start_pipeline_sse2_lowp+0x11a>
   DB  72,139,133,24,255,255,255           ; mov           -0xe8(%rbp),%rax
   DB  72,141,64,8                         ; lea           0x8(%rax),%rax
   DB  72,137,133,16,255,255,255           ; mov           %rax,-0xf0(%rbp)
   DB  76,57,189,16,255,255,255            ; cmp           %r15,-0xf0(%rbp)
   DB  72,139,149,24,255,255,255           ; mov           -0xe8(%rbp),%rdx
-  DB  119,62                              ; ja            29035 <_sk_start_pipeline_sse2_lowp+0xe1>
+  DB  119,62                              ; ja            29eed <_sk_start_pipeline_sse2_lowp+0xe1>
   DB  76,139,165,24,255,255,255           ; mov           -0xe8(%rbp),%r12
   DB  49,255                              ; xor           %edi,%edi
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -39278,10 +40145,10 @@
   DB  73,131,196,16                       ; add           $0x10,%r12
   DB  77,57,252                           ; cmp           %r15,%r12
   DB  73,137,212                          ; mov           %rdx,%r12
-  DB  118,201                             ; jbe           28ffe <_sk_start_pipeline_sse2_lowp+0xaa>
+  DB  118,201                             ; jbe           29eb6 <_sk_start_pipeline_sse2_lowp+0xaa>
   DB  76,137,255                          ; mov           %r15,%rdi
   DB  72,41,215                           ; sub           %rdx,%rdi
-  DB  116,33                              ; je            2905e <_sk_start_pipeline_sse2_lowp+0x10a>
+  DB  116,33                              ; je            29f16 <_sk_start_pipeline_sse2_lowp+0x10a>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -39295,7 +40162,7 @@
   DB  65,255,213                          ; callq         *%r13
   DB  72,255,195                          ; inc           %rbx
   DB  72,59,157,8,255,255,255             ; cmp           -0xf8(%rbp),%rbx
-  DB  15,133,121,255,255,255              ; jne           28fe7 <_sk_start_pipeline_sse2_lowp+0x93>
+  DB  15,133,121,255,255,255              ; jne           29e9f <_sk_start_pipeline_sse2_lowp+0x93>
   DB  15,40,181,32,255,255,255            ; movaps        -0xe0(%rbp),%xmm6
   DB  15,40,189,48,255,255,255            ; movaps        -0xd0(%rbp),%xmm7
   DB  68,15,40,133,64,255,255,255         ; movaps        -0xc0(%rbp),%xmm8
@@ -39334,11 +40201,47 @@
   DB  102,15,110,209                      ; movd          %ecx,%xmm2
   DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,88,21,45,51,0,0                  ; addps         0x332d(%rip),%xmm2        # 2c420 <_sk_srcover_rgba_8888_sse2_lowp+0xa18>
+  DB  15,88,21,85,59,0,0                  ; addps         0x3b55(%rip),%xmm2        # 2db00 <_sk_srcover_rgba_8888_sse2_lowp+0xa4c>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_translate_sse2_lowp
+_sk_matrix_translate_sse2_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
+  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  65,15,88,192                        ; addps         %xmm8,%xmm0
+  DB  65,15,88,200                        ; addps         %xmm8,%xmm1
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  65,15,88,209                        ; addps         %xmm9,%xmm2
+  DB  65,15,88,217                        ; addps         %xmm9,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
+PUBLIC _sk_matrix_scale_translate_sse2_lowp
+_sk_matrix_scale_translate_sse2_lowp LABEL PROC
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
+  DB  243,68,15,16,72,4                   ; movss         0x4(%rax),%xmm9
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  243,68,15,16,80,8                   ; movss         0x8(%rax),%xmm10
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
+  DB  65,15,89,192                        ; mulps         %xmm8,%xmm0
+  DB  65,15,88,194                        ; addps         %xmm10,%xmm0
+  DB  65,15,88,202                        ; addps         %xmm10,%xmm1
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  65,15,89,217                        ; mulps         %xmm9,%xmm3
+  DB  65,15,89,209                        ; mulps         %xmm9,%xmm2
+  DB  65,15,88,208                        ; addps         %xmm8,%xmm2
+  DB  65,15,88,216                        ; addps         %xmm8,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_matrix_2x3_sse2_lowp
 _sk_matrix_2x3_sse2_lowp LABEL PROC
   DB  68,15,40,192                        ; movaps        %xmm0,%xmm8
@@ -39378,6 +40281,68 @@
   DB  65,15,40,218                        ; movaps        %xmm10,%xmm3
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_matrix_perspective_sse2_lowp
+_sk_matrix_perspective_sse2_lowp LABEL PROC
+  DB  68,15,40,193                        ; movaps        %xmm1,%xmm8
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,15,16,8                         ; movss         (%rax),%xmm1
+  DB  243,68,15,16,80,4                   ; movss         0x4(%rax),%xmm10
+  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  243,68,15,16,72,8                   ; movss         0x8(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
+  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
+  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
+  DB  69,15,88,209                        ; addps         %xmm9,%xmm10
+  DB  69,15,88,217                        ; addps         %xmm9,%xmm11
+  DB  68,15,40,200                        ; movaps        %xmm0,%xmm9
+  DB  68,15,89,201                        ; mulps         %xmm1,%xmm9
+  DB  69,15,88,202                        ; addps         %xmm10,%xmm9
+  DB  65,15,89,200                        ; mulps         %xmm8,%xmm1
+  DB  65,15,88,203                        ; addps         %xmm11,%xmm1
+  DB  243,68,15,16,96,12                  ; movss         0xc(%rax),%xmm12
+  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
+  DB  243,68,15,16,80,16                  ; movss         0x10(%rax),%xmm10
+  DB  69,15,198,210,0                     ; shufps        $0x0,%xmm10,%xmm10
+  DB  243,68,15,16,104,20                 ; movss         0x14(%rax),%xmm13
+  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
+  DB  68,15,40,219                        ; movaps        %xmm3,%xmm11
+  DB  69,15,89,218                        ; mulps         %xmm10,%xmm11
+  DB  68,15,89,210                        ; mulps         %xmm2,%xmm10
+  DB  69,15,88,213                        ; addps         %xmm13,%xmm10
+  DB  69,15,88,221                        ; addps         %xmm13,%xmm11
+  DB  68,15,40,232                        ; movaps        %xmm0,%xmm13
+  DB  69,15,89,236                        ; mulps         %xmm12,%xmm13
+  DB  69,15,88,213                        ; addps         %xmm13,%xmm10
+  DB  69,15,89,224                        ; mulps         %xmm8,%xmm12
+  DB  69,15,88,220                        ; addps         %xmm12,%xmm11
+  DB  243,68,15,16,96,24                  ; movss         0x18(%rax),%xmm12
+  DB  69,15,198,228,0                     ; shufps        $0x0,%xmm12,%xmm12
+  DB  243,68,15,16,104,28                 ; movss         0x1c(%rax),%xmm13
+  DB  69,15,198,237,0                     ; shufps        $0x0,%xmm13,%xmm13
+  DB  243,68,15,16,112,32                 ; movss         0x20(%rax),%xmm14
+  DB  69,15,198,246,0                     ; shufps        $0x0,%xmm14,%xmm14
+  DB  65,15,89,221                        ; mulps         %xmm13,%xmm3
+  DB  68,15,89,234                        ; mulps         %xmm2,%xmm13
+  DB  69,15,88,238                        ; addps         %xmm14,%xmm13
+  DB  65,15,88,222                        ; addps         %xmm14,%xmm3
+  DB  65,15,89,196                        ; mulps         %xmm12,%xmm0
+  DB  65,15,88,197                        ; addps         %xmm13,%xmm0
+  DB  69,15,89,224                        ; mulps         %xmm8,%xmm12
+  DB  68,15,88,227                        ; addps         %xmm3,%xmm12
+  DB  15,83,192                           ; rcpps         %xmm0,%xmm0
+  DB  65,15,83,212                        ; rcpps         %xmm12,%xmm2
+  DB  68,15,89,200                        ; mulps         %xmm0,%xmm9
+  DB  15,89,202                           ; mulps         %xmm2,%xmm1
+  DB  68,15,89,208                        ; mulps         %xmm0,%xmm10
+  DB  68,15,89,218                        ; mulps         %xmm2,%xmm11
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  65,15,40,193                        ; movaps        %xmm9,%xmm0
+  DB  65,15,40,210                        ; movaps        %xmm10,%xmm2
+  DB  65,15,40,219                        ; movaps        %xmm11,%xmm3
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_uniform_color_sse2_lowp
 _sk_uniform_color_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -39401,7 +40366,7 @@
 PUBLIC _sk_black_color_sse2_lowp
 _sk_black_color_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,234,49,0,0                 ; movaps        0x31ea(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,154,56,0,0                 ; movaps        0x389a(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -39410,7 +40375,7 @@
 PUBLIC _sk_white_color_sse2_lowp
 _sk_white_color_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,5,214,49,0,0                  ; movaps        0x31d6(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,5,134,56,0,0                  ; movaps        0x3886(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -39419,10 +40384,10 @@
 PUBLIC _sk_set_rgb_sse2_lowp
 _sk_set_rgb_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  243,15,16,21,213,42,0,0             ; movss         0x2ad5(%rip),%xmm2        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,15,16,21,81,49,0,0              ; movss         0x3151(%rip),%xmm2        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
   DB  243,15,16,0                         ; movss         (%rax),%xmm0
   DB  243,15,89,194                       ; mulss         %xmm2,%xmm0
-  DB  243,68,15,16,5,148,42,0,0           ; movss         0x2a94(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,16,5,16,49,0,0            ; movss         0x3110(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,88,192                    ; addss         %xmm8,%xmm0
   DB  243,68,15,44,192                    ; cvttss2si     %xmm0,%r8d
   DB  102,65,15,110,192                   ; movd          %r8d,%xmm0
@@ -39447,7 +40412,7 @@
 PUBLIC _sk_premul_sse2_lowp
 _sk_premul_sse2_lowp LABEL PROC
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,68,15,111,5,72,49,0,0           ; movdqa        0x3148(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,248,55,0,0          ; movdqa        0x37f8(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
@@ -39487,7 +40452,7 @@
 
 PUBLIC _sk_invert_sse2_lowp
 _sk_invert_sse2_lowp LABEL PROC
-  DB  102,68,15,111,5,230,48,0,0          ; movdqa        0x30e6(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,150,55,0,0          ; movdqa        0x3796(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,200                   ; psubw         %xmm0,%xmm9
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -39515,7 +40480,7 @@
 _sk_srcatop_sse2_lowp LABEL PROC
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,15,111,29,138,48,0,0            ; movdqa        0x308a(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,29,58,55,0,0             ; movdqa        0x373a(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,111,203                   ; movdqa        %xmm3,%xmm9
   DB  102,69,15,249,200                   ; psubw         %xmm8,%xmm9
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
@@ -39545,7 +40510,7 @@
 _sk_dstatop_sse2_lowp LABEL PROC
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
   DB  102,68,15,213,195                   ; pmullw        %xmm3,%xmm8
-  DB  102,68,15,111,13,10,48,0,0          ; movdqa        0x300a(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,186,54,0,0         ; movdqa        0x36ba(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,253,193                   ; paddw         %xmm9,%xmm8
   DB  102,68,15,111,213                   ; movdqa        %xmm5,%xmm10
   DB  102,68,15,213,211                   ; pmullw        %xmm3,%xmm10
@@ -39572,7 +40537,7 @@
 PUBLIC _sk_srcin_sse2_lowp
 _sk_srcin_sse2_lowp LABEL PROC
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,68,15,111,5,149,47,0,0          ; movdqa        0x2f95(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,69,54,0,0           ; movdqa        0x3645(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,207                      ; pmullw        %xmm7,%xmm1
@@ -39591,7 +40556,7 @@
 _sk_dstin_sse2_lowp LABEL PROC
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,68,15,111,5,76,47,0,0           ; movdqa        0x2f4c(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,252,53,0,0          ; movdqa        0x35fc(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
@@ -39610,7 +40575,7 @@
 
 PUBLIC _sk_srcout_sse2_lowp
 _sk_srcout_sse2_lowp LABEL PROC
-  DB  102,68,15,111,5,3,47,0,0            ; movdqa        0x2f03(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,179,53,0,0          ; movdqa        0x35b3(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,207                   ; psubw         %xmm7,%xmm9
   DB  102,65,15,213,193                   ; pmullw        %xmm9,%xmm0
@@ -39631,7 +40596,7 @@
 PUBLIC _sk_dstout_sse2_lowp
 _sk_dstout_sse2_lowp LABEL PROC
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  102,68,15,111,5,172,46,0,0          ; movdqa        0x2eac(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,92,53,0,0           ; movdqa        0x355c(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  102,15,249,216                      ; psubw         %xmm0,%xmm3
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
@@ -39654,7 +40619,7 @@
 
 PUBLIC _sk_srcover_sse2_lowp
 _sk_srcover_sse2_lowp LABEL PROC
-  DB  102,68,15,111,13,82,46,0,0          ; movdqa        0x2e52(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,2,53,0,0           ; movdqa        0x3502(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,68,15,249,195                   ; psubw         %xmm3,%xmm8
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -39681,7 +40646,7 @@
 
 PUBLIC _sk_dstover_sse2_lowp
 _sk_dstover_sse2_lowp LABEL PROC
-  DB  102,68,15,111,5,216,45,0,0          ; movdqa        0x2dd8(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,136,52,0,0          ; movdqa        0x3488(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,68,15,249,207                   ; psubw         %xmm7,%xmm9
   DB  102,65,15,213,193                   ; pmullw        %xmm9,%xmm0
@@ -39706,7 +40671,7 @@
 PUBLIC _sk_modulate_sse2_lowp
 _sk_modulate_sse2_lowp LABEL PROC
   DB  102,15,213,196                      ; pmullw        %xmm4,%xmm0
-  DB  102,68,15,111,5,113,45,0,0          ; movdqa        0x2d71(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,33,52,0,0           ; movdqa        0x3421(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,192                   ; paddw         %xmm8,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,205                      ; pmullw        %xmm5,%xmm1
@@ -39723,7 +40688,7 @@
 
 PUBLIC _sk_multiply_sse2_lowp
 _sk_multiply_sse2_lowp LABEL PROC
-  DB  102,68,15,111,13,48,45,0,0          ; movdqa        0x2d30(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,224,51,0,0         ; movdqa        0x33e0(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,193                   ; movdqa        %xmm9,%xmm8
   DB  102,68,15,249,195                   ; psubw         %xmm3,%xmm8
   DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
@@ -39766,12 +40731,12 @@
   DB  102,68,15,111,193                   ; movdqa        %xmm1,%xmm8
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,253,204                      ; paddw         %xmm4,%xmm1
-  DB  102,68,15,111,21,156,52,0,0         ; movdqa        0x349c(%rip),%xmm10        # 2cc00 <_sk_srcover_rgba_8888_sse2_lowp+0x11f8>
+  DB  102,68,15,111,21,76,59,0,0          ; movdqa        0x3b4c(%rip),%xmm10        # 2e2e0 <_sk_srcover_rgba_8888_sse2_lowp+0x122c>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,217,194                   ; psubusw       %xmm10,%xmm0
   DB  102,69,15,239,228                   ; pxor          %xmm12,%xmm12
   DB  102,65,15,117,196                   ; pcmpeqw       %xmm12,%xmm0
-  DB  102,68,15,111,29,80,44,0,0          ; movdqa        0x2c50(%rip),%xmm11        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,29,0,51,0,0           ; movdqa        0x3300(%rip),%xmm11        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
   DB  102,65,15,223,195                   ; pandn         %xmm11,%xmm0
   DB  102,15,235,193                      ; por           %xmm1,%xmm0
@@ -39808,7 +40773,7 @@
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,204                      ; pmullw        %xmm4,%xmm1
-  DB  102,68,15,111,21,176,43,0,0         ; movdqa        0x2bb0(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,21,96,50,0,0          ; movdqa        0x3260(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,202                   ; paddw         %xmm10,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -39838,7 +40803,7 @@
 PUBLIC _sk_xor__sse2_lowp
 _sk_xor__sse2_lowp LABEL PROC
   DB  102,68,15,111,195                   ; movdqa        %xmm3,%xmm8
-  DB  102,68,15,111,13,46,43,0,0          ; movdqa        0x2b2e(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,222,49,0,0         ; movdqa        0x31de(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,209                   ; movdqa        %xmm9,%xmm10
   DB  102,68,15,249,215                   ; psubw         %xmm7,%xmm10
   DB  102,65,15,213,194                   ; pmullw        %xmm10,%xmm0
@@ -39878,7 +40843,7 @@
   DB  102,65,15,253,194                   ; paddw         %xmm10,%xmm0
   DB  102,68,15,213,215                   ; pmullw        %xmm7,%xmm10
   DB  102,68,15,213,227                   ; pmullw        %xmm3,%xmm12
-  DB  102,68,15,111,13,233,49,0,0         ; movdqa        0x31e9(%rip),%xmm9        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,13,153,56,0,0         ; movdqa        0x3899(%rip),%xmm9        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,69,15,111,218                   ; movdqa        %xmm10,%xmm11
   DB  102,69,15,239,217                   ; pxor          %xmm9,%xmm11
   DB  102,65,15,111,204                   ; movdqa        %xmm12,%xmm1
@@ -39887,7 +40852,7 @@
   DB  102,68,15,219,225                   ; pand          %xmm1,%xmm12
   DB  102,65,15,223,202                   ; pandn         %xmm10,%xmm1
   DB  102,65,15,235,204                   ; por           %xmm12,%xmm1
-  DB  102,68,15,111,21,72,42,0,0          ; movdqa        0x2a48(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,21,248,48,0,0         ; movdqa        0x30f8(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,202                   ; paddw         %xmm10,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -39941,7 +40906,7 @@
   DB  102,65,15,253,196                   ; paddw         %xmm12,%xmm0
   DB  102,68,15,213,231                   ; pmullw        %xmm7,%xmm12
   DB  102,68,15,213,211                   ; pmullw        %xmm3,%xmm10
-  DB  102,68,15,111,13,192,48,0,0         ; movdqa        0x30c0(%rip),%xmm9        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,13,112,55,0,0         ; movdqa        0x3770(%rip),%xmm9        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,69,15,111,220                   ; movdqa        %xmm12,%xmm11
   DB  102,69,15,239,217                   ; pxor          %xmm9,%xmm11
   DB  102,65,15,111,202                   ; movdqa        %xmm10,%xmm1
@@ -39950,7 +40915,7 @@
   DB  102,68,15,219,225                   ; pand          %xmm1,%xmm12
   DB  102,65,15,223,202                   ; pandn         %xmm10,%xmm1
   DB  102,65,15,235,204                   ; por           %xmm12,%xmm1
-  DB  102,68,15,111,21,31,41,0,0          ; movdqa        0x291f(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,21,207,47,0,0         ; movdqa        0x2fcf(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,202                   ; paddw         %xmm10,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -40004,7 +40969,7 @@
   DB  102,65,15,253,196                   ; paddw         %xmm12,%xmm0
   DB  102,68,15,213,231                   ; pmullw        %xmm7,%xmm12
   DB  102,68,15,213,211                   ; pmullw        %xmm3,%xmm10
-  DB  102,68,15,111,13,151,47,0,0         ; movdqa        0x2f97(%rip),%xmm9        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,13,71,54,0,0          ; movdqa        0x3647(%rip),%xmm9        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,69,15,111,220                   ; movdqa        %xmm12,%xmm11
   DB  102,69,15,239,217                   ; pxor          %xmm9,%xmm11
   DB  102,65,15,111,202                   ; movdqa        %xmm10,%xmm1
@@ -40013,10 +40978,10 @@
   DB  102,68,15,219,225                   ; pand          %xmm1,%xmm12
   DB  102,65,15,223,202                   ; pandn         %xmm10,%xmm1
   DB  102,65,15,235,204                   ; por           %xmm12,%xmm1
-  DB  102,68,15,111,21,246,39,0,0         ; movdqa        0x27f6(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,21,166,46,0,0         ; movdqa        0x2ea6(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,202                   ; paddw         %xmm10,%xmm1
   DB  102,15,113,209,7                    ; psrlw         $0x7,%xmm1
-  DB  102,68,15,111,29,67,47,0,0          ; movdqa        0x2f43(%rip),%xmm11        # 2cb30 <_sk_srcover_rgba_8888_sse2_lowp+0x1128>
+  DB  102,68,15,111,29,243,53,0,0         ; movdqa        0x35f3(%rip),%xmm11        # 2e210 <_sk_srcover_rgba_8888_sse2_lowp+0x115c>
   DB  102,65,15,219,203                   ; pand          %xmm11,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
   DB  102,68,15,111,229                   ; movdqa        %xmm5,%xmm12
@@ -40069,10 +41034,10 @@
   DB  102,15,111,196                      ; movdqa        %xmm4,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,204                      ; pmullw        %xmm4,%xmm1
-  DB  102,68,15,111,13,243,38,0,0         ; movdqa        0x26f3(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,163,45,0,0         ; movdqa        0x2da3(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,201                   ; paddw         %xmm9,%xmm1
   DB  102,15,113,209,7                    ; psrlw         $0x7,%xmm1
-  DB  102,68,15,111,21,64,46,0,0          ; movdqa        0x2e40(%rip),%xmm10        # 2cb30 <_sk_srcover_rgba_8888_sse2_lowp+0x1128>
+  DB  102,68,15,111,21,240,52,0,0         ; movdqa        0x34f0(%rip),%xmm10        # 2e210 <_sk_srcover_rgba_8888_sse2_lowp+0x115c>
   DB  102,65,15,219,202                   ; pand          %xmm10,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
@@ -40107,7 +41072,7 @@
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,68,15,111,194                   ; movdqa        %xmm2,%xmm8
   DB  102,68,15,111,200                   ; movdqa        %xmm0,%xmm9
-  DB  102,15,111,21,76,38,0,0             ; movdqa        0x264c(%rip),%xmm2        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,21,252,44,0,0            ; movdqa        0x2cfc(%rip),%xmm2        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,111,226                   ; movdqa        %xmm2,%xmm12
   DB  102,68,15,249,231                   ; psubw         %xmm7,%xmm12
   DB  102,65,15,111,196                   ; movdqa        %xmm12,%xmm0
@@ -40120,7 +41085,7 @@
   DB  102,68,15,111,243                   ; movdqa        %xmm3,%xmm14
   DB  102,69,15,249,241                   ; psubw         %xmm9,%xmm14
   DB  102,69,15,253,201                   ; paddw         %xmm9,%xmm9
-  DB  102,15,111,21,120,45,0,0            ; movdqa        0x2d78(%rip),%xmm2        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,15,111,21,40,52,0,0             ; movdqa        0x3428(%rip),%xmm2        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,15,111,199                      ; movdqa        %xmm7,%xmm0
   DB  102,68,15,111,215                   ; movdqa        %xmm7,%xmm10
   DB  102,15,111,231                      ; movdqa        %xmm7,%xmm4
@@ -40139,7 +41104,7 @@
   DB  102,68,15,219,232                   ; pand          %xmm0,%xmm13
   DB  102,65,15,223,193                   ; pandn         %xmm9,%xmm0
   DB  102,65,15,235,197                   ; por           %xmm13,%xmm0
-  DB  102,15,253,5,171,37,0,0             ; paddw         0x25ab(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,253,5,91,44,0,0              ; paddw         0x2c5b(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,199                   ; paddw         %xmm15,%xmm0
   DB  102,69,15,111,204                   ; movdqa        %xmm12,%xmm9
   DB  102,68,15,213,201                   ; pmullw        %xmm1,%xmm9
@@ -40162,7 +41127,7 @@
   DB  102,65,15,219,249                   ; pand          %xmm9,%xmm7
   DB  102,68,15,223,201                   ; pandn         %xmm1,%xmm9
   DB  102,68,15,235,207                   ; por           %xmm7,%xmm9
-  DB  102,68,15,253,13,54,37,0,0          ; paddw         0x2536(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,253,13,230,43,0,0         ; paddw         0x2be6(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,253,205                   ; paddw         %xmm13,%xmm9
   DB  102,69,15,213,224                   ; pmullw        %xmm8,%xmm12
   DB  102,69,15,111,251                   ; movdqa        %xmm11,%xmm15
@@ -40184,7 +41149,7 @@
   DB  102,68,15,219,210                   ; pand          %xmm2,%xmm10
   DB  102,65,15,223,208                   ; pandn         %xmm8,%xmm2
   DB  102,65,15,235,210                   ; por           %xmm10,%xmm2
-  DB  102,68,15,111,5,199,36,0,0          ; movdqa        0x24c7(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,119,43,0,0          ; movdqa        0x2b77(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,208                   ; paddw         %xmm8,%xmm2
   DB  102,65,15,253,215                   ; paddw         %xmm15,%xmm2
   DB  102,68,15,213,220                   ; pmullw        %xmm4,%xmm11
@@ -40212,7 +41177,7 @@
   DB  102,15,127,124,36,16                ; movdqa        %xmm7,0x10(%rsp)
   DB  102,68,15,111,193                   ; movdqa        %xmm1,%xmm8
   DB  102,68,15,111,248                   ; movdqa        %xmm0,%xmm15
-  DB  102,15,111,53,81,36,0,0             ; movdqa        0x2451(%rip),%xmm6        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,53,1,43,0,0              ; movdqa        0x2b01(%rip),%xmm6        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,111,222                   ; movdqa        %xmm6,%xmm11
   DB  102,69,15,249,222                   ; psubw         %xmm14,%xmm11
   DB  102,65,15,111,203                   ; movdqa        %xmm11,%xmm1
@@ -40234,7 +41199,7 @@
   DB  102,15,253,201                      ; paddw         %xmm1,%xmm1
   DB  102,69,15,111,225                   ; movdqa        %xmm9,%xmm12
   DB  102,68,15,249,225                   ; psubw         %xmm1,%xmm12
-  DB  102,15,111,13,85,43,0,0             ; movdqa        0x2b55(%rip),%xmm1        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,15,111,13,5,50,0,0              ; movdqa        0x3205(%rip),%xmm1        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,69,15,111,238                   ; movdqa        %xmm14,%xmm13
   DB  102,68,15,239,233                   ; pxor          %xmm1,%xmm13
   DB  102,15,239,249                      ; pxor          %xmm1,%xmm7
@@ -40261,7 +41226,7 @@
   DB  102,68,15,213,197                   ; pmullw        %xmm5,%xmm8
   DB  102,15,111,229                      ; movdqa        %xmm5,%xmm4
   DB  102,15,253,228                      ; paddw         %xmm4,%xmm4
-  DB  102,15,111,61,212,42,0,0            ; movdqa        0x2ad4(%rip),%xmm7        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,15,111,61,132,49,0,0            ; movdqa        0x3184(%rip),%xmm7        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,15,239,231                      ; pxor          %xmm7,%xmm4
   DB  102,65,15,101,229                   ; pcmpgtw       %xmm13,%xmm4
   DB  102,69,15,253,192                   ; paddw         %xmm8,%xmm8
@@ -40320,27 +41285,27 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            2a185 <_sk_load_8888_sse2_lowp+0x44>
+  DB  119,41                              ; ja            2b1b5 <_sk_load_8888_sse2_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,25,1,0,0                  ; lea           0x119(%rip),%r9        # 2a284 <_sk_load_8888_sse2_lowp+0x143>
+  DB  76,141,13,25,1,0,0                  ; lea           0x119(%rip),%r9        # 2b2b4 <_sk_load_8888_sse2_lowp+0x143>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  255,224                             ; jmpq          *%rax
   DB  243,65,15,16,28,144                 ; movss         (%r8,%rdx,4),%xmm3
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,102                             ; jmp           2a1eb <_sk_load_8888_sse2_lowp+0xaa>
+  DB  235,102                             ; jmp           2b21b <_sk_load_8888_sse2_lowp+0xaa>
   DB  102,65,15,16,28,144                 ; movupd        (%r8,%rdx,4),%xmm3
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,87                              ; jmp           2a1eb <_sk_load_8888_sse2_lowp+0xaa>
+  DB  235,87                              ; jmp           2b21b <_sk_load_8888_sse2_lowp+0xaa>
   DB  102,65,15,110,68,144,8              ; movd          0x8(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,69,15,18,4,144                  ; movlpd        (%r8,%rdx,4),%xmm8
   DB  102,65,15,40,216                    ; movapd        %xmm8,%xmm3
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
-  DB  235,52                              ; jmp           2a1eb <_sk_load_8888_sse2_lowp+0xaa>
+  DB  235,52                              ; jmp           2b21b <_sk_load_8888_sse2_lowp+0xaa>
   DB  102,65,15,110,68,144,24             ; movd          0x18(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  243,65,15,16,68,144,20              ; movss         0x14(%r8,%rdx,4),%xmm0
@@ -40357,7 +41322,7 @@
   DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
   DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
   DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
-  DB  102,68,15,111,13,187,33,0,0         ; movdqa        0x21bb(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,107,40,0,0         ; movdqa        0x286b(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
@@ -40412,27 +41377,27 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            2a2e4 <_sk_load_8888_dst_sse2_lowp+0x44>
+  DB  119,41                              ; ja            2b314 <_sk_load_8888_dst_sse2_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2a3e0 <_sk_load_8888_dst_sse2_lowp+0x140>
+  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2b410 <_sk_load_8888_dst_sse2_lowp+0x140>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  255,224                             ; jmpq          *%rax
   DB  243,65,15,16,60,144                 ; movss         (%r8,%rdx,4),%xmm7
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,102                             ; jmp           2a34a <_sk_load_8888_dst_sse2_lowp+0xaa>
+  DB  235,102                             ; jmp           2b37a <_sk_load_8888_dst_sse2_lowp+0xaa>
   DB  102,65,15,16,60,144                 ; movupd        (%r8,%rdx,4),%xmm7
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,87                              ; jmp           2a34a <_sk_load_8888_dst_sse2_lowp+0xaa>
+  DB  235,87                              ; jmp           2b37a <_sk_load_8888_dst_sse2_lowp+0xaa>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,69,15,18,4,144                  ; movlpd        (%r8,%rdx,4),%xmm8
   DB  102,65,15,40,248                    ; movapd        %xmm8,%xmm7
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
-  DB  235,52                              ; jmp           2a34a <_sk_load_8888_dst_sse2_lowp+0xaa>
+  DB  235,52                              ; jmp           2b37a <_sk_load_8888_dst_sse2_lowp+0xaa>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  243,65,15,16,100,144,20             ; movss         0x14(%r8,%rdx,4),%xmm4
@@ -40449,7 +41414,7 @@
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
   DB  102,15,114,228,16                   ; psrad         $0x10,%xmm4
   DB  102,15,107,236                      ; packssdw      %xmm4,%xmm5
-  DB  102,68,15,111,13,92,32,0,0          ; movdqa        0x205c(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,12,39,0,0          ; movdqa        0x270c(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,229                      ; movdqa        %xmm5,%xmm4
   DB  102,65,15,219,225                   ; pand          %xmm9,%xmm4
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
@@ -40515,21 +41480,21 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,27                              ; ja            2a47f <_sk_store_8888_sse2_lowp+0x83>
+  DB  119,27                              ; ja            2b4af <_sk_store_8888_sse2_lowp+0x83>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,98,0,0,0                  ; lea           0x62(%rip),%r9        # 2a4d0 <_sk_store_8888_sse2_lowp+0xd4>
+  DB  76,141,13,98,0,0,0                  ; lea           0x62(%rip),%r9        # 2b500 <_sk_store_8888_sse2_lowp+0xd4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,126,12,144                ; movd          %xmm9,(%r8,%rdx,4)
-  DB  235,75                              ; jmp           2a4ca <_sk_store_8888_sse2_lowp+0xce>
+  DB  235,75                              ; jmp           2b4fa <_sk_store_8888_sse2_lowp+0xce>
   DB  243,69,15,127,12,144                ; movdqu        %xmm9,(%r8,%rdx,4)
   DB  243,69,15,127,68,144,16             ; movdqu        %xmm8,0x10(%r8,%rdx,4)
-  DB  235,60                              ; jmp           2a4ca <_sk_store_8888_sse2_lowp+0xce>
+  DB  235,60                              ; jmp           2b4fa <_sk_store_8888_sse2_lowp+0xce>
   DB  102,69,15,112,193,78                ; pshufd        $0x4e,%xmm9,%xmm8
   DB  102,69,15,126,68,144,8              ; movd          %xmm8,0x8(%r8,%rdx,4)
   DB  102,69,15,214,12,144                ; movq          %xmm9,(%r8,%rdx,4)
-  DB  235,39                              ; jmp           2a4ca <_sk_store_8888_sse2_lowp+0xce>
+  DB  235,39                              ; jmp           2b4fa <_sk_store_8888_sse2_lowp+0xce>
   DB  102,69,15,112,208,78                ; pshufd        $0x4e,%xmm8,%xmm10
   DB  102,69,15,126,84,144,24             ; movd          %xmm10,0x18(%r8,%rdx,4)
   DB  102,69,15,112,208,229               ; pshufd        $0xe5,%xmm8,%xmm10
@@ -40572,27 +41537,27 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            2a530 <_sk_load_bgra_sse2_lowp+0x44>
+  DB  119,41                              ; ja            2b560 <_sk_load_bgra_sse2_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2a62c <_sk_load_bgra_sse2_lowp+0x140>
+  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2b65c <_sk_load_bgra_sse2_lowp+0x140>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  255,224                             ; jmpq          *%rax
   DB  243,65,15,16,28,144                 ; movss         (%r8,%rdx,4),%xmm3
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,102                             ; jmp           2a596 <_sk_load_bgra_sse2_lowp+0xaa>
+  DB  235,102                             ; jmp           2b5c6 <_sk_load_bgra_sse2_lowp+0xaa>
   DB  102,65,15,16,28,144                 ; movupd        (%r8,%rdx,4),%xmm3
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,87                              ; jmp           2a596 <_sk_load_bgra_sse2_lowp+0xaa>
+  DB  235,87                              ; jmp           2b5c6 <_sk_load_bgra_sse2_lowp+0xaa>
   DB  102,65,15,110,68,144,8              ; movd          0x8(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,69,15,18,4,144                  ; movlpd        (%r8,%rdx,4),%xmm8
   DB  102,65,15,40,216                    ; movapd        %xmm8,%xmm3
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
-  DB  235,52                              ; jmp           2a596 <_sk_load_bgra_sse2_lowp+0xaa>
+  DB  235,52                              ; jmp           2b5c6 <_sk_load_bgra_sse2_lowp+0xaa>
   DB  102,65,15,110,68,144,24             ; movd          0x18(%r8,%rdx,4),%xmm0
   DB  102,68,15,112,192,69                ; pshufd        $0x45,%xmm0,%xmm8
   DB  243,65,15,16,68,144,20              ; movss         0x14(%r8,%rdx,4),%xmm0
@@ -40609,7 +41574,7 @@
   DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
   DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
   DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
-  DB  102,68,15,111,13,16,30,0,0          ; movdqa        0x1e10(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,192,36,0,0         ; movdqa        0x24c0(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
@@ -40660,27 +41625,27 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,41                              ; ja            2a68c <_sk_load_bgra_dst_sse2_lowp+0x44>
+  DB  119,41                              ; ja            2b6bc <_sk_load_bgra_dst_sse2_lowp+0x44>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2a788 <_sk_load_bgra_dst_sse2_lowp+0x140>
+  DB  76,141,13,22,1,0,0                  ; lea           0x116(%rip),%r9        # 2b7b8 <_sk_load_bgra_dst_sse2_lowp+0x140>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  255,224                             ; jmpq          *%rax
   DB  243,65,15,16,60,144                 ; movss         (%r8,%rdx,4),%xmm7
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,102                             ; jmp           2a6f2 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  DB  235,102                             ; jmp           2b722 <_sk_load_bgra_dst_sse2_lowp+0xaa>
   DB  102,65,15,16,60,144                 ; movupd        (%r8,%rdx,4),%xmm7
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,87                              ; jmp           2a6f2 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  DB  235,87                              ; jmp           2b722 <_sk_load_bgra_dst_sse2_lowp+0xaa>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,69,15,18,4,144                  ; movlpd        (%r8,%rdx,4),%xmm8
   DB  102,65,15,40,248                    ; movapd        %xmm8,%xmm7
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
-  DB  235,52                              ; jmp           2a6f2 <_sk_load_bgra_dst_sse2_lowp+0xaa>
+  DB  235,52                              ; jmp           2b722 <_sk_load_bgra_dst_sse2_lowp+0xaa>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  243,65,15,16,100,144,20             ; movss         0x14(%r8,%rdx,4),%xmm4
@@ -40697,7 +41662,7 @@
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
   DB  102,15,114,228,16                   ; psrad         $0x10,%xmm4
   DB  102,15,107,236                      ; packssdw      %xmm4,%xmm5
-  DB  102,68,15,111,13,180,28,0,0         ; movdqa        0x1cb4(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,100,35,0,0         ; movdqa        0x2364(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,245                      ; movdqa        %xmm5,%xmm6
   DB  102,65,15,219,241                   ; pand          %xmm9,%xmm6
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
@@ -40763,21 +41728,21 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,27                              ; ja            2a827 <_sk_store_bgra_sse2_lowp+0x83>
+  DB  119,27                              ; ja            2b857 <_sk_store_bgra_sse2_lowp+0x83>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,98,0,0,0                  ; lea           0x62(%rip),%r9        # 2a878 <_sk_store_bgra_sse2_lowp+0xd4>
+  DB  76,141,13,98,0,0,0                  ; lea           0x62(%rip),%r9        # 2b8a8 <_sk_store_bgra_sse2_lowp+0xd4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,126,12,144                ; movd          %xmm9,(%r8,%rdx,4)
-  DB  235,75                              ; jmp           2a872 <_sk_store_bgra_sse2_lowp+0xce>
+  DB  235,75                              ; jmp           2b8a2 <_sk_store_bgra_sse2_lowp+0xce>
   DB  243,69,15,127,12,144                ; movdqu        %xmm9,(%r8,%rdx,4)
   DB  243,69,15,127,68,144,16             ; movdqu        %xmm8,0x10(%r8,%rdx,4)
-  DB  235,60                              ; jmp           2a872 <_sk_store_bgra_sse2_lowp+0xce>
+  DB  235,60                              ; jmp           2b8a2 <_sk_store_bgra_sse2_lowp+0xce>
   DB  102,69,15,112,193,78                ; pshufd        $0x4e,%xmm9,%xmm8
   DB  102,69,15,126,68,144,8              ; movd          %xmm8,0x8(%r8,%rdx,4)
   DB  102,69,15,214,12,144                ; movq          %xmm9,(%r8,%rdx,4)
-  DB  235,39                              ; jmp           2a872 <_sk_store_bgra_sse2_lowp+0xce>
+  DB  235,39                              ; jmp           2b8a2 <_sk_store_bgra_sse2_lowp+0xce>
   DB  102,69,15,112,208,78                ; pshufd        $0x4e,%xmm8,%xmm10
   DB  102,69,15,126,84,144,24             ; movd          %xmm10,0x18(%r8,%rdx,4)
   DB  102,69,15,112,208,229               ; pshufd        $0xe5,%xmm8,%xmm10
@@ -40889,7 +41854,7 @@
   DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
   DB  102,15,114,225,16                   ; psrad         $0x10,%xmm1
   DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
-  DB  102,68,15,111,13,208,25,0,0         ; movdqa        0x19d0(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,128,32,0,0         ; movdqa        0x2080(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
@@ -40917,6 +41882,114 @@
   DB  65,95                               ; pop           %r15
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_bgra_sse2_lowp
+_sk_gather_bgra_sse2_lowp LABEL PROC
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  243,68,15,91,195                    ; cvttps2dq     %xmm3,%xmm8
+  DB  243,68,15,91,202                    ; cvttps2dq     %xmm2,%xmm9
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,69,15,112,209,245               ; pshufd        $0xf5,%xmm9,%xmm10
+  DB  102,68,15,244,211                   ; pmuludq       %xmm3,%xmm10
+  DB  102,68,15,111,219                   ; movdqa        %xmm3,%xmm11
+  DB  102,69,15,112,224,245               ; pshufd        $0xf5,%xmm8,%xmm12
+  DB  102,68,15,244,227                   ; pmuludq       %xmm3,%xmm12
+  DB  102,65,15,244,217                   ; pmuludq       %xmm9,%xmm3
+  DB  72,139,0                            ; mov           (%rax),%rax
+  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
+  DB  102,65,15,112,210,232               ; pshufd        $0xe8,%xmm10,%xmm2
+  DB  102,15,98,218                       ; punpckldq     %xmm2,%xmm3
+  DB  102,69,15,244,216                   ; pmuludq       %xmm8,%xmm11
+  DB  102,69,15,112,195,232               ; pshufd        $0xe8,%xmm11,%xmm8
+  DB  102,65,15,112,212,232               ; pshufd        $0xe8,%xmm12,%xmm2
+  DB  102,68,15,98,194                    ; punpckldq     %xmm2,%xmm8
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,200                   ; paddd         %xmm8,%xmm1
+  DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
+  DB  102,73,15,126,208                   ; movq          %xmm2,%r8
+  DB  69,137,193                          ; mov           %r8d,%r9d
+  DB  73,193,232,32                       ; shr           $0x20,%r8
+  DB  102,73,15,126,202                   ; movq          %xmm1,%r10
+  DB  69,137,211                          ; mov           %r10d,%r11d
+  DB  73,193,234,32                       ; shr           $0x20,%r10
+  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
+  DB  102,72,15,126,203                   ; movq          %xmm1,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  102,66,15,110,4,184                 ; movd          (%rax,%r15,4),%xmm0
+  DB  102,66,15,110,28,160                ; movd          (%rax,%r12,4),%xmm3
+  DB  102,15,98,216                       ; punpckldq     %xmm0,%xmm3
+  DB  102,15,110,4,152                    ; movd          (%rax,%rbx,4),%xmm0
+  DB  102,66,15,110,12,176                ; movd          (%rax,%r14,4),%xmm1
+  DB  102,15,98,200                       ; punpckldq     %xmm0,%xmm1
+  DB  102,15,108,217                      ; punpcklqdq    %xmm1,%xmm3
+  DB  102,66,15,110,4,144                 ; movd          (%rax,%r10,4),%xmm0
+  DB  102,70,15,110,4,152                 ; movd          (%rax,%r11,4),%xmm8
+  DB  102,68,15,98,192                    ; punpckldq     %xmm0,%xmm8
+  DB  102,66,15,110,4,128                 ; movd          (%rax,%r8,4),%xmm0
+  DB  102,66,15,110,12,136                ; movd          (%rax,%r9,4),%xmm1
+  DB  102,15,98,200                       ; punpckldq     %xmm0,%xmm1
+  DB  102,68,15,108,193                   ; punpcklqdq    %xmm1,%xmm8
+  DB  102,65,15,111,192                   ; movdqa        %xmm8,%xmm0
+  DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
+  DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
+  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
+  DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
+  DB  102,15,114,225,16                   ; psrad         $0x10,%xmm1
+  DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
+  DB  102,68,15,111,13,161,30,0,0         ; movdqa        0x1ea1(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
+  DB  102,65,15,219,209                   ; pand          %xmm9,%xmm2
+  DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
+  DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
+  DB  102,15,114,208,16                   ; psrld         $0x10,%xmm0
+  DB  102,69,15,111,208                   ; movdqa        %xmm8,%xmm10
+  DB  102,65,15,114,210,16                ; psrld         $0x10,%xmm10
+  DB  102,65,15,114,208,24                ; psrld         $0x18,%xmm8
+  DB  102,65,15,114,240,16                ; pslld         $0x10,%xmm8
+  DB  102,65,15,114,224,16                ; psrad         $0x10,%xmm8
+  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
+  DB  102,15,114,243,16                   ; pslld         $0x10,%xmm3
+  DB  102,15,114,227,16                   ; psrad         $0x10,%xmm3
+  DB  102,65,15,107,216                   ; packssdw      %xmm8,%xmm3
+  DB  102,65,15,114,242,16                ; pslld         $0x10,%xmm10
+  DB  102,65,15,114,226,16                ; psrad         $0x10,%xmm10
+  DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
+  DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
+  DB  102,65,15,107,194                   ; packssdw      %xmm10,%xmm0
+  DB  102,65,15,219,193                   ; pand          %xmm9,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_565_sse2_lowp
 _sk_load_565_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -40928,23 +42001,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2aaaf <_sk_load_565_sse2_lowp+0x3c>
+  DB  119,34                              ; ja            2bcbe <_sk_load_565_sse2_lowp+0x3c>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,189,0,0,0                 ; lea           0xbd(%rip),%r9        # 2ab58 <_sk_load_565_sse2_lowp+0xe5>
+  DB  76,141,13,190,0,0,0                 ; lea           0xbe(%rip),%r9        # 2bd68 <_sk_load_565_sse2_lowp+0xe6>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,192                      ; movd          %eax,%xmm0
-  DB  235,66                              ; jmp           2aaf1 <_sk_load_565_sse2_lowp+0x7e>
+  DB  235,66                              ; jmp           2bd00 <_sk_load_565_sse2_lowp+0x7e>
   DB  243,65,15,111,4,80                  ; movdqu        (%r8,%rdx,2),%xmm0
-  DB  235,58                              ; jmp           2aaf1 <_sk_load_565_sse2_lowp+0x7e>
+  DB  235,58                              ; jmp           2bd00 <_sk_load_565_sse2_lowp+0x7e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,4,2             ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm0
   DB  243,65,15,16,12,80                  ; movss         (%r8,%rdx,2),%xmm1
   DB  243,15,16,193                       ; movss         %xmm1,%xmm0
-  DB  235,34                              ; jmp           2aaf1 <_sk_load_565_sse2_lowp+0x7e>
+  DB  235,34                              ; jmp           2bd00 <_sk_load_565_sse2_lowp+0x7e>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,65,15,196,68,80,12,6            ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm0
   DB  102,65,15,196,68,80,10,5            ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm0
@@ -40952,11 +42025,11 @@
   DB  102,65,15,18,4,80                   ; movlpd        (%r8,%rdx,2),%xmm0
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
   DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
-  DB  102,15,219,29,126,32,0,0            ; pand          0x207e(%rip),%xmm3        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,29,79,37,0,0             ; pand          0x254f(%rip),%xmm3        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
-  DB  102,15,219,13,125,32,0,0            ; pand          0x207d(%rip),%xmm1        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,15,111,21,133,32,0,0            ; movdqa        0x2085(%rip),%xmm2        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,15,219,13,78,37,0,0             ; pand          0x254e(%rip),%xmm1        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,21,86,37,0,0             ; movdqa        0x2556(%rip),%xmm2        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,15,219,208                      ; pand          %xmm0,%xmm2
   DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
   DB  102,15,235,195                      ; por           %xmm3,%xmm0
@@ -40969,19 +42042,19 @@
   DB  102,15,113,210,2                    ; psrlw         $0x2,%xmm2
   DB  102,15,235,211                      ; por           %xmm3,%xmm2
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,123,24,0,0                 ; movaps        0x187b(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,76,29,0,0                  ; movaps        0x1d4c(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  255,224                             ; jmpq          *%rax
-  DB  144                                 ; nop
-  DB  76,255                              ; rex.WR        (bad)
+  DB  102,144                             ; xchg          %ax,%ax
+  DB  75,255                              ; rex.WXB       (bad)
   DB  255                                 ; (bad)
-  DB  255,107,255                         ; ljmp          *-0x1(%rbx)
+  DB  255,106,255                         ; ljmp          *-0x1(%rdx)
   DB  255                                 ; (bad)
-  DB  255,95,255                          ; lcall         *-0x1(%rdi)
+  DB  255,94,255                          ; lcall         *-0x1(%rsi)
   DB  255                                 ; (bad)
-  DB  255,147,255,255,255,139             ; callq         *-0x74000001(%rbx)
+  DB  255,146,255,255,255,138             ; callq         *-0x75000001(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,131,255,255,255,119             ; incl          0x77ffffff(%rbx)
+  DB  255,130,255,255,255,118             ; incl          0x76ffffff(%rdx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
@@ -40997,23 +42070,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2abb0 <_sk_load_565_dst_sse2_lowp+0x3c>
+  DB  119,34                              ; ja            2bdc0 <_sk_load_565_dst_sse2_lowp+0x3c>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,188,0,0,0                 ; lea           0xbc(%rip),%r9        # 2ac58 <_sk_load_565_dst_sse2_lowp+0xe4>
+  DB  76,141,13,188,0,0,0                 ; lea           0xbc(%rip),%r9        # 2be68 <_sk_load_565_dst_sse2_lowp+0xe4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,224                      ; movd          %eax,%xmm4
-  DB  235,66                              ; jmp           2abf2 <_sk_load_565_dst_sse2_lowp+0x7e>
+  DB  235,66                              ; jmp           2be02 <_sk_load_565_dst_sse2_lowp+0x7e>
   DB  243,65,15,111,36,80                 ; movdqu        (%r8,%rdx,2),%xmm4
-  DB  235,58                              ; jmp           2abf2 <_sk_load_565_dst_sse2_lowp+0x7e>
+  DB  235,58                              ; jmp           2be02 <_sk_load_565_dst_sse2_lowp+0x7e>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,65,15,196,100,80,4,2            ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm4
   DB  243,65,15,16,44,80                  ; movss         (%r8,%rdx,2),%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,34                              ; jmp           2abf2 <_sk_load_565_dst_sse2_lowp+0x7e>
+  DB  235,34                              ; jmp           2be02 <_sk_load_565_dst_sse2_lowp+0x7e>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,65,15,196,100,80,12,6           ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm4
   DB  102,65,15,196,100,80,10,5           ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm4
@@ -41021,11 +42094,11 @@
   DB  102,65,15,18,36,80                  ; movlpd        (%r8,%rdx,2),%xmm4
   DB  102,15,111,252                      ; movdqa        %xmm4,%xmm7
   DB  102,15,113,215,8                    ; psrlw         $0x8,%xmm7
-  DB  102,15,219,61,125,31,0,0            ; pand          0x1f7d(%rip),%xmm7        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,15,219,61,77,36,0,0             ; pand          0x244d(%rip),%xmm7        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,113,213,5                    ; psrlw         $0x5,%xmm5
-  DB  102,15,219,45,124,31,0,0            ; pand          0x1f7c(%rip),%xmm5        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,15,111,53,132,31,0,0            ; movdqa        0x1f84(%rip),%xmm6        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,15,219,45,76,36,0,0             ; pand          0x244c(%rip),%xmm5        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,53,84,36,0,0             ; movdqa        0x2454(%rip),%xmm6        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,15,219,244                      ; pand          %xmm4,%xmm6
   DB  102,15,113,212,13                   ; psrlw         $0xd,%xmm4
   DB  102,15,235,231                      ; por           %xmm7,%xmm4
@@ -41038,7 +42111,7 @@
   DB  102,15,113,214,2                    ; psrlw         $0x2,%xmm6
   DB  102,15,235,247                      ; por           %xmm7,%xmm6
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,122,23,0,0                 ; movaps        0x177a(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,61,74,28,0,0                  ; movaps        0x1c4a(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  255,224                             ; jmpq          *%rax
   DB  77,255                              ; rex.WRB       (bad)
   DB  255                                 ; (bad)
@@ -41060,10 +42133,10 @@
   DB  76,3,0                              ; add           (%rax),%r8
   DB  102,68,15,111,192                   ; movdqa        %xmm0,%xmm8
   DB  102,65,15,113,240,8                 ; psllw         $0x8,%xmm8
-  DB  102,68,15,219,5,24,31,0,0           ; pand          0x1f18(%rip),%xmm8        # 2cbb0 <_sk_srcover_rgba_8888_sse2_lowp+0x11a8>
+  DB  102,68,15,219,5,232,35,0,0          ; pand          0x23e8(%rip),%xmm8        # 2e290 <_sk_srcover_rgba_8888_sse2_lowp+0x11dc>
   DB  102,68,15,111,201                   ; movdqa        %xmm1,%xmm9
   DB  102,65,15,113,241,3                 ; psllw         $0x3,%xmm9
-  DB  102,68,15,219,13,20,31,0,0          ; pand          0x1f14(%rip),%xmm9        # 2cbc0 <_sk_srcover_rgba_8888_sse2_lowp+0x11b8>
+  DB  102,68,15,219,13,228,35,0,0         ; pand          0x23e4(%rip),%xmm9        # 2e2a0 <_sk_srcover_rgba_8888_sse2_lowp+0x11ec>
   DB  102,69,15,235,200                   ; por           %xmm8,%xmm9
   DB  102,68,15,111,194                   ; movdqa        %xmm2,%xmm8
   DB  102,65,15,113,208,3                 ; psrlw         $0x3,%xmm8
@@ -41072,21 +42145,21 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,31                              ; ja            2acea <_sk_store_565_sse2_lowp+0x76>
+  DB  119,31                              ; ja            2befa <_sk_store_565_sse2_lowp+0x76>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,95,0,0,0                  ; lea           0x5f(%rip),%r9        # 2ad34 <_sk_store_565_sse2_lowp+0xc0>
+  DB  76,141,13,95,0,0,0                  ; lea           0x5f(%rip),%r9        # 2bf44 <_sk_store_565_sse2_lowp+0xc0>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,68,15,126,192                   ; movd          %xmm8,%eax
   DB  102,65,137,4,80                     ; mov           %ax,(%r8,%rdx,2)
-  DB  235,70                              ; jmp           2ad30 <_sk_store_565_sse2_lowp+0xbc>
+  DB  235,70                              ; jmp           2bf40 <_sk_store_565_sse2_lowp+0xbc>
   DB  243,69,15,127,4,80                  ; movdqu        %xmm8,(%r8,%rdx,2)
-  DB  235,62                              ; jmp           2ad30 <_sk_store_565_sse2_lowp+0xbc>
+  DB  235,62                              ; jmp           2bf40 <_sk_store_565_sse2_lowp+0xbc>
   DB  102,65,15,197,192,2                 ; pextrw        $0x2,%xmm8,%eax
   DB  102,65,137,68,80,4                  ; mov           %ax,0x4(%r8,%rdx,2)
   DB  102,69,15,126,4,80                  ; movd          %xmm8,(%r8,%rdx,2)
-  DB  235,42                              ; jmp           2ad30 <_sk_store_565_sse2_lowp+0xbc>
+  DB  235,42                              ; jmp           2bf40 <_sk_store_565_sse2_lowp+0xbc>
   DB  102,65,15,197,192,6                 ; pextrw        $0x6,%xmm8,%eax
   DB  102,65,137,68,80,12                 ; mov           %ax,0xc(%r8,%rdx,2)
   DB  102,65,15,197,192,5                 ; pextrw        $0x5,%xmm8,%eax
@@ -41118,6 +42191,116 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_565_sse2_lowp
+_sk_gather_565_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,64,12                  ; movss         0xc(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,118,201                   ; pcmpeqd       %xmm9,%xmm9
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  69,15,87,210                        ; xorps         %xmm10,%xmm10
+  DB  65,15,95,202                        ; maxps         %xmm10,%xmm1
+  DB  65,15,95,194                        ; maxps         %xmm10,%xmm0
+  DB  65,15,93,192                        ; minps         %xmm8,%xmm0
+  DB  65,15,93,200                        ; minps         %xmm8,%xmm1
+  DB  243,68,15,16,64,16                  ; movss         0x10(%rax),%xmm8
+  DB  69,15,198,192,0                     ; shufps        $0x0,%xmm8,%xmm8
+  DB  102,69,15,254,193                   ; paddd         %xmm9,%xmm8
+  DB  65,15,95,210                        ; maxps         %xmm10,%xmm2
+  DB  65,15,95,218                        ; maxps         %xmm10,%xmm3
+  DB  65,15,93,216                        ; minps         %xmm8,%xmm3
+  DB  65,15,93,208                        ; minps         %xmm8,%xmm2
+  DB  243,68,15,91,194                    ; cvttps2dq     %xmm2,%xmm8
+  DB  243,68,15,91,203                    ; cvttps2dq     %xmm3,%xmm9
+  DB  102,15,110,80,8                     ; movd          0x8(%rax),%xmm2
+  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
+  DB  102,69,15,112,209,245               ; pshufd        $0xf5,%xmm9,%xmm10
+  DB  102,68,15,244,210                   ; pmuludq       %xmm2,%xmm10
+  DB  102,68,15,111,218                   ; movdqa        %xmm2,%xmm11
+  DB  102,69,15,112,224,245               ; pshufd        $0xf5,%xmm8,%xmm12
+  DB  102,68,15,244,226                   ; pmuludq       %xmm2,%xmm12
+  DB  102,65,15,244,209                   ; pmuludq       %xmm9,%xmm2
+  DB  76,139,16                           ; mov           (%rax),%r10
+  DB  102,15,112,210,232                  ; pshufd        $0xe8,%xmm2,%xmm2
+  DB  102,65,15,112,218,232               ; pshufd        $0xe8,%xmm10,%xmm3
+  DB  102,15,98,211                       ; punpckldq     %xmm3,%xmm2
+  DB  102,69,15,244,216                   ; pmuludq       %xmm8,%xmm11
+  DB  102,69,15,112,195,232               ; pshufd        $0xe8,%xmm11,%xmm8
+  DB  102,65,15,112,220,232               ; pshufd        $0xe8,%xmm12,%xmm3
+  DB  102,68,15,98,195                    ; punpckldq     %xmm3,%xmm8
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,15,254,202                      ; paddd         %xmm2,%xmm1
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,65,15,254,192                   ; paddd         %xmm8,%xmm0
+  DB  102,15,112,208,78                   ; pshufd        $0x4e,%xmm0,%xmm2
+  DB  102,73,15,126,208                   ; movq          %xmm2,%r8
+  DB  69,137,193                          ; mov           %r8d,%r9d
+  DB  73,193,232,32                       ; shr           $0x20,%r8
+  DB  102,72,15,126,192                   ; movq          %xmm0,%rax
+  DB  65,137,195                          ; mov           %eax,%r11d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,15,112,193,78                   ; pshufd        $0x4e,%xmm1,%xmm0
+  DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,73,15,126,207                   ; movq          %xmm1,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  67,15,183,44,122                    ; movzwl        (%r10,%r15,2),%ebp
+  DB  102,15,110,197                      ; movd          %ebp,%xmm0
+  DB  67,15,183,44,98                     ; movzwl        (%r10,%r12,2),%ebp
+  DB  102,15,110,205                      ; movd          %ebp,%xmm1
+  DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
+  DB  65,15,183,28,90                     ; movzwl        (%r10,%rbx,2),%ebx
+  DB  102,15,110,195                      ; movd          %ebx,%xmm0
+  DB  67,15,183,28,114                    ; movzwl        (%r10,%r14,2),%ebx
+  DB  102,15,110,211                      ; movd          %ebx,%xmm2
+  DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
+  DB  102,15,98,202                       ; punpckldq     %xmm2,%xmm1
+  DB  65,15,183,4,66                      ; movzwl        (%r10,%rax,2),%eax
+  DB  102,15,110,208                      ; movd          %eax,%xmm2
+  DB  67,15,183,4,90                      ; movzwl        (%r10,%r11,2),%eax
+  DB  102,15,110,192                      ; movd          %eax,%xmm0
+  DB  102,15,97,194                       ; punpcklwd     %xmm2,%xmm0
+  DB  67,15,183,4,66                      ; movzwl        (%r10,%r8,2),%eax
+  DB  102,15,110,208                      ; movd          %eax,%xmm2
+  DB  67,15,183,4,74                      ; movzwl        (%r10,%r9,2),%eax
+  DB  102,15,110,216                      ; movd          %eax,%xmm3
+  DB  102,15,97,218                       ; punpcklwd     %xmm2,%xmm3
+  DB  102,15,98,195                       ; punpckldq     %xmm3,%xmm0
+  DB  102,15,108,193                      ; punpcklqdq    %xmm1,%xmm0
+  DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
+  DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
+  DB  102,15,219,29,145,33,0,0            ; pand          0x2191(%rip),%xmm3        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
+  DB  102,15,219,13,144,33,0,0            ; pand          0x2190(%rip),%xmm1        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,15,111,21,152,33,0,0            ; movdqa        0x2198(%rip),%xmm2        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
+  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
+  DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
+  DB  102,15,235,195                      ; por           %xmm3,%xmm0
+  DB  102,15,111,217                      ; movdqa        %xmm1,%xmm3
+  DB  102,15,113,243,2                    ; psllw         $0x2,%xmm3
+  DB  102,15,113,209,4                    ; psrlw         $0x4,%xmm1
+  DB  102,15,235,203                      ; por           %xmm3,%xmm1
+  DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
+  DB  102,15,113,243,3                    ; psllw         $0x3,%xmm3
+  DB  102,15,113,210,2                    ; psrlw         $0x2,%xmm2
+  DB  102,15,235,211                      ; por           %xmm3,%xmm2
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  15,40,29,142,25,0,0                 ; movaps        0x198e(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_a8_sse2_lowp
 _sk_load_a8_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -41128,19 +42311,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2ad89 <_sk_load_a8_sse2_lowp+0x39>
+  DB  119,34                              ; ja            2c165 <_sk_load_a8_sse2_lowp+0x39>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,139,0,0,0                 ; lea           0x8b(%rip),%r9        # 2ae00 <_sk_load_a8_sse2_lowp+0xb0>
+  DB  76,141,13,139,0,0,0                 ; lea           0x8b(%rip),%r9        # 2c1dc <_sk_load_a8_sse2_lowp+0xb0>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,216                      ; movd          %eax,%xmm3
-  DB  235,97                              ; jmp           2adea <_sk_load_a8_sse2_lowp+0x9a>
+  DB  235,97                              ; jmp           2c1c6 <_sk_load_a8_sse2_lowp+0x9a>
   DB  243,65,15,126,28,16                 ; movq          (%r8,%rdx,1),%xmm3
   DB  102,15,96,216                       ; punpcklbw     %xmm0,%xmm3
-  DB  235,85                              ; jmp           2adea <_sk_load_a8_sse2_lowp+0x9a>
+  DB  235,85                              ; jmp           2c1c6 <_sk_load_a8_sse2_lowp+0x9a>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,216,2                    ; pinsrw        $0x2,%eax,%xmm3
@@ -41148,7 +42331,7 @@
   DB  102,15,110,192                      ; movd          %eax,%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  243,15,16,216                       ; movss         %xmm0,%xmm3
-  DB  235,51                              ; jmp           2adea <_sk_load_a8_sse2_lowp+0x9a>
+  DB  235,51                              ; jmp           2c1c6 <_sk_load_a8_sse2_lowp+0x9a>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,216,6                    ; pinsrw        $0x6,%eax,%xmm3
@@ -41159,14 +42342,14 @@
   DB  102,65,15,110,4,16                  ; movd          (%r8,%rdx,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  242,15,16,216                       ; movsd         %xmm0,%xmm3
-  DB  102,15,219,29,222,21,0,0            ; pand          0x15de(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,29,226,24,0,0            ; pand          0x18e2(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
   DB  255,224                             ; jmpq          *%rax
   DB  144                                 ; nop
-  DB  126,255                             ; jle           2ae01 <_sk_load_a8_sse2_lowp+0xb1>
+  DB  126,255                             ; jle           2c1dd <_sk_load_a8_sse2_lowp+0xb1>
   DB  255                                 ; (bad)
   DB  255,164,255,255,255,149,255         ; jmpq          *-0x6a0001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
@@ -41194,19 +42377,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2ae55 <_sk_load_a8_dst_sse2_lowp+0x39>
+  DB  119,34                              ; ja            2c231 <_sk_load_a8_dst_sse2_lowp+0x39>
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,139,0,0,0                 ; lea           0x8b(%rip),%r9        # 2aecc <_sk_load_a8_dst_sse2_lowp+0xb0>
+  DB  76,141,13,139,0,0,0                 ; lea           0x8b(%rip),%r9        # 2c2a8 <_sk_load_a8_dst_sse2_lowp+0xb0>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,248                      ; movd          %eax,%xmm7
-  DB  235,97                              ; jmp           2aeb6 <_sk_load_a8_dst_sse2_lowp+0x9a>
+  DB  235,97                              ; jmp           2c292 <_sk_load_a8_dst_sse2_lowp+0x9a>
   DB  243,65,15,126,60,16                 ; movq          (%r8,%rdx,1),%xmm7
   DB  102,15,96,248                       ; punpcklbw     %xmm0,%xmm7
-  DB  235,85                              ; jmp           2aeb6 <_sk_load_a8_dst_sse2_lowp+0x9a>
+  DB  235,85                              ; jmp           2c292 <_sk_load_a8_dst_sse2_lowp+0x9a>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  102,15,196,248,2                    ; pinsrw        $0x2,%eax,%xmm7
@@ -41214,7 +42397,7 @@
   DB  102,15,110,224                      ; movd          %eax,%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  243,15,16,252                       ; movss         %xmm4,%xmm7
-  DB  235,51                              ; jmp           2aeb6 <_sk_load_a8_dst_sse2_lowp+0x9a>
+  DB  235,51                              ; jmp           2c292 <_sk_load_a8_dst_sse2_lowp+0x9a>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  102,15,196,248,6                    ; pinsrw        $0x6,%eax,%xmm7
@@ -41225,14 +42408,14 @@
   DB  102,65,15,110,36,16                 ; movd          (%r8,%rdx,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  242,15,16,252                       ; movsd         %xmm4,%xmm7
-  DB  102,15,219,61,18,21,0,0             ; pand          0x1512(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,61,22,24,0,0             ; pand          0x1816(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,87,237                           ; xorps         %xmm5,%xmm5
   DB  15,87,246                           ; xorps         %xmm6,%xmm6
   DB  255,224                             ; jmpq          *%rax
   DB  144                                 ; nop
-  DB  126,255                             ; jle           2aecd <_sk_load_a8_dst_sse2_lowp+0xb1>
+  DB  126,255                             ; jle           2c2a9 <_sk_load_a8_dst_sse2_lowp+0xb1>
   DB  255                                 ; (bad)
   DB  255,164,255,255,255,149,255         ; jmpq          *-0x6a0001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
@@ -41260,42 +42443,42 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,29                              ; ja            2af1c <_sk_store_a8_sse2_lowp+0x34>
+  DB  119,29                              ; ja            2c2f8 <_sk_store_a8_sse2_lowp+0x34>
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2af9c <_sk_store_a8_sse2_lowp+0xb4>
+  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2c378 <_sk_store_a8_sse2_lowp+0xb4>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,15,126,216                      ; movd          %xmm3,%eax
   DB  65,136,4,16                         ; mov           %al,(%r8,%rdx,1)
-  DB  235,123                             ; jmp           2af97 <_sk_store_a8_sse2_lowp+0xaf>
-  DB  102,68,15,111,5,171,20,0,0          ; movdqa        0x14ab(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  235,123                             ; jmp           2c373 <_sk_store_a8_sse2_lowp+0xaf>
+  DB  102,68,15,111,5,175,23,0,0          ; movdqa        0x17af(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,219,195                   ; pand          %xmm3,%xmm8
   DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
   DB  102,69,15,214,4,16                  ; movq          %xmm8,(%r8,%rdx,1)
-  DB  235,96                              ; jmp           2af97 <_sk_store_a8_sse2_lowp+0xaf>
+  DB  235,96                              ; jmp           2c373 <_sk_store_a8_sse2_lowp+0xaf>
   DB  102,15,197,195,2                    ; pextrw        $0x2,%xmm3,%eax
   DB  65,136,68,16,2                      ; mov           %al,0x2(%r8,%rdx,1)
-  DB  102,68,15,111,5,134,20,0,0          ; movdqa        0x1486(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,138,23,0,0          ; movdqa        0x178a(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,219,195                   ; pand          %xmm3,%xmm8
   DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
   DB  102,68,15,126,192                   ; movd          %xmm8,%eax
   DB  102,65,137,4,16                     ; mov           %ax,(%r8,%rdx,1)
-  DB  235,55                              ; jmp           2af97 <_sk_store_a8_sse2_lowp+0xaf>
+  DB  235,55                              ; jmp           2c373 <_sk_store_a8_sse2_lowp+0xaf>
   DB  102,15,197,195,6                    ; pextrw        $0x6,%xmm3,%eax
   DB  65,136,68,16,6                      ; mov           %al,0x6(%r8,%rdx,1)
   DB  102,15,197,195,5                    ; pextrw        $0x5,%xmm3,%eax
   DB  65,136,68,16,5                      ; mov           %al,0x5(%r8,%rdx,1)
   DB  102,15,197,195,4                    ; pextrw        $0x4,%xmm3,%eax
   DB  65,136,68,16,4                      ; mov           %al,0x4(%r8,%rdx,1)
-  DB  102,68,15,111,5,73,20,0,0           ; movdqa        0x1449(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,77,23,0,0           ; movdqa        0x174d(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,68,15,219,195                   ; pand          %xmm3,%xmm8
   DB  102,69,15,103,192                   ; packuswb      %xmm8,%xmm8
   DB  102,69,15,126,4,16                  ; movd          %xmm8,(%r8,%rdx,1)
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  255,224                             ; jmpq          *%rax
   DB  144                                 ; nop
-  DB  118,255                             ; jbe           2af9d <_sk_store_a8_sse2_lowp+0xb5>
+  DB  118,255                             ; jbe           2c379 <_sk_store_a8_sse2_lowp+0xb5>
   DB  255                                 ; (bad)
   DB  255,165,255,255,255,155             ; jmpq          *-0x64000001(%rbp)
   DB  255                                 ; (bad)
@@ -41314,6 +42497,98 @@
   DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
+PUBLIC _sk_gather_a8_sse2_lowp
+_sk_gather_a8_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  102,69,15,118,210                   ; pcmpeqd       %xmm10,%xmm10
+  DB  102,69,15,254,202                   ; paddd         %xmm10,%xmm9
+  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
+  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
+  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
+  DB  65,15,93,201                        ; minps         %xmm9,%xmm1
+  DB  65,15,93,193                        ; minps         %xmm9,%xmm0
+  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  102,69,15,254,202                   ; paddd         %xmm10,%xmm9
+  DB  65,15,95,216                        ; maxps         %xmm8,%xmm3
+  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
+  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
+  DB  65,15,93,217                        ; minps         %xmm9,%xmm3
+  DB  243,68,15,91,203                    ; cvttps2dq     %xmm3,%xmm9
+  DB  243,68,15,91,210                    ; cvttps2dq     %xmm2,%xmm10
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,69,15,112,218,245               ; pshufd        $0xf5,%xmm10,%xmm11
+  DB  102,68,15,244,219                   ; pmuludq       %xmm3,%xmm11
+  DB  102,68,15,111,227                   ; movdqa        %xmm3,%xmm12
+  DB  102,69,15,112,233,245               ; pshufd        $0xf5,%xmm9,%xmm13
+  DB  102,68,15,244,235                   ; pmuludq       %xmm3,%xmm13
+  DB  102,65,15,244,218                   ; pmuludq       %xmm10,%xmm3
+  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
+  DB  102,65,15,112,211,232               ; pshufd        $0xe8,%xmm11,%xmm2
+  DB  102,15,98,218                       ; punpckldq     %xmm2,%xmm3
+  DB  76,139,16                           ; mov           (%rax),%r10
+  DB  102,69,15,244,225                   ; pmuludq       %xmm9,%xmm12
+  DB  102,69,15,112,204,232               ; pshufd        $0xe8,%xmm12,%xmm9
+  DB  102,65,15,112,213,232               ; pshufd        $0xe8,%xmm13,%xmm2
+  DB  102,68,15,98,202                    ; punpckldq     %xmm2,%xmm9
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,201                   ; paddd         %xmm9,%xmm1
+  DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
+  DB  102,73,15,126,208                   ; movq          %xmm2,%r8
+  DB  69,137,193                          ; mov           %r8d,%r9d
+  DB  73,193,232,32                       ; shr           $0x20,%r8
+  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
+  DB  65,137,195                          ; mov           %eax,%r11d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  71,15,182,36,34                     ; movzbl        (%r10,%r12,1),%r12d
+  DB  67,15,182,44,58                     ; movzbl        (%r10,%r15,1),%ebp
+  DB  193,229,8                           ; shl           $0x8,%ebp
+  DB  68,9,229                            ; or            %r12d,%ebp
+  DB  71,15,182,52,50                     ; movzbl        (%r10,%r14,1),%r14d
+  DB  65,15,182,28,26                     ; movzbl        (%r10,%rbx,1),%ebx
+  DB  193,227,8                           ; shl           $0x8,%ebx
+  DB  68,9,243                            ; or            %r14d,%ebx
+  DB  102,15,110,219                      ; movd          %ebx,%xmm3
+  DB  102,15,196,221,1                    ; pinsrw        $0x1,%ebp,%xmm3
+  DB  67,15,182,44,26                     ; movzbl        (%r10,%r11,1),%ebp
+  DB  65,15,182,4,2                       ; movzbl        (%r10,%rax,1),%eax
+  DB  193,224,8                           ; shl           $0x8,%eax
+  DB  9,232                               ; or            %ebp,%eax
+  DB  102,15,196,216,2                    ; pinsrw        $0x2,%eax,%xmm3
+  DB  67,15,182,4,10                      ; movzbl        (%r10,%r9,1),%eax
+  DB  67,15,182,44,2                      ; movzbl        (%r10,%r8,1),%ebp
+  DB  193,229,8                           ; shl           $0x8,%ebp
+  DB  9,197                               ; or            %eax,%ebp
+  DB  102,15,196,221,3                    ; pinsrw        $0x3,%ebp,%xmm3
+  DB  102,65,15,96,216                    ; punpcklbw     %xmm8,%xmm3
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
+  DB  102,15,239,201                      ; pxor          %xmm1,%xmm1
+  DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_load_g8_sse2_lowp
 _sk_load_g8_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -41324,19 +42599,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2aff1 <_sk_load_g8_sse2_lowp+0x39>
+  DB  119,34                              ; ja            2c535 <_sk_load_g8_sse2_lowp+0x39>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2b070 <_sk_load_g8_sse2_lowp+0xb8>
+  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2c5b4 <_sk_load_g8_sse2_lowp+0xb8>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,192                      ; movd          %eax,%xmm0
-  DB  235,97                              ; jmp           2b052 <_sk_load_g8_sse2_lowp+0x9a>
+  DB  235,97                              ; jmp           2c596 <_sk_load_g8_sse2_lowp+0x9a>
   DB  243,65,15,126,4,16                  ; movq          (%r8,%rdx,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
-  DB  235,85                              ; jmp           2b052 <_sk_load_g8_sse2_lowp+0x9a>
+  DB  235,85                              ; jmp           2c596 <_sk_load_g8_sse2_lowp+0x9a>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,192,2                    ; pinsrw        $0x2,%eax,%xmm0
@@ -41344,7 +42619,7 @@
   DB  102,15,110,200                      ; movd          %eax,%xmm1
   DB  102,15,96,200                       ; punpcklbw     %xmm0,%xmm1
   DB  243,15,16,193                       ; movss         %xmm1,%xmm0
-  DB  235,51                              ; jmp           2b052 <_sk_load_g8_sse2_lowp+0x9a>
+  DB  235,51                              ; jmp           2c596 <_sk_load_g8_sse2_lowp+0x9a>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,192,6                    ; pinsrw        $0x6,%eax,%xmm0
@@ -41355,14 +42630,14 @@
   DB  102,65,15,110,12,16                 ; movd          (%r8,%rdx,1),%xmm1
   DB  102,15,96,200                       ; punpcklbw     %xmm0,%xmm1
   DB  242,15,16,193                       ; movsd         %xmm1,%xmm0
-  DB  102,15,219,5,118,19,0,0             ; pand          0x1376(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,5,18,21,0,0              ; pand          0x1512(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,29,109,19,0,0                 ; movaps        0x136d(%rip),%xmm3        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,29,9,21,0,0                   ; movaps        0x1509(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  118,255                             ; jbe           2b071 <_sk_load_g8_sse2_lowp+0xb9>
+  DB  118,255                             ; jbe           2c5b5 <_sk_load_g8_sse2_lowp+0xb9>
   DB  255                                 ; (bad)
   DB  255,156,255,255,255,141,255         ; lcall         *-0x720001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
@@ -41388,19 +42663,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2b0c5 <_sk_load_g8_dst_sse2_lowp+0x39>
+  DB  119,34                              ; ja            2c609 <_sk_load_g8_dst_sse2_lowp+0x39>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2b144 <_sk_load_g8_dst_sse2_lowp+0xb8>
+  DB  76,141,13,147,0,0,0                 ; lea           0x93(%rip),%r9        # 2c688 <_sk_load_g8_dst_sse2_lowp+0xb8>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,15,110,224                      ; movd          %eax,%xmm4
-  DB  235,97                              ; jmp           2b126 <_sk_load_g8_dst_sse2_lowp+0x9a>
+  DB  235,97                              ; jmp           2c66a <_sk_load_g8_dst_sse2_lowp+0x9a>
   DB  243,65,15,126,36,16                 ; movq          (%r8,%rdx,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
-  DB  235,85                              ; jmp           2b126 <_sk_load_g8_dst_sse2_lowp+0x9a>
+  DB  235,85                              ; jmp           2c66a <_sk_load_g8_dst_sse2_lowp+0x9a>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,224,2                    ; pinsrw        $0x2,%eax,%xmm4
@@ -41408,7 +42683,7 @@
   DB  102,15,110,232                      ; movd          %eax,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,51                              ; jmp           2b126 <_sk_load_g8_dst_sse2_lowp+0x9a>
+  DB  235,51                              ; jmp           2c66a <_sk_load_g8_dst_sse2_lowp+0x9a>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,224,6                    ; pinsrw        $0x6,%eax,%xmm4
@@ -41419,14 +42694,14 @@
   DB  102,65,15,110,44,16                 ; movd          (%r8,%rdx,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,229                       ; movsd         %xmm5,%xmm4
-  DB  102,15,219,37,162,18,0,0            ; pand          0x12a2(%rip),%xmm4        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,219,37,62,20,0,0             ; pand          0x143e(%rip),%xmm4        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  72,173                              ; lods          %ds:(%rsi),%rax
-  DB  15,40,61,153,18,0,0                 ; movaps        0x1299(%rip),%xmm7        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  15,40,61,53,20,0,0                  ; movaps        0x1435(%rip),%xmm7        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  118,255                             ; jbe           2b145 <_sk_load_g8_dst_sse2_lowp+0xb9>
+  DB  118,255                             ; jbe           2c689 <_sk_load_g8_dst_sse2_lowp+0xb9>
   DB  255                                 ; (bad)
   DB  255,156,255,255,255,141,255         ; lcall         *-0x720001(%rdi,%rdi,8)
   DB  255                                 ; (bad)
@@ -41445,10 +42720,10 @@
 PUBLIC _sk_luminance_to_alpha_sse2_lowp
 _sk_luminance_to_alpha_sse2_lowp LABEL PROC
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
-  DB  102,15,213,5,100,26,0,0             ; pmullw        0x1a64(%rip),%xmm0        # 2cbd0 <_sk_srcover_rgba_8888_sse2_lowp+0x11c8>
-  DB  102,15,213,13,108,26,0,0            ; pmullw        0x1a6c(%rip),%xmm1        # 2cbe0 <_sk_srcover_rgba_8888_sse2_lowp+0x11d8>
+  DB  102,15,213,5,0,28,0,0               ; pmullw        0x1c00(%rip),%xmm0        # 2e2b0 <_sk_srcover_rgba_8888_sse2_lowp+0x11fc>
+  DB  102,15,213,13,8,28,0,0              ; pmullw        0x1c08(%rip),%xmm1        # 2e2c0 <_sk_srcover_rgba_8888_sse2_lowp+0x120c>
   DB  102,15,253,200                      ; paddw         %xmm0,%xmm1
-  DB  102,15,213,29,112,26,0,0            ; pmullw        0x1a70(%rip),%xmm3        # 2cbf0 <_sk_srcover_rgba_8888_sse2_lowp+0x11e8>
+  DB  102,15,213,29,12,28,0,0             ; pmullw        0x1c0c(%rip),%xmm3        # 2e2d0 <_sk_srcover_rgba_8888_sse2_lowp+0x121c>
   DB  102,15,253,217                      ; paddw         %xmm1,%xmm3
   DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
   DB  72,173                              ; lods          %ds:(%rsi),%rax
@@ -41457,18 +42732,110 @@
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  255,224                             ; jmpq          *%rax
 
+PUBLIC _sk_gather_g8_sse2_lowp
+_sk_gather_g8_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %rbp
+  DB  65,87                               ; push          %r15
+  DB  65,86                               ; push          %r14
+  DB  65,84                               ; push          %r12
+  DB  83                                  ; push          %rbx
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  243,68,15,16,72,12                  ; movss         0xc(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  102,69,15,118,210                   ; pcmpeqd       %xmm10,%xmm10
+  DB  102,69,15,254,202                   ; paddd         %xmm10,%xmm9
+  DB  69,15,87,192                        ; xorps         %xmm8,%xmm8
+  DB  65,15,95,192                        ; maxps         %xmm8,%xmm0
+  DB  65,15,95,200                        ; maxps         %xmm8,%xmm1
+  DB  65,15,93,201                        ; minps         %xmm9,%xmm1
+  DB  65,15,93,193                        ; minps         %xmm9,%xmm0
+  DB  243,68,15,16,72,16                  ; movss         0x10(%rax),%xmm9
+  DB  69,15,198,201,0                     ; shufps        $0x0,%xmm9,%xmm9
+  DB  102,69,15,254,202                   ; paddd         %xmm10,%xmm9
+  DB  65,15,95,216                        ; maxps         %xmm8,%xmm3
+  DB  65,15,95,208                        ; maxps         %xmm8,%xmm2
+  DB  65,15,93,209                        ; minps         %xmm9,%xmm2
+  DB  65,15,93,217                        ; minps         %xmm9,%xmm3
+  DB  243,68,15,91,203                    ; cvttps2dq     %xmm3,%xmm9
+  DB  243,68,15,91,210                    ; cvttps2dq     %xmm2,%xmm10
+  DB  102,15,110,88,8                     ; movd          0x8(%rax),%xmm3
+  DB  102,15,112,219,0                    ; pshufd        $0x0,%xmm3,%xmm3
+  DB  102,69,15,112,218,245               ; pshufd        $0xf5,%xmm10,%xmm11
+  DB  102,68,15,244,219                   ; pmuludq       %xmm3,%xmm11
+  DB  102,68,15,111,227                   ; movdqa        %xmm3,%xmm12
+  DB  102,69,15,112,233,245               ; pshufd        $0xf5,%xmm9,%xmm13
+  DB  102,68,15,244,235                   ; pmuludq       %xmm3,%xmm13
+  DB  102,65,15,244,218                   ; pmuludq       %xmm10,%xmm3
+  DB  102,15,112,219,232                  ; pshufd        $0xe8,%xmm3,%xmm3
+  DB  102,65,15,112,211,232               ; pshufd        $0xe8,%xmm11,%xmm2
+  DB  102,15,98,218                       ; punpckldq     %xmm2,%xmm3
+  DB  76,139,16                           ; mov           (%rax),%r10
+  DB  102,69,15,244,225                   ; pmuludq       %xmm9,%xmm12
+  DB  102,69,15,112,204,232               ; pshufd        $0xe8,%xmm12,%xmm9
+  DB  102,65,15,112,213,232               ; pshufd        $0xe8,%xmm13,%xmm2
+  DB  102,68,15,98,202                    ; punpckldq     %xmm2,%xmm9
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,65,15,254,201                   ; paddd         %xmm9,%xmm1
+  DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
+  DB  102,73,15,126,208                   ; movq          %xmm2,%r8
+  DB  69,137,193                          ; mov           %r8d,%r9d
+  DB  73,193,232,32                       ; shr           $0x20,%r8
+  DB  102,72,15,126,200                   ; movq          %xmm1,%rax
+  DB  65,137,195                          ; mov           %eax,%r11d
+  DB  72,193,232,32                       ; shr           $0x20,%rax
+  DB  102,72,15,126,195                   ; movq          %xmm0,%rbx
+  DB  65,137,222                          ; mov           %ebx,%r14d
+  DB  72,193,235,32                       ; shr           $0x20,%rbx
+  DB  102,15,112,192,78                   ; pshufd        $0x4e,%xmm0,%xmm0
+  DB  102,73,15,126,199                   ; movq          %xmm0,%r15
+  DB  69,137,252                          ; mov           %r15d,%r12d
+  DB  73,193,239,32                       ; shr           $0x20,%r15
+  DB  71,15,182,36,34                     ; movzbl        (%r10,%r12,1),%r12d
+  DB  67,15,182,44,58                     ; movzbl        (%r10,%r15,1),%ebp
+  DB  193,229,8                           ; shl           $0x8,%ebp
+  DB  68,9,229                            ; or            %r12d,%ebp
+  DB  71,15,182,52,50                     ; movzbl        (%r10,%r14,1),%r14d
+  DB  65,15,182,28,26                     ; movzbl        (%r10,%rbx,1),%ebx
+  DB  193,227,8                           ; shl           $0x8,%ebx
+  DB  68,9,243                            ; or            %r14d,%ebx
+  DB  102,15,110,195                      ; movd          %ebx,%xmm0
+  DB  102,15,196,197,1                    ; pinsrw        $0x1,%ebp,%xmm0
+  DB  67,15,182,44,26                     ; movzbl        (%r10,%r11,1),%ebp
+  DB  65,15,182,4,2                       ; movzbl        (%r10,%rax,1),%eax
+  DB  193,224,8                           ; shl           $0x8,%eax
+  DB  9,232                               ; or            %ebp,%eax
+  DB  102,15,196,192,2                    ; pinsrw        $0x2,%eax,%xmm0
+  DB  67,15,182,4,10                      ; movzbl        (%r10,%r9,1),%eax
+  DB  67,15,182,44,2                      ; movzbl        (%r10,%r8,1),%ebp
+  DB  193,229,8                           ; shl           $0x8,%ebp
+  DB  9,197                               ; or            %eax,%ebp
+  DB  102,15,196,197,3                    ; pinsrw        $0x3,%ebp,%xmm0
+  DB  102,65,15,96,192                    ; punpcklbw     %xmm8,%xmm0
+  DB  72,173                              ; lods          %ds:(%rsi),%rax
+  DB  15,40,29,122,18,0,0                 ; movaps        0x127a(%rip),%xmm3        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
+  DB  91                                  ; pop           %rbx
+  DB  65,92                               ; pop           %r12
+  DB  65,94                               ; pop           %r14
+  DB  65,95                               ; pop           %r15
+  DB  93                                  ; pop           %rbp
+  DB  255,224                             ; jmpq          *%rax
+
 PUBLIC _sk_scale_1_float_sse2_lowp
 _sk_scale_1_float_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
-  DB  243,68,15,89,5,59,11,0,0            ; mulss         0xb3b(%rip),%xmm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  243,68,15,88,5,2,11,0,0             ; addss         0xb02(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,5,56,11,0,0            ; mulss         0xb38(%rip),%xmm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,68,15,88,5,255,10,0,0           ; addss         0xaff(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,44,192                    ; cvttss2si     %xmm8,%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  242,69,15,112,192,0                 ; pshuflw       $0x0,%xmm8,%xmm8
   DB  102,69,15,112,192,80                ; pshufd        $0x50,%xmm8,%xmm8
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
-  DB  102,68,15,111,13,250,17,0,0         ; movdqa        0x11fa(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,43,18,0,0          ; movdqa        0x122b(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -41487,13 +42854,13 @@
 _sk_lerp_1_float_sse2_lowp LABEL PROC
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  243,68,15,16,0                      ; movss         (%rax),%xmm8
-  DB  243,68,15,89,5,195,10,0,0           ; mulss         0xac3(%rip),%xmm8        # 2bce4 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
-  DB  243,68,15,88,5,138,10,0,0           ; addss         0xa8a(%rip),%xmm8        # 2bcb4 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
+  DB  243,68,15,89,5,192,10,0,0           ; mulss         0xac0(%rip),%xmm8        # 2d390 <_sk_srcover_rgba_8888_sse2_lowp+0x2dc>
+  DB  243,68,15,88,5,135,10,0,0           ; addss         0xa87(%rip),%xmm8        # 2d360 <_sk_srcover_rgba_8888_sse2_lowp+0x2ac>
   DB  243,65,15,44,192                    ; cvttss2si     %xmm8,%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
   DB  242,69,15,112,192,0                 ; pshuflw       $0x0,%xmm8,%xmm8
   DB  102,69,15,112,192,80                ; pshufd        $0x50,%xmm8,%xmm8
-  DB  102,68,15,111,13,135,17,0,0         ; movdqa        0x1187(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,184,17,0,0         ; movdqa        0x11b8(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -41531,19 +42898,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,36                              ; ja            2b300 <_sk_scale_u8_sse2_lowp+0x3b>
+  DB  119,36                              ; ja            2c9af <_sk_scale_u8_sse2_lowp+0x3b>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,221,0,0,0                 ; lea           0xdd(%rip),%r9        # 2b3c8 <_sk_scale_u8_sse2_lowp+0x103>
+  DB  76,141,13,218,0,0,0                 ; lea           0xda(%rip),%r9        # 2ca74 <_sk_scale_u8_sse2_lowp+0x100>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
-  DB  235,109                             ; jmp           2b36d <_sk_scale_u8_sse2_lowp+0xa8>
+  DB  235,109                             ; jmp           2ca1c <_sk_scale_u8_sse2_lowp+0xa8>
   DB  243,69,15,126,4,16                  ; movq          (%r8,%rdx,1),%xmm8
   DB  102,68,15,96,192                    ; punpcklbw     %xmm0,%xmm8
-  DB  235,96                              ; jmp           2b36d <_sk_scale_u8_sse2_lowp+0xa8>
+  DB  235,96                              ; jmp           2ca1c <_sk_scale_u8_sse2_lowp+0xa8>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,68,15,196,192,2                 ; pinsrw        $0x2,%eax,%xmm8
@@ -41551,7 +42918,7 @@
   DB  102,68,15,110,200                   ; movd          %eax,%xmm9
   DB  102,68,15,96,200                    ; punpcklbw     %xmm0,%xmm9
   DB  243,69,15,16,193                    ; movss         %xmm9,%xmm8
-  DB  235,57                              ; jmp           2b36d <_sk_scale_u8_sse2_lowp+0xa8>
+  DB  235,57                              ; jmp           2ca1c <_sk_scale_u8_sse2_lowp+0xa8>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,68,15,196,192,6                 ; pinsrw        $0x6,%eax,%xmm8
@@ -41562,9 +42929,9 @@
   DB  102,69,15,110,12,16                 ; movd          (%r8,%rdx,1),%xmm9
   DB  102,68,15,96,200                    ; punpcklbw     %xmm0,%xmm9
   DB  242,69,15,16,193                    ; movsd         %xmm9,%xmm8
-  DB  102,68,15,219,5,90,16,0,0           ; pand          0x105a(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,219,5,139,16,0,0          ; pand          0x108b(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,213,192                   ; pmullw        %xmm8,%xmm0
-  DB  102,68,15,111,13,76,16,0,0          ; movdqa        0x104c(%rip),%xmm9        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,13,125,16,0,0         ; movdqa        0x107d(%rip),%xmm9        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,253,193                   ; paddw         %xmm9,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,65,15,213,200                   ; pmullw        %xmm8,%xmm1
@@ -41579,20 +42946,19 @@
   DB  72,173                              ; lods          %ds:(%rsi),%rax
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  255,224                             ; jmpq          *%rax
-  DB  15,31,0                             ; nopl          (%rax)
-  DB  44,255                              ; sub           $0xff,%al
-  DB  255                                 ; (bad)
-  DB  255,86,255                          ; callq         *-0x1(%rsi)
-  DB  255                                 ; (bad)
-  DB  255,69,255                          ; incl          -0x1(%rbp)
-  DB  255                                 ; (bad)
-  DB  255,149,255,255,255,137             ; callq         *-0x76000001(%rbp)
+  DB  47                                  ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
+  DB  255,89,255                          ; lcall         *-0x1(%rcx)
   DB  255                                 ; (bad)
-  DB  125,255                             ; jge           2b3dd <_sk_scale_u8_sse2_lowp+0x118>
+  DB  255,72,255                          ; decl          -0x1(%rax)
   DB  255                                 ; (bad)
-  DB  255,108,255,255                     ; ljmp          *-0x1(%rdi,%rdi,8)
+  DB  255,152,255,255,255,140             ; lcall         *-0x73000001(%rax)
+  DB  255                                 ; (bad)
+  DB  255                                 ; (bad)
+  DB  255,128,255,255,255,111             ; incl          0x6fffffff(%rax)
+  DB  255                                 ; (bad)
+  DB  255                                 ; (bad)
   DB  255                                 ; .byte         0xff
 
 PUBLIC _sk_lerp_u8_sse2_lowp
@@ -41605,19 +42971,19 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,36                              ; ja            2b41f <_sk_lerp_u8_sse2_lowp+0x3b>
+  DB  119,36                              ; ja            2cacb <_sk_lerp_u8_sse2_lowp+0x3b>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,30,1,0,0                  ; lea           0x11e(%rip),%r9        # 2b528 <_sk_lerp_u8_sse2_lowp+0x144>
+  DB  76,141,13,30,1,0,0                  ; lea           0x11e(%rip),%r9        # 2cbd4 <_sk_lerp_u8_sse2_lowp+0x144>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,182,4,16                      ; movzbl        (%r8,%rdx,1),%eax
   DB  102,68,15,110,192                   ; movd          %eax,%xmm8
-  DB  235,109                             ; jmp           2b48c <_sk_lerp_u8_sse2_lowp+0xa8>
+  DB  235,109                             ; jmp           2cb38 <_sk_lerp_u8_sse2_lowp+0xa8>
   DB  243,69,15,126,4,16                  ; movq          (%r8,%rdx,1),%xmm8
   DB  102,68,15,96,192                    ; punpcklbw     %xmm0,%xmm8
-  DB  235,96                              ; jmp           2b48c <_sk_lerp_u8_sse2_lowp+0xa8>
+  DB  235,96                              ; jmp           2cb38 <_sk_lerp_u8_sse2_lowp+0xa8>
   DB  65,15,182,68,16,2                   ; movzbl        0x2(%r8,%rdx,1),%eax
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,68,15,196,192,2                 ; pinsrw        $0x2,%eax,%xmm8
@@ -41625,7 +42991,7 @@
   DB  102,68,15,110,200                   ; movd          %eax,%xmm9
   DB  102,68,15,96,200                    ; punpcklbw     %xmm0,%xmm9
   DB  243,69,15,16,193                    ; movss         %xmm9,%xmm8
-  DB  235,57                              ; jmp           2b48c <_sk_lerp_u8_sse2_lowp+0xa8>
+  DB  235,57                              ; jmp           2cb38 <_sk_lerp_u8_sse2_lowp+0xa8>
   DB  65,15,182,68,16,6                   ; movzbl        0x6(%r8,%rdx,1),%eax
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
   DB  102,68,15,196,192,6                 ; pinsrw        $0x6,%eax,%xmm8
@@ -41636,8 +43002,8 @@
   DB  102,69,15,110,12,16                 ; movd          (%r8,%rdx,1),%xmm9
   DB  102,68,15,96,200                    ; punpcklbw     %xmm0,%xmm9
   DB  242,69,15,16,193                    ; movsd         %xmm9,%xmm8
-  DB  102,68,15,219,5,59,15,0,0           ; pand          0xf3b(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
-  DB  102,68,15,111,21,50,15,0,0          ; movdqa        0xf32(%rip),%xmm10        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,219,5,111,15,0,0          ; pand          0xf6f(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
+  DB  102,68,15,111,21,102,15,0,0         ; movdqa        0xf66(%rip),%xmm10        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,69,15,111,200                   ; movdqa        %xmm8,%xmm9
   DB  102,69,15,239,202                   ; pxor          %xmm10,%xmm9
   DB  102,69,15,111,217                   ; movdqa        %xmm9,%xmm11
@@ -41667,9 +43033,9 @@
   DB  102,65,15,111,216                   ; movdqa        %xmm8,%xmm3
   DB  255,224                             ; jmpq          *%rax
   DB  15,31,0                             ; nopl          (%rax)
-  DB  235,254                             ; jmp           2b528 <_sk_lerp_u8_sse2_lowp+0x144>
+  DB  235,254                             ; jmp           2cbd4 <_sk_lerp_u8_sse2_lowp+0x144>
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,4                ; callq         *0x4ffffff(%rip)        # 502b530 <_sk_srcover_rgba_8888_sse2_lowp+0x4fffb28>
+  DB  255,21,255,255,255,4                ; callq         *0x4ffffff(%rip)        # 502cbdc <_sk_srcover_rgba_8888_sse2_lowp+0x4fffb28>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,84,255,255                      ; callq         *-0x1(%rdi,%rdi,8)
@@ -41698,23 +43064,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2b592 <_sk_scale_565_sse2_lowp+0x4e>
+  DB  119,34                              ; ja            2cc3e <_sk_scale_565_sse2_lowp+0x4e>
   DB  102,15,239,246                      ; pxor          %xmm6,%xmm6
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,234,1,0,0                 ; lea           0x1ea(%rip),%r9        # 2b768 <_sk_scale_565_sse2_lowp+0x224>
+  DB  76,141,13,234,1,0,0                 ; lea           0x1ea(%rip),%r9        # 2ce14 <_sk_scale_565_sse2_lowp+0x224>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,240                      ; movd          %eax,%xmm6
-  DB  235,67                              ; jmp           2b5d5 <_sk_scale_565_sse2_lowp+0x91>
+  DB  235,67                              ; jmp           2cc81 <_sk_scale_565_sse2_lowp+0x91>
   DB  243,65,15,111,52,80                 ; movdqu        (%r8,%rdx,2),%xmm6
-  DB  235,59                              ; jmp           2b5d5 <_sk_scale_565_sse2_lowp+0x91>
+  DB  235,59                              ; jmp           2cc81 <_sk_scale_565_sse2_lowp+0x91>
   DB  102,15,239,246                      ; pxor          %xmm6,%xmm6
   DB  102,65,15,196,116,80,4,2            ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm6
   DB  243,69,15,16,12,80                  ; movss         (%r8,%rdx,2),%xmm9
   DB  243,65,15,16,241                    ; movss         %xmm9,%xmm6
-  DB  235,34                              ; jmp           2b5d5 <_sk_scale_565_sse2_lowp+0x91>
+  DB  235,34                              ; jmp           2cc81 <_sk_scale_565_sse2_lowp+0x91>
   DB  102,15,239,246                      ; pxor          %xmm6,%xmm6
   DB  102,65,15,196,116,80,12,6           ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm6
   DB  102,65,15,196,116,80,10,5           ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm6
@@ -41722,11 +43088,11 @@
   DB  102,65,15,18,52,80                  ; movlpd        (%r8,%rdx,2),%xmm6
   DB  102,68,15,111,222                   ; movdqa        %xmm6,%xmm11
   DB  102,65,15,113,211,8                 ; psrlw         $0x8,%xmm11
-  DB  102,68,15,219,29,151,21,0,0         ; pand          0x1597(%rip),%xmm11        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,68,15,219,29,203,21,0,0         ; pand          0x15cb(%rip),%xmm11        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,68,15,111,206                   ; movdqa        %xmm6,%xmm9
   DB  102,65,15,113,209,5                 ; psrlw         $0x5,%xmm9
-  DB  102,68,15,219,13,147,21,0,0         ; pand          0x1593(%rip),%xmm9        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,68,15,111,21,154,21,0,0         ; movdqa        0x159a(%rip),%xmm10        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,68,15,219,13,199,21,0,0         ; pand          0x15c7(%rip),%xmm9        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,68,15,111,21,206,21,0,0         ; movdqa        0x15ce(%rip),%xmm10        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,68,15,219,214                   ; pand          %xmm6,%xmm10
   DB  102,15,113,214,13                   ; psrlw         $0xd,%xmm6
   DB  102,65,15,235,243                   ; por           %xmm11,%xmm6
@@ -41738,7 +43104,7 @@
   DB  102,65,15,113,243,3                 ; psllw         $0x3,%xmm11
   DB  102,65,15,113,210,2                 ; psrlw         $0x2,%xmm10
   DB  102,69,15,235,211                   ; por           %xmm11,%xmm10
-  DB  102,68,15,111,37,246,20,0,0         ; movdqa        0x14f6(%rip),%xmm12        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,37,42,21,0,0          ; movdqa        0x152a(%rip),%xmm12        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,68,15,111,235                   ; movdqa        %xmm3,%xmm13
   DB  102,69,15,239,236                   ; pxor          %xmm12,%xmm13
   DB  102,68,15,111,223                   ; movdqa        %xmm7,%xmm11
@@ -41779,7 +43145,7 @@
   DB  102,68,15,213,201                   ; pmullw        %xmm1,%xmm9
   DB  102,68,15,213,210                   ; pmullw        %xmm2,%xmm10
   DB  102,68,15,213,219                   ; pmullw        %xmm3,%xmm11
-  DB  102,15,111,5,184,12,0,0             ; movdqa        0xcb8(%rip),%xmm0        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,15,111,5,236,12,0,0             ; movdqa        0xcec(%rip),%xmm0        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,253,240                      ; paddw         %xmm0,%xmm6
   DB  102,68,15,253,200                   ; paddw         %xmm0,%xmm9
   DB  102,68,15,253,208                   ; paddw         %xmm0,%xmm10
@@ -41832,23 +43198,23 @@
   DB  36,7                                ; and           $0x7,%al
   DB  254,200                             ; dec           %al
   DB  60,6                                ; cmp           $0x6,%al
-  DB  119,34                              ; ja            2b7d2 <_sk_lerp_565_sse2_lowp+0x4e>
+  DB  119,34                              ; ja            2ce7e <_sk_lerp_565_sse2_lowp+0x4e>
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  15,182,192                          ; movzbl        %al,%eax
-  DB  76,141,13,46,2,0,0                  ; lea           0x22e(%rip),%r9        # 2b9ec <_sk_lerp_565_sse2_lowp+0x268>
+  DB  76,141,13,46,2,0,0                  ; lea           0x22e(%rip),%r9        # 2d098 <_sk_lerp_565_sse2_lowp+0x268>
   DB  73,99,4,129                         ; movslq        (%r9,%rax,4),%rax
   DB  76,1,200                            ; add           %r9,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  65,15,183,4,80                      ; movzwl        (%r8,%rdx,2),%eax
   DB  102,15,110,208                      ; movd          %eax,%xmm2
-  DB  235,67                              ; jmp           2b815 <_sk_lerp_565_sse2_lowp+0x91>
+  DB  235,67                              ; jmp           2cec1 <_sk_lerp_565_sse2_lowp+0x91>
   DB  243,65,15,111,20,80                 ; movdqu        (%r8,%rdx,2),%xmm2
-  DB  235,59                              ; jmp           2b815 <_sk_lerp_565_sse2_lowp+0x91>
+  DB  235,59                              ; jmp           2cec1 <_sk_lerp_565_sse2_lowp+0x91>
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  102,65,15,196,84,80,4,2             ; pinsrw        $0x2,0x4(%r8,%rdx,2),%xmm2
   DB  243,69,15,16,12,80                  ; movss         (%r8,%rdx,2),%xmm9
   DB  243,65,15,16,209                    ; movss         %xmm9,%xmm2
-  DB  235,34                              ; jmp           2b815 <_sk_lerp_565_sse2_lowp+0x91>
+  DB  235,34                              ; jmp           2cec1 <_sk_lerp_565_sse2_lowp+0x91>
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  102,65,15,196,84,80,12,6            ; pinsrw        $0x6,0xc(%r8,%rdx,2),%xmm2
   DB  102,65,15,196,84,80,10,5            ; pinsrw        $0x5,0xa(%r8,%rdx,2),%xmm2
@@ -41856,11 +43222,11 @@
   DB  102,65,15,18,20,80                  ; movlpd        (%r8,%rdx,2),%xmm2
   DB  102,68,15,111,218                   ; movdqa        %xmm2,%xmm11
   DB  102,65,15,113,211,8                 ; psrlw         $0x8,%xmm11
-  DB  102,68,15,219,29,87,19,0,0          ; pand          0x1357(%rip),%xmm11        # 2cb80 <_sk_srcover_rgba_8888_sse2_lowp+0x1178>
+  DB  102,68,15,219,29,139,19,0,0         ; pand          0x138b(%rip),%xmm11        # 2e260 <_sk_srcover_rgba_8888_sse2_lowp+0x11ac>
   DB  102,68,15,111,202                   ; movdqa        %xmm2,%xmm9
   DB  102,65,15,113,209,5                 ; psrlw         $0x5,%xmm9
-  DB  102,68,15,219,13,83,19,0,0          ; pand          0x1353(%rip),%xmm9        # 2cb90 <_sk_srcover_rgba_8888_sse2_lowp+0x1188>
-  DB  102,68,15,111,21,90,19,0,0          ; movdqa        0x135a(%rip),%xmm10        # 2cba0 <_sk_srcover_rgba_8888_sse2_lowp+0x1198>
+  DB  102,68,15,219,13,135,19,0,0         ; pand          0x1387(%rip),%xmm9        # 2e270 <_sk_srcover_rgba_8888_sse2_lowp+0x11bc>
+  DB  102,68,15,111,21,142,19,0,0         ; movdqa        0x138e(%rip),%xmm10        # 2e280 <_sk_srcover_rgba_8888_sse2_lowp+0x11cc>
   DB  102,68,15,219,210                   ; pand          %xmm2,%xmm10
   DB  102,15,113,210,13                   ; psrlw         $0xd,%xmm2
   DB  102,65,15,235,211                   ; por           %xmm11,%xmm2
@@ -41872,7 +43238,7 @@
   DB  102,65,15,113,243,3                 ; psllw         $0x3,%xmm11
   DB  102,65,15,113,210,2                 ; psrlw         $0x2,%xmm10
   DB  102,69,15,235,211                   ; por           %xmm11,%xmm10
-  DB  102,68,15,111,37,182,18,0,0         ; movdqa        0x12b6(%rip),%xmm12        # 2cb40 <_sk_srcover_rgba_8888_sse2_lowp+0x1138>
+  DB  102,68,15,111,37,234,18,0,0         ; movdqa        0x12ea(%rip),%xmm12        # 2e220 <_sk_srcover_rgba_8888_sse2_lowp+0x116c>
   DB  102,68,15,111,235                   ; movdqa        %xmm3,%xmm13
   DB  102,69,15,239,236                   ; pxor          %xmm12,%xmm13
   DB  102,68,15,111,223                   ; movdqa        %xmm7,%xmm11
@@ -41911,7 +43277,7 @@
   DB  102,69,15,235,222                   ; por           %xmm14,%xmm11
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
   DB  102,15,213,209                      ; pmullw        %xmm1,%xmm2
-  DB  102,68,15,111,5,130,10,0,0          ; movdqa        0xa82(%rip),%xmm8        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,5,182,10,0,0          ; movdqa        0xab6(%rip),%xmm8        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,65,15,239,192                   ; pxor          %xmm8,%xmm0
   DB  102,15,213,196                      ; pmullw        %xmm4,%xmm0
   DB  102,65,15,253,208                   ; paddw         %xmm8,%xmm2
@@ -41981,26 +43347,26 @@
   DB  65,254,202                          ; dec           %r10b
   DB  69,15,182,202                       ; movzbl        %r10b,%r9d
   DB  65,128,249,6                        ; cmp           $0x6,%r9b
-  DB  119,38                              ; ja            2ba53 <_sk_srcover_rgba_8888_sse2_lowp+0x4b>
+  DB  119,38                              ; ja            2d0ff <_sk_srcover_rgba_8888_sse2_lowp+0x4b>
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  76,141,29,67,2,0,0                  ; lea           0x243(%rip),%r11        # 2bc7c <_sk_srcover_rgba_8888_sse2_lowp+0x274>
+  DB  76,141,29,67,2,0,0                  ; lea           0x243(%rip),%r11        # 2d328 <_sk_srcover_rgba_8888_sse2_lowp+0x274>
   DB  75,99,4,139                         ; movslq        (%r11,%r9,4),%rax
   DB  76,1,216                            ; add           %r11,%rax
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  255,224                             ; jmpq          *%rax
   DB  243,65,15,16,60,144                 ; movss         (%r8,%rdx,4),%xmm7
   DB  102,69,15,239,192                   ; pxor          %xmm8,%xmm8
-  DB  235,102                             ; jmp           2bab9 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  DB  235,102                             ; jmp           2d165 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   DB  102,65,15,16,60,144                 ; movupd        (%r8,%rdx,4),%xmm7
   DB  243,69,15,111,68,144,16             ; movdqu        0x10(%r8,%rdx,4),%xmm8
-  DB  235,87                              ; jmp           2bab9 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  DB  235,87                              ; jmp           2d165 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   DB  102,65,15,110,100,144,8             ; movd          0x8(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,69,15,18,4,144                  ; movlpd        (%r8,%rdx,4),%xmm8
   DB  102,65,15,40,248                    ; movapd        %xmm8,%xmm7
   DB  102,68,15,111,196                   ; movdqa        %xmm4,%xmm8
-  DB  235,52                              ; jmp           2bab9 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
+  DB  235,52                              ; jmp           2d165 <_sk_srcover_rgba_8888_sse2_lowp+0xb1>
   DB  102,65,15,110,100,144,24            ; movd          0x18(%r8,%rdx,4),%xmm4
   DB  102,68,15,112,196,69                ; pshufd        $0x45,%xmm4,%xmm8
   DB  243,65,15,16,100,144,20             ; movss         0x14(%r8,%rdx,4),%xmm4
@@ -42017,7 +43383,7 @@
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
   DB  102,15,114,228,16                   ; psrad         $0x10,%xmm4
   DB  102,15,107,236                      ; packssdw      %xmm4,%xmm5
-  DB  102,68,15,111,37,237,8,0,0          ; movdqa        0x8ed(%rip),%xmm12        # 2c3d0 <_sk_srcover_rgba_8888_sse2_lowp+0x9c8>
+  DB  102,68,15,111,37,33,9,0,0           ; movdqa        0x921(%rip),%xmm12        # 2dab0 <_sk_srcover_rgba_8888_sse2_lowp+0x9fc>
   DB  102,15,111,229                      ; movdqa        %xmm5,%xmm4
   DB  102,65,15,219,228                   ; pand          %xmm12,%xmm4
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
@@ -42075,20 +43441,20 @@
   DB  102,15,105,193                      ; punpckhwd     %xmm1,%xmm0
   DB  102,15,235,194                      ; por           %xmm2,%xmm0
   DB  65,128,250,6                        ; cmp           $0x6,%r10b
-  DB  119,24                              ; ja            2bc1b <_sk_srcover_rgba_8888_sse2_lowp+0x213>
-  DB  76,141,21,142,0,0,0                 ; lea           0x8e(%rip),%r10        # 2bc98 <_sk_srcover_rgba_8888_sse2_lowp+0x290>
+  DB  119,24                              ; ja            2d2c7 <_sk_srcover_rgba_8888_sse2_lowp+0x213>
+  DB  76,141,21,142,0,0,0                 ; lea           0x8e(%rip),%r10        # 2d344 <_sk_srcover_rgba_8888_sse2_lowp+0x290>
   DB  75,99,4,138                         ; movslq        (%r10,%r9,4),%rax
   DB  76,1,208                            ; add           %r10,%rax
   DB  255,224                             ; jmpq          *%rax
   DB  102,69,15,126,36,144                ; movd          %xmm12,(%r8,%rdx,4)
-  DB  235,73                              ; jmp           2bc64 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  DB  235,73                              ; jmp           2d310 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   DB  243,69,15,127,36,144                ; movdqu        %xmm12,(%r8,%rdx,4)
   DB  243,65,15,127,68,144,16             ; movdqu        %xmm0,0x10(%r8,%rdx,4)
-  DB  235,58                              ; jmp           2bc64 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  DB  235,58                              ; jmp           2d310 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   DB  102,65,15,112,196,78                ; pshufd        $0x4e,%xmm12,%xmm0
   DB  102,65,15,126,68,144,8              ; movd          %xmm0,0x8(%r8,%rdx,4)
   DB  102,69,15,214,36,144                ; movq          %xmm12,(%r8,%rdx,4)
-  DB  235,37                              ; jmp           2bc64 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
+  DB  235,37                              ; jmp           2d310 <_sk_srcover_rgba_8888_sse2_lowp+0x25c>
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
   DB  102,65,15,126,76,144,24             ; movd          %xmm1,0x18(%r8,%rdx,4)
   DB  102,15,112,200,229                  ; pshufd        $0xe5,%xmm0,%xmm1
@@ -42121,7 +43487,7 @@
   DB  254                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  123,255                             ; jnp           2bc99 <_sk_srcover_rgba_8888_sse2_lowp+0x291>
+  DB  123,255                             ; jnp           2d345 <_sk_srcover_rgba_8888_sse2_lowp+0x291>
   DB  255                                 ; (bad)
   DB  255,159,255,255,255,146             ; lcall         *-0x6d000001(%rdi)
   DB  255                                 ; (bad)
@@ -42154,14 +43520,14 @@
   DB  252                                 ; cld
   DB  190,0,0,128,191                     ; mov           $0xbf800000,%esi
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        2bd18 <.literal4+0x64>
+  DB  224,64                              ; loopne        2d3c4 <.literal4+0x64>
   DB  154                                 ; (bad)
   DB  153                                 ; cltd
   DB  153                                 ; cltd
   DB  62,61,10,23,63,174                  ; ds            cmp $0xae3f170a,%eax
-  DB  71,225,61                           ; rex.RXB       loope 2bd21 <.literal4+0x6d>
+  DB  71,225,61                           ; rex.RXB       loope 2d3cd <.literal4+0x6d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            2bd2b <.literal4+0x77>
+  DB  127,67                              ; jg            2d3d7 <.literal4+0x77>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,127,145,131                     ; cmpb          $0x83,-0x6f(%rdi)
   DB  158                                 ; sahf
@@ -42196,7 +43562,7 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        2bd45 <.literal4+0x91>
+  DB  224,7                               ; loopne        2d3f1 <.literal4+0x91>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -42207,8 +43573,8 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,52,255                            ; add           %dh,(%rdi,%rdi,8)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2bd58 <.literal4+0xa4>
-  DB  119,115                             ; ja            2bdcd <.literal4+0x119>
+  DB  127,0                               ; jg            2d404 <.literal4+0xa4>
+  DB  119,115                             ; ja            2d479 <.literal4+0x119>
   DB  248                                 ; clc
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,249,68,180                   ; mov           $0xb444f93f,%edi
@@ -42231,12 +43597,12 @@
   DB  190,80,128,3,62                     ; mov           $0x3e038050,%esi
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           2bde3 <.literal4+0x12f>
+  DB  118,63                              ; jbe           2d48f <.literal4+0x12f>
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            2bdf2 <.literal4+0x13e>
+  DB  124,66                              ; jl            2d49e <.literal4+0x13e>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  137,136,136,55,0,15                 ; mov           %ecx,0xf003788(%rax)
@@ -42246,9 +43612,9 @@
   DB  137,136,136,59,15,0                 ; mov           %ecx,0xf3b88(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  137,136,136,61,0,0                  ; mov           %ecx,0x3d88(%rax)
-  DB  112,65                              ; jo            2be15 <.literal4+0x161>
+  DB  112,65                              ; jo            2d4c1 <.literal4+0x161>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            2be1f <.literal4+0x16b>
+  DB  127,71                              ; jg            2d4cb <.literal4+0x16b>
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,89                               ; ds            pop %rcx
@@ -42275,15 +43641,15 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,0,0                             ; addb          $0x0,(%rax)
   DB  0,191,0,0,192,191                   ; add           %bh,-0x40400000(%rdi)
-  DB  114,28                              ; jb            2be3e <.literal4+0x18a>
+  DB  114,28                              ; jb            2d4ea <.literal4+0x18a>
   DB  199                                 ; (bad)
   DB  62,0,0                              ; add           %al,%ds:(%rax)
   DB  192,63,85                           ; sarb          $0x55,(%rdi)
   DB  85                                  ; push          %rbp
   DB  149                                 ; xchg          %eax,%ebp
   DB  191,57,142,99,61                    ; mov           $0x3d638e39,%edi
-  DB  114,249                             ; jb            2be2b <.literal4+0x177>
-  DB  127,63                              ; jg            2be73 <_sk_srcover_rgba_8888_sse2_lowp+0x46b>
+  DB  114,249                             ; jb            2d4d7 <.literal4+0x177>
+  DB  127,63                              ; jg            2d51f <_sk_srcover_rgba_8888_sse2_lowp+0x46b>
   DB  3,0                                 ; add           (%rax),%eax
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,4,0                               ; add           %al,(%rax,%rax,1)
@@ -42292,7 +43658,7 @@
   DB  56,255                              ; cmp           %bh,%bh
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2be45 <.literal4+0x191>
+  DB  127,0                               ; jg            2d4f1 <.literal4+0x191>
   DB  0,128,56,0,64,254                   ; add           %al,-0x1bfffc8(%rax)
   DB  255                                 ; .byte         0xff
 
@@ -42330,16 +43696,16 @@
   DB  0,1                                 ; add           %al,(%rcx)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02bea8 <_sk_srcover_rgba_8888_sse2_lowp+0xa0004a0>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d548 <_sk_srcover_rgba_8888_sse2_lowp+0xa000494>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202beb0 <_sk_srcover_rgba_8888_sse2_lowp+0x120004a8>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d550 <_sk_srcover_rgba_8888_sse2_lowp+0x1200049c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02beb8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0004b0>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d558 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0004a4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302bec0 <_sk_srcover_rgba_8888_sse2_lowp+0x30004b8>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d560 <_sk_srcover_rgba_8888_sse2_lowp+0x30004ac>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42382,16 +43748,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02bf08 <_sk_srcover_rgba_8888_sse2_lowp+0xa000500>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d5a8 <_sk_srcover_rgba_8888_sse2_lowp+0xa0004f4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202bf10 <_sk_srcover_rgba_8888_sse2_lowp+0x12000508>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d5b0 <_sk_srcover_rgba_8888_sse2_lowp+0x120004fc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02bf18 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000510>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d5b8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000504>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302bf20 <_sk_srcover_rgba_8888_sse2_lowp+0x3000518>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d5c0 <_sk_srcover_rgba_8888_sse2_lowp+0x300050c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42434,16 +43800,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02bf68 <_sk_srcover_rgba_8888_sse2_lowp+0xa000560>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d608 <_sk_srcover_rgba_8888_sse2_lowp+0xa000554>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202bf70 <_sk_srcover_rgba_8888_sse2_lowp+0x12000568>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d610 <_sk_srcover_rgba_8888_sse2_lowp+0x1200055c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02bf78 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000570>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d618 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000564>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302bf80 <_sk_srcover_rgba_8888_sse2_lowp+0x3000578>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d620 <_sk_srcover_rgba_8888_sse2_lowp+0x300056c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42486,16 +43852,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02bfc8 <_sk_srcover_rgba_8888_sse2_lowp+0xa0005c0>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d668 <_sk_srcover_rgba_8888_sse2_lowp+0xa0005b4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202bfd0 <_sk_srcover_rgba_8888_sse2_lowp+0x120005c8>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d670 <_sk_srcover_rgba_8888_sse2_lowp+0x120005bc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02bfd8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0005d0>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d678 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0005c4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302bfe0 <_sk_srcover_rgba_8888_sse2_lowp+0x30005d8>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d680 <_sk_srcover_rgba_8888_sse2_lowp+0x30005cc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42538,16 +43904,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02c028 <_sk_srcover_rgba_8888_sse2_lowp+0xa000620>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d6c8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000614>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202c030 <_sk_srcover_rgba_8888_sse2_lowp+0x12000628>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d6d0 <_sk_srcover_rgba_8888_sse2_lowp+0x1200061c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02c038 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000630>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d6d8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000624>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302c040 <_sk_srcover_rgba_8888_sse2_lowp+0x3000638>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d6e0 <_sk_srcover_rgba_8888_sse2_lowp+0x300062c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42590,16 +43956,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02c088 <_sk_srcover_rgba_8888_sse2_lowp+0xa000680>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d728 <_sk_srcover_rgba_8888_sse2_lowp+0xa000674>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202c090 <_sk_srcover_rgba_8888_sse2_lowp+0x12000688>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d730 <_sk_srcover_rgba_8888_sse2_lowp+0x1200067c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02c098 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000690>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d738 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000684>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302c0a0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000698>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d740 <_sk_srcover_rgba_8888_sse2_lowp+0x300068c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42642,16 +44008,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02c0e8 <_sk_srcover_rgba_8888_sse2_lowp+0xa0006e0>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d788 <_sk_srcover_rgba_8888_sse2_lowp+0xa0006d4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202c0f0 <_sk_srcover_rgba_8888_sse2_lowp+0x120006e8>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d790 <_sk_srcover_rgba_8888_sse2_lowp+0x120006dc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02c0f8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0006f0>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d798 <_sk_srcover_rgba_8888_sse2_lowp+0x1a0006e4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302c100 <_sk_srcover_rgba_8888_sse2_lowp+0x30006f8>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d7a0 <_sk_srcover_rgba_8888_sse2_lowp+0x30006ec>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42694,16 +44060,16 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02c148 <_sk_srcover_rgba_8888_sse2_lowp+0xa000740>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02d7e8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000734>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202c150 <_sk_srcover_rgba_8888_sse2_lowp+0x12000748>
+  DB  255,13,255,255,255,17               ; decl          0x11ffffff(%rip)        # 1202d7f0 <_sk_srcover_rgba_8888_sse2_lowp+0x1200073c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02c158 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000750>
+  DB  255,21,255,255,255,25               ; callq         *0x19ffffff(%rip)        # 1a02d7f8 <_sk_srcover_rgba_8888_sse2_lowp+0x1a000744>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302c160 <_sk_srcover_rgba_8888_sse2_lowp+0x3000758>
+  DB  255,29,255,255,255,2                ; lcall         *0x2ffffff(%rip)        # 302d800 <_sk_srcover_rgba_8888_sse2_lowp+0x300074c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -42966,6 +44332,22 @@
   DB  255,0                               ; incl          (%rax)
   DB  255,0                               ; incl          (%rax)
   DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
+  DB  255,0                               ; incl          (%rax)
   DB  248                                 ; clc
   DB  0,63                                ; add           %bh,(%rdi)
   DB  0,31                                ; add           %bl,(%rdi)
@@ -42976,10 +44358,16 @@
   DB  0,255                               ; add           %bh,%bh
   DB  0,0                                 ; add           %al,(%rax)
   DB  248                                 ; clc
-  DB  224,255                             ; loopne        2c373 <.const+0x513>
-  DB  255,0                               ; incl          (%rax)
-  DB  255,0                               ; incl          (%rax)
-  DB  54,0,183,0,19,0,255                 ; add           %dh,%ss:-0xffed00(%rdi)
+  DB  224,255                             ; loopne        2da33 <.const+0x533>
+  DB  248                                 ; clc
+  DB  0,63                                ; add           %bh,(%rdi)
+  DB  0,31                                ; add           %bl,(%rdi)
+  DB  0,255                               ; add           %bh,%bh
+  DB  0,255                               ; add           %bh,%bh
+  DB  0,255                               ; add           %bh,%bh
+  DB  0,54                                ; add           %dh,(%rsi)
+  DB  0,183,0,19,0,255                    ; add           %dh,-0xffed00(%rdi)
+  DB  0,255                               ; add           %bh,%bh
   DB  0,255                               ; add           %bh,%bh
   DB  0,255                               ; add           %bh,%bh
   DB  0,255                               ; add           %bh,%bh
@@ -42996,6 +44384,17 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
+  DB  0,0                                 ; add           %al,(%rax)
   DB  255,0                               ; incl          (%rax)
   DB  255,0                               ; incl          (%rax)
   DB  255,0                               ; incl          (%rax)
@@ -43106,11 +44505,11 @@
   DB  0,128,191,0,0,128                   ; add           %al,-0x7fffff41(%rax)
   DB  191,0,0,224,64                      ; mov           $0x40e00000,%edi
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        2c4e8 <.literal16+0x118>
+  DB  224,64                              ; loopne        2dbc8 <.literal16+0x118>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        2c4ec <.literal16+0x11c>
+  DB  224,64                              ; loopne        2dbcc <.literal16+0x11c>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,64                              ; loopne        2c4f0 <.literal16+0x120>
+  DB  224,64                              ; loopne        2dbd0 <.literal16+0x120>
   DB  154                                 ; (bad)
   DB  153                                 ; cltd
   DB  153                                 ; cltd
@@ -43130,13 +44529,13 @@
   DB  10,23                               ; or            (%rdi),%dl
   DB  63                                  ; (bad)
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 2c511 <.literal16+0x141>
+  DB  71,225,61                           ; rex.RXB       loope 2dbf1 <.literal16+0x141>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 2c515 <.literal16+0x145>
+  DB  71,225,61                           ; rex.RXB       loope 2dbf5 <.literal16+0x145>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 2c519 <.literal16+0x149>
+  DB  71,225,61                           ; rex.RXB       loope 2dbf9 <.literal16+0x149>
   DB  174                                 ; scas          %es:(%rdi),%al
-  DB  71,225,61                           ; rex.RXB       loope 2c51d <.literal16+0x14d>
+  DB  71,225,61                           ; rex.RXB       loope 2dbfd <.literal16+0x14d>
   DB  255,0                               ; incl          (%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  255,0                               ; incl          (%rax)
@@ -43147,10 +44546,10 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  1,255                               ; add           %edi,%edi
   DB  255                                 ; (bad)
-  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02c4f8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000af0>
+  DB  255,5,255,255,255,9                 ; incl          0x9ffffff(%rip)        # a02dbd8 <_sk_srcover_rgba_8888_sse2_lowp+0xa000b24>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 302c500 <_sk_srcover_rgba_8888_sse2_lowp+0x3000af8>
+  DB  255,13,255,255,255,2                ; decl          0x2ffffff(%rip)        # 302dbe0 <_sk_srcover_rgba_8888_sse2_lowp+0x3000b2c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255,6                               ; incl          (%rsi)
@@ -43165,11 +44564,11 @@
   DB  255,0                               ; incl          (%rax)
   DB  0,127,67                            ; add           %bh,0x43(%rdi)
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            2c55b <.literal16+0x18b>
+  DB  127,67                              ; jg            2dc3b <.literal16+0x18b>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            2c55f <.literal16+0x18f>
+  DB  127,67                              ; jg            2dc3f <.literal16+0x18f>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  127,67                              ; jg            2c563 <.literal16+0x193>
+  DB  127,67                              ; jg            2dc43 <.literal16+0x193>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,127,0,0                         ; cmpb          $0x0,0x0(%rdi)
   DB  128,127,0,0                         ; cmpb          $0x0,0x0(%rdi)
@@ -43325,13 +44724,13 @@
   DB  132,55                              ; test          %dh,(%rdi)
   DB  8,33                                ; or            %ah,(%rcx)
   DB  132,55                              ; test          %dh,(%rdi)
-  DB  224,7                               ; loopne        2c679 <.literal16+0x2a9>
+  DB  224,7                               ; loopne        2dd59 <.literal16+0x2a9>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        2c67d <.literal16+0x2ad>
+  DB  224,7                               ; loopne        2dd5d <.literal16+0x2ad>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        2c681 <.literal16+0x2b1>
+  DB  224,7                               ; loopne        2dd61 <.literal16+0x2b1>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  224,7                               ; loopne        2c685 <.literal16+0x2b5>
+  DB  224,7                               ; loopne        2dd65 <.literal16+0x2b5>
   DB  0,0                                 ; add           %al,(%rax)
   DB  33,8                                ; and           %ecx,(%rax)
   DB  2,58                                ; add           (%rdx),%bh
@@ -43385,17 +44784,17 @@
   DB  0,0                                 ; add           %al,(%rax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c6f4 <.literal16+0x324>
+  DB  127,0                               ; jg            2ddd4 <.literal16+0x324>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c6f8 <.literal16+0x328>
+  DB  127,0                               ; jg            2ddd8 <.literal16+0x328>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c6fc <.literal16+0x32c>
+  DB  127,0                               ; jg            2dddc <.literal16+0x32c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c700 <.literal16+0x330>
-  DB  119,115                             ; ja            2c775 <.literal16+0x3a5>
+  DB  127,0                               ; jg            2dde0 <.literal16+0x330>
+  DB  119,115                             ; ja            2de55 <.literal16+0x3a5>
   DB  248                                 ; clc
   DB  194,119,115                         ; retq          $0x7377
   DB  248                                 ; clc
@@ -43406,7 +44805,7 @@
   DB  194,117,191                         ; retq          $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
-  DB  117,191                             ; jne           2c6d9 <.literal16+0x309>
+  DB  117,191                             ; jne           2ddb9 <.literal16+0x309>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; (bad)
   DB  249                                 ; stc
@@ -43418,7 +44817,7 @@
   DB  249                                 ; stc
   DB  68,180,62                           ; rex.R         mov $0x3e,%spl
   DB  163,233,220,63,163,233,220,63,163   ; movabs        %eax,0xa33fdce9a33fdce9
-  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a6071a <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe9a34d12>
+  DB  233,220,63,163,233                  ; jmpq          ffffffffe9a61dfa <_sk_srcover_rgba_8888_sse2_lowp+0xffffffffe9a34d46>
   DB  220,63                              ; fdivrl        (%rdi)
   DB  81                                  ; push          %rcx
   DB  140,242                             ; mov           %?,%edx
@@ -43508,16 +44907,16 @@
   DB  128,3,62                            ; addb          $0x3e,(%rbx)
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           2c863 <.literal16+0x493>
+  DB  118,63                              ; jbe           2df43 <.literal16+0x493>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           2c867 <.literal16+0x497>
+  DB  118,63                              ; jbe           2df47 <.literal16+0x497>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           2c86b <.literal16+0x49b>
+  DB  118,63                              ; jbe           2df4b <.literal16+0x49b>
   DB  31                                  ; (bad)
   DB  215                                 ; xlat          %ds:(%rbx)
-  DB  118,63                              ; jbe           2c86f <.literal16+0x49f>
+  DB  118,63                              ; jbe           2df4f <.literal16+0x49f>
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%rax)
@@ -43538,13 +44937,13 @@
   DB  65,0,0                              ; add           %al,(%r8)
   DB  248                                 ; clc
   DB  65,0,0                              ; add           %al,(%r8)
-  DB  124,66                              ; jl            2c8a6 <.literal16+0x4d6>
+  DB  124,66                              ; jl            2df86 <.literal16+0x4d6>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            2c8aa <.literal16+0x4da>
+  DB  124,66                              ; jl            2df8a <.literal16+0x4da>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            2c8ae <.literal16+0x4de>
+  DB  124,66                              ; jl            2df8e <.literal16+0x4de>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  124,66                              ; jl            2c8b2 <.literal16+0x4e2>
+  DB  124,66                              ; jl            2df92 <.literal16+0x4e2>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,240                               ; add           %dh,%al
@@ -43590,25 +44989,25 @@
   DB  136,136,61,137,136,136              ; mov           %cl,-0x777776c3(%rax)
   DB  61,137,136,136,61                   ; cmp           $0x3d888889,%eax
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            2c935 <.literal16+0x565>
+  DB  112,65                              ; jo            2e015 <.literal16+0x565>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            2c939 <.literal16+0x569>
+  DB  112,65                              ; jo            2e019 <.literal16+0x569>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            2c93d <.literal16+0x56d>
+  DB  112,65                              ; jo            2e01d <.literal16+0x56d>
   DB  0,0                                 ; add           %al,(%rax)
-  DB  112,65                              ; jo            2c941 <.literal16+0x571>
+  DB  112,65                              ; jo            2e021 <.literal16+0x571>
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%rax)
   DB  0,0                                 ; add           %al,(%rax)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c913 <.literal16+0x543>
+  DB  127,0                               ; jg            2dff3 <.literal16+0x543>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            2c917 <.literal16+0x547>
+  DB  127,0                               ; jg            2dff7 <.literal16+0x547>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            2c91b <.literal16+0x54b>
+  DB  127,0                               ; jg            2dffb <.literal16+0x54b>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            2c91f <.literal16+0x54f>
+  DB  127,0                               ; jg            2dfff <.literal16+0x54f>
   DB  0,0                                 ; add           %al,(%rax)
   DB  4,0                                 ; add           $0x0,%al
   DB  0,0                                 ; add           %al,(%rax)
@@ -43634,13 +45033,13 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,255                             ; jg            2c958 <.literal16+0x588>
+  DB  127,255                             ; jg            2e038 <.literal16+0x588>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,255                             ; jg            2c95c <.literal16+0x58c>
+  DB  127,255                             ; jg            2e03c <.literal16+0x58c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2c961 <.literal16+0x591>
+  DB  127,0                               ; jg            2e041 <.literal16+0x591>
   DB  0,128,56,0,0,128                    ; add           %al,-0x7fffffc8(%rax)
   DB  56,0                                ; cmp           %al,(%rax)
   DB  0,128,56,0,0,128                    ; add           %al,-0x7fffffc8(%rax)
@@ -43654,13 +45053,13 @@
   DB  64,254                              ; rex           (bad)
   DB  255,0                               ; incl          (%rax)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            2c9cb <.literal16+0x5fb>
+  DB  127,71                              ; jg            2e0ab <.literal16+0x5fb>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            2c9cf <.literal16+0x5ff>
+  DB  127,71                              ; jg            2e0af <.literal16+0x5ff>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            2c9d3 <.literal16+0x603>
+  DB  127,71                              ; jg            2e0b3 <.literal16+0x603>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            2c9d7 <.literal16+0x607>
+  DB  127,71                              ; jg            2e0b7 <.literal16+0x607>
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,208                              ; ds            (bad)
@@ -43690,10 +45089,10 @@
   DB  61,152,221,147,61                   ; cmp           $0x3d93dd98,%eax
   DB  152                                 ; cwtl
   DB  221,147,61,45,16,17                 ; fstl          0x11102d3d(%rbx)
-  DB  192,45,16,17,192,45,16              ; shrb          $0x10,0x2dc01110(%rip)        # 2dc2dada <_sk_srcover_rgba_8888_sse2_lowp+0x2dc020d2>
+  DB  192,45,16,17,192,45,16              ; shrb          $0x10,0x2dc01110(%rip)        # 2dc2f1ba <_sk_srcover_rgba_8888_sse2_lowp+0x2dc02106>
   DB  17,192                              ; adc           %eax,%eax
   DB  45,16,17,192,18                     ; sub           $0x12c01110,%eax
-  DB  120,57                              ; js            2ca0c <.literal16+0x63c>
+  DB  120,57                              ; js            2e0ec <.literal16+0x63c>
   DB  64,18,120,57                        ; adc           0x39(%rax),%dil
   DB  64,18,120,57                        ; adc           0x39(%rax),%dil
   DB  64,18,120,57                        ; adc           0x39(%rax),%dil
@@ -43758,13 +45157,13 @@
   DB  192,191,0,0,192,191,0               ; sarb          $0x0,-0x40400000(%rdi)
   DB  0,192                               ; add           %al,%al
   DB  191,0,0,192,191                     ; mov           $0xbfc00000,%edi
-  DB  114,28                              ; jb            2cabe <.literal16+0x6ee>
+  DB  114,28                              ; jb            2e19e <.literal16+0x6ee>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         2cac2 <.literal16+0x6f2>
+  DB  62,114,28                           ; jb,pt         2e1a2 <.literal16+0x6f2>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         2cac6 <.literal16+0x6f6>
+  DB  62,114,28                           ; jb,pt         2e1a6 <.literal16+0x6f6>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         2caca <.literal16+0x6fa>
+  DB  62,114,28                           ; jb,pt         2e1aa <.literal16+0x6fa>
   DB  199                                 ; (bad)
   DB  62,85                               ; ds            push %rbp
   DB  85                                  ; push          %rbp
@@ -43783,15 +45182,15 @@
   DB  0,192                               ; add           %al,%al
   DB  63                                  ; (bad)
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%rsi)
-  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d665915 <_sk_srcover_rgba_8888_sse2_lowp+0x3d639f0d>
+  DB  99,61,57,142,99,61                  ; movslq        0x3d638e39(%rip),%edi        # 3d666ff5 <_sk_srcover_rgba_8888_sse2_lowp+0x3d639f41>
   DB  57,142,99,61,114,249                ; cmp           %ecx,-0x68dc29d(%rsi)
-  DB  127,63                              ; jg            2cb23 <.literal16+0x753>
-  DB  114,249                             ; jb            2cadf <.literal16+0x70f>
-  DB  127,63                              ; jg            2cb27 <.literal16+0x757>
-  DB  114,249                             ; jb            2cae3 <.literal16+0x713>
-  DB  127,63                              ; jg            2cb2b <.literal16+0x75b>
-  DB  114,249                             ; jb            2cae7 <.literal16+0x717>
-  DB  127,63                              ; jg            2cb2f <.literal16+0x75f>
+  DB  127,63                              ; jg            2e203 <.literal16+0x753>
+  DB  114,249                             ; jb            2e1bf <.literal16+0x70f>
+  DB  127,63                              ; jg            2e207 <.literal16+0x757>
+  DB  114,249                             ; jb            2e1c3 <.literal16+0x713>
+  DB  127,63                              ; jg            2e20b <.literal16+0x75b>
+  DB  114,249                             ; jb            2e1c7 <.literal16+0x717>
+  DB  127,63                              ; jg            2e20f <.literal16+0x75f>
   DB  3,0                                 ; add           (%rax),%eax
   DB  0,0                                 ; add           %al,(%rax)
   DB  3,0                                 ; add           (%rax),%eax
@@ -43901,14 +45300,14 @@
   DB  0,248                               ; add           %bh,%al
   DB  0,248                               ; add           %bh,%al
   DB  0,248                               ; add           %bh,%al
-  DB  224,255                             ; loopne        2cbc1 <.literal16+0x7f1>
-  DB  224,255                             ; loopne        2cbc3 <.literal16+0x7f3>
-  DB  224,255                             ; loopne        2cbc5 <.literal16+0x7f5>
-  DB  224,255                             ; loopne        2cbc7 <.literal16+0x7f7>
-  DB  224,255                             ; loopne        2cbc9 <.literal16+0x7f9>
-  DB  224,255                             ; loopne        2cbcb <.literal16+0x7fb>
-  DB  224,255                             ; loopne        2cbcd <.literal16+0x7fd>
-  DB  224,255                             ; loopne        2cbcf <.literal16+0x7ff>
+  DB  224,255                             ; loopne        2e2a1 <.literal16+0x7f1>
+  DB  224,255                             ; loopne        2e2a3 <.literal16+0x7f3>
+  DB  224,255                             ; loopne        2e2a5 <.literal16+0x7f5>
+  DB  224,255                             ; loopne        2e2a7 <.literal16+0x7f7>
+  DB  224,255                             ; loopne        2e2a9 <.literal16+0x7f9>
+  DB  224,255                             ; loopne        2e2ab <.literal16+0x7fb>
+  DB  224,255                             ; loopne        2e2ad <.literal16+0x7fd>
+  DB  224,255                             ; loopne        2e2af <.literal16+0x7ff>
   DB  54,0,54                             ; add           %dh,%ss:(%rsi)
   DB  0,54                                ; add           %dh,(%rsi)
   DB  0,54                                ; add           %dh,(%rsi)
@@ -43950,9 +45349,9 @@
   DB  4,0                                 ; add           $0x0,%al
   DB  0,0                                 ; add           %al,(%rax)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            2cc2b <.literal8+0x1b>
+  DB  127,0                               ; jg            2e30b <.literal8+0x1b>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            2cc2f <.literal8+0x1f>
+  DB  127,0                               ; jg            2e30f <.literal8+0x1f>
   DB  0,0                                 ; add           %al,(%rax)
   DB  128,0,0                             ; addb          $0x0,(%rax)
   DB  0                                   ; .byte         0x0
@@ -44061,7 +45460,7 @@
   DB  102,15,110,73,4                     ; movd          0x4(%ecx),%xmm1
   DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,88,136,119,207,0,0               ; addps         0xcf77(%eax),%xmm1
+  DB  15,88,136,151,216,0,0               ; addps         0xd897(%eax),%xmm1
   DB  141,114,8                           ; lea           0x8(%edx),%esi
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
   DB  15,41,81,64                         ; movaps        %xmm2,0x40(%ecx)
@@ -44069,7 +45468,7 @@
   DB  15,41,81,32                         ; movaps        %xmm2,0x20(%ecx)
   DB  15,41,81,16                         ; movaps        %xmm2,0x10(%ecx)
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,144,135,207,0,0               ; movaps        0xcf87(%eax),%xmm2
+  DB  15,40,144,167,216,0,0               ; movaps        0xd8a7(%eax),%xmm2
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
   DB  86                                  ; push          %esi
   DB  81                                  ; push          %ecx
@@ -44094,15 +45493,15 @@
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  102,15,110,32                       ; movd          (%eax),%xmm4
   DB  102,15,112,228,0                    ; pshufd        $0x0,%xmm4,%xmm4
-  DB  102,15,254,161,32,207,0,0           ; paddd         0xcf20(%ecx),%xmm4
+  DB  102,15,254,161,64,216,0,0           ; paddd         0xd840(%ecx),%xmm4
   DB  102,15,110,104,4                    ; movd          0x4(%eax),%xmm5
   DB  102,15,112,253,0                    ; pshufd        $0x0,%xmm5,%xmm7
   DB  102,15,239,252                      ; pxor          %xmm4,%xmm7
-  DB  102,15,111,177,48,207,0,0           ; movdqa        0xcf30(%ecx),%xmm6
+  DB  102,15,111,177,80,216,0,0           ; movdqa        0xd850(%ecx),%xmm6
   DB  102,15,111,207                      ; movdqa        %xmm7,%xmm1
   DB  102,15,219,206                      ; pand          %xmm6,%xmm1
   DB  102,15,114,241,5                    ; pslld         $0x5,%xmm1
-  DB  102,15,111,129,80,207,0,0           ; movdqa        0xcf50(%ecx),%xmm0
+  DB  102,15,111,129,112,216,0,0          ; movdqa        0xd870(%ecx),%xmm0
   DB  102,15,111,239                      ; movdqa        %xmm7,%xmm5
   DB  102,15,219,232                      ; pand          %xmm0,%xmm5
   DB  102,15,114,245,2                    ; pslld         $0x2,%xmm5
@@ -44112,7 +45511,7 @@
   DB  102,15,219,196                      ; pand          %xmm4,%xmm0
   DB  102,15,254,192                      ; paddd         %xmm0,%xmm0
   DB  102,15,235,198                      ; por           %xmm6,%xmm0
-  DB  102,15,111,137,64,207,0,0           ; movdqa        0xcf40(%ecx),%xmm1
+  DB  102,15,111,137,96,216,0,0           ; movdqa        0xd860(%ecx),%xmm1
   DB  102,15,219,249                      ; pand          %xmm1,%xmm7
   DB  102,15,219,225                      ; pand          %xmm1,%xmm4
   DB  102,15,114,212,2                    ; psrld         $0x2,%xmm4
@@ -44123,8 +45522,8 @@
   DB  102,15,114,215,1                    ; psrld         $0x1,%xmm7
   DB  102,15,235,231                      ; por           %xmm7,%xmm4
   DB  15,91,196                           ; cvtdq2ps      %xmm4,%xmm0
-  DB  15,89,129,96,207,0,0                ; mulps         0xcf60(%ecx),%xmm0
-  DB  15,88,129,112,207,0,0               ; addps         0xcf70(%ecx),%xmm0
+  DB  15,89,129,128,216,0,0               ; mulps         0xd880(%ecx),%xmm0
+  DB  15,88,129,144,216,0,0               ; addps         0xd890(%ecx),%xmm0
   DB  243,15,16,14                        ; movss         (%esi),%xmm1
   DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
@@ -44188,7 +45587,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  141,81,4                            ; lea           0x4(%ecx),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,152,212,205,0,0               ; movaps        0xcdd4(%eax),%xmm3
+  DB  15,40,152,244,214,0,0               ; movaps        0xd6f4(%eax),%xmm3
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -44209,7 +45608,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  141,81,4                            ; lea           0x4(%ecx),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,128,164,205,0,0               ; movaps        0xcda4(%eax),%xmm0
+  DB  15,40,128,196,214,0,0               ; movaps        0xd6c4(%eax),%xmm0
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -44291,7 +45690,7 @@
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  15,40,97,64                         ; movaps        0x40(%ecx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
-  DB  15,40,170,245,204,0,0               ; movaps        0xccf5(%edx),%xmm5
+  DB  15,40,170,21,214,0,0                ; movaps        0xd615(%edx),%xmm5
   DB  15,92,235                           ; subps         %xmm3,%xmm5
   DB  15,40,113,16                        ; movaps        0x10(%ecx),%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
@@ -44328,7 +45727,7 @@
   DB  15,40,105,64                        ; movaps        0x40(%ecx),%xmm5
   DB  15,40,113,16                        ; movaps        0x10(%ecx),%xmm6
   DB  15,89,243                           ; mulps         %xmm3,%xmm6
-  DB  15,40,162,150,204,0,0               ; movaps        0xcc96(%edx),%xmm4
+  DB  15,40,162,182,213,0,0               ; movaps        0xd5b6(%edx),%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  15,88,198                           ; addps         %xmm6,%xmm0
@@ -44406,7 +45805,7 @@
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
-  DB  15,40,160,213,203,0,0               ; movaps        0xcbd5(%eax),%xmm4
+  DB  15,40,160,245,212,0,0               ; movaps        0xd4f5(%eax),%xmm4
   DB  15,92,98,64                         ; subps         0x40(%edx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
@@ -44430,7 +45829,7 @@
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
-  DB  15,40,160,157,203,0,0               ; movaps        0xcb9d(%eax),%xmm4
+  DB  15,40,160,189,212,0,0               ; movaps        0xd4bd(%eax),%xmm4
   DB  15,92,227                           ; subps         %xmm3,%xmm4
   DB  15,40,66,16                         ; movaps        0x10(%edx),%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
@@ -44458,7 +45857,7 @@
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
-  DB  15,40,160,86,203,0,0                ; movaps        0xcb56(%eax),%xmm4
+  DB  15,40,160,118,212,0,0               ; movaps        0xd476(%eax),%xmm4
   DB  15,92,227                           ; subps         %xmm3,%xmm4
   DB  15,40,106,16                        ; movaps        0x10(%edx),%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
@@ -44490,7 +45889,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
   DB  15,40,106,64                        ; movaps        0x40(%edx),%xmm5
-  DB  15,40,160,6,203,0,0                 ; movaps        0xcb06(%eax),%xmm4
+  DB  15,40,160,38,212,0,0                ; movaps        0xd426(%eax),%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  15,88,66,16                         ; addps         0x10(%edx),%xmm0
@@ -44542,7 +45941,7 @@
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  15,40,80,64                         ; movaps        0x40(%eax),%xmm2
   DB  15,41,85,200                        ; movaps        %xmm2,-0x38(%ebp)
-  DB  15,40,129,135,202,0,0               ; movaps        0xca87(%ecx),%xmm0
+  DB  15,40,129,167,211,0,0               ; movaps        0xd3a7(%ecx),%xmm0
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,92,242                           ; subps         %xmm2,%xmm6
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
@@ -44600,7 +45999,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
   DB  15,88,66,16                         ; addps         0x10(%edx),%xmm0
-  DB  15,40,160,225,201,0,0               ; movaps        0xc9e1(%eax),%xmm4
+  DB  15,40,160,1,211,0,0                 ; movaps        0xd301(%eax),%xmm4
   DB  15,93,196                           ; minps         %xmm4,%xmm0
   DB  15,88,74,32                         ; addps         0x20(%edx),%xmm1
   DB  15,93,204                           ; minps         %xmm4,%xmm1
@@ -44668,7 +46067,7 @@
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  15,40,105,64                        ; movaps        0x40(%ecx),%xmm5
-  DB  15,40,154,51,201,0,0                ; movaps        0xc933(%edx),%xmm3
+  DB  15,40,154,83,210,0,0                ; movaps        0xd253(%edx),%xmm3
   DB  15,40,243                           ; movaps        %xmm3,%xmm6
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
@@ -44728,7 +46127,7 @@
   DB  15,89,203                           ; mulps         %xmm3,%xmm1
   DB  15,95,209                           ; maxps         %xmm1,%xmm2
   DB  15,92,234                           ; subps         %xmm2,%xmm5
-  DB  15,40,138,203,200,0,0               ; movaps        0xc8cb(%edx),%xmm1
+  DB  15,40,138,235,209,0,0               ; movaps        0xd1eb(%edx),%xmm1
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -44775,7 +46174,7 @@
   DB  15,89,203                           ; mulps         %xmm3,%xmm1
   DB  15,93,209                           ; minps         %xmm1,%xmm2
   DB  15,92,234                           ; subps         %xmm2,%xmm5
-  DB  15,40,138,75,200,0,0                ; movaps        0xc84b(%edx),%xmm1
+  DB  15,40,138,107,209,0,0               ; movaps        0xd16b(%edx),%xmm1
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -44826,7 +46225,7 @@
   DB  15,93,214                           ; minps         %xmm6,%xmm2
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  15,92,226                           ; subps         %xmm2,%xmm4
-  DB  15,40,146,200,199,0,0               ; movaps        0xc7c8(%edx),%xmm2
+  DB  15,40,146,232,208,0,0               ; movaps        0xd0e8(%edx),%xmm2
   DB  15,92,211                           ; subps         %xmm3,%xmm2
   DB  15,89,213                           ; mulps         %xmm5,%xmm2
   DB  15,88,218                           ; addps         %xmm2,%xmm3
@@ -44868,7 +46267,7 @@
   DB  15,89,202                           ; mulps         %xmm2,%xmm1
   DB  15,88,201                           ; addps         %xmm1,%xmm1
   DB  15,92,233                           ; subps         %xmm1,%xmm5
-  DB  15,40,138,66,199,0,0                ; movaps        0xc742(%edx),%xmm1
+  DB  15,40,138,98,208,0,0                ; movaps        0xd062(%edx),%xmm1
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,89,73,64                         ; mulps         0x40(%ecx),%xmm1
   DB  15,88,217                           ; addps         %xmm1,%xmm3
@@ -44907,7 +46306,7 @@
   DB  15,93,198                           ; minps         %xmm6,%xmm0
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  15,92,232                           ; subps         %xmm0,%xmm5
-  DB  15,40,177,195,198,0,0               ; movaps        0xc6c3(%ecx),%xmm6
+  DB  15,40,177,227,207,0,0               ; movaps        0xcfe3(%ecx),%xmm6
   DB  15,40,198                           ; movaps        %xmm6,%xmm0
   DB  15,92,193                           ; subps         %xmm1,%xmm0
   DB  15,41,69,200                        ; movaps        %xmm0,-0x38(%ebp)
@@ -45022,7 +46421,7 @@
   DB  89                                  ; pop           %ecx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  15,40,72,16                         ; movaps        0x10(%eax),%xmm1
-  DB  15,40,145,58,197,0,0                ; movaps        0xc53a(%ecx),%xmm2
+  DB  15,40,145,90,206,0,0                ; movaps        0xce5a(%ecx),%xmm2
   DB  15,40,226                           ; movaps        %xmm2,%xmm4
   DB  15,92,214                           ; subps         %xmm6,%xmm2
   DB  15,40,250                           ; movaps        %xmm2,%xmm7
@@ -45137,7 +46536,7 @@
   DB  89                                  ; pop           %ecx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  15,40,72,64                         ; movaps        0x40(%eax),%xmm1
-  DB  15,40,145,217,195,0,0               ; movaps        0xc3d9(%ecx),%xmm2
+  DB  15,40,145,249,204,0,0               ; movaps        0xccf9(%ecx),%xmm2
   DB  15,40,234                           ; movaps        %xmm2,%xmm5
   DB  15,92,233                           ; subps         %xmm1,%xmm5
   DB  15,40,225                           ; movaps        %xmm1,%xmm4
@@ -45245,7 +46644,7 @@
   DB  89                                  ; pop           %ecx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  15,40,80,64                         ; movaps        0x40(%eax),%xmm2
-  DB  15,40,161,142,194,0,0               ; movaps        0xc28e(%ecx),%xmm4
+  DB  15,40,161,174,203,0,0               ; movaps        0xcbae(%ecx),%xmm4
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  15,92,242                           ; subps         %xmm2,%xmm6
   DB  15,40,206                           ; movaps        %xmm6,%xmm1
@@ -45371,17 +46770,17 @@
   DB  232,0,0,0,0                         ; call          f62 <_sk_softlight_sse2+0x52>
   DB  89                                  ; pop           %ecx
   DB  15,40,194                           ; movaps        %xmm2,%xmm0
-  DB  15,40,137,110,193,0,0               ; movaps        0xc16e(%ecx),%xmm1
+  DB  15,40,137,142,202,0,0               ; movaps        0xca8e(%ecx),%xmm1
   DB  15,41,141,104,255,255,255           ; movaps        %xmm1,-0x98(%ebp)
   DB  15,88,193                           ; addps         %xmm1,%xmm0
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
-  DB  15,40,185,254,192,0,0               ; movaps        0xc0fe(%ecx),%xmm7
+  DB  15,40,185,30,202,0,0                ; movaps        0xca1e(%ecx),%xmm7
   DB  15,41,125,232                       ; movaps        %xmm7,-0x18(%ebp)
   DB  15,92,250                           ; subps         %xmm2,%xmm7
   DB  15,82,242                           ; rsqrtps       %xmm2,%xmm6
   DB  15,83,206                           ; rcpps         %xmm6,%xmm1
   DB  15,92,202                           ; subps         %xmm2,%xmm1
-  DB  15,40,161,126,193,0,0               ; movaps        0xc17e(%ecx),%xmm4
+  DB  15,40,161,158,202,0,0               ; movaps        0xca9e(%ecx),%xmm4
   DB  15,41,101,152                       ; movaps        %xmm4,-0x68(%ebp)
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  15,88,208                           ; addps         %xmm0,%xmm2
@@ -45585,7 +46984,7 @@
   DB  15,40,226                           ; movaps        %xmm2,%xmm4
   DB  15,40,72,64                         ; movaps        0x40(%eax),%xmm1
   DB  15,41,77,216                        ; movaps        %xmm1,-0x28(%ebp)
-  DB  15,40,169,61,190,0,0                ; movaps        0xbe3d(%ecx),%xmm5
+  DB  15,40,169,93,199,0,0                ; movaps        0xc75d(%ecx),%xmm5
   DB  15,92,233                           ; subps         %xmm1,%xmm5
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
   DB  15,41,133,72,255,255,255            ; movaps        %xmm0,-0xb8(%ebp)
@@ -45618,17 +47017,17 @@
   DB  15,89,233                           ; mulps         %xmm1,%xmm5
   DB  15,41,173,120,255,255,255           ; movaps        %xmm5,-0x88(%ebp)
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,153,205,190,0,0               ; movaps        0xbecd(%ecx),%xmm3
+  DB  15,40,153,237,199,0,0               ; movaps        0xc7ed(%ecx),%xmm3
   DB  15,41,93,232                        ; movaps        %xmm3,-0x18(%ebp)
   DB  15,40,117,152                       ; movaps        -0x68(%ebp),%xmm6
   DB  15,40,206                           ; movaps        %xmm6,%xmm1
   DB  15,89,203                           ; mulps         %xmm3,%xmm1
-  DB  15,40,153,221,190,0,0               ; movaps        0xbedd(%ecx),%xmm3
+  DB  15,40,153,253,199,0,0               ; movaps        0xc7fd(%ecx),%xmm3
   DB  15,40,109,136                       ; movaps        -0x78(%ebp),%xmm5
   DB  15,40,253                           ; movaps        %xmm5,%xmm7
   DB  15,89,251                           ; mulps         %xmm3,%xmm7
   DB  15,88,249                           ; addps         %xmm1,%xmm7
-  DB  15,40,137,61,190,0,0                ; movaps        0xbe3d(%ecx),%xmm1
+  DB  15,40,137,93,199,0,0                ; movaps        0xc75d(%ecx),%xmm1
   DB  15,92,77,184                        ; subps         -0x48(%ebp),%xmm1
   DB  15,89,241                           ; mulps         %xmm1,%xmm6
   DB  15,41,117,152                       ; movaps        %xmm6,-0x68(%ebp)
@@ -45638,7 +47037,7 @@
   DB  15,40,77,200                        ; movaps        -0x38(%ebp),%xmm1
   DB  15,89,241                           ; mulps         %xmm1,%xmm6
   DB  15,41,117,168                       ; movaps        %xmm6,-0x58(%ebp)
-  DB  15,40,169,237,190,0,0               ; movaps        0xbeed(%ecx),%xmm5
+  DB  15,40,169,13,200,0,0                ; movaps        0xc80d(%ecx),%xmm5
   DB  15,89,205                           ; mulps         %xmm5,%xmm1
   DB  15,88,207                           ; addps         %xmm7,%xmm1
   DB  15,94,212                           ; divps         %xmm4,%xmm2
@@ -45829,17 +47228,17 @@
   DB  15,89,253                           ; mulps         %xmm5,%xmm7
   DB  232,0,0,0,0                         ; call          1577 <_sk_saturation_sse2+0xb0>
   DB  89                                  ; pop           %ecx
-  DB  15,40,153,121,187,0,0               ; movaps        0xbb79(%ecx),%xmm3
+  DB  15,40,153,153,196,0,0               ; movaps        0xc499(%ecx),%xmm3
   DB  15,41,157,88,255,255,255            ; movaps        %xmm3,-0xa8(%ebp)
   DB  15,40,77,184                        ; movaps        -0x48(%ebp),%xmm1
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,40,153,137,187,0,0               ; movaps        0xbb89(%ecx),%xmm3
+  DB  15,40,153,169,196,0,0               ; movaps        0xc4a9(%ecx),%xmm3
   DB  15,41,93,232                        ; movaps        %xmm3,-0x18(%ebp)
   DB  15,40,238                           ; movaps        %xmm6,%xmm5
   DB  15,89,235                           ; mulps         %xmm3,%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  15,40,129,233,186,0,0               ; movaps        0xbae9(%ecx),%xmm0
+  DB  15,40,129,9,196,0,0                 ; movaps        0xc409(%ecx),%xmm0
   DB  15,41,133,104,255,255,255           ; movaps        %xmm0,-0x98(%ebp)
   DB  15,92,69,216                        ; subps         -0x28(%ebp),%xmm0
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
@@ -45851,7 +47250,7 @@
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  15,41,141,120,255,255,255           ; movaps        %xmm1,-0x88(%ebp)
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
-  DB  15,40,129,153,187,0,0               ; movaps        0xbb99(%ecx),%xmm0
+  DB  15,40,129,185,196,0,0               ; movaps        0xc4b9(%ecx),%xmm0
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  15,88,205                           ; addps         %xmm5,%xmm1
   DB  15,40,109,168                       ; movaps        -0x58(%ebp),%xmm5
@@ -46008,16 +47407,16 @@
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  15,40,80,16                         ; movaps        0x10(%eax),%xmm2
   DB  15,40,72,32                         ; movaps        0x20(%eax),%xmm1
-  DB  15,40,177,30,185,0,0                ; movaps        0xb91e(%ecx),%xmm6
+  DB  15,40,177,62,194,0,0                ; movaps        0xc23e(%ecx),%xmm6
   DB  15,41,117,216                       ; movaps        %xmm6,-0x28(%ebp)
   DB  15,40,194                           ; movaps        %xmm2,%xmm0
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
-  DB  15,40,153,46,185,0,0                ; movaps        0xb92e(%ecx),%xmm3
+  DB  15,40,153,78,194,0,0                ; movaps        0xc24e(%ecx),%xmm3
   DB  15,41,93,184                        ; movaps        %xmm3,-0x48(%ebp)
   DB  15,40,241                           ; movaps        %xmm1,%xmm6
   DB  15,89,243                           ; mulps         %xmm3,%xmm6
   DB  15,88,240                           ; addps         %xmm0,%xmm6
-  DB  15,40,129,142,184,0,0               ; movaps        0xb88e(%ecx),%xmm0
+  DB  15,40,129,174,193,0,0               ; movaps        0xc1ae(%ecx),%xmm0
   DB  15,41,69,168                        ; movaps        %xmm0,-0x58(%ebp)
   DB  15,92,69,200                        ; subps         -0x38(%ebp),%xmm0
   DB  15,89,208                           ; mulps         %xmm0,%xmm2
@@ -46027,7 +47426,7 @@
   DB  15,40,80,48                         ; movaps        0x30(%eax),%xmm2
   DB  15,89,194                           ; mulps         %xmm2,%xmm0
   DB  15,41,133,72,255,255,255            ; movaps        %xmm0,-0xb8(%ebp)
-  DB  15,40,137,62,185,0,0                ; movaps        0xb93e(%ecx),%xmm1
+  DB  15,40,137,94,194,0,0                ; movaps        0xc25e(%ecx),%xmm1
   DB  15,89,209                           ; mulps         %xmm1,%xmm2
   DB  15,88,214                           ; addps         %xmm6,%xmm2
   DB  15,40,88,64                         ; movaps        0x40(%eax),%xmm3
@@ -46184,18 +47583,18 @@
   DB  232,0,0,0,0                         ; call          1a35 <_sk_luminosity_sse2+0x11>
   DB  89                                  ; pop           %ecx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
-  DB  15,40,161,187,182,0,0               ; movaps        0xb6bb(%ecx),%xmm4
+  DB  15,40,161,219,191,0,0               ; movaps        0xbfdb(%ecx),%xmm4
   DB  15,41,101,232                       ; movaps        %xmm4,-0x18(%ebp)
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
-  DB  15,40,161,203,182,0,0               ; movaps        0xb6cb(%ecx),%xmm4
+  DB  15,40,161,235,191,0,0               ; movaps        0xbfeb(%ecx),%xmm4
   DB  15,41,101,216                       ; movaps        %xmm4,-0x28(%ebp)
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
   DB  15,88,234                           ; addps         %xmm2,%xmm5
   DB  15,40,80,64                         ; movaps        0x40(%eax),%xmm2
   DB  15,41,85,184                        ; movaps        %xmm2,-0x48(%ebp)
-  DB  15,40,161,43,182,0,0                ; movaps        0xb62b(%ecx),%xmm4
+  DB  15,40,161,75,191,0,0                ; movaps        0xbf4b(%ecx),%xmm4
   DB  15,40,252                           ; movaps        %xmm4,%xmm7
   DB  15,92,250                           ; subps         %xmm2,%xmm7
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
@@ -46205,7 +47604,7 @@
   DB  15,89,254                           ; mulps         %xmm6,%xmm7
   DB  15,41,189,104,255,255,255           ; movaps        %xmm7,-0x98(%ebp)
   DB  15,40,214                           ; movaps        %xmm6,%xmm2
-  DB  15,40,137,219,182,0,0               ; movaps        0xb6db(%ecx),%xmm1
+  DB  15,40,137,251,191,0,0               ; movaps        0xbffb(%ecx),%xmm1
   DB  15,89,209                           ; mulps         %xmm1,%xmm2
   DB  15,88,213                           ; addps         %xmm5,%xmm2
   DB  15,40,104,16                        ; movaps        0x10(%eax),%xmm5
@@ -46379,7 +47778,7 @@
   DB  139,49                              ; mov           (%ecx),%esi
   DB  15,133,224,0,0,0                    ; jne           1da2 <_sk_srcover_rgba_8888_sse2+0x119>
   DB  243,15,111,20,178                   ; movdqu        (%edx,%esi,4),%xmm2
-  DB  102,15,111,167,128,180,0,0          ; movdqa        0xb480(%edi),%xmm4
+  DB  102,15,111,167,160,189,0,0          ; movdqa        0xbda0(%edi),%xmm4
   DB  102,15,111,234                      ; movdqa        %xmm2,%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  102,15,111,242                      ; movdqa        %xmm2,%xmm6
@@ -46390,9 +47789,9 @@
   DB  102,15,219,252                      ; pand          %xmm4,%xmm7
   DB  15,91,229                           ; cvtdq2ps      %xmm5,%xmm4
   DB  15,41,97,16                         ; movaps        %xmm4,0x10(%ecx)
-  DB  15,40,175,192,179,0,0               ; movaps        0xb3c0(%edi),%xmm5
+  DB  15,40,175,224,188,0,0               ; movaps        0xbce0(%edi),%xmm5
   DB  15,92,235                           ; subps         %xmm3,%xmm5
-  DB  15,40,143,144,180,0,0               ; movaps        0xb490(%edi),%xmm1
+  DB  15,40,143,176,189,0,0               ; movaps        0xbdb0(%edi),%xmm1
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,88,224                           ; addps         %xmm0,%xmm4
@@ -46502,7 +47901,7 @@
   DB  232,0,0,0,0                         ; call          1e4c <_sk_clamp_1_sse2+0xb>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,40,160,20,178,0,0                ; movaps        0xb214(%eax),%xmm4
+  DB  15,40,160,52,187,0,0                ; movaps        0xbb34(%eax),%xmm4
   DB  15,93,196                           ; minps         %xmm4,%xmm0
   DB  15,93,204                           ; minps         %xmm4,%xmm1
   DB  15,93,212                           ; minps         %xmm4,%xmm2
@@ -46524,7 +47923,7 @@
   DB  232,0,0,0,0                         ; call          1e7f <_sk_clamp_a_sse2+0xb>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,93,152,225,177,0,0               ; minps         0xb1e1(%eax),%xmm3
+  DB  15,93,152,1,187,0,0                 ; minps         0xbb01(%eax),%xmm3
   DB  15,93,195                           ; minps         %xmm3,%xmm0
   DB  15,93,203                           ; minps         %xmm3,%xmm1
   DB  15,93,211                           ; minps         %xmm3,%xmm2
@@ -46547,7 +47946,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,85,8                            ; mov           0x8(%ebp),%edx
   DB  15,40,98,64                         ; movaps        0x40(%edx),%xmm4
-  DB  15,93,160,177,177,0,0               ; minps         0xb1b1(%eax),%xmm4
+  DB  15,93,160,209,186,0,0               ; minps         0xbad1(%eax),%xmm4
   DB  15,41,98,64                         ; movaps        %xmm4,0x40(%edx)
   DB  15,40,106,16                        ; movaps        0x10(%edx),%xmm5
   DB  15,40,114,32                        ; movaps        0x20(%edx),%xmm6
@@ -46615,7 +48014,7 @@
   DB  232,0,0,0,0                         ; call          1f5a <_sk_invert_sse2+0xb>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,40,160,6,177,0,0                 ; movaps        0xb106(%eax),%xmm4
+  DB  15,40,160,38,186,0,0                ; movaps        0xba26(%eax),%xmm4
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,92,232                           ; subps         %xmm0,%xmm5
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
@@ -46727,10 +48126,10 @@
   DB  232,0,0,0,0                         ; call          2058 <_sk_unpremul_sse2+0xb>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,40,160,8,176,0,0                 ; movaps        0xb008(%eax),%xmm4
+  DB  15,40,160,40,185,0,0                ; movaps        0xb928(%eax),%xmm4
   DB  15,94,227                           ; divps         %xmm3,%xmm4
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
-  DB  15,194,168,232,176,0,0,1            ; cmpltps       0xb0e8(%eax),%xmm5
+  DB  15,194,168,8,186,0,0,1              ; cmpltps       0xba08(%eax),%xmm5
   DB  15,84,236                           ; andps         %xmm4,%xmm5
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
   DB  15,89,205                           ; mulps         %xmm5,%xmm1
@@ -46757,19 +48156,19 @@
   DB  88                                  ; pop           %eax
   DB  15,40,202                           ; movaps        %xmm2,%xmm1
   DB  15,89,201                           ; mulps         %xmm1,%xmm1
-  DB  15,40,160,74,176,0,0                ; movaps        0xb04a(%eax),%xmm4
+  DB  15,40,160,106,185,0,0               ; movaps        0xb96a(%eax),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
-  DB  15,40,176,186,176,0,0               ; movaps        0xb0ba(%eax),%xmm6
+  DB  15,40,176,218,185,0,0               ; movaps        0xb9da(%eax),%xmm6
   DB  15,41,117,216                       ; movaps        %xmm6,-0x28(%ebp)
   DB  15,88,198                           ; addps         %xmm6,%xmm0
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,176,170,176,0,0               ; movaps        0xb0aa(%eax),%xmm6
+  DB  15,40,176,202,185,0,0               ; movaps        0xb9ca(%eax),%xmm6
   DB  15,40,202                           ; movaps        %xmm2,%xmm1
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
-  DB  15,40,184,202,176,0,0               ; movaps        0xb0ca(%eax),%xmm7
+  DB  15,40,184,234,185,0,0               ; movaps        0xb9ea(%eax),%xmm7
   DB  15,41,125,232                       ; movaps        %xmm7,-0x18(%ebp)
   DB  15,88,199                           ; addps         %xmm7,%xmm0
-  DB  15,40,184,218,176,0,0               ; movaps        0xb0da(%eax),%xmm7
+  DB  15,40,184,250,185,0,0               ; movaps        0xb9fa(%eax),%xmm7
   DB  15,194,215,1                        ; cmpltps       %xmm7,%xmm2
   DB  15,84,202                           ; andps         %xmm2,%xmm1
   DB  15,85,208                           ; andnps        %xmm0,%xmm2
@@ -46827,18 +48226,18 @@
   DB  15,40,72,16                         ; movaps        0x10(%eax),%xmm1
   DB  15,40,233                           ; movaps        %xmm1,%xmm5
   DB  15,89,237                           ; mulps         %xmm5,%xmm5
-  DB  15,40,161,111,175,0,0               ; movaps        0xaf6f(%ecx),%xmm4
+  DB  15,40,161,143,184,0,0               ; movaps        0xb88f(%ecx),%xmm4
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
-  DB  15,40,177,223,175,0,0               ; movaps        0xafdf(%ecx),%xmm6
+  DB  15,40,177,255,184,0,0               ; movaps        0xb8ff(%ecx),%xmm6
   DB  15,88,198                           ; addps         %xmm6,%xmm0
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
-  DB  15,40,169,207,175,0,0               ; movaps        0xafcf(%ecx),%xmm5
+  DB  15,40,169,239,184,0,0               ; movaps        0xb8ef(%ecx),%xmm5
   DB  15,40,209                           ; movaps        %xmm1,%xmm2
   DB  15,89,213                           ; mulps         %xmm5,%xmm2
-  DB  15,40,185,239,175,0,0               ; movaps        0xafef(%ecx),%xmm7
+  DB  15,40,185,15,185,0,0                ; movaps        0xb90f(%ecx),%xmm7
   DB  15,88,199                           ; addps         %xmm7,%xmm0
-  DB  15,40,153,255,175,0,0               ; movaps        0xafff(%ecx),%xmm3
+  DB  15,40,153,31,185,0,0                ; movaps        0xb91f(%ecx),%xmm3
   DB  15,194,203,1                        ; cmpltps       %xmm3,%xmm1
   DB  15,84,209                           ; andps         %xmm1,%xmm2
   DB  15,85,200                           ; andnps        %xmm0,%xmm1
@@ -46898,25 +48297,25 @@
   DB  232,0,0,0,0                         ; call          226c <_sk_to_srgb_sse2+0x18>
   DB  88                                  ; pop           %eax
   DB  15,82,218                           ; rsqrtps       %xmm2,%xmm3
-  DB  15,40,160,52,175,0,0                ; movaps        0xaf34(%eax),%xmm4
+  DB  15,40,160,84,184,0,0                ; movaps        0xb854(%eax),%xmm4
   DB  15,40,235                           ; movaps        %xmm3,%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
-  DB  15,40,128,68,175,0,0                ; movaps        0xaf44(%eax),%xmm0
+  DB  15,40,128,100,184,0,0               ; movaps        0xb864(%eax),%xmm0
   DB  15,41,69,232                        ; movaps        %xmm0,-0x18(%ebp)
   DB  15,88,232                           ; addps         %xmm0,%xmm5
   DB  15,89,235                           ; mulps         %xmm3,%xmm5
-  DB  15,40,128,84,175,0,0                ; movaps        0xaf54(%eax),%xmm0
+  DB  15,40,128,116,184,0,0               ; movaps        0xb874(%eax),%xmm0
   DB  15,41,69,200                        ; movaps        %xmm0,-0x38(%ebp)
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  15,40,128,100,175,0,0               ; movaps        0xaf64(%eax),%xmm0
+  DB  15,40,128,132,184,0,0               ; movaps        0xb884(%eax),%xmm0
   DB  15,41,69,216                        ; movaps        %xmm0,-0x28(%ebp)
   DB  15,88,216                           ; addps         %xmm0,%xmm3
   DB  15,83,195                           ; rcpps         %xmm3,%xmm0
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
-  DB  15,40,168,36,175,0,0                ; movaps        0xaf24(%eax),%xmm5
+  DB  15,40,168,68,184,0,0                ; movaps        0xb844(%eax),%xmm5
   DB  15,40,202                           ; movaps        %xmm2,%xmm1
   DB  15,89,205                           ; mulps         %xmm5,%xmm1
-  DB  15,40,152,116,175,0,0               ; movaps        0xaf74(%eax),%xmm3
+  DB  15,40,152,148,184,0,0               ; movaps        0xb894(%eax),%xmm3
   DB  15,194,211,1                        ; cmpltps       %xmm3,%xmm2
   DB  15,84,202                           ; andps         %xmm2,%xmm1
   DB  15,85,208                           ; andnps        %xmm0,%xmm2
@@ -46985,14 +48384,14 @@
   DB  15,40,195                           ; movaps        %xmm3,%xmm0
   DB  15,92,193                           ; subps         %xmm1,%xmm0
   DB  15,41,69,216                        ; movaps        %xmm0,-0x28(%ebp)
-  DB  15,40,136,253,172,0,0               ; movaps        0xacfd(%eax),%xmm1
+  DB  15,40,136,29,182,0,0                ; movaps        0xb61d(%eax),%xmm1
   DB  15,94,200                           ; divps         %xmm0,%xmm1
   DB  15,40,199                           ; movaps        %xmm7,%xmm0
   DB  15,92,194                           ; subps         %xmm2,%xmm0
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
   DB  15,40,239                           ; movaps        %xmm7,%xmm5
   DB  15,194,234,1                        ; cmpltps       %xmm2,%xmm5
-  DB  15,84,168,141,174,0,0               ; andps         0xae8d(%eax),%xmm5
+  DB  15,84,168,173,183,0,0               ; andps         0xb7ad(%eax),%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
   DB  15,40,195                           ; movaps        %xmm3,%xmm0
   DB  15,194,196,0                        ; cmpeqps       %xmm4,%xmm0
@@ -47002,9 +48401,9 @@
   DB  15,92,231                           ; subps         %xmm7,%xmm4
   DB  15,89,209                           ; mulps         %xmm1,%xmm2
   DB  15,89,225                           ; mulps         %xmm1,%xmm4
-  DB  15,40,136,157,174,0,0               ; movaps        0xae9d(%eax),%xmm1
+  DB  15,40,136,189,183,0,0               ; movaps        0xb7bd(%eax),%xmm1
   DB  15,88,209                           ; addps         %xmm1,%xmm2
-  DB  15,88,160,173,174,0,0               ; addps         0xaead(%eax),%xmm4
+  DB  15,88,160,205,183,0,0               ; addps         0xb7cd(%eax),%xmm4
   DB  15,84,214                           ; andps         %xmm6,%xmm2
   DB  15,85,244                           ; andnps        %xmm4,%xmm6
   DB  15,86,242                           ; orps          %xmm2,%xmm6
@@ -47016,7 +48415,7 @@
   DB  15,194,226,4                        ; cmpneqps      %xmm2,%xmm4
   DB  15,92,203                           ; subps         %xmm3,%xmm1
   DB  15,88,218                           ; addps         %xmm2,%xmm3
-  DB  15,40,168,237,172,0,0               ; movaps        0xaced(%eax),%xmm5
+  DB  15,40,168,13,182,0,0                ; movaps        0xb60d(%eax),%xmm5
   DB  15,92,202                           ; subps         %xmm2,%xmm1
   DB  15,40,211                           ; movaps        %xmm3,%xmm2
   DB  15,89,213                           ; mulps         %xmm5,%xmm2
@@ -47026,7 +48425,7 @@
   DB  15,86,233                           ; orps          %xmm1,%xmm5
   DB  15,40,77,216                        ; movaps        -0x28(%ebp),%xmm1
   DB  15,94,205                           ; divps         %xmm5,%xmm1
-  DB  15,89,128,189,174,0,0               ; mulps         0xaebd(%eax),%xmm0
+  DB  15,89,128,221,183,0,0               ; mulps         0xb7dd(%eax),%xmm0
   DB  15,84,196                           ; andps         %xmm4,%xmm0
   DB  15,84,204                           ; andps         %xmm4,%xmm1
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
@@ -47049,7 +48448,7 @@
   DB  15,41,69,184                        ; movaps        %xmm0,-0x48(%ebp)
   DB  232,0,0,0,0                         ; call          245e <_sk_hsl_to_rgb_sse2+0x19>
   DB  88                                  ; pop           %eax
-  DB  15,40,160,242,171,0,0               ; movaps        0xabf2(%eax),%xmm4
+  DB  15,40,160,18,181,0,0                ; movaps        0xb512(%eax),%xmm4
   DB  15,40,220                           ; movaps        %xmm4,%xmm3
   DB  15,41,165,104,255,255,255           ; movaps        %xmm4,-0x98(%ebp)
   DB  15,194,218,2                        ; cmpleps       %xmm2,%xmm3
@@ -47062,13 +48461,13 @@
   DB  15,84,243                           ; andps         %xmm3,%xmm6
   DB  15,85,217                           ; andnps        %xmm1,%xmm3
   DB  15,86,222                           ; orps          %xmm6,%xmm3
-  DB  15,40,168,210,173,0,0               ; movaps        0xadd2(%eax),%xmm5
+  DB  15,40,168,242,182,0,0               ; movaps        0xb6f2(%eax),%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
   DB  243,15,91,205                       ; cvttps2dq     %xmm5,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,194,193,1                        ; cmpltps       %xmm1,%xmm0
-  DB  15,40,176,2,172,0,0                 ; movaps        0xac02(%eax),%xmm6
+  DB  15,40,176,34,181,0,0                ; movaps        0xb522(%eax),%xmm6
   DB  15,41,117,136                       ; movaps        %xmm6,-0x78(%ebp)
   DB  15,84,198                           ; andps         %xmm6,%xmm0
   DB  15,92,200                           ; subps         %xmm0,%xmm1
@@ -47078,19 +48477,19 @@
   DB  15,41,93,216                        ; movaps        %xmm3,-0x28(%ebp)
   DB  15,88,210                           ; addps         %xmm2,%xmm2
   DB  15,92,211                           ; subps         %xmm3,%xmm2
-  DB  15,40,184,226,173,0,0               ; movaps        0xade2(%eax),%xmm7
+  DB  15,40,184,2,183,0,0                 ; movaps        0xb702(%eax),%xmm7
   DB  15,41,189,88,255,255,255            ; movaps        %xmm7,-0xa8(%ebp)
   DB  15,194,253,2                        ; cmpleps       %xmm5,%xmm7
   DB  15,92,218                           ; subps         %xmm2,%xmm3
   DB  15,41,93,200                        ; movaps        %xmm3,-0x38(%ebp)
   DB  15,40,204                           ; movaps        %xmm4,%xmm1
   DB  15,194,205,2                        ; cmpleps       %xmm5,%xmm1
-  DB  15,40,160,194,173,0,0               ; movaps        0xadc2(%eax),%xmm4
+  DB  15,40,160,226,182,0,0               ; movaps        0xb6e2(%eax),%xmm4
   DB  15,41,165,120,255,255,255           ; movaps        %xmm4,-0x88(%ebp)
   DB  15,194,229,2                        ; cmpleps       %xmm5,%xmm4
-  DB  15,40,176,146,173,0,0               ; movaps        0xad92(%eax),%xmm6
+  DB  15,40,176,178,182,0,0               ; movaps        0xb6b2(%eax),%xmm6
   DB  15,89,238                           ; mulps         %xmm6,%xmm5
-  DB  15,40,128,178,173,0,0               ; movaps        0xadb2(%eax),%xmm0
+  DB  15,40,128,210,182,0,0               ; movaps        0xb6d2(%eax),%xmm0
   DB  15,41,69,168                        ; movaps        %xmm0,-0x58(%ebp)
   DB  15,92,197                           ; subps         %xmm5,%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
@@ -47116,18 +48515,18 @@
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,203                           ; movaps        %xmm3,%xmm1
   DB  15,194,200,1                        ; cmpltps       %xmm0,%xmm1
-  DB  15,84,136,2,172,0,0                 ; andps         0xac02(%eax),%xmm1
+  DB  15,84,136,34,181,0,0                ; andps         0xb522(%eax),%xmm1
   DB  15,92,193                           ; subps         %xmm1,%xmm0
   DB  15,40,203                           ; movaps        %xmm3,%xmm1
   DB  15,92,200                           ; subps         %xmm0,%xmm1
-  DB  15,40,128,226,173,0,0               ; movaps        0xade2(%eax),%xmm0
+  DB  15,40,128,2,183,0,0                 ; movaps        0xb702(%eax),%xmm0
   DB  15,194,193,2                        ; cmpleps       %xmm1,%xmm0
-  DB  15,40,152,242,171,0,0               ; movaps        0xabf2(%eax),%xmm3
+  DB  15,40,152,18,181,0,0                ; movaps        0xb512(%eax),%xmm3
   DB  15,194,217,2                        ; cmpleps       %xmm1,%xmm3
-  DB  15,40,160,194,173,0,0               ; movaps        0xadc2(%eax),%xmm4
+  DB  15,40,160,226,182,0,0               ; movaps        0xb6e2(%eax),%xmm4
   DB  15,194,225,2                        ; cmpleps       %xmm1,%xmm4
   DB  15,89,206                           ; mulps         %xmm6,%xmm1
-  DB  15,40,168,178,173,0,0               ; movaps        0xadb2(%eax),%xmm5
+  DB  15,40,168,210,182,0,0               ; movaps        0xb6d2(%eax),%xmm5
   DB  15,92,233                           ; subps         %xmm1,%xmm5
   DB  15,89,239                           ; mulps         %xmm7,%xmm5
   DB  15,88,234                           ; addps         %xmm2,%xmm5
@@ -47147,7 +48546,7 @@
   DB  15,40,125,232                       ; movaps        -0x18(%ebp),%xmm7
   DB  15,85,252                           ; andnps        %xmm4,%xmm7
   DB  15,40,93,184                        ; movaps        -0x48(%ebp),%xmm3
-  DB  15,88,152,242,173,0,0               ; addps         0xadf2(%eax),%xmm3
+  DB  15,88,152,18,183,0,0                ; addps         0xb712(%eax),%xmm3
   DB  243,15,91,195                       ; cvttps2dq     %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
   DB  15,40,203                           ; movaps        %xmm3,%xmm1
@@ -47245,9 +48644,9 @@
   DB  102,15,110,36,62                    ; movd          (%esi,%edi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,162,72,170,0,0           ; pand          0xaa48(%edx),%xmm4
+  DB  102,15,219,162,104,179,0,0          ; pand          0xb368(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,136,171,0,0               ; mulps         0xab88(%edx),%xmm4
+  DB  15,89,162,168,180,0,0               ; mulps         0xb4a8(%edx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
@@ -47311,17 +48710,17 @@
   DB  15,133,157,0,0,0                    ; jne           2850 <_sk_scale_565_sse2+0xd8>
   DB  243,15,126,52,126                   ; movq          (%esi,%edi,2),%xmm6
   DB  102,15,97,240                       ; punpcklwd     %xmm0,%xmm6
-  DB  102,15,111,162,222,170,0,0          ; movdqa        0xaade(%edx),%xmm4
+  DB  102,15,111,162,254,179,0,0          ; movdqa        0xb3fe(%edx),%xmm4
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
   DB  15,91,212                           ; cvtdq2ps      %xmm4,%xmm2
-  DB  15,89,146,238,170,0,0               ; mulps         0xaaee(%edx),%xmm2
-  DB  102,15,111,170,254,170,0,0          ; movdqa        0xaafe(%edx),%xmm5
+  DB  15,89,146,14,180,0,0                ; mulps         0xb40e(%edx),%xmm2
+  DB  102,15,111,170,30,180,0,0           ; movdqa        0xb41e(%edx),%xmm5
   DB  102,15,219,238                      ; pand          %xmm6,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,14,171,0,0                ; mulps         0xab0e(%edx),%xmm5
-  DB  102,15,219,178,30,171,0,0           ; pand          0xab1e(%edx),%xmm6
+  DB  15,89,170,46,180,0,0                ; mulps         0xb42e(%edx),%xmm5
+  DB  102,15,219,178,62,180,0,0           ; pand          0xb43e(%edx),%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,178,46,171,0,0                ; mulps         0xab2e(%edx),%xmm6
+  DB  15,89,178,78,180,0,0                ; mulps         0xb44e(%edx),%xmm6
   DB  15,40,253                           ; movaps        %xmm5,%xmm7
   DB  15,93,254                           ; minps         %xmm6,%xmm7
   DB  15,40,226                           ; movaps        %xmm2,%xmm4
@@ -47431,9 +48830,9 @@
   DB  102,15,110,36,62                    ; movd          (%esi,%edi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,162,32,168,0,0           ; pand          0xa820(%edx),%xmm4
+  DB  102,15,219,162,64,177,0,0           ; pand          0xb140(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,96,169,0,0                ; mulps         0xa960(%edx),%xmm4
+  DB  15,89,162,128,178,0,0               ; mulps         0xb280(%edx),%xmm4
   DB  15,40,105,16                        ; movaps        0x10(%ecx),%xmm5
   DB  15,40,113,32                        ; movaps        0x20(%ecx),%xmm6
   DB  15,92,197                           ; subps         %xmm5,%xmm0
@@ -47509,17 +48908,17 @@
   DB  15,133,211,0,0,0                    ; jne           2ade <_sk_lerp_565_sse2+0x113>
   DB  243,15,126,44,126                   ; movq          (%esi,%edi,2),%xmm5
   DB  102,15,97,232                       ; punpcklwd     %xmm0,%xmm5
-  DB  102,15,111,162,134,168,0,0          ; movdqa        0xa886(%edx),%xmm4
+  DB  102,15,111,162,166,177,0,0          ; movdqa        0xb1a6(%edx),%xmm4
   DB  102,15,219,229                      ; pand          %xmm5,%xmm4
   DB  15,91,204                           ; cvtdq2ps      %xmm4,%xmm1
-  DB  15,89,138,150,168,0,0               ; mulps         0xa896(%edx),%xmm1
-  DB  102,15,111,178,166,168,0,0          ; movdqa        0xa8a6(%edx),%xmm6
+  DB  15,89,138,182,177,0,0               ; mulps         0xb1b6(%edx),%xmm1
+  DB  102,15,111,178,198,177,0,0          ; movdqa        0xb1c6(%edx),%xmm6
   DB  102,15,219,245                      ; pand          %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,89,178,182,168,0,0               ; mulps         0xa8b6(%edx),%xmm6
-  DB  102,15,219,170,198,168,0,0          ; pand          0xa8c6(%edx),%xmm5
+  DB  15,89,178,214,177,0,0               ; mulps         0xb1d6(%edx),%xmm6
+  DB  102,15,219,170,230,177,0,0          ; pand          0xb1e6(%edx),%xmm5
   DB  15,91,213                           ; cvtdq2ps      %xmm5,%xmm2
-  DB  15,89,146,214,168,0,0               ; mulps         0xa8d6(%edx),%xmm2
+  DB  15,89,146,246,177,0,0               ; mulps         0xb1f6(%edx),%xmm2
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  15,41,93,200                        ; movaps        %xmm3,-0x38(%ebp)
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
@@ -47605,7 +49004,7 @@
   DB  139,30                              ; mov           (%esi),%ebx
   DB  15,133,22,1,0,0                     ; jne           2c62 <_sk_load_tables_sse2+0x13c>
   DB  243,15,111,28,187                   ; movdqu        (%ebx,%edi,4),%xmm3
-  DB  102,15,111,146,236,165,0,0          ; movdqa        0xa5ec(%edx),%xmm2
+  DB  102,15,111,146,12,175,0,0           ; movdqa        0xaf0c(%edx),%xmm2
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
@@ -47661,7 +49060,7 @@
   DB  102,15,20,215                       ; unpcklpd      %xmm7,%xmm2
   DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,154,44,167,0,0                ; mulps         0xa72c(%edx),%xmm3
+  DB  15,89,154,76,176,0,0                ; mulps         0xb04c(%edx),%xmm3
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  137,194                             ; mov           %eax,%edx
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -47734,7 +49133,7 @@
   DB  242,15,112,192,232                  ; pshuflw       $0xe8,%xmm0,%xmm0
   DB  243,15,112,192,232                  ; pshufhw       $0xe8,%xmm0,%xmm0
   DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
-  DB  102,15,111,146,43,166,0,0           ; movdqa        0xa62b(%edx),%xmm2
+  DB  102,15,111,146,75,175,0,0           ; movdqa        0xaf4b(%edx),%xmm2
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,97,196                       ; punpcklwd     %xmm4,%xmm0
@@ -47806,7 +49205,7 @@
   DB  243,15,126,93,232                   ; movq          -0x18(%ebp),%xmm3
   DB  102,15,97,220                       ; punpcklwd     %xmm4,%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,154,59,166,0,0                ; mulps         0xa63b(%edx),%xmm3
+  DB  15,89,154,91,175,0,0                ; mulps         0xaf5b(%edx),%xmm3
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  137,194                             ; mov           %eax,%edx
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -47877,7 +49276,7 @@
   DB  242,15,112,194,232                  ; pshuflw       $0xe8,%xmm2,%xmm0
   DB  243,15,112,192,232                  ; pshufhw       $0xe8,%xmm0,%xmm0
   DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
-  DB  102,15,111,146,4,164,0,0            ; movdqa        0xa404(%edx),%xmm2
+  DB  102,15,111,146,36,173,0,0           ; movdqa        0xad24(%edx),%xmm2
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,97,195                       ; punpcklwd     %xmm3,%xmm0
@@ -47942,7 +49341,7 @@
   DB  137,198                             ; mov           %eax,%esi
   DB  141,70,8                            ; lea           0x8(%esi),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,154,148,161,0,0               ; movaps        0xa194(%edx),%xmm3
+  DB  15,40,154,180,170,0,0               ; movaps        0xaab4(%edx),%xmm3
   DB  80                                  ; push          %eax
   DB  81                                  ; push          %ecx
   DB  255,86,4                            ; call          *0x4(%esi)
@@ -47983,7 +49382,7 @@
   DB  90                                  ; pop           %edx
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  139,8                               ; mov           (%eax),%ecx
-  DB  15,40,162,55,160,0,0                ; movaps        0xa037(%edx),%xmm4
+  DB  15,40,162,87,169,0,0                ; movaps        0xa957(%edx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  102,15,112,232,78                   ; pshufd        $0x4e,%xmm0,%xmm5
@@ -48005,7 +49404,7 @@
   DB  102,15,110,192                      ; movd          %eax,%xmm0
   DB  139,65,4                            ; mov           0x4(%ecx),%eax
   DB  102,15,196,195,1                    ; pinsrw        $0x1,%ebx,%xmm0
-  DB  15,40,170,103,161,0,0               ; movaps        0xa167(%edx),%xmm5
+  DB  15,40,170,135,170,0,0               ; movaps        0xaa87(%edx),%xmm5
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
   DB  102,15,112,241,78                   ; pshufd        $0x4e,%xmm1,%xmm6
@@ -48133,7 +49532,7 @@
   DB  102,15,110,199                      ; movd          %edi,%xmm0
   DB  139,121,4                           ; mov           0x4(%ecx),%edi
   DB  102,15,196,198,1                    ; pinsrw        $0x1,%esi,%xmm0
-  DB  15,40,170,107,159,0,0               ; movaps        0x9f6b(%edx),%xmm5
+  DB  15,40,170,139,168,0,0               ; movaps        0xa88b(%edx),%xmm5
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
   DB  102,15,91,201                       ; cvtps2dq      %xmm1,%xmm1
   DB  102,15,112,241,78                   ; pshufd        $0x4e,%xmm1,%xmm6
@@ -48379,15 +49778,15 @@
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,88,245                           ; addps         %xmm5,%xmm6
   DB  15,91,238                           ; cvtdq2ps      %xmm6,%xmm5
-  DB  15,89,170,43,157,0,0                ; mulps         0x9d2b(%edx),%xmm5
-  DB  15,84,178,59,157,0,0                ; andps         0x9d3b(%edx),%xmm6
-  DB  15,86,178,139,154,0,0               ; orps          0x9a8b(%edx),%xmm6
-  DB  15,88,170,75,157,0,0                ; addps         0x9d4b(%edx),%xmm5
-  DB  15,40,186,91,157,0,0                ; movaps        0x9d5b(%edx),%xmm7
+  DB  15,89,170,75,166,0,0                ; mulps         0xa64b(%edx),%xmm5
+  DB  15,84,178,91,166,0,0                ; andps         0xa65b(%edx),%xmm6
+  DB  15,86,178,171,163,0,0               ; orps          0xa3ab(%edx),%xmm6
+  DB  15,88,170,107,166,0,0               ; addps         0xa66b(%edx),%xmm5
+  DB  15,40,186,123,166,0,0               ; movaps        0xa67b(%edx),%xmm7
   DB  15,89,254                           ; mulps         %xmm6,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,88,178,107,157,0,0               ; addps         0x9d6b(%edx),%xmm6
-  DB  15,40,186,123,157,0,0               ; movaps        0x9d7b(%edx),%xmm7
+  DB  15,88,178,139,166,0,0               ; addps         0xa68b(%edx),%xmm6
+  DB  15,40,186,155,166,0,0               ; movaps        0xa69b(%edx),%xmm7
   DB  15,94,254                           ; divps         %xmm6,%xmm7
   DB  243,15,16,49                        ; movss         (%ecx),%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
@@ -48397,21 +49796,21 @@
   DB  15,91,254                           ; cvtdq2ps      %xmm6,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
-  DB  15,40,178,155,154,0,0               ; movaps        0x9a9b(%edx),%xmm6
+  DB  15,40,178,187,163,0,0               ; movaps        0xa3bb(%edx),%xmm6
   DB  15,84,198                           ; andps         %xmm6,%xmm0
   DB  15,92,248                           ; subps         %xmm0,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,199                           ; subps         %xmm7,%xmm0
-  DB  15,88,170,139,157,0,0               ; addps         0x9d8b(%edx),%xmm5
-  DB  15,40,186,155,157,0,0               ; movaps        0x9d9b(%edx),%xmm7
+  DB  15,88,170,171,166,0,0               ; addps         0xa6ab(%edx),%xmm5
+  DB  15,40,186,187,166,0,0               ; movaps        0xa6bb(%edx),%xmm7
   DB  15,89,248                           ; mulps         %xmm0,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,40,186,171,157,0,0               ; movaps        0x9dab(%edx),%xmm7
+  DB  15,40,186,203,166,0,0               ; movaps        0xa6cb(%edx),%xmm7
   DB  15,92,248                           ; subps         %xmm0,%xmm7
-  DB  15,40,130,187,157,0,0               ; movaps        0x9dbb(%edx),%xmm0
+  DB  15,40,130,219,166,0,0               ; movaps        0xa6db(%edx),%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  15,88,197                           ; addps         %xmm5,%xmm0
-  DB  15,89,130,203,157,0,0               ; mulps         0x9dcb(%edx),%xmm0
+  DB  15,89,130,235,166,0,0               ; mulps         0xa6eb(%edx),%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  243,15,16,105,20                    ; movss         0x14(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
@@ -48463,15 +49862,15 @@
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,88,245                           ; addps         %xmm5,%xmm6
   DB  15,91,238                           ; cvtdq2ps      %xmm6,%xmm5
-  DB  15,89,170,244,155,0,0               ; mulps         0x9bf4(%edx),%xmm5
-  DB  15,84,178,4,156,0,0                 ; andps         0x9c04(%edx),%xmm6
-  DB  15,86,178,84,153,0,0                ; orps          0x9954(%edx),%xmm6
-  DB  15,88,170,20,156,0,0                ; addps         0x9c14(%edx),%xmm5
-  DB  15,40,186,36,156,0,0                ; movaps        0x9c24(%edx),%xmm7
+  DB  15,89,170,20,165,0,0                ; mulps         0xa514(%edx),%xmm5
+  DB  15,84,178,36,165,0,0                ; andps         0xa524(%edx),%xmm6
+  DB  15,86,178,116,162,0,0               ; orps          0xa274(%edx),%xmm6
+  DB  15,88,170,52,165,0,0                ; addps         0xa534(%edx),%xmm5
+  DB  15,40,186,68,165,0,0                ; movaps        0xa544(%edx),%xmm7
   DB  15,89,254                           ; mulps         %xmm6,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,88,178,52,156,0,0                ; addps         0x9c34(%edx),%xmm6
-  DB  15,40,186,68,156,0,0                ; movaps        0x9c44(%edx),%xmm7
+  DB  15,88,178,84,165,0,0                ; addps         0xa554(%edx),%xmm6
+  DB  15,40,186,100,165,0,0               ; movaps        0xa564(%edx),%xmm7
   DB  15,94,254                           ; divps         %xmm6,%xmm7
   DB  243,15,16,49                        ; movss         (%ecx),%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
@@ -48481,21 +49880,21 @@
   DB  15,91,254                           ; cvtdq2ps      %xmm6,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
-  DB  15,40,178,100,153,0,0               ; movaps        0x9964(%edx),%xmm6
+  DB  15,40,178,132,162,0,0               ; movaps        0xa284(%edx),%xmm6
   DB  15,84,198                           ; andps         %xmm6,%xmm0
   DB  15,92,248                           ; subps         %xmm0,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,199                           ; subps         %xmm7,%xmm0
-  DB  15,88,170,84,156,0,0                ; addps         0x9c54(%edx),%xmm5
-  DB  15,40,186,100,156,0,0               ; movaps        0x9c64(%edx),%xmm7
+  DB  15,88,170,116,165,0,0               ; addps         0xa574(%edx),%xmm5
+  DB  15,40,186,132,165,0,0               ; movaps        0xa584(%edx),%xmm7
   DB  15,89,248                           ; mulps         %xmm0,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,40,186,116,156,0,0               ; movaps        0x9c74(%edx),%xmm7
+  DB  15,40,186,148,165,0,0               ; movaps        0xa594(%edx),%xmm7
   DB  15,92,248                           ; subps         %xmm0,%xmm7
-  DB  15,40,130,132,156,0,0               ; movaps        0x9c84(%edx),%xmm0
+  DB  15,40,130,164,165,0,0               ; movaps        0xa5a4(%edx),%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  15,88,197                           ; addps         %xmm5,%xmm0
-  DB  15,89,130,148,156,0,0               ; mulps         0x9c94(%edx),%xmm0
+  DB  15,89,130,180,165,0,0               ; mulps         0xa5b4(%edx),%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  243,15,16,105,20                    ; movss         0x14(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
@@ -48546,15 +49945,15 @@
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,88,245                           ; addps         %xmm5,%xmm6
   DB  15,91,238                           ; cvtdq2ps      %xmm6,%xmm5
-  DB  15,89,170,192,154,0,0               ; mulps         0x9ac0(%edx),%xmm5
-  DB  15,84,178,208,154,0,0               ; andps         0x9ad0(%edx),%xmm6
-  DB  15,86,178,32,152,0,0                ; orps          0x9820(%edx),%xmm6
-  DB  15,88,170,224,154,0,0               ; addps         0x9ae0(%edx),%xmm5
-  DB  15,40,186,240,154,0,0               ; movaps        0x9af0(%edx),%xmm7
+  DB  15,89,170,224,163,0,0               ; mulps         0xa3e0(%edx),%xmm5
+  DB  15,84,178,240,163,0,0               ; andps         0xa3f0(%edx),%xmm6
+  DB  15,86,178,64,161,0,0                ; orps          0xa140(%edx),%xmm6
+  DB  15,88,170,0,164,0,0                 ; addps         0xa400(%edx),%xmm5
+  DB  15,40,186,16,164,0,0                ; movaps        0xa410(%edx),%xmm7
   DB  15,89,254                           ; mulps         %xmm6,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,88,178,0,155,0,0                 ; addps         0x9b00(%edx),%xmm6
-  DB  15,40,186,16,155,0,0                ; movaps        0x9b10(%edx),%xmm7
+  DB  15,88,178,32,164,0,0                ; addps         0xa420(%edx),%xmm6
+  DB  15,40,186,48,164,0,0                ; movaps        0xa430(%edx),%xmm7
   DB  15,94,254                           ; divps         %xmm6,%xmm7
   DB  243,15,16,49                        ; movss         (%ecx),%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
@@ -48564,21 +49963,21 @@
   DB  15,91,254                           ; cvtdq2ps      %xmm6,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
-  DB  15,40,178,48,152,0,0                ; movaps        0x9830(%edx),%xmm6
+  DB  15,40,178,80,161,0,0                ; movaps        0xa150(%edx),%xmm6
   DB  15,84,198                           ; andps         %xmm6,%xmm0
   DB  15,92,248                           ; subps         %xmm0,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,199                           ; subps         %xmm7,%xmm0
-  DB  15,88,170,32,155,0,0                ; addps         0x9b20(%edx),%xmm5
-  DB  15,40,186,48,155,0,0                ; movaps        0x9b30(%edx),%xmm7
+  DB  15,88,170,64,164,0,0                ; addps         0xa440(%edx),%xmm5
+  DB  15,40,186,80,164,0,0                ; movaps        0xa450(%edx),%xmm7
   DB  15,89,248                           ; mulps         %xmm0,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,40,186,64,155,0,0                ; movaps        0x9b40(%edx),%xmm7
+  DB  15,40,186,96,164,0,0                ; movaps        0xa460(%edx),%xmm7
   DB  15,92,248                           ; subps         %xmm0,%xmm7
-  DB  15,40,130,80,155,0,0                ; movaps        0x9b50(%edx),%xmm0
+  DB  15,40,130,112,164,0,0               ; movaps        0xa470(%edx),%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  15,88,197                           ; addps         %xmm5,%xmm0
-  DB  15,89,130,96,155,0,0                ; mulps         0x9b60(%edx),%xmm0
+  DB  15,89,130,128,164,0,0               ; mulps         0xa480(%edx),%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  243,15,16,105,20                    ; movss         0x14(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
@@ -48629,15 +50028,15 @@
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,88,245                           ; addps         %xmm5,%xmm6
   DB  15,91,238                           ; cvtdq2ps      %xmm6,%xmm5
-  DB  15,89,170,140,153,0,0               ; mulps         0x998c(%edx),%xmm5
-  DB  15,84,178,156,153,0,0               ; andps         0x999c(%edx),%xmm6
-  DB  15,86,178,236,150,0,0               ; orps          0x96ec(%edx),%xmm6
-  DB  15,88,170,172,153,0,0               ; addps         0x99ac(%edx),%xmm5
-  DB  15,40,186,188,153,0,0               ; movaps        0x99bc(%edx),%xmm7
+  DB  15,89,170,172,162,0,0               ; mulps         0xa2ac(%edx),%xmm5
+  DB  15,84,178,188,162,0,0               ; andps         0xa2bc(%edx),%xmm6
+  DB  15,86,178,12,160,0,0                ; orps          0xa00c(%edx),%xmm6
+  DB  15,88,170,204,162,0,0               ; addps         0xa2cc(%edx),%xmm5
+  DB  15,40,186,220,162,0,0               ; movaps        0xa2dc(%edx),%xmm7
   DB  15,89,254                           ; mulps         %xmm6,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,88,178,204,153,0,0               ; addps         0x99cc(%edx),%xmm6
-  DB  15,40,186,220,153,0,0               ; movaps        0x99dc(%edx),%xmm7
+  DB  15,88,178,236,162,0,0               ; addps         0xa2ec(%edx),%xmm6
+  DB  15,40,186,252,162,0,0               ; movaps        0xa2fc(%edx),%xmm7
   DB  15,94,254                           ; divps         %xmm6,%xmm7
   DB  243,15,16,49                        ; movss         (%ecx),%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
@@ -48647,21 +50046,21 @@
   DB  15,91,254                           ; cvtdq2ps      %xmm6,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,194,199,1                        ; cmpltps       %xmm7,%xmm0
-  DB  15,40,178,252,150,0,0               ; movaps        0x96fc(%edx),%xmm6
+  DB  15,40,178,28,160,0,0                ; movaps        0xa01c(%edx),%xmm6
   DB  15,84,198                           ; andps         %xmm6,%xmm0
   DB  15,92,248                           ; subps         %xmm0,%xmm7
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,199                           ; subps         %xmm7,%xmm0
-  DB  15,88,170,236,153,0,0               ; addps         0x99ec(%edx),%xmm5
-  DB  15,40,186,252,153,0,0               ; movaps        0x99fc(%edx),%xmm7
+  DB  15,88,170,12,163,0,0                ; addps         0xa30c(%edx),%xmm5
+  DB  15,40,186,28,163,0,0                ; movaps        0xa31c(%edx),%xmm7
   DB  15,89,248                           ; mulps         %xmm0,%xmm7
   DB  15,92,239                           ; subps         %xmm7,%xmm5
-  DB  15,40,186,12,154,0,0                ; movaps        0x9a0c(%edx),%xmm7
+  DB  15,40,186,44,163,0,0                ; movaps        0xa32c(%edx),%xmm7
   DB  15,92,248                           ; subps         %xmm0,%xmm7
-  DB  15,40,130,28,154,0,0                ; movaps        0x9a1c(%edx),%xmm0
+  DB  15,40,130,60,163,0,0                ; movaps        0xa33c(%edx),%xmm0
   DB  15,94,199                           ; divps         %xmm7,%xmm0
   DB  15,88,197                           ; addps         %xmm5,%xmm0
-  DB  15,89,130,44,154,0,0                ; mulps         0x9a2c(%edx),%xmm0
+  DB  15,89,130,76,163,0,0                ; mulps         0xa34c(%edx),%xmm0
   DB  102,15,91,192                       ; cvtps2dq      %xmm0,%xmm0
   DB  243,15,16,105,20                    ; movss         0x14(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
@@ -48694,15 +50093,15 @@
   DB  232,0,0,0,0                         ; call          3a95 <_sk_gamma_sse2+0x12>
   DB  88                                  ; pop           %eax
   DB  15,91,195                           ; cvtdq2ps      %xmm3,%xmm0
-  DB  15,89,128,91,152,0,0                ; mulps         0x985b(%eax),%xmm0
-  DB  15,84,152,107,152,0,0               ; andps         0x986b(%eax),%xmm3
-  DB  15,86,152,187,149,0,0               ; orps          0x95bb(%eax),%xmm3
-  DB  15,88,128,123,152,0,0               ; addps         0x987b(%eax),%xmm0
+  DB  15,89,128,123,161,0,0               ; mulps         0xa17b(%eax),%xmm0
+  DB  15,84,152,139,161,0,0               ; andps         0xa18b(%eax),%xmm3
+  DB  15,86,152,219,158,0,0               ; orps          0x9edb(%eax),%xmm3
+  DB  15,88,128,155,161,0,0               ; addps         0xa19b(%eax),%xmm0
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
-  DB  15,89,160,139,152,0,0               ; mulps         0x988b(%eax),%xmm4
+  DB  15,89,160,171,161,0,0               ; mulps         0xa1ab(%eax),%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  15,88,152,155,152,0,0               ; addps         0x989b(%eax),%xmm3
-  DB  15,40,160,171,152,0,0               ; movaps        0x98ab(%eax),%xmm4
+  DB  15,88,152,187,161,0,0               ; addps         0xa1bb(%eax),%xmm3
+  DB  15,40,160,203,161,0,0               ; movaps        0xa1cb(%eax),%xmm4
   DB  15,94,227                           ; divps         %xmm3,%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
@@ -48715,30 +50114,30 @@
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
   DB  15,194,227,1                        ; cmpltps       %xmm3,%xmm4
-  DB  15,84,160,203,149,0,0               ; andps         0x95cb(%eax),%xmm4
+  DB  15,84,160,235,158,0,0               ; andps         0x9eeb(%eax),%xmm4
   DB  15,92,220                           ; subps         %xmm4,%xmm3
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
   DB  15,92,227                           ; subps         %xmm3,%xmm4
-  DB  15,88,128,187,152,0,0               ; addps         0x98bb(%eax),%xmm0
-  DB  15,40,176,219,152,0,0               ; movaps        0x98db(%eax),%xmm6
+  DB  15,88,128,219,161,0,0               ; addps         0xa1db(%eax),%xmm0
+  DB  15,40,176,251,161,0,0               ; movaps        0xa1fb(%eax),%xmm6
   DB  15,92,244                           ; subps         %xmm4,%xmm6
-  DB  15,40,168,203,152,0,0               ; movaps        0x98cb(%eax),%xmm5
+  DB  15,40,168,235,161,0,0               ; movaps        0xa1eb(%eax),%xmm5
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  15,40,152,235,152,0,0               ; movaps        0x98eb(%eax),%xmm3
+  DB  15,40,152,11,162,0,0                ; movaps        0xa20b(%eax),%xmm3
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  15,94,230                           ; divps         %xmm6,%xmm4
   DB  15,88,224                           ; addps         %xmm0,%xmm4
   DB  15,91,193                           ; cvtdq2ps      %xmm1,%xmm0
-  DB  15,89,128,91,152,0,0                ; mulps         0x985b(%eax),%xmm0
-  DB  15,84,136,107,152,0,0               ; andps         0x986b(%eax),%xmm1
-  DB  15,86,136,187,149,0,0               ; orps          0x95bb(%eax),%xmm1
-  DB  15,88,128,123,152,0,0               ; addps         0x987b(%eax),%xmm0
+  DB  15,89,128,123,161,0,0               ; mulps         0xa17b(%eax),%xmm0
+  DB  15,84,136,139,161,0,0               ; andps         0xa18b(%eax),%xmm1
+  DB  15,86,136,219,158,0,0               ; orps          0x9edb(%eax),%xmm1
+  DB  15,88,128,155,161,0,0               ; addps         0xa19b(%eax),%xmm0
   DB  15,40,241                           ; movaps        %xmm1,%xmm6
-  DB  15,89,176,139,152,0,0               ; mulps         0x988b(%eax),%xmm6
+  DB  15,89,176,171,161,0,0               ; mulps         0xa1ab(%eax),%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
-  DB  15,88,136,155,152,0,0               ; addps         0x989b(%eax),%xmm1
-  DB  15,40,176,171,152,0,0               ; movaps        0x98ab(%eax),%xmm6
+  DB  15,88,136,187,161,0,0               ; addps         0xa1bb(%eax),%xmm1
+  DB  15,40,176,203,161,0,0               ; movaps        0xa1cb(%eax),%xmm6
   DB  15,94,241                           ; divps         %xmm1,%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
@@ -48746,12 +50145,12 @@
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,194,241,1                        ; cmpltps       %xmm1,%xmm6
-  DB  15,84,176,203,149,0,0               ; andps         0x95cb(%eax),%xmm6
+  DB  15,84,176,235,158,0,0               ; andps         0x9eeb(%eax),%xmm6
   DB  15,92,206                           ; subps         %xmm6,%xmm1
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,92,241                           ; subps         %xmm1,%xmm6
-  DB  15,88,128,187,152,0,0               ; addps         0x98bb(%eax),%xmm0
-  DB  15,40,184,219,152,0,0               ; movaps        0x98db(%eax),%xmm7
+  DB  15,88,128,219,161,0,0               ; addps         0xa1db(%eax),%xmm0
+  DB  15,40,184,251,161,0,0               ; movaps        0xa1fb(%eax),%xmm7
   DB  15,92,254                           ; subps         %xmm6,%xmm7
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
@@ -48759,15 +50158,15 @@
   DB  15,94,207                           ; divps         %xmm7,%xmm1
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  15,91,194                           ; cvtdq2ps      %xmm2,%xmm0
-  DB  15,89,128,91,152,0,0                ; mulps         0x985b(%eax),%xmm0
-  DB  15,84,144,107,152,0,0               ; andps         0x986b(%eax),%xmm2
-  DB  15,86,144,187,149,0,0               ; orps          0x95bb(%eax),%xmm2
-  DB  15,88,128,123,152,0,0               ; addps         0x987b(%eax),%xmm0
-  DB  15,40,176,139,152,0,0               ; movaps        0x988b(%eax),%xmm6
+  DB  15,89,128,123,161,0,0               ; mulps         0xa17b(%eax),%xmm0
+  DB  15,84,144,139,161,0,0               ; andps         0xa18b(%eax),%xmm2
+  DB  15,86,144,219,158,0,0               ; orps          0x9edb(%eax),%xmm2
+  DB  15,88,128,155,161,0,0               ; addps         0xa19b(%eax),%xmm0
+  DB  15,40,176,171,161,0,0               ; movaps        0xa1ab(%eax),%xmm6
   DB  15,89,242                           ; mulps         %xmm2,%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
-  DB  15,88,144,155,152,0,0               ; addps         0x989b(%eax),%xmm2
-  DB  15,40,176,171,152,0,0               ; movaps        0x98ab(%eax),%xmm6
+  DB  15,88,144,187,161,0,0               ; addps         0xa1bb(%eax),%xmm2
+  DB  15,40,176,203,161,0,0               ; movaps        0xa1cb(%eax),%xmm6
   DB  15,94,242                           ; divps         %xmm2,%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
   DB  15,89,69,232                        ; mulps         -0x18(%ebp),%xmm0
@@ -48775,18 +50174,18 @@
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,194,242,1                        ; cmpltps       %xmm2,%xmm6
-  DB  15,84,176,203,149,0,0               ; andps         0x95cb(%eax),%xmm6
+  DB  15,84,176,235,158,0,0               ; andps         0x9eeb(%eax),%xmm6
   DB  15,92,214                           ; subps         %xmm6,%xmm2
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,92,242                           ; subps         %xmm2,%xmm6
-  DB  15,88,128,187,152,0,0               ; addps         0x98bb(%eax),%xmm0
+  DB  15,88,128,219,161,0,0               ; addps         0xa1db(%eax),%xmm0
   DB  15,89,238                           ; mulps         %xmm6,%xmm5
   DB  15,92,197                           ; subps         %xmm5,%xmm0
-  DB  15,40,144,219,152,0,0               ; movaps        0x98db(%eax),%xmm2
+  DB  15,40,144,251,161,0,0               ; movaps        0xa1fb(%eax),%xmm2
   DB  15,92,214                           ; subps         %xmm6,%xmm2
   DB  15,94,218                           ; divps         %xmm2,%xmm3
   DB  15,88,216                           ; addps         %xmm0,%xmm3
-  DB  15,40,128,251,152,0,0               ; movaps        0x98fb(%eax),%xmm0
+  DB  15,40,128,27,162,0,0                ; movaps        0xa21b(%eax),%xmm0
   DB  15,89,224                           ; mulps         %xmm0,%xmm4
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  15,89,216                           ; mulps         %xmm0,%xmm3
@@ -48813,29 +50212,29 @@
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  232,0,0,0,0                         ; call          3c7d <_sk_lab_to_xyz_sse2+0x15>
   DB  88                                  ; pop           %eax
-  DB  15,89,144,35,151,0,0                ; mulps         0x9723(%eax),%xmm2
-  DB  15,40,160,179,148,0,0               ; movaps        0x94b3(%eax),%xmm4
+  DB  15,89,144,67,160,0,0                ; mulps         0xa043(%eax),%xmm2
+  DB  15,40,160,211,157,0,0               ; movaps        0x9dd3(%eax),%xmm4
   DB  15,89,204                           ; mulps         %xmm4,%xmm1
-  DB  15,40,168,51,151,0,0                ; movaps        0x9733(%eax),%xmm5
+  DB  15,40,168,83,160,0,0                ; movaps        0xa053(%eax),%xmm5
   DB  15,88,205                           ; addps         %xmm5,%xmm1
   DB  15,89,220                           ; mulps         %xmm4,%xmm3
   DB  15,88,221                           ; addps         %xmm5,%xmm3
-  DB  15,88,144,67,151,0,0                ; addps         0x9743(%eax),%xmm2
-  DB  15,89,144,83,151,0,0                ; mulps         0x9753(%eax),%xmm2
-  DB  15,89,136,99,151,0,0                ; mulps         0x9763(%eax),%xmm1
+  DB  15,88,144,99,160,0,0                ; addps         0xa063(%eax),%xmm2
+  DB  15,89,144,115,160,0,0               ; mulps         0xa073(%eax),%xmm2
+  DB  15,89,136,131,160,0,0               ; mulps         0xa083(%eax),%xmm1
   DB  15,88,202                           ; addps         %xmm2,%xmm1
-  DB  15,89,152,115,151,0,0               ; mulps         0x9773(%eax),%xmm3
+  DB  15,89,152,147,160,0,0               ; mulps         0xa093(%eax),%xmm3
   DB  15,40,234                           ; movaps        %xmm2,%xmm5
   DB  15,92,235                           ; subps         %xmm3,%xmm5
   DB  15,40,193                           ; movaps        %xmm1,%xmm0
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  15,89,193                           ; mulps         %xmm1,%xmm0
-  DB  15,40,152,131,151,0,0               ; movaps        0x9783(%eax),%xmm3
+  DB  15,40,152,163,160,0,0               ; movaps        0xa0a3(%eax),%xmm3
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  15,194,224,1                        ; cmpltps       %xmm0,%xmm4
-  DB  15,40,176,147,151,0,0               ; movaps        0x9793(%eax),%xmm6
+  DB  15,40,176,179,160,0,0               ; movaps        0xa0b3(%eax),%xmm6
   DB  15,88,206                           ; addps         %xmm6,%xmm1
-  DB  15,40,184,163,151,0,0               ; movaps        0x97a3(%eax),%xmm7
+  DB  15,40,184,195,160,0,0               ; movaps        0xa0c3(%eax),%xmm7
   DB  15,89,207                           ; mulps         %xmm7,%xmm1
   DB  15,84,196                           ; andps         %xmm4,%xmm0
   DB  15,85,225                           ; andnps        %xmm1,%xmm4
@@ -48860,8 +50259,8 @@
   DB  15,85,221                           ; andnps        %xmm5,%xmm3
   DB  15,86,216                           ; orps          %xmm0,%xmm3
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,89,160,179,151,0,0               ; mulps         0x97b3(%eax),%xmm4
-  DB  15,89,152,195,151,0,0               ; mulps         0x97c3(%eax),%xmm3
+  DB  15,89,160,211,160,0,0               ; mulps         0xa0d3(%eax),%xmm4
+  DB  15,89,152,227,160,0,0               ; mulps         0xa0e3(%eax),%xmm3
   DB  141,65,4                            ; lea           0x4(%ecx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  15,40,196                           ; movaps        %xmm4,%xmm0
@@ -48897,9 +50296,9 @@
   DB  102,15,110,4,62                     ; movd          (%esi,%edi,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  102,15,97,192                       ; punpcklwd     %xmm0,%xmm0
-  DB  102,15,219,130,179,147,0,0          ; pand          0x93b3(%edx),%xmm0
+  DB  102,15,219,130,211,156,0,0          ; pand          0x9cd3(%edx),%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,154,243,148,0,0               ; mulps         0x94f3(%edx),%xmm3
+  DB  15,89,154,19,158,0,0                ; mulps         0x9e13(%edx),%xmm3
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -48962,9 +50361,9 @@
   DB  102,15,110,36,55                    ; movd          (%edi,%esi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,162,223,146,0,0          ; pand          0x92df(%edx),%xmm4
+  DB  102,15,219,162,255,155,0,0          ; pand          0x9bff(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,31,148,0,0                ; mulps         0x941f(%edx),%xmm4
+  DB  15,89,162,63,157,0,0                ; mulps         0x9d3f(%edx),%xmm4
   DB  15,41,97,64                         ; movaps        %xmm4,0x40(%ecx)
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -49053,7 +50452,7 @@
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,216                           ; cvtdq2ps      %xmm0,%xmm3
-  DB  15,89,153,242,146,0,0               ; mulps         0x92f2(%ecx),%xmm3
+  DB  15,89,153,18,156,0,0                ; mulps         0x9c12(%ecx),%xmm3
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
@@ -49085,7 +50484,7 @@
   DB  15,175,81,4                         ; imul          0x4(%ecx),%edx
   DB  3,23                                ; add           (%edi),%edx
   DB  139,89,8                            ; mov           0x8(%ecx),%ebx
-  DB  15,40,166,126,145,0,0               ; movaps        0x917e(%esi),%xmm4
+  DB  15,40,166,158,154,0,0               ; movaps        0x9a9e(%esi),%xmm4
   DB  15,89,227                           ; mulps         %xmm3,%xmm4
   DB  102,15,91,228                       ; cvtps2dq      %xmm4,%xmm4
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
@@ -49120,7 +50519,7 @@
   DB  102,15,197,220,4                    ; pextrw        $0x4,%xmm4,%ebx
   DB  136,92,58,2                         ; mov           %bl,0x2(%edx,%edi,1)
   DB  102,15,112,228,212                  ; pshufd        $0xd4,%xmm4,%xmm4
-  DB  102,15,219,166,158,148,0,0          ; pand          0x949e(%esi),%xmm4
+  DB  102,15,219,166,190,157,0,0          ; pand          0x9dbe(%esi),%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
@@ -49154,12 +50553,12 @@
   DB  102,15,110,4,62                     ; movd          (%esi,%edi,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
   DB  102,15,97,192                       ; punpcklwd     %xmm0,%xmm0
-  DB  102,15,219,130,188,144,0,0          ; pand          0x90bc(%edx),%xmm0
+  DB  102,15,219,130,220,153,0,0          ; pand          0x99dc(%edx),%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,130,252,145,0,0               ; mulps         0x91fc(%edx),%xmm0
+  DB  15,89,130,28,155,0,0                ; mulps         0x9b1c(%edx),%xmm0
   DB  141,112,8                           ; lea           0x8(%eax),%esi
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,154,252,143,0,0               ; movaps        0x8ffc(%edx),%xmm3
+  DB  15,40,154,28,153,0,0                ; movaps        0x991c(%edx),%xmm3
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  86                                  ; push          %esi
@@ -49215,13 +50614,13 @@
   DB  102,15,110,36,62                    ; movd          (%esi,%edi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,219,162,16,144,0,0           ; pand          0x9010(%edx),%xmm4
+  DB  102,15,219,162,48,153,0,0           ; pand          0x9930(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,80,145,0,0                ; mulps         0x9150(%edx),%xmm4
+  DB  15,89,162,112,154,0,0               ; mulps         0x9a70(%edx),%xmm4
   DB  15,41,97,48                         ; movaps        %xmm4,0x30(%ecx)
   DB  15,41,97,32                         ; movaps        %xmm4,0x20(%ecx)
   DB  15,41,97,16                         ; movaps        %xmm4,0x10(%ecx)
-  DB  15,40,162,80,143,0,0                ; movaps        0x8f50(%edx),%xmm4
+  DB  15,40,162,112,152,0,0               ; movaps        0x9870(%edx),%xmm4
   DB  15,41,97,64                         ; movaps        %xmm4,0x40(%ecx)
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -49311,10 +50710,10 @@
   DB  102,15,96,193                       ; punpcklbw     %xmm1,%xmm0
   DB  102,15,97,193                       ; punpcklwd     %xmm1,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,129,249,143,0,0               ; mulps         0x8ff9(%ecx),%xmm0
+  DB  15,89,129,25,153,0,0                ; mulps         0x9919(%ecx),%xmm0
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,153,249,141,0,0               ; movaps        0x8df9(%ecx),%xmm3
+  DB  15,40,153,25,151,0,0                ; movaps        0x9719(%ecx),%xmm3
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  82                                  ; push          %edx
@@ -49349,20 +50748,20 @@
   DB  117,97                              ; jne           4330 <_sk_load_565_sse2+0x8c>
   DB  243,15,126,20,126                   ; movq          (%esi,%edi,2),%xmm2
   DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
-  DB  102,15,111,130,190,143,0,0          ; movdqa        0x8fbe(%edx),%xmm0
+  DB  102,15,111,130,222,152,0,0          ; movdqa        0x98de(%edx),%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,130,206,143,0,0               ; mulps         0x8fce(%edx),%xmm0
-  DB  102,15,111,138,222,143,0,0          ; movdqa        0x8fde(%edx),%xmm1
+  DB  15,89,130,238,152,0,0               ; mulps         0x98ee(%edx),%xmm0
+  DB  102,15,111,138,254,152,0,0          ; movdqa        0x98fe(%edx),%xmm1
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,138,238,143,0,0               ; mulps         0x8fee(%edx),%xmm1
-  DB  102,15,219,146,254,143,0,0          ; pand          0x8ffe(%edx),%xmm2
+  DB  15,89,138,14,153,0,0                ; mulps         0x990e(%edx),%xmm1
+  DB  102,15,219,146,30,153,0,0           ; pand          0x991e(%edx),%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,146,14,144,0,0                ; mulps         0x900e(%edx),%xmm2
+  DB  15,89,146,46,153,0,0                ; mulps         0x992e(%edx),%xmm2
   DB  141,112,8                           ; lea           0x8(%eax),%esi
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,154,174,141,0,0               ; movaps        0x8dae(%edx),%xmm3
+  DB  15,40,154,206,150,0,0               ; movaps        0x96ce(%edx),%xmm3
   DB  86                                  ; push          %esi
   DB  81                                  ; push          %ecx
   DB  255,80,4                            ; call          *0x4(%eax)
@@ -49414,21 +50813,21 @@
   DB  117,113                             ; jne           4410 <_sk_load_565_dst_sse2+0x9c>
   DB  243,15,126,36,126                   ; movq          (%esi,%edi,2),%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,111,170,238,142,0,0          ; movdqa        0x8eee(%edx),%xmm5
+  DB  102,15,111,170,14,152,0,0           ; movdqa        0x980e(%edx),%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,254,142,0,0               ; mulps         0x8efe(%edx),%xmm5
+  DB  15,89,170,30,152,0,0                ; mulps         0x981e(%edx),%xmm5
   DB  15,41,105,16                        ; movaps        %xmm5,0x10(%ecx)
-  DB  102,15,111,170,14,143,0,0           ; movdqa        0x8f0e(%edx),%xmm5
+  DB  102,15,111,170,46,152,0,0           ; movdqa        0x982e(%edx),%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,30,143,0,0                ; mulps         0x8f1e(%edx),%xmm5
+  DB  15,89,170,62,152,0,0                ; mulps         0x983e(%edx),%xmm5
   DB  15,41,105,32                        ; movaps        %xmm5,0x20(%ecx)
-  DB  102,15,219,162,46,143,0,0           ; pand          0x8f2e(%edx),%xmm4
+  DB  102,15,219,162,78,152,0,0           ; pand          0x984e(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,62,143,0,0                ; mulps         0x8f3e(%edx),%xmm4
+  DB  15,89,162,94,152,0,0                ; mulps         0x985e(%edx),%xmm4
   DB  15,41,97,48                         ; movaps        %xmm4,0x30(%ecx)
-  DB  15,40,162,222,140,0,0               ; movaps        0x8cde(%edx),%xmm4
+  DB  15,40,162,254,149,0,0               ; movaps        0x95fe(%edx),%xmm4
   DB  15,41,97,64                         ; movaps        %xmm4,0x40(%ecx)
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -49511,20 +50910,20 @@
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,209,3                    ; pinsrw        $0x3,%ecx,%xmm2
   DB  102,15,97,208                       ; punpcklwd     %xmm0,%xmm2
-  DB  102,15,111,130,108,141,0,0          ; movdqa        0x8d6c(%edx),%xmm0
+  DB  102,15,111,130,140,150,0,0          ; movdqa        0x968c(%edx),%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,130,124,141,0,0               ; mulps         0x8d7c(%edx),%xmm0
-  DB  102,15,111,138,140,141,0,0          ; movdqa        0x8d8c(%edx),%xmm1
+  DB  15,89,130,156,150,0,0               ; mulps         0x969c(%edx),%xmm0
+  DB  102,15,111,138,172,150,0,0          ; movdqa        0x96ac(%edx),%xmm1
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,138,156,141,0,0               ; mulps         0x8d9c(%edx),%xmm1
-  DB  102,15,219,146,172,141,0,0          ; pand          0x8dac(%edx),%xmm2
+  DB  15,89,138,188,150,0,0               ; mulps         0x96bc(%edx),%xmm1
+  DB  102,15,219,146,204,150,0,0          ; pand          0x96cc(%edx),%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,146,188,141,0,0               ; mulps         0x8dbc(%edx),%xmm2
+  DB  15,89,146,220,150,0,0               ; mulps         0x96dc(%edx),%xmm2
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,154,92,139,0,0                ; movaps        0x8b5c(%edx),%xmm3
+  DB  15,40,154,124,148,0,0               ; movaps        0x947c(%edx),%xmm3
   DB  81                                  ; push          %ecx
   DB  255,117,8                           ; pushl         0x8(%ebp)
   DB  255,80,4                            ; call          *0x4(%eax)
@@ -49550,8 +50949,8 @@
   DB  1,210                               ; add           %edx,%edx
   DB  3,22                                ; add           (%esi),%edx
   DB  139,49                              ; mov           (%ecx),%esi
-  DB  15,40,167,233,142,0,0               ; movaps        0x8ee9(%edi),%xmm4
-  DB  15,40,175,249,142,0,0               ; movaps        0x8ef9(%edi),%xmm5
+  DB  15,40,167,9,152,0,0                 ; movaps        0x9809(%edi),%xmm4
+  DB  15,40,175,25,152,0,0                ; movaps        0x9819(%edi),%xmm5
   DB  139,89,8                            ; mov           0x8(%ecx),%ebx
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,89,244                           ; mulps         %xmm4,%xmm6
@@ -49629,21 +51028,21 @@
   DB  117,112                             ; jne           46e6 <_sk_load_4444_sse2+0x9b>
   DB  243,15,126,28,126                   ; movq          (%esi,%edi,2),%xmm3
   DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
-  DB  102,15,111,130,39,142,0,0           ; movdqa        0x8e27(%edx),%xmm0
+  DB  102,15,111,130,71,151,0,0           ; movdqa        0x9747(%edx),%xmm0
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,130,55,142,0,0                ; mulps         0x8e37(%edx),%xmm0
-  DB  102,15,111,138,71,142,0,0           ; movdqa        0x8e47(%edx),%xmm1
+  DB  15,89,130,87,151,0,0                ; mulps         0x9757(%edx),%xmm0
+  DB  102,15,111,138,103,151,0,0          ; movdqa        0x9767(%edx),%xmm1
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,138,87,142,0,0                ; mulps         0x8e57(%edx),%xmm1
-  DB  102,15,111,146,103,142,0,0          ; movdqa        0x8e67(%edx),%xmm2
+  DB  15,89,138,119,151,0,0               ; mulps         0x9777(%edx),%xmm1
+  DB  102,15,111,146,135,151,0,0          ; movdqa        0x9787(%edx),%xmm2
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,146,119,142,0,0               ; mulps         0x8e77(%edx),%xmm2
-  DB  102,15,219,154,135,142,0,0          ; pand          0x8e87(%edx),%xmm3
+  DB  15,89,146,151,151,0,0               ; mulps         0x9797(%edx),%xmm2
+  DB  102,15,219,154,167,151,0,0          ; pand          0x97a7(%edx),%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,154,151,142,0,0               ; mulps         0x8e97(%edx),%xmm3
+  DB  15,89,154,183,151,0,0               ; mulps         0x97b7(%edx),%xmm3
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  82                                  ; push          %edx
@@ -49697,24 +51096,24 @@
   DB  15,133,128,0,0,0                    ; jne           47d9 <_sk_load_4444_dst_sse2+0xaf>
   DB  243,15,126,36,126                   ; movq          (%esi,%edi,2),%xmm4
   DB  102,15,97,224                       ; punpcklwd     %xmm0,%xmm4
-  DB  102,15,111,170,72,141,0,0           ; movdqa        0x8d48(%edx),%xmm5
+  DB  102,15,111,170,104,150,0,0          ; movdqa        0x9668(%edx),%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,88,141,0,0                ; mulps         0x8d58(%edx),%xmm5
+  DB  15,89,170,120,150,0,0               ; mulps         0x9678(%edx),%xmm5
   DB  15,41,105,16                        ; movaps        %xmm5,0x10(%ecx)
-  DB  102,15,111,170,104,141,0,0          ; movdqa        0x8d68(%edx),%xmm5
+  DB  102,15,111,170,136,150,0,0          ; movdqa        0x9688(%edx),%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,120,141,0,0               ; mulps         0x8d78(%edx),%xmm5
+  DB  15,89,170,152,150,0,0               ; mulps         0x9698(%edx),%xmm5
   DB  15,41,105,32                        ; movaps        %xmm5,0x20(%ecx)
-  DB  102,15,111,170,136,141,0,0          ; movdqa        0x8d88(%edx),%xmm5
+  DB  102,15,111,170,168,150,0,0          ; movdqa        0x96a8(%edx),%xmm5
   DB  102,15,219,236                      ; pand          %xmm4,%xmm5
   DB  15,91,237                           ; cvtdq2ps      %xmm5,%xmm5
-  DB  15,89,170,152,141,0,0               ; mulps         0x8d98(%edx),%xmm5
+  DB  15,89,170,184,150,0,0               ; mulps         0x96b8(%edx),%xmm5
   DB  15,41,105,48                        ; movaps        %xmm5,0x30(%ecx)
-  DB  102,15,219,162,168,141,0,0          ; pand          0x8da8(%edx),%xmm4
+  DB  102,15,219,162,200,150,0,0          ; pand          0x96c8(%edx),%xmm4
   DB  15,91,228                           ; cvtdq2ps      %xmm4,%xmm4
-  DB  15,89,162,184,141,0,0               ; mulps         0x8db8(%edx),%xmm4
+  DB  15,89,162,216,150,0,0               ; mulps         0x96d8(%edx),%xmm4
   DB  15,41,97,64                         ; movaps        %xmm4,0x40(%ecx)
   DB  141,80,8                            ; lea           0x8(%eax),%edx
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -49797,21 +51196,21 @@
   DB  89                                  ; pop           %ecx
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,216                       ; punpcklwd     %xmm0,%xmm3
-  DB  102,15,111,129,170,139,0,0          ; movdqa        0x8baa(%ecx),%xmm0
+  DB  102,15,111,129,202,148,0,0          ; movdqa        0x94ca(%ecx),%xmm0
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,89,129,186,139,0,0               ; mulps         0x8bba(%ecx),%xmm0
-  DB  102,15,111,137,202,139,0,0          ; movdqa        0x8bca(%ecx),%xmm1
+  DB  15,89,129,218,148,0,0               ; mulps         0x94da(%ecx),%xmm0
+  DB  102,15,111,137,234,148,0,0          ; movdqa        0x94ea(%ecx),%xmm1
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
   DB  15,91,201                           ; cvtdq2ps      %xmm1,%xmm1
-  DB  15,89,137,218,139,0,0               ; mulps         0x8bda(%ecx),%xmm1
-  DB  102,15,111,145,234,139,0,0          ; movdqa        0x8bea(%ecx),%xmm2
+  DB  15,89,137,250,148,0,0               ; mulps         0x94fa(%ecx),%xmm1
+  DB  102,15,111,145,10,149,0,0           ; movdqa        0x950a(%ecx),%xmm2
   DB  102,15,219,211                      ; pand          %xmm3,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,89,145,250,139,0,0               ; mulps         0x8bfa(%ecx),%xmm2
-  DB  102,15,219,153,10,140,0,0           ; pand          0x8c0a(%ecx),%xmm3
+  DB  15,89,145,26,149,0,0                ; mulps         0x951a(%ecx),%xmm2
+  DB  102,15,219,153,42,149,0,0           ; pand          0x952a(%ecx),%xmm3
   DB  15,91,219                           ; cvtdq2ps      %xmm3,%xmm3
-  DB  15,89,153,26,140,0,0                ; mulps         0x8c1a(%ecx),%xmm3
+  DB  15,89,153,58,149,0,0                ; mulps         0x953a(%ecx),%xmm3
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  81                                  ; push          %ecx
@@ -49839,7 +51238,7 @@
   DB  1,210                               ; add           %edx,%edx
   DB  3,22                                ; add           (%esi),%edx
   DB  139,49                              ; mov           (%ecx),%esi
-  DB  15,40,167,173,139,0,0               ; movaps        0x8bad(%edi),%xmm4
+  DB  15,40,167,205,148,0,0               ; movaps        0x94cd(%edi),%xmm4
   DB  139,89,8                            ; mov           0x8(%ecx),%ebx
   DB  15,40,232                           ; movaps        %xmm0,%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
@@ -49922,11 +51321,11 @@
   DB  139,57                              ; mov           (%ecx),%edi
   DB  117,102                             ; jne           4ac8 <_sk_load_8888_sse2+0x92>
   DB  243,15,111,28,190                   ; movdqu        (%esi,%edi,4),%xmm3
-  DB  102,15,111,146,220,134,0,0          ; movdqa        0x86dc(%edx),%xmm2
+  DB  102,15,111,146,252,143,0,0          ; movdqa        0x8ffc(%edx),%xmm2
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,40,162,28,136,0,0                ; movaps        0x881c(%edx),%xmm4
+  DB  15,40,162,60,145,0,0                ; movaps        0x913c(%edx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -49989,11 +51388,11 @@
   DB  139,57                              ; mov           (%ecx),%edi
   DB  117,118                             ; jne           4b9f <_sk_load_8888_dst_sse2+0xa2>
   DB  243,15,111,36,190                   ; movdqu        (%esi,%edi,4),%xmm4
-  DB  102,15,111,170,21,134,0,0           ; movdqa        0x8615(%edx),%xmm5
+  DB  102,15,111,170,53,143,0,0           ; movdqa        0x8f35(%edx),%xmm5
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  102,15,219,245                      ; pand          %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,40,186,85,135,0,0                ; movaps        0x8755(%edx),%xmm7
+  DB  15,40,186,117,144,0,0               ; movaps        0x9075(%edx),%xmm7
   DB  15,89,247                           ; mulps         %xmm7,%xmm6
   DB  15,41,113,16                        ; movaps        %xmm6,0x10(%ecx)
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
@@ -50086,11 +51485,11 @@
   DB  102,15,98,194                       ; punpckldq     %xmm2,%xmm0
   DB  102,15,98,217                       ; punpckldq     %xmm1,%xmm3
   DB  102,15,108,216                      ; punpcklqdq    %xmm0,%xmm3
-  DB  102,15,111,145,162,132,0,0          ; movdqa        0x84a2(%ecx),%xmm2
+  DB  102,15,111,145,194,141,0,0          ; movdqa        0x8dc2(%ecx),%xmm2
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,219,194                      ; pand          %xmm2,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,40,161,226,133,0,0               ; movaps        0x85e2(%ecx),%xmm4
+  DB  15,40,161,2,143,0,0                 ; movaps        0x8f02(%ecx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -50132,7 +51531,7 @@
   DB  193,226,2                           ; shl           $0x2,%edx
   DB  3,22                                ; add           (%esi),%edx
   DB  139,49                              ; mov           (%ecx),%esi
-  DB  15,40,167,55,132,0,0                ; movaps        0x8437(%edi),%xmm4
+  DB  15,40,167,87,141,0,0                ; movaps        0x8d57(%edi),%xmm4
   DB  139,89,8                            ; mov           0x8(%ecx),%ebx
   DB  15,40,232                           ; movaps        %xmm0,%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
@@ -50201,11 +51600,11 @@
   DB  139,57                              ; mov           (%ecx),%edi
   DB  117,102                             ; jne           4e33 <_sk_load_bgra_sse2+0x92>
   DB  243,15,111,28,190                   ; movdqu        (%esi,%edi,4),%xmm3
-  DB  102,15,111,130,113,131,0,0          ; movdqa        0x8371(%edx),%xmm0
+  DB  102,15,111,130,145,140,0,0          ; movdqa        0x8c91(%edx),%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
   DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
-  DB  15,40,162,177,132,0,0               ; movaps        0x84b1(%edx),%xmm4
+  DB  15,40,162,209,141,0,0               ; movaps        0x8dd1(%edx),%xmm4
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -50268,11 +51667,11 @@
   DB  139,57                              ; mov           (%ecx),%edi
   DB  117,118                             ; jne           4f0a <_sk_load_bgra_dst_sse2+0xa2>
   DB  243,15,111,36,190                   ; movdqu        (%esi,%edi,4),%xmm4
-  DB  102,15,111,170,170,130,0,0          ; movdqa        0x82aa(%edx),%xmm5
+  DB  102,15,111,170,202,139,0,0          ; movdqa        0x8bca(%edx),%xmm5
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  102,15,219,245                      ; pand          %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
-  DB  15,40,186,234,131,0,0               ; movaps        0x83ea(%edx),%xmm7
+  DB  15,40,186,10,141,0,0                ; movaps        0x8d0a(%edx),%xmm7
   DB  15,89,247                           ; mulps         %xmm7,%xmm6
   DB  15,41,113,48                        ; movaps        %xmm6,0x30(%ecx)
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
@@ -50365,11 +51764,11 @@
   DB  102,15,98,194                       ; punpckldq     %xmm2,%xmm0
   DB  102,15,98,217                       ; punpckldq     %xmm1,%xmm3
   DB  102,15,108,216                      ; punpcklqdq    %xmm0,%xmm3
-  DB  102,15,111,129,55,129,0,0           ; movdqa        0x8137(%ecx),%xmm0
+  DB  102,15,111,129,87,138,0,0           ; movdqa        0x8a57(%ecx),%xmm0
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
   DB  15,91,209                           ; cvtdq2ps      %xmm1,%xmm2
-  DB  15,40,161,119,130,0,0               ; movaps        0x8277(%ecx),%xmm4
+  DB  15,40,161,151,139,0,0               ; movaps        0x8b97(%ecx),%xmm4
   DB  15,89,212                           ; mulps         %xmm4,%xmm2
   DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
   DB  102,15,114,209,8                    ; psrld         $0x8,%xmm1
@@ -50411,7 +51810,7 @@
   DB  193,226,2                           ; shl           $0x2,%edx
   DB  3,22                                ; add           (%esi),%edx
   DB  139,49                              ; mov           (%ecx),%esi
-  DB  15,40,167,204,128,0,0               ; movaps        0x80cc(%edi),%xmm4
+  DB  15,40,167,236,137,0,0               ; movaps        0x89ec(%edi),%xmm4
   DB  139,89,8                            ; mov           0x8(%ecx),%ebx
   DB  15,40,234                           ; movaps        %xmm2,%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
@@ -50508,11 +51907,11 @@
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,97,200                       ; punpcklwd     %xmm0,%xmm1
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
-  DB  102,15,111,170,246,131,0,0          ; movdqa        0x83f6(%edx),%xmm5
+  DB  102,15,111,170,22,141,0,0           ; movdqa        0x8d16(%edx),%xmm5
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
   DB  102,15,219,245                      ; pand          %xmm5,%xmm6
-  DB  102,15,219,138,6,132,0,0            ; pand          0x8406(%edx),%xmm1
-  DB  102,15,111,130,22,132,0,0           ; movdqa        0x8416(%edx),%xmm0
+  DB  102,15,219,138,38,141,0,0           ; pand          0x8d26(%edx),%xmm1
+  DB  102,15,111,130,54,141,0,0           ; movdqa        0x8d36(%edx),%xmm0
   DB  102,15,127,69,184                   ; movdqa        %xmm0,-0x48(%ebp)
   DB  102,15,114,246,16                   ; pslld         $0x10,%xmm6
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
@@ -50524,7 +51923,7 @@
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  102,15,111,212                      ; movdqa        %xmm4,%xmm2
   DB  102,15,219,213                      ; pand          %xmm5,%xmm2
-  DB  102,15,111,154,6,132,0,0            ; movdqa        0x8406(%edx),%xmm3
+  DB  102,15,111,154,38,141,0,0           ; movdqa        0x8d26(%edx),%xmm3
   DB  102,15,219,227                      ; pand          %xmm3,%xmm4
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,102,196                      ; pcmpgtd       %xmm4,%xmm0
@@ -50537,7 +51936,7 @@
   DB  102,15,219,253                      ; pand          %xmm5,%xmm7
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  102,15,114,247,16                   ; pslld         $0x10,%xmm7
-  DB  102,15,111,146,22,132,0,0           ; movdqa        0x8416(%edx),%xmm2
+  DB  102,15,111,146,54,141,0,0           ; movdqa        0x8d36(%edx),%xmm2
   DB  102,15,102,208                      ; pcmpgtd       %xmm0,%xmm2
   DB  102,15,114,240,13                   ; pslld         $0xd,%xmm0
   DB  102,15,235,199                      ; por           %xmm7,%xmm0
@@ -50552,7 +51951,7 @@
   DB  102,15,127,93,184                   ; movdqa        %xmm3,-0x48(%ebp)
   DB  102,15,114,247,13                   ; pslld         $0xd,%xmm7
   DB  102,15,235,253                      ; por           %xmm5,%xmm7
-  DB  102,15,111,154,38,132,0,0           ; movdqa        0x8426(%edx),%xmm3
+  DB  102,15,111,154,70,141,0,0           ; movdqa        0x8d46(%edx),%xmm3
   DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
   DB  102,15,254,227                      ; paddd         %xmm3,%xmm4
   DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
@@ -50639,18 +52038,18 @@
   DB  102,15,126,85,232                   ; movd          %xmm2,-0x18(%ebp)
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
   DB  102,15,97,202                       ; punpcklwd     %xmm2,%xmm1
-  DB  102,15,111,186,217,129,0,0          ; movdqa        0x81d9(%edx),%xmm7
+  DB  102,15,111,186,249,138,0,0          ; movdqa        0x8af9(%edx),%xmm7
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,219,215                      ; pand          %xmm7,%xmm2
-  DB  102,15,111,154,233,129,0,0          ; movdqa        0x81e9(%edx),%xmm3
+  DB  102,15,111,154,9,139,0,0            ; movdqa        0x8b09(%edx),%xmm3
   DB  102,15,219,203                      ; pand          %xmm3,%xmm1
-  DB  102,15,111,162,249,129,0,0          ; movdqa        0x81f9(%edx),%xmm4
+  DB  102,15,111,162,25,139,0,0           ; movdqa        0x8b19(%edx),%xmm4
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,102,233                      ; pcmpgtd       %xmm1,%xmm5
   DB  102,15,114,241,13                   ; pslld         $0xd,%xmm1
   DB  102,15,235,202                      ; por           %xmm2,%xmm1
-  DB  102,15,111,146,9,130,0,0            ; movdqa        0x8209(%edx),%xmm2
+  DB  102,15,111,146,41,139,0,0           ; movdqa        0x8b29(%edx),%xmm2
   DB  102,15,127,85,184                   ; movdqa        %xmm2,-0x48(%ebp)
   DB  102,15,254,202                      ; paddd         %xmm2,%xmm1
   DB  102,15,223,233                      ; pandn         %xmm1,%xmm5
@@ -50794,11 +52193,11 @@
   DB  102,15,126,69,248                   ; movd          %xmm0,-0x8(%ebp)
   DB  102,15,97,206                       ; punpcklwd     %xmm6,%xmm1
   DB  102,15,239,210                      ; pxor          %xmm2,%xmm2
-  DB  102,15,111,177,61,127,0,0           ; movdqa        0x7f3d(%ecx),%xmm6
+  DB  102,15,111,177,93,136,0,0           ; movdqa        0x885d(%ecx),%xmm6
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,219,230                      ; pand          %xmm6,%xmm4
-  DB  102,15,219,137,77,127,0,0           ; pand          0x7f4d(%ecx),%xmm1
-  DB  102,15,111,129,93,127,0,0           ; movdqa        0x7f5d(%ecx),%xmm0
+  DB  102,15,219,137,109,136,0,0          ; pand          0x886d(%ecx),%xmm1
+  DB  102,15,111,129,125,136,0,0          ; movdqa        0x887d(%ecx),%xmm0
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
   DB  102,15,102,217                      ; pcmpgtd       %xmm1,%xmm3
@@ -50809,7 +52208,7 @@
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  102,15,111,213                      ; movdqa        %xmm5,%xmm2
   DB  102,15,219,214                      ; pand          %xmm6,%xmm2
-  DB  102,15,111,153,77,127,0,0           ; movdqa        0x7f4d(%ecx),%xmm3
+  DB  102,15,111,153,109,136,0,0          ; movdqa        0x886d(%ecx),%xmm3
   DB  102,15,219,235                      ; pand          %xmm3,%xmm5
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,111,224                      ; movdqa        %xmm0,%xmm4
@@ -50823,7 +52222,7 @@
   DB  102,15,219,254                      ; pand          %xmm6,%xmm7
   DB  102,15,219,195                      ; pand          %xmm3,%xmm0
   DB  102,15,114,247,16                   ; pslld         $0x10,%xmm7
-  DB  102,15,111,145,93,127,0,0           ; movdqa        0x7f5d(%ecx),%xmm2
+  DB  102,15,111,145,125,136,0,0          ; movdqa        0x887d(%ecx),%xmm2
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
   DB  102,15,102,216                      ; pcmpgtd       %xmm0,%xmm3
   DB  102,15,114,240,13                   ; pslld         $0xd,%xmm0
@@ -50832,12 +52231,12 @@
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,97,252                       ; punpcklwd     %xmm4,%xmm7
   DB  102,15,219,247                      ; pand          %xmm7,%xmm6
-  DB  102,15,219,185,77,127,0,0           ; pand          0x7f4d(%ecx),%xmm7
+  DB  102,15,219,185,109,136,0,0          ; pand          0x886d(%ecx),%xmm7
   DB  102,15,114,246,16                   ; pslld         $0x10,%xmm6
   DB  102,15,102,215                      ; pcmpgtd       %xmm7,%xmm2
   DB  102,15,114,247,13                   ; pslld         $0xd,%xmm7
   DB  102,15,235,254                      ; por           %xmm6,%xmm7
-  DB  102,15,111,161,109,127,0,0          ; movdqa        0x7f6d(%ecx),%xmm4
+  DB  102,15,111,161,141,136,0,0          ; movdqa        0x888d(%ecx),%xmm4
   DB  102,15,254,204                      ; paddd         %xmm4,%xmm1
   DB  102,15,254,236                      ; paddd         %xmm4,%xmm5
   DB  102,15,254,196                      ; paddd         %xmm4,%xmm0
@@ -50874,22 +52273,22 @@
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  232,0,0,0,0                         ; call          5797 <_sk_store_f16_sse2+0x19>
   DB  88                                  ; pop           %eax
-  DB  102,15,111,152,185,125,0,0          ; movdqa        0x7db9(%eax),%xmm3
+  DB  102,15,111,152,217,134,0,0          ; movdqa        0x86d9(%eax),%xmm3
   DB  102,15,111,224                      ; movdqa        %xmm0,%xmm4
   DB  102,15,219,227                      ; pand          %xmm3,%xmm4
-  DB  102,15,111,136,233,125,0,0          ; movdqa        0x7de9(%eax),%xmm1
+  DB  102,15,111,136,9,135,0,0            ; movdqa        0x8709(%eax),%xmm1
   DB  102,15,127,77,232                   ; movdqa        %xmm1,-0x18(%ebp)
   DB  102,15,111,240                      ; movdqa        %xmm0,%xmm6
   DB  102,15,114,246,3                    ; pslld         $0x3,%xmm6
   DB  102,15,219,241                      ; pand          %xmm1,%xmm6
   DB  102,15,254,244                      ; paddd         %xmm4,%xmm6
-  DB  102,15,111,184,201,125,0,0          ; movdqa        0x7dc9(%eax),%xmm7
+  DB  102,15,111,184,233,134,0,0          ; movdqa        0x86e9(%eax),%xmm7
   DB  102,15,127,69,168                   ; movdqa        %xmm0,-0x58(%ebp)
   DB  102,15,219,199                      ; pand          %xmm7,%xmm0
-  DB  102,15,111,168,217,125,0,0          ; movdqa        0x7dd9(%eax),%xmm5
+  DB  102,15,111,168,249,134,0,0          ; movdqa        0x86f9(%eax),%xmm5
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
   DB  102,15,102,200                      ; pcmpgtd       %xmm0,%xmm1
-  DB  102,15,111,128,105,122,0,0          ; movdqa        0x7a69(%eax),%xmm0
+  DB  102,15,111,128,137,131,0,0          ; movdqa        0x8389(%eax),%xmm0
   DB  102,15,127,69,216                   ; movdqa        %xmm0,-0x28(%ebp)
   DB  102,15,254,240                      ; paddd         %xmm0,%xmm6
   DB  102,15,114,230,16                   ; psrad         $0x10,%xmm6
@@ -51021,7 +52420,7 @@
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,97,195                       ; punpcklwd     %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,40,162,124,121,0,0               ; movaps        0x797c(%edx),%xmm4
+  DB  15,40,162,156,130,0,0               ; movaps        0x829c(%edx),%xmm4
   DB  15,89,196                           ; mulps         %xmm4,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -51123,7 +52522,7 @@
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,97,195                       ; punpcklwd     %xmm3,%xmm0
   DB  15,91,192                           ; cvtdq2ps      %xmm0,%xmm0
-  DB  15,40,170,21,120,0,0                ; movaps        0x7815(%edx),%xmm5
+  DB  15,40,170,53,129,0,0                ; movaps        0x8135(%edx),%xmm5
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
   DB  102,15,111,204                      ; movdqa        %xmm4,%xmm1
   DB  102,15,113,241,8                    ; psllw         $0x8,%xmm1
@@ -51149,7 +52548,7 @@
   DB  15,89,213                           ; mulps         %xmm5,%xmm2
   DB  141,112,8                           ; lea           0x8(%eax),%esi
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,154,149,117,0,0               ; movaps        0x7595(%edx),%xmm3
+  DB  15,40,154,181,126,0,0               ; movaps        0x7eb5(%edx),%xmm3
   DB  86                                  ; push          %esi
   DB  81                                  ; push          %ecx
   DB  255,80,4                            ; call          *0x4(%eax)
@@ -51194,7 +52593,7 @@
   DB  1,210                               ; add           %edx,%edx
   DB  3,22                                ; add           (%esi),%edx
   DB  139,49                              ; mov           (%ecx),%esi
-  DB  15,40,183,113,121,0,0               ; movaps        0x7971(%edi),%xmm6
+  DB  15,40,183,145,130,0,0               ; movaps        0x8291(%edi),%xmm6
   DB  139,121,8                           ; mov           0x8(%ecx),%edi
   DB  193,230,2                           ; shl           $0x2,%esi
   DB  15,40,224                           ; movaps        %xmm0,%xmm4
@@ -51466,7 +52865,7 @@
   DB  243,15,91,245                       ; cvttps2dq     %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  15,194,238,1                        ; cmpltps       %xmm6,%xmm5
-  DB  15,84,168,29,113,0,0                ; andps         0x711d(%eax),%xmm5
+  DB  15,84,168,61,122,0,0                ; andps         0x7a3d(%eax),%xmm5
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
   DB  15,89,230                           ; mulps         %xmm6,%xmm4
@@ -51496,7 +52895,7 @@
   DB  243,15,91,245                       ; cvttps2dq     %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  15,194,238,1                        ; cmpltps       %xmm6,%xmm5
-  DB  15,84,168,203,112,0,0               ; andps         0x70cb(%eax),%xmm5
+  DB  15,84,168,235,121,0,0               ; andps         0x79eb(%eax),%xmm5
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
   DB  15,89,230                           ; mulps         %xmm6,%xmm4
@@ -51524,13 +52923,13 @@
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
   DB  15,92,198                           ; subps         %xmm6,%xmm0
-  DB  243,15,89,169,25,120,0,0            ; mulss         0x7819(%ecx),%xmm5
+  DB  243,15,89,169,57,129,0,0            ; mulss         0x8139(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,89,232                           ; mulps         %xmm0,%xmm5
   DB  243,15,91,253                       ; cvttps2dq     %xmm5,%xmm7
   DB  15,91,255                           ; cvtdq2ps      %xmm7,%xmm7
   DB  15,194,239,1                        ; cmpltps       %xmm7,%xmm5
-  DB  15,84,169,121,112,0,0               ; andps         0x7079(%ecx),%xmm5
+  DB  15,84,169,153,121,0,0               ; andps         0x7999(%ecx),%xmm5
   DB  15,92,253                           ; subps         %xmm5,%xmm7
   DB  243,15,88,228                       ; addss         %xmm4,%xmm4
   DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
@@ -51563,13 +52962,13 @@
   DB  15,40,244                           ; movaps        %xmm4,%xmm6
   DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
   DB  15,92,206                           ; subps         %xmm6,%xmm1
-  DB  243,15,89,169,165,119,0,0           ; mulss         0x77a5(%ecx),%xmm5
+  DB  243,15,89,169,197,128,0,0           ; mulss         0x80c5(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,89,233                           ; mulps         %xmm1,%xmm5
   DB  243,15,91,253                       ; cvttps2dq     %xmm5,%xmm7
   DB  15,91,255                           ; cvtdq2ps      %xmm7,%xmm7
   DB  15,194,239,1                        ; cmpltps       %xmm7,%xmm5
-  DB  15,84,169,5,112,0,0                 ; andps         0x7005(%ecx),%xmm5
+  DB  15,84,169,37,121,0,0                ; andps         0x7925(%ecx),%xmm5
   DB  15,92,253                           ; subps         %xmm5,%xmm7
   DB  243,15,88,228                       ; addss         %xmm4,%xmm4
   DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
@@ -51598,7 +52997,7 @@
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,95,224                           ; maxps         %xmm0,%xmm4
-  DB  15,93,160,145,111,0,0               ; minps         0x6f91(%eax),%xmm4
+  DB  15,93,160,177,120,0,0               ; minps         0x78b1(%eax),%xmm4
   DB  141,65,4                            ; lea           0x4(%ecx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  15,40,196                           ; movaps        %xmm4,%xmm0
@@ -51621,7 +53020,7 @@
   DB  15,91,236                           ; cvtdq2ps      %xmm4,%xmm5
   DB  15,40,240                           ; movaps        %xmm0,%xmm6
   DB  15,194,245,1                        ; cmpltps       %xmm5,%xmm6
-  DB  15,40,184,97,111,0,0                ; movaps        0x6f61(%eax),%xmm7
+  DB  15,40,184,129,120,0,0               ; movaps        0x7881(%eax),%xmm7
   DB  15,84,247                           ; andps         %xmm7,%xmm6
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,92,238                           ; subps         %xmm6,%xmm5
@@ -51645,14 +53044,14 @@
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  232,0,0,0,0                         ; call          6149 <_sk_mirror_x_1_sse2+0xb>
   DB  88                                  ; pop           %eax
-  DB  15,40,160,135,111,0,0               ; movaps        0x6f87(%eax),%xmm4
+  DB  15,40,160,167,120,0,0               ; movaps        0x78a7(%eax),%xmm4
   DB  15,88,196                           ; addps         %xmm4,%xmm0
-  DB  15,40,168,7,111,0,0                 ; movaps        0x6f07(%eax),%xmm5
+  DB  15,40,168,39,120,0,0                ; movaps        0x7827(%eax),%xmm5
   DB  15,89,232                           ; mulps         %xmm0,%xmm5
   DB  243,15,91,245                       ; cvttps2dq     %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  15,194,238,1                        ; cmpltps       %xmm6,%xmm5
-  DB  15,40,184,23,111,0,0                ; movaps        0x6f17(%eax),%xmm7
+  DB  15,40,184,55,120,0,0                ; movaps        0x7837(%eax),%xmm7
   DB  15,84,239                           ; andps         %xmm7,%xmm5
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  15,88,246                           ; addps         %xmm6,%xmm6
@@ -51683,10 +53082,10 @@
   DB  232,0,0,0,0                         ; call          61b3 <_sk_luminance_to_alpha_sse2+0xe>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,89,128,237,115,0,0               ; mulps         0x73ed(%eax),%xmm0
-  DB  15,89,136,253,115,0,0               ; mulps         0x73fd(%eax),%xmm1
+  DB  15,89,128,13,125,0,0                ; mulps         0x7d0d(%eax),%xmm0
+  DB  15,89,136,29,125,0,0                ; mulps         0x7d1d(%eax),%xmm1
   DB  15,88,200                           ; addps         %xmm0,%xmm1
-  DB  15,89,152,13,116,0,0                ; mulps         0x740d(%eax),%xmm3
+  DB  15,89,152,45,125,0,0                ; mulps         0x7d2d(%eax),%xmm3
   DB  15,88,217                           ; addps         %xmm1,%xmm3
   DB  141,65,4                            ; lea           0x4(%ecx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52066,12 +53465,12 @@
   DB  78                                  ; dec           %esi
   DB  102,15,110,206                      ; movd          %esi,%xmm1
   DB  102,15,112,201,0                    ; pshufd        $0x0,%xmm1,%xmm1
-  DB  102,15,111,146,135,111,0,0          ; movdqa        0x6f87(%edx),%xmm2
+  DB  102,15,111,146,167,120,0,0          ; movdqa        0x78a7(%edx),%xmm2
   DB  102,15,219,209                      ; pand          %xmm1,%xmm2
-  DB  102,15,235,146,71,109,0,0           ; por           0x6d47(%edx),%xmm2
+  DB  102,15,235,146,103,118,0,0          ; por           0x7667(%edx),%xmm2
   DB  102,15,114,209,16                   ; psrld         $0x10,%xmm1
-  DB  102,15,235,138,151,111,0,0          ; por           0x6f97(%edx),%xmm1
-  DB  15,88,138,167,111,0,0               ; addps         0x6fa7(%edx),%xmm1
+  DB  102,15,235,138,183,120,0,0          ; por           0x78b7(%edx),%xmm1
+  DB  15,88,138,199,120,0,0               ; addps         0x78c7(%edx),%xmm1
   DB  15,88,202                           ; addps         %xmm2,%xmm1
   DB  15,89,200                           ; mulps         %xmm0,%xmm1
   DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
@@ -52178,15 +53577,15 @@
   DB  232,0,0,0,0                         ; call          6802 <_sk_gauss_a_to_rgba_sse2+0xb>
   DB  88                                  ; pop           %eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
-  DB  15,40,128,254,109,0,0               ; movaps        0x6dfe(%eax),%xmm0
+  DB  15,40,128,30,119,0,0                ; movaps        0x771e(%eax),%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,128,14,110,0,0                ; addps         0x6e0e(%eax),%xmm0
+  DB  15,88,128,46,119,0,0                ; addps         0x772e(%eax),%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,128,30,110,0,0                ; addps         0x6e1e(%eax),%xmm0
+  DB  15,88,128,62,119,0,0                ; addps         0x773e(%eax),%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,128,46,110,0,0                ; addps         0x6e2e(%eax),%xmm0
+  DB  15,88,128,78,119,0,0                ; addps         0x774e(%eax),%xmm0
   DB  15,89,195                           ; mulps         %xmm3,%xmm0
-  DB  15,88,128,62,110,0,0                ; addps         0x6e3e(%eax),%xmm0
+  DB  15,88,128,94,119,0,0                ; addps         0x775e(%eax),%xmm0
   DB  141,65,4                            ; lea           0x4(%ecx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
@@ -52387,30 +53786,30 @@
   DB  15,94,247                           ; divps         %xmm7,%xmm6
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
   DB  15,89,255                           ; mulps         %xmm7,%xmm7
-  DB  15,40,128,191,107,0,0               ; movaps        0x6bbf(%eax),%xmm0
+  DB  15,40,128,223,116,0,0               ; movaps        0x74df(%eax),%xmm0
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  15,88,128,207,107,0,0               ; addps         0x6bcf(%eax),%xmm0
+  DB  15,88,128,239,116,0,0               ; addps         0x74ef(%eax),%xmm0
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  15,88,128,223,107,0,0               ; addps         0x6bdf(%eax),%xmm0
+  DB  15,88,128,255,116,0,0               ; addps         0x74ff(%eax),%xmm0
   DB  15,89,199                           ; mulps         %xmm7,%xmm0
-  DB  15,88,128,239,107,0,0               ; addps         0x6bef(%eax),%xmm0
+  DB  15,88,128,15,117,0,0                ; addps         0x750f(%eax),%xmm0
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
   DB  15,194,236,1                        ; cmpltps       %xmm4,%xmm5
-  DB  15,40,176,255,107,0,0               ; movaps        0x6bff(%eax),%xmm6
+  DB  15,40,176,31,117,0,0                ; movaps        0x751f(%eax),%xmm6
   DB  15,92,240                           ; subps         %xmm0,%xmm6
   DB  15,84,245                           ; andps         %xmm5,%xmm6
   DB  15,85,232                           ; andnps        %xmm0,%xmm5
   DB  15,87,228                           ; xorps         %xmm4,%xmm4
   DB  15,86,238                           ; orps          %xmm6,%xmm5
   DB  15,194,204,1                        ; cmpltps       %xmm4,%xmm1
-  DB  15,40,128,191,101,0,0               ; movaps        0x65bf(%eax),%xmm0
+  DB  15,40,128,223,110,0,0               ; movaps        0x6edf(%eax),%xmm0
   DB  15,92,197                           ; subps         %xmm5,%xmm0
   DB  15,84,193                           ; andps         %xmm1,%xmm0
   DB  15,85,205                           ; andnps        %xmm5,%xmm1
   DB  15,86,200                           ; orps          %xmm0,%xmm1
   DB  15,40,194                           ; movaps        %xmm2,%xmm0
   DB  15,194,196,1                        ; cmpltps       %xmm4,%xmm0
-  DB  15,40,168,207,101,0,0               ; movaps        0x65cf(%eax),%xmm5
+  DB  15,40,168,239,110,0,0               ; movaps        0x6eef(%eax),%xmm5
   DB  15,92,233                           ; subps         %xmm1,%xmm5
   DB  15,84,232                           ; andps         %xmm0,%xmm5
   DB  15,85,193                           ; andnps        %xmm1,%xmm0
@@ -52474,8 +53873,8 @@
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  243,15,16,65,64                     ; movss         0x40(%ecx),%xmm0
   DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
-  DB  15,89,162,242,106,0,0               ; mulps         0x6af2(%edx),%xmm4
-  DB  15,89,130,2,107,0,0                 ; mulps         0x6b02(%edx),%xmm0
+  DB  15,89,162,18,116,0,0                ; mulps         0x7412(%edx),%xmm4
+  DB  15,89,130,34,116,0,0                ; mulps         0x7422(%edx),%xmm0
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,237                           ; mulps         %xmm5,%xmm5
@@ -52483,11 +53882,11 @@
   DB  15,81,237                           ; sqrtps        %xmm5,%xmm5
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  15,87,162,162,105,0,0               ; xorps         0x69a2(%edx),%xmm4
+  DB  15,87,162,194,114,0,0               ; xorps         0x72c2(%edx),%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  243,15,16,105,68                    ; movss         0x44(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
-  DB  15,89,170,162,100,0,0               ; mulps         0x64a2(%edx),%xmm5
+  DB  15,89,170,194,109,0,0               ; mulps         0x6dc2(%edx),%xmm5
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,95,196                           ; maxps         %xmm4,%xmm0
@@ -52523,8 +53922,8 @@
   DB  15,92,245                           ; subps         %xmm5,%xmm6
   DB  243,15,16,65,64                     ; movss         0x40(%ecx),%xmm0
   DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
-  DB  15,89,162,84,106,0,0                ; mulps         0x6a54(%edx),%xmm4
-  DB  15,89,130,100,106,0,0               ; mulps         0x6a64(%edx),%xmm0
+  DB  15,89,162,116,115,0,0               ; mulps         0x7374(%edx),%xmm4
+  DB  15,89,130,132,115,0,0               ; mulps         0x7384(%edx),%xmm0
   DB  15,89,198                           ; mulps         %xmm6,%xmm0
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,237                           ; mulps         %xmm5,%xmm5
@@ -52532,11 +53931,11 @@
   DB  15,81,237                           ; sqrtps        %xmm5,%xmm5
   DB  15,40,197                           ; movaps        %xmm5,%xmm0
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  15,87,162,4,105,0,0                 ; xorps         0x6904(%edx),%xmm4
+  DB  15,87,162,36,114,0,0                ; xorps         0x7224(%edx),%xmm4
   DB  15,92,229                           ; subps         %xmm5,%xmm4
   DB  243,15,16,105,68                    ; movss         0x44(%ecx),%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
-  DB  15,89,170,4,100,0,0                 ; mulps         0x6404(%edx),%xmm5
+  DB  15,89,170,36,109,0,0                ; mulps         0x6d24(%edx),%xmm5
   DB  15,89,197                           ; mulps         %xmm5,%xmm0
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,93,196                           ; minps         %xmm4,%xmm0
@@ -52563,7 +53962,7 @@
   DB  243,15,89,236                       ; mulss         %xmm4,%xmm5
   DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
   DB  15,88,232                           ; addps         %xmm0,%xmm5
-  DB  15,89,169,220,105,0,0               ; mulps         0x69dc(%ecx),%xmm5
+  DB  15,89,169,252,114,0,0               ; mulps         0x72fc(%ecx),%xmm5
   DB  15,89,192                           ; mulps         %xmm0,%xmm0
   DB  15,40,241                           ; movaps        %xmm1,%xmm6
   DB  15,89,246                           ; mulps         %xmm6,%xmm6
@@ -52571,7 +53970,7 @@
   DB  243,15,89,228                       ; mulss         %xmm4,%xmm4
   DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
   DB  15,92,196                           ; subps         %xmm4,%xmm0
-  DB  15,87,129,140,104,0,0               ; xorps         0x688c(%ecx),%xmm0
+  DB  15,87,129,172,113,0,0               ; xorps         0x71ac(%ecx),%xmm0
   DB  15,94,197                           ; divps         %xmm5,%xmm0
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52644,14 +54043,14 @@
   DB  15,40,218                           ; movaps        %xmm2,%xmm3
   DB  232,0,0,0,0                         ; call          6dad <_sk_save_xy_sse2+0x12>
   DB  88                                  ; pop           %eax
-  DB  15,40,160,163,98,0,0                ; movaps        0x62a3(%eax),%xmm4
+  DB  15,40,160,195,107,0,0               ; movaps        0x6bc3(%eax),%xmm4
   DB  15,40,232                           ; movaps        %xmm0,%xmm5
   DB  15,88,236                           ; addps         %xmm4,%xmm5
   DB  243,15,91,245                       ; cvttps2dq     %xmm5,%xmm6
   DB  15,91,246                           ; cvtdq2ps      %xmm6,%xmm6
   DB  15,40,253                           ; movaps        %xmm5,%xmm7
   DB  15,194,254,1                        ; cmpltps       %xmm6,%xmm7
-  DB  15,40,144,179,98,0,0                ; movaps        0x62b3(%eax),%xmm2
+  DB  15,40,144,211,107,0,0               ; movaps        0x6bd3(%eax),%xmm2
   DB  15,84,250                           ; andps         %xmm2,%xmm7
   DB  15,92,247                           ; subps         %xmm7,%xmm6
   DB  15,92,238                           ; subps         %xmm6,%xmm5
@@ -52729,8 +54128,8 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,166,128,0,0,0                 ; movups        0x80(%esi),%xmm4
-  DB  15,88,128,43,104,0,0                ; addps         0x682b(%eax),%xmm0
-  DB  15,40,168,203,97,0,0                ; movaps        0x61cb(%eax),%xmm5
+  DB  15,88,128,75,113,0,0                ; addps         0x714b(%eax),%xmm0
+  DB  15,40,168,235,106,0,0               ; movaps        0x6aeb(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,17,174,0,1,0,0                   ; movups        %xmm5,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -52756,7 +54155,7 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,166,128,0,0,0                 ; movups        0x80(%esi),%xmm4
-  DB  15,88,128,117,97,0,0                ; addps         0x6175(%eax),%xmm0
+  DB  15,88,128,149,106,0,0               ; addps         0x6a95(%eax),%xmm0
   DB  15,17,166,0,1,0,0                   ; movups        %xmm4,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52781,8 +54180,8 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,166,192,0,0,0                 ; movups        0xc0(%esi),%xmm4
-  DB  15,88,136,169,103,0,0               ; addps         0x67a9(%eax),%xmm1
-  DB  15,40,168,73,97,0,0                 ; movaps        0x6149(%eax),%xmm5
+  DB  15,88,136,201,112,0,0               ; addps         0x70c9(%eax),%xmm1
+  DB  15,40,168,105,106,0,0               ; movaps        0x6a69(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,17,174,64,1,0,0                  ; movups        %xmm5,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -52808,7 +54207,7 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,166,192,0,0,0                 ; movups        0xc0(%esi),%xmm4
-  DB  15,88,136,242,96,0,0                ; addps         0x60f2(%eax),%xmm1
+  DB  15,88,136,18,106,0,0                ; addps         0x6a12(%eax),%xmm1
   DB  15,17,166,64,1,0,0                  ; movups        %xmm4,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52833,13 +54232,13 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,166,128,0,0,0                 ; movups        0x80(%esi),%xmm4
-  DB  15,88,128,53,103,0,0                ; addps         0x6735(%eax),%xmm0
-  DB  15,40,168,197,96,0,0                ; movaps        0x60c5(%eax),%xmm5
+  DB  15,88,128,85,112,0,0                ; addps         0x7055(%eax),%xmm0
+  DB  15,40,168,229,105,0,0               ; movaps        0x69e5(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,40,229                           ; movaps        %xmm5,%xmm4
   DB  15,89,228                           ; mulps         %xmm4,%xmm4
-  DB  15,89,168,69,103,0,0                ; mulps         0x6745(%eax),%xmm5
-  DB  15,88,168,181,98,0,0                ; addps         0x62b5(%eax),%xmm5
+  DB  15,89,168,101,112,0,0               ; mulps         0x7065(%eax),%xmm5
+  DB  15,88,168,213,107,0,0               ; addps         0x6bd5(%eax),%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
   DB  15,17,174,0,1,0,0                   ; movups        %xmm5,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -52865,16 +54264,16 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,166,128,0,0,0                 ; movups        0x80(%esi),%xmm4
-  DB  15,88,128,200,102,0,0               ; addps         0x66c8(%eax),%xmm0
-  DB  15,40,168,104,96,0,0                ; movaps        0x6068(%eax),%xmm5
+  DB  15,88,128,232,111,0,0               ; addps         0x6fe8(%eax),%xmm0
+  DB  15,40,168,136,105,0,0               ; movaps        0x6988(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
-  DB  15,40,160,248,102,0,0               ; movaps        0x66f8(%eax),%xmm4
+  DB  15,40,160,24,112,0,0                ; movaps        0x7018(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,8,103,0,0                 ; addps         0x6708(%eax),%xmm4
+  DB  15,88,160,40,112,0,0                ; addps         0x7028(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,88,96,0,0                 ; addps         0x6058(%eax),%xmm4
+  DB  15,88,160,120,105,0,0               ; addps         0x6978(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,24,103,0,0                ; addps         0x6718(%eax),%xmm4
+  DB  15,88,160,56,112,0,0                ; addps         0x7038(%eax),%xmm4
   DB  15,17,166,0,1,0,0                   ; movups        %xmm4,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52897,17 +54296,17 @@
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,12                           ; mov           0xc(%ebp),%edx
   DB  139,50                              ; mov           (%edx),%esi
-  DB  15,40,160,237,95,0,0                ; movaps        0x5fed(%eax),%xmm4
+  DB  15,40,160,13,105,0,0                ; movaps        0x690d(%eax),%xmm4
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,174,128,0,0,0                 ; movups        0x80(%esi),%xmm5
   DB  15,88,196                           ; addps         %xmm4,%xmm0
-  DB  15,40,176,141,102,0,0               ; movaps        0x668d(%eax),%xmm6
+  DB  15,40,176,173,111,0,0               ; movaps        0x6fad(%eax),%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
-  DB  15,88,176,157,102,0,0               ; addps         0x669d(%eax),%xmm6
+  DB  15,88,176,189,111,0,0               ; addps         0x6fbd(%eax),%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
   DB  15,88,244                           ; addps         %xmm4,%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
-  DB  15,88,176,173,102,0,0               ; addps         0x66ad(%eax),%xmm6
+  DB  15,88,176,205,111,0,0               ; addps         0x6fcd(%eax),%xmm6
   DB  15,17,182,0,1,0,0                   ; movups        %xmm6,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -52932,11 +54331,11 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,6                             ; movups        (%esi),%xmm0
   DB  15,16,166,128,0,0,0                 ; movups        0x80(%esi),%xmm4
-  DB  15,88,128,61,102,0,0                ; addps         0x663d(%eax),%xmm0
+  DB  15,88,128,93,111,0,0                ; addps         0x6f5d(%eax),%xmm0
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,237                           ; mulps         %xmm5,%xmm5
-  DB  15,89,160,29,102,0,0                ; mulps         0x661d(%eax),%xmm4
-  DB  15,88,160,141,97,0,0                ; addps         0x618d(%eax),%xmm4
+  DB  15,89,160,61,111,0,0                ; mulps         0x6f3d(%eax),%xmm4
+  DB  15,88,160,173,106,0,0               ; addps         0x6aad(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,17,166,0,1,0,0                   ; movups        %xmm4,0x100(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -52962,13 +54361,13 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,166,192,0,0,0                 ; movups        0xc0(%esi),%xmm4
-  DB  15,88,136,186,101,0,0               ; addps         0x65ba(%eax),%xmm1
-  DB  15,40,168,74,95,0,0                 ; movaps        0x5f4a(%eax),%xmm5
+  DB  15,88,136,218,110,0,0               ; addps         0x6eda(%eax),%xmm1
+  DB  15,40,168,106,104,0,0               ; movaps        0x686a(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
   DB  15,40,229                           ; movaps        %xmm5,%xmm4
   DB  15,89,228                           ; mulps         %xmm4,%xmm4
-  DB  15,89,168,202,101,0,0               ; mulps         0x65ca(%eax),%xmm5
-  DB  15,88,168,58,97,0,0                 ; addps         0x613a(%eax),%xmm5
+  DB  15,89,168,234,110,0,0               ; mulps         0x6eea(%eax),%xmm5
+  DB  15,88,168,90,106,0,0                ; addps         0x6a5a(%eax),%xmm5
   DB  15,89,236                           ; mulps         %xmm4,%xmm5
   DB  15,17,174,64,1,0,0                  ; movups        %xmm5,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -52994,16 +54393,16 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,166,192,0,0,0                 ; movups        0xc0(%esi),%xmm4
-  DB  15,88,136,76,101,0,0                ; addps         0x654c(%eax),%xmm1
-  DB  15,40,168,236,94,0,0                ; movaps        0x5eec(%eax),%xmm5
+  DB  15,88,136,108,110,0,0               ; addps         0x6e6c(%eax),%xmm1
+  DB  15,40,168,12,104,0,0                ; movaps        0x680c(%eax),%xmm5
   DB  15,92,236                           ; subps         %xmm4,%xmm5
-  DB  15,40,160,124,101,0,0               ; movaps        0x657c(%eax),%xmm4
+  DB  15,40,160,156,110,0,0               ; movaps        0x6e9c(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,140,101,0,0               ; addps         0x658c(%eax),%xmm4
+  DB  15,88,160,172,110,0,0               ; addps         0x6eac(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,220,94,0,0                ; addps         0x5edc(%eax),%xmm4
+  DB  15,88,160,252,103,0,0               ; addps         0x67fc(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
-  DB  15,88,160,156,101,0,0               ; addps         0x659c(%eax),%xmm4
+  DB  15,88,160,188,110,0,0               ; addps         0x6ebc(%eax),%xmm4
   DB  15,17,166,64,1,0,0                  ; movups        %xmm4,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -53026,17 +54425,17 @@
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,12                           ; mov           0xc(%ebp),%edx
   DB  139,50                              ; mov           (%edx),%esi
-  DB  15,40,160,112,94,0,0                ; movaps        0x5e70(%eax),%xmm4
+  DB  15,40,160,144,103,0,0               ; movaps        0x6790(%eax),%xmm4
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,174,192,0,0,0                 ; movups        0xc0(%esi),%xmm5
   DB  15,88,204                           ; addps         %xmm4,%xmm1
-  DB  15,40,176,16,101,0,0                ; movaps        0x6510(%eax),%xmm6
+  DB  15,40,176,48,110,0,0                ; movaps        0x6e30(%eax),%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
-  DB  15,88,176,32,101,0,0                ; addps         0x6520(%eax),%xmm6
+  DB  15,88,176,64,110,0,0                ; addps         0x6e40(%eax),%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
   DB  15,88,244                           ; addps         %xmm4,%xmm6
   DB  15,89,245                           ; mulps         %xmm5,%xmm6
-  DB  15,88,176,48,101,0,0                ; addps         0x6530(%eax),%xmm6
+  DB  15,88,176,80,110,0,0                ; addps         0x6e50(%eax),%xmm6
   DB  15,17,182,64,1,0,0                  ; movups        %xmm6,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
@@ -53061,11 +54460,11 @@
   DB  139,50                              ; mov           (%edx),%esi
   DB  15,16,78,64                         ; movups        0x40(%esi),%xmm1
   DB  15,16,166,192,0,0,0                 ; movups        0xc0(%esi),%xmm4
-  DB  15,88,136,191,100,0,0               ; addps         0x64bf(%eax),%xmm1
+  DB  15,88,136,223,109,0,0               ; addps         0x6ddf(%eax),%xmm1
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
   DB  15,89,237                           ; mulps         %xmm5,%xmm5
-  DB  15,89,160,159,100,0,0               ; mulps         0x649f(%eax),%xmm4
-  DB  15,88,160,15,96,0,0                 ; addps         0x600f(%eax),%xmm4
+  DB  15,89,160,191,109,0,0               ; mulps         0x6dbf(%eax),%xmm4
+  DB  15,88,160,47,105,0,0                ; addps         0x692f(%eax),%xmm4
   DB  15,89,229                           ; mulps         %xmm5,%xmm4
   DB  15,17,166,64,1,0,0                  ; movups        %xmm4,0x140(%esi)
   DB  141,66,8                            ; lea           0x8(%edx),%eax
@@ -53212,7 +54611,7 @@
   DB  102,15,254,216                      ; paddd         %xmm0,%xmm3
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,254,195                      ; paddd         %xmm3,%xmm0
-  DB  102,15,111,162,9,99,0,0             ; movdqa        0x6309(%edx),%xmm4
+  DB  102,15,111,162,41,108,0,0           ; movdqa        0x6c29(%edx),%xmm4
   DB  102,15,127,165,120,255,255,255      ; movdqa        %xmm4,-0x88(%ebp)
   DB  102,15,112,200,245                  ; pshufd        $0xf5,%xmm0,%xmm1
   DB  102,15,244,196                      ; pmuludq       %xmm4,%xmm0
@@ -53252,7 +54651,7 @@
   DB  243,15,16,4,129                     ; movss         (%ecx,%eax,4),%xmm0
   DB  15,20,248                           ; unpcklps      %xmm0,%xmm7
   DB  102,15,20,252                       ; unpcklpd      %xmm4,%xmm7
-  DB  102,15,111,130,121,92,0,0           ; movdqa        0x5c79(%edx),%xmm0
+  DB  102,15,111,130,153,101,0,0          ; movdqa        0x6599(%edx),%xmm0
   DB  102,15,127,133,104,255,255,255      ; movdqa        %xmm0,-0x98(%ebp)
   DB  102,15,254,200                      ; paddd         %xmm0,%xmm1
   DB  102,15,112,193,229                  ; pshufd        $0xe5,%xmm1,%xmm0
@@ -53269,7 +54668,7 @@
   DB  243,15,16,4,129                     ; movss         (%ecx,%eax,4),%xmm0
   DB  15,20,240                           ; unpcklps      %xmm0,%xmm6
   DB  102,15,20,241                       ; unpcklpd      %xmm1,%xmm6
-  DB  15,40,130,249,98,0,0                ; movaps        0x62f9(%edx),%xmm0
+  DB  15,40,130,25,108,0,0                ; movaps        0x6c19(%edx),%xmm0
   DB  15,40,141,248,254,255,255           ; movaps        -0x108(%ebp),%xmm1
   DB  15,88,200                           ; addps         %xmm0,%xmm1
   DB  15,41,77,200                        ; movaps        %xmm1,-0x38(%ebp)
@@ -53898,7 +55297,7 @@
   DB  102,15,111,212                      ; movdqa        %xmm4,%xmm2
   DB  102,15,254,214                      ; paddd         %xmm6,%xmm2
   DB  102,15,254,218                      ; paddd         %xmm2,%xmm3
-  DB  102,15,111,161,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm4
+  DB  102,15,111,161,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm4
   DB  102,15,112,203,245                  ; pshufd        $0xf5,%xmm3,%xmm1
   DB  102,15,244,220                      ; pmuludq       %xmm4,%xmm3
   DB  102,15,244,204                      ; pmuludq       %xmm4,%xmm1
@@ -53940,7 +55339,7 @@
   DB  15,20,223                           ; unpcklps      %xmm7,%xmm3
   DB  102,15,20,222                       ; unpcklpd      %xmm6,%xmm3
   DB  102,15,41,157,232,254,255,255       ; movapd        %xmm3,-0x118(%ebp)
-  DB  102,15,254,137,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm1
+  DB  102,15,254,137,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm1
   DB  102,15,112,241,229                  ; pshufd        $0xe5,%xmm1,%xmm6
   DB  102,15,126,240                      ; movd          %xmm6,%eax
   DB  102,15,112,241,78                   ; pshufd        $0x4e,%xmm1,%xmm6
@@ -53955,7 +55354,7 @@
   DB  243,15,16,60,130                    ; movss         (%edx,%eax,4),%xmm7
   DB  15,20,223                           ; unpcklps      %xmm7,%xmm3
   DB  102,15,20,222                       ; unpcklpd      %xmm6,%xmm3
-  DB  15,40,185,109,87,0,0                ; movaps        0x576d(%ecx),%xmm7
+  DB  15,40,185,141,96,0,0                ; movaps        0x608d(%ecx),%xmm7
   DB  15,88,199                           ; addps         %xmm7,%xmm0
   DB  15,41,133,120,255,255,255           ; movaps        %xmm0,-0x88(%ebp)
   DB  15,40,133,248,254,255,255           ; movaps        -0x108(%ebp),%xmm0
@@ -53976,7 +55375,7 @@
   DB  102,15,127,101,168                  ; movdqa        %xmm4,-0x58(%ebp)
   DB  102,15,254,212                      ; paddd         %xmm4,%xmm2
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,161,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm4
+  DB  102,15,111,161,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm4
   DB  102,15,244,212                      ; pmuludq       %xmm4,%xmm2
   DB  102,15,244,196                      ; pmuludq       %xmm4,%xmm0
   DB  102,15,112,240,232                  ; pshufd        $0xe8,%xmm0,%xmm6
@@ -54012,7 +55411,7 @@
   DB  243,15,16,44,130                    ; movss         (%edx,%eax,4),%xmm5
   DB  15,20,213                           ; unpcklps      %xmm5,%xmm2
   DB  102,15,20,214                       ; unpcklpd      %xmm6,%xmm2
-  DB  102,15,254,129,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm0
+  DB  102,15,254,129,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm0
   DB  102,15,112,232,229                  ; pshufd        $0xe5,%xmm0,%xmm5
   DB  102,15,126,232                      ; movd          %xmm5,%eax
   DB  102,15,112,232,78                   ; pshufd        $0x4e,%xmm0,%xmm5
@@ -54096,7 +55495,7 @@
   DB  243,15,16,44,130                    ; movss         (%edx,%eax,4),%xmm5
   DB  15,20,245                           ; unpcklps      %xmm5,%xmm6
   DB  102,15,20,240                       ; unpcklpd      %xmm0,%xmm6
-  DB  102,15,254,137,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm1
+  DB  102,15,254,137,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm1
   DB  102,15,112,193,229                  ; pshufd        $0xe5,%xmm1,%xmm0
   DB  102,15,126,192                      ; movd          %xmm0,%eax
   DB  102,15,112,193,78                   ; pshufd        $0x4e,%xmm1,%xmm0
@@ -54114,7 +55513,7 @@
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,15,254,77,168                   ; paddd         -0x58(%ebp),%xmm1
   DB  102,15,112,193,245                  ; pshufd        $0xf5,%xmm1,%xmm0
-  DB  102,15,111,145,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm2
+  DB  102,15,111,145,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm2
   DB  102,15,244,202                      ; pmuludq       %xmm2,%xmm1
   DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
   DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
@@ -54150,7 +55549,7 @@
   DB  243,15,16,4,130                     ; movss         (%edx,%eax,4),%xmm0
   DB  15,20,216                           ; unpcklps      %xmm0,%xmm3
   DB  102,15,20,217                       ; unpcklpd      %xmm1,%xmm3
-  DB  102,15,254,169,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm5
+  DB  102,15,254,169,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm5
   DB  102,15,112,197,229                  ; pshufd        $0xe5,%xmm5,%xmm0
   DB  102,15,126,192                      ; movd          %xmm0,%eax
   DB  102,15,112,197,78                   ; pshufd        $0x4e,%xmm5,%xmm0
@@ -54210,7 +55609,7 @@
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,254,85,152                   ; paddd         -0x68(%ebp),%xmm2
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,153,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm3
+  DB  102,15,111,153,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm3
   DB  102,15,244,211                      ; pmuludq       %xmm3,%xmm2
   DB  102,15,244,195                      ; pmuludq       %xmm3,%xmm0
   DB  102,15,111,251                      ; movdqa        %xmm3,%xmm7
@@ -54249,7 +55648,7 @@
   DB  243,15,16,44,130                    ; movss         (%edx,%eax,4),%xmm5
   DB  15,20,229                           ; unpcklps      %xmm5,%xmm4
   DB  102,15,20,227                       ; unpcklpd      %xmm3,%xmm4
-  DB  102,15,254,129,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm0
+  DB  102,15,254,129,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm0
   DB  102,15,112,216,229                  ; pshufd        $0xe5,%xmm0,%xmm3
   DB  102,15,126,216                      ; movd          %xmm3,%eax
   DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
@@ -54301,7 +55700,7 @@
   DB  243,15,16,52,130                    ; movss         (%edx,%eax,4),%xmm6
   DB  15,20,238                           ; unpcklps      %xmm6,%xmm5
   DB  102,15,20,235                       ; unpcklpd      %xmm3,%xmm5
-  DB  102,15,111,137,237,80,0,0           ; movdqa        0x50ed(%ecx),%xmm1
+  DB  102,15,111,137,13,90,0,0            ; movdqa        0x5a0d(%ecx),%xmm1
   DB  102,15,254,249                      ; paddd         %xmm1,%xmm7
   DB  102,15,112,223,229                  ; pshufd        $0xe5,%xmm7,%xmm3
   DB  102,15,126,216                      ; movd          %xmm3,%eax
@@ -54335,7 +55734,7 @@
   DB  102,15,111,226                      ; movdqa        %xmm2,%xmm4
   DB  102,15,254,69,152                   ; paddd         -0x68(%ebp),%xmm0
   DB  102,15,112,208,245                  ; pshufd        $0xf5,%xmm0,%xmm2
-  DB  102,15,111,153,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm3
+  DB  102,15,111,153,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm3
   DB  102,15,244,195                      ; pmuludq       %xmm3,%xmm0
   DB  102,15,244,211                      ; pmuludq       %xmm3,%xmm2
   DB  102,15,112,218,232                  ; pshufd        $0xe8,%xmm2,%xmm3
@@ -54392,7 +55791,7 @@
   DB  102,15,111,212                      ; movdqa        %xmm4,%xmm2
   DB  102,15,254,85,168                   ; paddd         -0x58(%ebp),%xmm2
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,137,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm1
+  DB  102,15,111,137,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm1
   DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
   DB  102,15,244,193                      ; pmuludq       %xmm1,%xmm0
   DB  102,15,112,240,232                  ; pshufd        $0xe8,%xmm0,%xmm6
@@ -54429,7 +55828,7 @@
   DB  243,15,16,36,130                    ; movss         (%edx,%eax,4),%xmm4
   DB  15,20,212                           ; unpcklps      %xmm4,%xmm2
   DB  102,15,20,214                       ; unpcklpd      %xmm6,%xmm2
-  DB  102,15,254,129,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm0
+  DB  102,15,254,129,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm0
   DB  102,15,112,224,229                  ; pshufd        $0xe5,%xmm0,%xmm4
   DB  102,15,126,224                      ; movd          %xmm4,%eax
   DB  102,15,112,224,78                   ; pshufd        $0x4e,%xmm0,%xmm4
@@ -54497,7 +55896,7 @@
   DB  102,15,111,93,152                   ; movdqa        -0x68(%ebp),%xmm3
   DB  102,15,254,217                      ; paddd         %xmm1,%xmm3
   DB  102,15,112,211,245                  ; pshufd        $0xf5,%xmm3,%xmm2
-  DB  102,15,111,129,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm0
+  DB  102,15,111,129,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm0
   DB  102,15,244,216                      ; pmuludq       %xmm0,%xmm3
   DB  102,15,244,208                      ; pmuludq       %xmm0,%xmm2
   DB  102,15,111,248                      ; movdqa        %xmm0,%xmm7
@@ -54536,7 +55935,7 @@
   DB  243,15,16,28,130                    ; movss         (%edx,%eax,4),%xmm3
   DB  15,20,235                           ; unpcklps      %xmm3,%xmm5
   DB  102,15,20,238                       ; unpcklpd      %xmm6,%xmm5
-  DB  102,15,254,145,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm2
+  DB  102,15,254,145,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm2
   DB  102,15,112,218,229                  ; pshufd        $0xe5,%xmm2,%xmm3
   DB  102,15,126,216                      ; movd          %xmm3,%eax
   DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
@@ -54588,7 +55987,7 @@
   DB  243,15,16,4,130                     ; movss         (%edx,%eax,4),%xmm0
   DB  15,20,208                           ; unpcklps      %xmm0,%xmm2
   DB  102,15,20,215                       ; unpcklpd      %xmm7,%xmm2
-  DB  102,15,111,137,237,80,0,0           ; movdqa        0x50ed(%ecx),%xmm1
+  DB  102,15,111,137,13,90,0,0            ; movdqa        0x5a0d(%ecx),%xmm1
   DB  102,15,254,217                      ; paddd         %xmm1,%xmm3
   DB  102,15,112,195,229                  ; pshufd        $0xe5,%xmm3,%xmm0
   DB  102,15,126,192                      ; movd          %xmm0,%eax
@@ -54623,7 +56022,7 @@
   DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
   DB  102,15,254,69,152                   ; paddd         -0x68(%ebp),%xmm0
   DB  102,15,112,224,245                  ; pshufd        $0xf5,%xmm0,%xmm4
-  DB  102,15,111,169,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm5
+  DB  102,15,111,169,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm5
   DB  102,15,111,221                      ; movdqa        %xmm5,%xmm3
   DB  102,15,244,195                      ; pmuludq       %xmm3,%xmm0
   DB  102,15,244,227                      ; pmuludq       %xmm3,%xmm4
@@ -54679,7 +56078,7 @@
   DB  102,15,111,202                      ; movdqa        %xmm2,%xmm1
   DB  102,15,254,77,168                   ; paddd         -0x58(%ebp),%xmm1
   DB  102,15,112,193,245                  ; pshufd        $0xf5,%xmm1,%xmm0
-  DB  102,15,111,145,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm2
+  DB  102,15,111,145,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm2
   DB  102,15,244,202                      ; pmuludq       %xmm2,%xmm1
   DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
   DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
@@ -54715,7 +56114,7 @@
   DB  243,15,16,4,130                     ; movss         (%edx,%eax,4),%xmm0
   DB  15,20,216                           ; unpcklps      %xmm0,%xmm3
   DB  102,15,20,217                       ; unpcklpd      %xmm1,%xmm3
-  DB  102,15,254,185,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm7
+  DB  102,15,254,185,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm7
   DB  102,15,112,199,229                  ; pshufd        $0xe5,%xmm7,%xmm0
   DB  102,15,126,192                      ; movd          %xmm0,%eax
   DB  102,15,112,199,78                   ; pshufd        $0x4e,%xmm7,%xmm0
@@ -54765,7 +56164,7 @@
   DB  102,15,111,233                      ; movdqa        %xmm1,%xmm5
   DB  102,15,254,69,152                   ; paddd         -0x68(%ebp),%xmm0
   DB  102,15,112,200,245                  ; pshufd        $0xf5,%xmm0,%xmm1
-  DB  102,15,111,145,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm2
+  DB  102,15,111,145,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm2
   DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
   DB  102,15,244,202                      ; pmuludq       %xmm2,%xmm1
   DB  102,15,111,250                      ; movdqa        %xmm2,%xmm7
@@ -54804,7 +56203,7 @@
   DB  243,15,16,36,130                    ; movss         (%edx,%eax,4),%xmm4
   DB  15,20,204                           ; unpcklps      %xmm4,%xmm1
   DB  102,15,20,203                       ; unpcklpd      %xmm3,%xmm1
-  DB  102,15,254,145,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm2
+  DB  102,15,254,145,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm2
   DB  102,15,112,218,229                  ; pshufd        $0xe5,%xmm2,%xmm3
   DB  102,15,126,216                      ; movd          %xmm3,%eax
   DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
@@ -54858,7 +56257,7 @@
   DB  243,15,16,36,130                    ; movss         (%edx,%eax,4),%xmm4
   DB  15,20,252                           ; unpcklps      %xmm4,%xmm7
   DB  102,15,20,254                       ; unpcklpd      %xmm6,%xmm7
-  DB  102,15,254,153,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm3
+  DB  102,15,254,153,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm3
   DB  102,15,112,227,229                  ; pshufd        $0xe5,%xmm3,%xmm4
   DB  102,15,126,224                      ; movd          %xmm4,%eax
   DB  102,15,112,227,78                   ; pshufd        $0x4e,%xmm3,%xmm4
@@ -54891,7 +56290,7 @@
   DB  102,15,254,208                      ; paddd         %xmm0,%xmm2
   DB  102,15,111,216                      ; movdqa        %xmm0,%xmm3
   DB  102,15,112,194,245                  ; pshufd        $0xf5,%xmm2,%xmm0
-  DB  102,15,111,137,125,87,0,0           ; movdqa        0x577d(%ecx),%xmm1
+  DB  102,15,111,137,157,96,0,0           ; movdqa        0x609d(%ecx),%xmm1
   DB  102,15,244,209                      ; pmuludq       %xmm1,%xmm2
   DB  102,15,244,193                      ; pmuludq       %xmm1,%xmm0
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
@@ -54931,7 +56330,7 @@
   DB  15,20,225                           ; unpcklps      %xmm1,%xmm4
   DB  102,15,20,226                       ; unpcklpd      %xmm2,%xmm4
   DB  102,15,41,165,88,255,255,255        ; movapd        %xmm4,-0xa8(%ebp)
-  DB  102,15,254,129,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm0
+  DB  102,15,254,129,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm0
   DB  102,15,112,200,229                  ; pshufd        $0xe5,%xmm0,%xmm1
   DB  102,15,126,200                      ; movd          %xmm1,%eax
   DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
@@ -54985,7 +56384,7 @@
   DB  243,15,16,52,130                    ; movss         (%edx,%eax,4),%xmm6
   DB  15,20,206                           ; unpcklps      %xmm6,%xmm1
   DB  102,15,20,203                       ; unpcklpd      %xmm3,%xmm1
-  DB  102,15,254,145,237,80,0,0           ; paddd         0x50ed(%ecx),%xmm2
+  DB  102,15,254,145,13,90,0,0            ; paddd         0x5a0d(%ecx),%xmm2
   DB  102,15,112,218,229                  ; pshufd        $0xe5,%xmm2,%xmm3
   DB  102,15,126,216                      ; movd          %xmm3,%eax
   DB  102,15,112,218,78                   ; pshufd        $0x4e,%xmm2,%xmm3
@@ -55054,7 +56453,7 @@
   DB  139,85,12                           ; mov           0xc(%ebp),%edx
   DB  141,66,8                            ; lea           0x8(%edx),%eax
   DB  131,236,8                           ; sub           $0x8,%esp
-  DB  15,40,153,173,80,0,0                ; movaps        0x50ad(%ecx),%xmm3
+  DB  15,40,153,205,89,0,0                ; movaps        0x59cd(%ecx),%xmm3
   DB  80                                  ; push          %eax
   DB  255,117,8                           ; pushl         0x8(%ebp)
   DB  255,82,4                            ; call          *0x4(%edx)
@@ -55179,7 +56578,7 @@
   DB  102,15,110,210                      ; movd          %edx,%xmm2
   DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
   DB  15,91,210                           ; cvtdq2ps      %xmm2,%xmm2
-  DB  15,88,145,92,59,0,0                 ; addps         0x3b5c(%ecx),%xmm2
+  DB  15,88,145,124,68,0,0                ; addps         0x447c(%ecx),%xmm2
   DB  141,78,8                            ; lea           0x8(%esi),%ecx
   DB  15,41,116,36,64                     ; movaps        %xmm6,0x40(%esp)
   DB  15,41,108,36,48                     ; movaps        %xmm5,0x30(%esp)
@@ -55197,6 +56596,94 @@
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
 
+PUBLIC _sk_matrix_translate_sse2_lowp
+_sk_matrix_translate_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,80                          ; sub           $0x50,%esp
+  DB  139,69,8                            ; mov           0x8(%ebp),%eax
+  DB  139,77,16                           ; mov           0x10(%ebp),%ecx
+  DB  139,85,20                           ; mov           0x14(%ebp),%edx
+  DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
+  DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
+  DB  15,40,117,56                        ; movaps        0x38(%ebp),%xmm6
+  DB  139,117,12                          ; mov           0xc(%ebp),%esi
+  DB  139,62                              ; mov           (%esi),%edi
+  DB  243,15,16,63                        ; movss         (%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,88,199                           ; addps         %xmm7,%xmm0
+  DB  15,88,207                           ; addps         %xmm7,%xmm1
+  DB  243,15,16,127,4                     ; movss         0x4(%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,88,215                           ; addps         %xmm7,%xmm2
+  DB  15,88,223                           ; addps         %xmm7,%xmm3
+  DB  15,40,125,72                        ; movaps        0x48(%ebp),%xmm7
+  DB  141,126,8                           ; lea           0x8(%esi),%edi
+  DB  15,41,124,36,64                     ; movaps        %xmm7,0x40(%esp)
+  DB  15,41,116,36,48                     ; movaps        %xmm6,0x30(%esp)
+  DB  15,41,108,36,32                     ; movaps        %xmm5,0x20(%esp)
+  DB  15,41,100,36,16                     ; movaps        %xmm4,0x10(%esp)
+  DB  137,84,36,12                        ; mov           %edx,0xc(%esp)
+  DB  137,76,36,8                         ; mov           %ecx,0x8(%esp)
+  DB  137,124,36,4                        ; mov           %edi,0x4(%esp)
+  DB  137,4,36                            ; mov           %eax,(%esp)
+  DB  255,86,4                            ; call          *0x4(%esi)
+  DB  131,196,80                          ; add           $0x50,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
+PUBLIC _sk_matrix_scale_translate_sse2_lowp
+_sk_matrix_scale_translate_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,80                          ; sub           $0x50,%esp
+  DB  139,69,8                            ; mov           0x8(%ebp),%eax
+  DB  139,85,16                           ; mov           0x10(%ebp),%edx
+  DB  139,117,20                          ; mov           0x14(%ebp),%esi
+  DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
+  DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
+  DB  15,40,117,56                        ; movaps        0x38(%ebp),%xmm6
+  DB  139,77,12                           ; mov           0xc(%ebp),%ecx
+  DB  139,57                              ; mov           (%ecx),%edi
+  DB  243,15,16,63                        ; movss         (%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,89,207                           ; mulps         %xmm7,%xmm1
+  DB  15,89,199                           ; mulps         %xmm7,%xmm0
+  DB  243,15,16,127,8                     ; movss         0x8(%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,88,199                           ; addps         %xmm7,%xmm0
+  DB  15,88,207                           ; addps         %xmm7,%xmm1
+  DB  243,15,16,127,4                     ; movss         0x4(%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,89,223                           ; mulps         %xmm7,%xmm3
+  DB  15,89,215                           ; mulps         %xmm7,%xmm2
+  DB  243,15,16,127,12                    ; movss         0xc(%edi),%xmm7
+  DB  15,198,255,0                        ; shufps        $0x0,%xmm7,%xmm7
+  DB  15,88,215                           ; addps         %xmm7,%xmm2
+  DB  15,88,223                           ; addps         %xmm7,%xmm3
+  DB  15,40,125,72                        ; movaps        0x48(%ebp),%xmm7
+  DB  141,121,8                           ; lea           0x8(%ecx),%edi
+  DB  15,41,124,36,64                     ; movaps        %xmm7,0x40(%esp)
+  DB  15,41,116,36,48                     ; movaps        %xmm6,0x30(%esp)
+  DB  15,41,108,36,32                     ; movaps        %xmm5,0x20(%esp)
+  DB  15,41,100,36,16                     ; movaps        %xmm4,0x10(%esp)
+  DB  137,116,36,12                       ; mov           %esi,0xc(%esp)
+  DB  137,84,36,8                         ; mov           %edx,0x8(%esp)
+  DB  137,124,36,4                        ; mov           %edi,0x4(%esp)
+  DB  137,4,36                            ; mov           %eax,(%esp)
+  DB  255,81,4                            ; call          *0x4(%ecx)
+  DB  131,196,80                          ; add           $0x50,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_matrix_2x3_sse2_lowp
 _sk_matrix_2x3_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -55264,6 +56751,97 @@
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
 
+PUBLIC _sk_matrix_perspective_sse2_lowp
+_sk_matrix_perspective_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,112                         ; sub           $0x70,%esp
+  DB  15,40,235                           ; movaps        %xmm3,%xmm5
+  DB  15,41,77,232                        ; movaps        %xmm1,-0x18(%ebp)
+  DB  15,40,216                           ; movaps        %xmm0,%xmm3
+  DB  139,69,12                           ; mov           0xc(%ebp),%eax
+  DB  139,8                               ; mov           (%eax),%ecx
+  DB  243,15,16,9                         ; movss         (%ecx),%xmm1
+  DB  243,15,16,113,4                     ; movss         0x4(%ecx),%xmm6
+  DB  15,198,201,0                        ; shufps        $0x0,%xmm1,%xmm1
+  DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
+  DB  243,15,16,97,8                      ; movss         0x8(%ecx),%xmm4
+  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
+  DB  15,40,253                           ; movaps        %xmm5,%xmm7
+  DB  15,89,254                           ; mulps         %xmm6,%xmm7
+  DB  15,89,242                           ; mulps         %xmm2,%xmm6
+  DB  15,41,85,216                        ; movaps        %xmm2,-0x28(%ebp)
+  DB  15,88,244                           ; addps         %xmm4,%xmm6
+  DB  15,88,252                           ; addps         %xmm4,%xmm7
+  DB  15,40,227                           ; movaps        %xmm3,%xmm4
+  DB  15,89,225                           ; mulps         %xmm1,%xmm4
+  DB  15,88,230                           ; addps         %xmm6,%xmm4
+  DB  15,89,77,232                        ; mulps         -0x18(%ebp),%xmm1
+  DB  15,88,207                           ; addps         %xmm7,%xmm1
+  DB  243,15,16,113,16                    ; movss         0x10(%ecx),%xmm6
+  DB  15,198,246,0                        ; shufps        $0x0,%xmm6,%xmm6
+  DB  243,15,16,65,20                     ; movss         0x14(%ecx),%xmm0
+  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
+  DB  15,40,253                           ; movaps        %xmm5,%xmm7
+  DB  15,89,254                           ; mulps         %xmm6,%xmm7
+  DB  15,89,242                           ; mulps         %xmm2,%xmm6
+  DB  15,88,240                           ; addps         %xmm0,%xmm6
+  DB  15,88,248                           ; addps         %xmm0,%xmm7
+  DB  243,15,16,65,12                     ; movss         0xc(%ecx),%xmm0
+  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
+  DB  15,40,211                           ; movaps        %xmm3,%xmm2
+  DB  15,89,208                           ; mulps         %xmm0,%xmm2
+  DB  15,88,242                           ; addps         %xmm2,%xmm6
+  DB  15,89,69,232                        ; mulps         -0x18(%ebp),%xmm0
+  DB  15,88,248                           ; addps         %xmm0,%xmm7
+  DB  243,15,16,65,28                     ; movss         0x1c(%ecx),%xmm0
+  DB  15,198,192,0                        ; shufps        $0x0,%xmm0,%xmm0
+  DB  15,89,232                           ; mulps         %xmm0,%xmm5
+  DB  15,89,69,216                        ; mulps         -0x28(%ebp),%xmm0
+  DB  243,15,16,81,32                     ; movss         0x20(%ecx),%xmm2
+  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
+  DB  15,88,194                           ; addps         %xmm2,%xmm0
+  DB  15,88,234                           ; addps         %xmm2,%xmm5
+  DB  243,15,16,81,24                     ; movss         0x18(%ecx),%xmm2
+  DB  15,198,210,0                        ; shufps        $0x0,%xmm2,%xmm2
+  DB  15,89,218                           ; mulps         %xmm2,%xmm3
+  DB  15,88,216                           ; addps         %xmm0,%xmm3
+  DB  139,77,8                            ; mov           0x8(%ebp),%ecx
+  DB  139,85,16                           ; mov           0x10(%ebp),%edx
+  DB  139,117,20                          ; mov           0x14(%ebp),%esi
+  DB  15,89,85,232                        ; mulps         -0x18(%ebp),%xmm2
+  DB  15,40,69,24                         ; movaps        0x18(%ebp),%xmm0
+  DB  15,88,213                           ; addps         %xmm5,%xmm2
+  DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
+  DB  15,83,219                           ; rcpps         %xmm3,%xmm3
+  DB  15,89,227                           ; mulps         %xmm3,%xmm4
+  DB  15,89,243                           ; mulps         %xmm3,%xmm6
+  DB  15,40,93,56                         ; movaps        0x38(%ebp),%xmm3
+  DB  15,83,210                           ; rcpps         %xmm2,%xmm2
+  DB  15,89,202                           ; mulps         %xmm2,%xmm1
+  DB  15,89,250                           ; mulps         %xmm2,%xmm7
+  DB  15,40,85,72                         ; movaps        0x48(%ebp),%xmm2
+  DB  141,120,8                           ; lea           0x8(%eax),%edi
+  DB  15,41,84,36,64                      ; movaps        %xmm2,0x40(%esp)
+  DB  15,41,92,36,48                      ; movaps        %xmm3,0x30(%esp)
+  DB  15,41,108,36,32                     ; movaps        %xmm5,0x20(%esp)
+  DB  15,41,68,36,16                      ; movaps        %xmm0,0x10(%esp)
+  DB  137,116,36,12                       ; mov           %esi,0xc(%esp)
+  DB  137,84,36,8                         ; mov           %edx,0x8(%esp)
+  DB  137,124,36,4                        ; mov           %edi,0x4(%esp)
+  DB  137,12,36                           ; mov           %ecx,(%esp)
+  DB  15,40,196                           ; movaps        %xmm4,%xmm0
+  DB  15,40,214                           ; movaps        %xmm6,%xmm2
+  DB  15,40,223                           ; movaps        %xmm7,%xmm3
+  DB  255,80,4                            ; call          *0x4(%eax)
+  DB  131,196,112                         ; add           $0x70,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_uniform_color_sse2_lowp
 _sk_uniform_color_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -55318,7 +56896,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,92                          ; sub           $0x5c,%esp
-  DB  232,0,0,0,0                         ; call          96e8 <_sk_black_color_sse2_lowp+0xe>
+  DB  232,0,0,0,0                         ; call          9913 <_sk_black_color_sse2_lowp+0xe>
   DB  88                                  ; pop           %eax
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
@@ -55337,7 +56915,7 @@
   DB  137,84,36,8                         ; mov           %edx,0x8(%esp)
   DB  137,92,36,4                         ; mov           %ebx,0x4(%esp)
   DB  137,12,36                           ; mov           %ecx,(%esp)
-  DB  15,40,152,88,64,0,0                 ; movaps        0x4058(%eax),%xmm3
+  DB  15,40,152,77,71,0,0                 ; movaps        0x474d(%eax),%xmm3
   DB  15,87,192                           ; xorps         %xmm0,%xmm0
   DB  15,87,201                           ; xorps         %xmm1,%xmm1
   DB  15,87,210                           ; xorps         %xmm2,%xmm2
@@ -55357,7 +56935,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,92                          ; sub           $0x5c,%esp
-  DB  232,0,0,0,0                         ; call          9753 <_sk_white_color_sse2_lowp+0xe>
+  DB  232,0,0,0,0                         ; call          997e <_sk_white_color_sse2_lowp+0xe>
   DB  88                                  ; pop           %eax
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
@@ -55376,7 +56954,7 @@
   DB  137,84,36,8                         ; mov           %edx,0x8(%esp)
   DB  137,92,36,4                         ; mov           %ebx,0x4(%esp)
   DB  137,12,36                           ; mov           %ecx,(%esp)
-  DB  15,40,128,237,63,0,0                ; movaps        0x3fed(%eax),%xmm0
+  DB  15,40,128,226,70,0,0                ; movaps        0x46e2(%eax),%xmm0
   DB  15,40,200                           ; movaps        %xmm0,%xmm1
   DB  15,40,208                           ; movaps        %xmm0,%xmm2
   DB  15,40,216                           ; movaps        %xmm0,%xmm3
@@ -55395,16 +56973,16 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          97bd <_sk_set_rgb_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          99e8 <_sk_set_rgb_sse2_lowp+0xd>
   DB  94                                  ; pop           %esi
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,57                              ; mov           (%ecx),%edi
-  DB  243,15,16,150,71,64,0,0             ; movss         0x4047(%esi),%xmm2
+  DB  243,15,16,150,60,71,0,0             ; movss         0x473c(%esi),%xmm2
   DB  243,15,16,7                         ; movss         (%edi),%xmm0
   DB  243,15,89,194                       ; mulss         %xmm2,%xmm0
-  DB  243,15,16,142,67,64,0,0             ; movss         0x4043(%esi),%xmm1
+  DB  243,15,16,142,56,71,0,0             ; movss         0x4738(%esi),%xmm1
   DB  243,15,88,193                       ; addss         %xmm1,%xmm0
   DB  243,15,44,240                       ; cvttss2si     %xmm0,%esi
   DB  102,15,110,198                      ; movd          %esi,%xmm0
@@ -55451,7 +57029,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9881 <_sk_premul_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9aac <_sk_premul_sse2_lowp+0xd>
   DB  90                                  ; pop           %edx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,77,16                           ; mov           0x10(%ebp),%ecx
@@ -55460,7 +57038,7 @@
   DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
   DB  15,40,117,56                        ; movaps        0x38(%ebp),%xmm6
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,15,111,186,191,62,0,0           ; movdqa        0x3ebf(%edx),%xmm7
+  DB  102,15,111,186,180,69,0,0           ; movdqa        0x45b4(%edx),%xmm7
   DB  102,15,253,199                      ; paddw         %xmm7,%xmm0
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
   DB  102,15,253,207                      ; paddw         %xmm7,%xmm1
@@ -55586,12 +57164,12 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          99fb <_sk_invert_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9c26 <_sk_invert_sse2_lowp+0xd>
   DB  88                                  ; pop           %eax
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
   DB  139,117,20                          ; mov           0x14(%ebp),%esi
-  DB  102,15,111,160,69,61,0,0            ; movdqa        0x3d45(%eax),%xmm4
+  DB  102,15,111,160,58,68,0,0            ; movdqa        0x443a(%eax),%xmm4
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,249,232                      ; psubw         %xmm0,%xmm5
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
@@ -55669,11 +57247,11 @@
   DB  102,15,111,234                      ; movdqa        %xmm2,%xmm5
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          9af3 <_sk_srcatop_sse2_lowp+0x19>
+  DB  232,0,0,0,0                         ; call          9d1e <_sk_srcatop_sse2_lowp+0x19>
   DB  88                                  ; pop           %eax
   DB  102,15,111,117,72                   ; movdqa        0x48(%ebp),%xmm6
   DB  102,15,213,206                      ; pmullw        %xmm6,%xmm1
-  DB  102,15,111,128,77,60,0,0            ; movdqa        0x3c4d(%eax),%xmm0
+  DB  102,15,111,128,66,67,0,0            ; movdqa        0x4342(%eax),%xmm0
   DB  102,15,253,200                      ; paddw         %xmm0,%xmm1
   DB  102,15,213,214                      ; pmullw        %xmm6,%xmm2
   DB  102,15,253,208                      ; paddw         %xmm0,%xmm2
@@ -55729,12 +57307,12 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9bc7 <_sk_dstatop_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9df2 <_sk_dstatop_sse2_lowp+0xd>
   DB  88                                  ; pop           %eax
   DB  102,15,111,101,40                   ; movdqa        0x28(%ebp),%xmm4
   DB  102,15,111,117,24                   ; movdqa        0x18(%ebp),%xmm6
   DB  102,15,213,243                      ; pmullw        %xmm3,%xmm6
-  DB  102,15,111,184,121,59,0,0           ; movdqa        0x3b79(%eax),%xmm7
+  DB  102,15,111,184,110,66,0,0           ; movdqa        0x426e(%eax),%xmm7
   DB  102,15,253,247                      ; paddw         %xmm7,%xmm6
   DB  102,15,213,227                      ; pmullw        %xmm3,%xmm4
   DB  102,15,253,231                      ; paddw         %xmm7,%xmm4
@@ -55785,7 +57363,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9c88 <_sk_srcin_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9eb3 <_sk_srcin_sse2_lowp+0xd>
   DB  94                                  ; pop           %esi
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,77,16                           ; mov           0x10(%ebp),%ecx
@@ -55794,7 +57372,7 @@
   DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
   DB  102,15,111,117,72                   ; movdqa        0x48(%ebp),%xmm6
   DB  102,15,213,198                      ; pmullw        %xmm6,%xmm0
-  DB  102,15,111,190,184,58,0,0           ; movdqa        0x3ab8(%esi),%xmm7
+  DB  102,15,111,190,173,65,0,0           ; movdqa        0x41ad(%esi),%xmm7
   DB  102,15,253,199                      ; paddw         %xmm7,%xmm0
   DB  102,15,213,206                      ; pmullw        %xmm6,%xmm1
   DB  102,15,253,207                      ; paddw         %xmm7,%xmm1
@@ -55831,14 +57409,14 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9d1f <_sk_dstin_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9f4a <_sk_dstin_sse2_lowp+0xd>
   DB  88                                  ; pop           %eax
   DB  102,15,111,85,56                    ; movdqa        0x38(%ebp),%xmm2
   DB  102,15,111,109,40                   ; movdqa        0x28(%ebp),%xmm5
   DB  102,15,111,117,24                   ; movdqa        0x18(%ebp),%xmm6
   DB  102,15,111,198                      ; movdqa        %xmm6,%xmm0
   DB  102,15,213,195                      ; pmullw        %xmm3,%xmm0
-  DB  102,15,111,184,33,58,0,0            ; movdqa        0x3a21(%eax),%xmm7
+  DB  102,15,111,184,22,65,0,0            ; movdqa        0x4116(%eax),%xmm7
   DB  102,15,253,199                      ; paddw         %xmm7,%xmm0
   DB  102,15,111,205                      ; movdqa        %xmm5,%xmm1
   DB  102,15,213,203                      ; pmullw        %xmm3,%xmm1
@@ -55880,14 +57458,14 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9dc7 <_sk_srcout_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          9ff2 <_sk_srcout_sse2_lowp+0xd>
   DB  94                                  ; pop           %esi
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,77,16                           ; mov           0x10(%ebp),%ecx
   DB  139,85,20                           ; mov           0x14(%ebp),%edx
   DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
   DB  102,15,111,109,72                   ; movdqa        0x48(%ebp),%xmm5
-  DB  102,15,111,182,121,57,0,0           ; movdqa        0x3979(%esi),%xmm6
+  DB  102,15,111,182,110,64,0,0           ; movdqa        0x406e(%esi),%xmm6
   DB  102,15,111,254                      ; movdqa        %xmm6,%xmm7
   DB  102,15,249,253                      ; psubw         %xmm5,%xmm7
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
@@ -55929,11 +57507,11 @@
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
-  DB  232,0,0,0,0                         ; call          9e6a <_sk_dstout_sse2_lowp+0x11>
+  DB  232,0,0,0,0                         ; call          a095 <_sk_dstout_sse2_lowp+0x11>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,56                   ; movdqa        0x38(%ebp),%xmm5
   DB  102,15,111,117,24                   ; movdqa        0x18(%ebp),%xmm6
-  DB  102,15,111,184,214,56,0,0           ; movdqa        0x38d6(%eax),%xmm7
+  DB  102,15,111,184,203,63,0,0           ; movdqa        0x3fcb(%eax),%xmm7
   DB  102,15,111,223                      ; movdqa        %xmm7,%xmm3
   DB  102,15,249,216                      ; psubw         %xmm0,%xmm3
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
@@ -55983,9 +57561,9 @@
   DB  15,41,85,232                        ; movaps        %xmm2,-0x18(%ebp)
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,111,224                      ; movdqa        %xmm0,%xmm4
-  DB  232,0,0,0,0                         ; call          9f26 <_sk_srcover_sse2_lowp+0x19>
+  DB  232,0,0,0,0                         ; call          a151 <_sk_srcover_sse2_lowp+0x19>
   DB  88                                  ; pop           %eax
-  DB  102,15,111,136,26,56,0,0            ; movdqa        0x381a(%eax),%xmm1
+  DB  102,15,111,136,15,63,0,0            ; movdqa        0x3f0f(%eax),%xmm1
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,249,195                      ; psubw         %xmm3,%xmm0
   DB  102,15,111,248                      ; movdqa        %xmm0,%xmm7
@@ -56042,7 +57620,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          9ff6 <_sk_dstover_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          a221 <_sk_dstover_sse2_lowp+0xd>
   DB  95                                  ; pop           %edi
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
@@ -56050,7 +57628,7 @@
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  102,15,111,101,56                   ; movdqa        0x38(%ebp),%xmm4
   DB  102,15,111,109,72                   ; movdqa        0x48(%ebp),%xmm5
-  DB  102,15,111,183,74,55,0,0            ; movdqa        0x374a(%edi),%xmm6
+  DB  102,15,111,183,63,62,0,0            ; movdqa        0x3e3f(%edi),%xmm6
   DB  102,15,111,254                      ; movdqa        %xmm6,%xmm7
   DB  102,15,249,253                      ; psubw         %xmm5,%xmm7
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
@@ -56094,12 +57672,12 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          a0ab <_sk_modulate_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          a2d6 <_sk_modulate_sse2_lowp+0xd>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,40                   ; movdqa        0x28(%ebp),%xmm5
   DB  102,15,111,117,24                   ; movdqa        0x18(%ebp),%xmm6
   DB  102,15,213,198                      ; pmullw        %xmm6,%xmm0
-  DB  102,15,111,184,149,54,0,0           ; movdqa        0x3695(%eax),%xmm7
+  DB  102,15,111,184,138,61,0,0           ; movdqa        0x3d8a(%eax),%xmm7
   DB  102,15,253,199                      ; paddw         %xmm7,%xmm0
   DB  102,15,213,205                      ; pmullw        %xmm5,%xmm1
   DB  102,15,253,207                      ; paddw         %xmm7,%xmm1
@@ -56140,10 +57718,10 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          a147 <_sk_multiply_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          a372 <_sk_multiply_sse2_lowp+0xd>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
-  DB  102,15,111,160,249,53,0,0           ; movdqa        0x35f9(%eax),%xmm4
+  DB  102,15,111,160,238,60,0,0           ; movdqa        0x3cee(%eax),%xmm4
   DB  102,15,111,245                      ; movdqa        %xmm5,%xmm6
   DB  102,15,253,244                      ; paddw         %xmm4,%xmm6
   DB  102,15,249,117,72                   ; psubw         0x48(%ebp),%xmm6
@@ -56211,17 +57789,17 @@
   DB  131,236,96                          ; sub           $0x60,%esp
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          a248 <_sk_plus__sse2_lowp+0x15>
+  DB  232,0,0,0,0                         ; call          a473 <_sk_plus__sse2_lowp+0x15>
   DB  88                                  ; pop           %eax
   DB  102,15,253,77,24                    ; paddw         0x18(%ebp),%xmm1
-  DB  102,15,111,168,8,53,0,0             ; movdqa        0x3508(%eax),%xmm5
+  DB  102,15,111,168,253,59,0,0           ; movdqa        0x3bfd(%eax),%xmm5
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,217,197                      ; psubusw       %xmm5,%xmm0
   DB  102,15,127,109,232                  ; movdqa        %xmm5,-0x18(%ebp)
   DB  102,15,239,255                      ; pxor          %xmm7,%xmm7
   DB  102,15,117,199                      ; pcmpeqw       %xmm7,%xmm0
   DB  102,15,219,200                      ; pand          %xmm0,%xmm1
-  DB  102,15,111,176,248,52,0,0           ; movdqa        0x34f8(%eax),%xmm6
+  DB  102,15,111,176,237,59,0,0           ; movdqa        0x3bed(%eax),%xmm6
   DB  102,15,223,198                      ; pandn         %xmm6,%xmm0
   DB  102,15,235,193                      ; por           %xmm1,%xmm0
   DB  102,15,253,101,40                   ; paddw         0x28(%ebp),%xmm4
@@ -56280,14 +57858,14 @@
   DB  131,236,80                          ; sub           $0x50,%esp
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          a340 <_sk_screen_sse2_lowp+0x15>
+  DB  232,0,0,0,0                         ; call          a56b <_sk_screen_sse2_lowp+0x15>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,40                   ; movdqa        0x28(%ebp),%xmm5
   DB  102,15,111,117,24                   ; movdqa        0x18(%ebp),%xmm6
   DB  102,15,111,198                      ; movdqa        %xmm6,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,206                      ; pmullw        %xmm6,%xmm1
-  DB  102,15,111,176,0,52,0,0             ; movdqa        0x3400(%eax),%xmm6
+  DB  102,15,111,176,245,58,0,0           ; movdqa        0x3af5(%eax),%xmm6
   DB  102,15,253,206                      ; paddw         %xmm6,%xmm1
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
@@ -56343,9 +57921,9 @@
   DB  86                                  ; push          %esi
   DB  131,236,96                          ; sub           $0x60,%esp
   DB  102,15,127,93,232                   ; movdqa        %xmm3,-0x18(%ebp)
-  DB  232,0,0,0,0                         ; call          a420 <_sk_xor__sse2_lowp+0x12>
+  DB  232,0,0,0,0                         ; call          a64b <_sk_xor__sse2_lowp+0x12>
   DB  88                                  ; pop           %eax
-  DB  102,15,111,176,32,51,0,0            ; movdqa        0x3320(%eax),%xmm6
+  DB  102,15,111,176,21,58,0,0            ; movdqa        0x3a15(%eax),%xmm6
   DB  102,15,111,254                      ; movdqa        %xmm6,%xmm7
   DB  102,15,249,125,72                   ; psubw         0x48(%ebp),%xmm7
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
@@ -56410,7 +57988,7 @@
   DB  15,41,85,200                        ; movaps        %xmm2,-0x38(%ebp)
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          a515 <_sk_darken_sse2_lowp+0x25>
+  DB  232,0,0,0,0                         ; call          a740 <_sk_darken_sse2_lowp+0x25>
   DB  88                                  ; pop           %eax
   DB  102,15,111,117,72                   ; movdqa        0x48(%ebp),%xmm6
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
@@ -56421,7 +57999,7 @@
   DB  102,15,111,222                      ; movdqa        %xmm6,%xmm3
   DB  102,15,111,197                      ; movdqa        %xmm5,%xmm0
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,15,111,184,75,50,0,0            ; movdqa        0x324b(%eax),%xmm7
+  DB  102,15,111,184,64,57,0,0            ; movdqa        0x3940(%eax),%xmm7
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
   DB  102,15,239,247                      ; pxor          %xmm7,%xmm6
   DB  102,15,111,232                      ; movdqa        %xmm0,%xmm5
@@ -56430,7 +58008,7 @@
   DB  102,15,219,197                      ; pand          %xmm5,%xmm0
   DB  102,15,223,233                      ; pandn         %xmm1,%xmm5
   DB  102,15,235,232                      ; por           %xmm0,%xmm5
-  DB  102,15,111,128,43,50,0,0            ; movdqa        0x322b(%eax),%xmm0
+  DB  102,15,111,128,32,57,0,0            ; movdqa        0x3920(%eax),%xmm0
   DB  102,15,127,69,232                   ; movdqa        %xmm0,-0x18(%ebp)
   DB  102,15,253,232                      ; paddw         %xmm0,%xmm5
   DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
@@ -56519,7 +58097,7 @@
   DB  15,41,85,200                        ; movaps        %xmm2,-0x38(%ebp)
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          a6be <_sk_lighten_sse2_lowp+0x25>
+  DB  232,0,0,0,0                         ; call          a8e9 <_sk_lighten_sse2_lowp+0x25>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
   DB  102,15,111,197                      ; movdqa        %xmm5,%xmm0
@@ -56528,7 +58106,7 @@
   DB  102,15,213,77,72                    ; pmullw        0x48(%ebp),%xmm1
   DB  102,15,111,197                      ; movdqa        %xmm5,%xmm0
   DB  102,15,213,198                      ; pmullw        %xmm6,%xmm0
-  DB  102,15,111,184,162,48,0,0           ; movdqa        0x30a2(%eax),%xmm7
+  DB  102,15,111,184,151,55,0,0           ; movdqa        0x3797(%eax),%xmm7
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
   DB  102,15,239,247                      ; pxor          %xmm7,%xmm6
   DB  102,15,111,232                      ; movdqa        %xmm0,%xmm5
@@ -56537,7 +58115,7 @@
   DB  102,15,219,205                      ; pand          %xmm5,%xmm1
   DB  102,15,223,232                      ; pandn         %xmm0,%xmm5
   DB  102,15,235,233                      ; por           %xmm1,%xmm5
-  DB  102,15,111,128,130,48,0,0           ; movdqa        0x3082(%eax),%xmm0
+  DB  102,15,111,128,119,55,0,0           ; movdqa        0x3777(%eax),%xmm0
   DB  102,15,253,232                      ; paddw         %xmm0,%xmm5
   DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
   DB  102,15,127,85,232                   ; movdqa        %xmm2,-0x18(%ebp)
@@ -56625,7 +58203,7 @@
   DB  15,41,85,184                        ; movaps        %xmm2,-0x48(%ebp)
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          a85c <_sk_difference_sse2_lowp+0x25>
+  DB  232,0,0,0,0                         ; call          aa87 <_sk_difference_sse2_lowp+0x25>
   DB  88                                  ; pop           %eax
   DB  102,15,111,117,72                   ; movdqa        0x48(%ebp),%xmm6
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
@@ -56636,7 +58214,7 @@
   DB  102,15,111,222                      ; movdqa        %xmm6,%xmm3
   DB  102,15,111,197                      ; movdqa        %xmm5,%xmm0
   DB  102,15,213,199                      ; pmullw        %xmm7,%xmm0
-  DB  102,15,111,184,4,47,0,0             ; movdqa        0x2f04(%eax),%xmm7
+  DB  102,15,111,184,249,53,0,0           ; movdqa        0x35f9(%eax),%xmm7
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
   DB  102,15,239,247                      ; pxor          %xmm7,%xmm6
   DB  102,15,111,232                      ; movdqa        %xmm0,%xmm5
@@ -56645,11 +58223,11 @@
   DB  102,15,219,205                      ; pand          %xmm5,%xmm1
   DB  102,15,223,232                      ; pandn         %xmm0,%xmm5
   DB  102,15,235,233                      ; por           %xmm1,%xmm5
-  DB  102,15,111,128,228,46,0,0           ; movdqa        0x2ee4(%eax),%xmm0
+  DB  102,15,111,128,217,53,0,0           ; movdqa        0x35d9(%eax),%xmm0
   DB  102,15,127,69,216                   ; movdqa        %xmm0,-0x28(%ebp)
   DB  102,15,253,232                      ; paddw         %xmm0,%xmm5
   DB  102,15,113,213,7                    ; psrlw         $0x7,%xmm5
-  DB  102,15,111,128,20,47,0,0            ; movdqa        0x2f14(%eax),%xmm0
+  DB  102,15,111,128,9,54,0,0             ; movdqa        0x3609(%eax),%xmm0
   DB  102,15,127,69,232                   ; movdqa        %xmm0,-0x18(%ebp)
   DB  102,15,219,232                      ; pand          %xmm0,%xmm5
   DB  102,15,249,213                      ; psubw         %xmm5,%xmm2
@@ -56736,16 +58314,16 @@
   DB  131,236,80                          ; sub           $0x50,%esp
   DB  102,15,111,225                      ; movdqa        %xmm1,%xmm4
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          aa10 <_sk_exclusion_sse2_lowp+0x15>
+  DB  232,0,0,0,0                         ; call          ac3b <_sk_exclusion_sse2_lowp+0x15>
   DB  88                                  ; pop           %eax
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
   DB  102,15,111,197                      ; movdqa        %xmm5,%xmm0
   DB  102,15,253,193                      ; paddw         %xmm1,%xmm0
   DB  102,15,213,205                      ; pmullw        %xmm5,%xmm1
-  DB  102,15,111,184,48,45,0,0            ; movdqa        0x2d30(%eax),%xmm7
+  DB  102,15,111,184,37,52,0,0            ; movdqa        0x3425(%eax),%xmm7
   DB  102,15,253,207                      ; paddw         %xmm7,%xmm1
   DB  102,15,113,209,7                    ; psrlw         $0x7,%xmm1
-  DB  102,15,111,176,96,45,0,0            ; movdqa        0x2d60(%eax),%xmm6
+  DB  102,15,111,176,85,52,0,0            ; movdqa        0x3455(%eax),%xmm6
   DB  102,15,219,206                      ; pand          %xmm6,%xmm1
   DB  102,15,249,193                      ; psubw         %xmm1,%xmm0
   DB  102,15,111,109,40                   ; movdqa        0x28(%ebp),%xmm5
@@ -56804,9 +58382,9 @@
   DB  15,41,85,152                        ; movaps        %xmm2,-0x68(%ebp)
   DB  102,15,111,233                      ; movdqa        %xmm1,%xmm5
   DB  102,15,111,224                      ; movdqa        %xmm0,%xmm4
-  DB  232,0,0,0,0                         ; call          ab0a <_sk_hardlight_sse2_lowp+0x1c>
+  DB  232,0,0,0,0                         ; call          ad35 <_sk_hardlight_sse2_lowp+0x1c>
   DB  88                                  ; pop           %eax
-  DB  102,15,111,184,54,44,0,0            ; movdqa        0x2c36(%eax),%xmm7
+  DB  102,15,111,184,43,51,0,0            ; movdqa        0x332b(%eax),%xmm7
   DB  102,15,127,125,200                  ; movdqa        %xmm7,-0x38(%ebp)
   DB  102,15,111,215                      ; movdqa        %xmm7,%xmm2
   DB  102,15,249,85,72                    ; psubw         0x48(%ebp),%xmm2
@@ -56833,7 +58411,7 @@
   DB  102,15,111,214                      ; movdqa        %xmm6,%xmm2
   DB  102,15,249,209                      ; psubw         %xmm1,%xmm2
   DB  102,15,253,228                      ; paddw         %xmm4,%xmm4
-  DB  102,15,111,128,86,44,0,0            ; movdqa        0x2c56(%eax),%xmm0
+  DB  102,15,111,128,75,51,0,0            ; movdqa        0x334b(%eax),%xmm0
   DB  102,15,239,216                      ; pxor          %xmm0,%xmm3
   DB  102,15,127,93,184                   ; movdqa        %xmm3,-0x48(%ebp)
   DB  102,15,111,204                      ; movdqa        %xmm4,%xmm1
@@ -56938,10 +58516,10 @@
   DB  15,41,85,152                        ; movaps        %xmm2,-0x68(%ebp)
   DB  15,41,141,120,255,255,255           ; movaps        %xmm1,-0x88(%ebp)
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          ad2c <_sk_overlay_sse2_lowp+0x1f>
+  DB  232,0,0,0,0                         ; call          af57 <_sk_overlay_sse2_lowp+0x1f>
   DB  88                                  ; pop           %eax
   DB  102,15,111,69,72                    ; movdqa        0x48(%ebp),%xmm0
-  DB  102,15,111,184,20,42,0,0            ; movdqa        0x2a14(%eax),%xmm7
+  DB  102,15,111,184,9,49,0,0             ; movdqa        0x3109(%eax),%xmm7
   DB  102,15,127,125,232                  ; movdqa        %xmm7,-0x18(%ebp)
   DB  102,15,111,215                      ; movdqa        %xmm7,%xmm2
   DB  102,15,249,208                      ; psubw         %xmm0,%xmm2
@@ -56968,7 +58546,7 @@
   DB  102,15,249,234                      ; psubw         %xmm2,%xmm5
   DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
   DB  102,15,253,192                      ; paddw         %xmm0,%xmm0
-  DB  102,15,111,144,52,42,0,0            ; movdqa        0x2a34(%eax),%xmm2
+  DB  102,15,111,144,41,49,0,0            ; movdqa        0x3129(%eax),%xmm2
   DB  102,15,127,85,184                   ; movdqa        %xmm2,-0x48(%ebp)
   DB  102,15,239,226                      ; pxor          %xmm2,%xmm4
   DB  102,15,239,194                      ; pxor          %xmm2,%xmm0
@@ -57096,26 +58674,26 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          af9b <_sk_load_8888_sse2_lowp+0x30>
+  DB  232,0,0,0,0                         ; call          b1c6 <_sk_load_8888_sse2_lowp+0x30>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,32                              ; ja            afc1 <_sk_load_8888_sse2_lowp+0x56>
+  DB  119,32                              ; ja            b1ec <_sk_load_8888_sse2_lowp+0x56>
   DB  102,15,239,246                      ; pxor          %xmm6,%xmm6
   DB  15,182,201                          ; movzbl        %cl,%ecx
-  DB  139,140,143,53,1,0,0                ; mov           0x135(%edi,%ecx,4),%ecx
+  DB  139,140,143,50,1,0,0                ; mov           0x132(%edi,%ecx,4),%ecx
   DB  1,249                               ; add           %edi,%ecx
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  255,225                             ; jmp           *%ecx
   DB  243,15,16,28,179                    ; movss         (%ebx,%esi,4),%xmm3
   DB  15,87,246                           ; xorps         %xmm6,%xmm6
-  DB  235,73                              ; jmp           b00a <_sk_load_8888_sse2_lowp+0x9f>
+  DB  235,73                              ; jmp           b235 <_sk_load_8888_sse2_lowp+0x9f>
   DB  102,15,16,28,179                    ; movupd        (%ebx,%esi,4),%xmm3
   DB  15,16,116,179,16                    ; movups        0x10(%ebx,%esi,4),%xmm6
-  DB  235,61                              ; jmp           b00a <_sk_load_8888_sse2_lowp+0x9f>
+  DB  235,61                              ; jmp           b235 <_sk_load_8888_sse2_lowp+0x9f>
   DB  102,15,110,68,179,8                 ; movd          0x8(%ebx,%esi,4),%xmm0
   DB  102,15,112,216,69                   ; pshufd        $0x45,%xmm0,%xmm3
   DB  102,15,18,28,179                    ; movlpd        (%ebx,%esi,4),%xmm3
-  DB  235,43                              ; jmp           b00a <_sk_load_8888_sse2_lowp+0x9f>
+  DB  235,43                              ; jmp           b235 <_sk_load_8888_sse2_lowp+0x9f>
   DB  102,15,110,68,179,24                ; movd          0x18(%ebx,%esi,4),%xmm0
   DB  102,15,112,240,69                   ; pshufd        $0x45,%xmm0,%xmm6
   DB  243,15,16,68,179,20                 ; movss         0x14(%ebx,%esi,4),%xmm0
@@ -57132,7 +58710,7 @@
   DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
   DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
   DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
-  DB  102,15,111,175,165,39,0,0           ; movdqa        0x27a5(%edi),%xmm5
+  DB  102,15,111,175,154,46,0,0           ; movdqa        0x2e9a(%edi),%xmm5
   DB  102,15,111,193                      ; movdqa        %xmm1,%xmm0
   DB  102,15,219,197                      ; pand          %xmm5,%xmm0
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
@@ -57173,7 +58751,6 @@
   DB  91                                  ; pop           %ebx
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
-  DB  15,31,0                             ; nopl          (%eax)
   DB  28,0                                ; sbb           $0x0,%al
   DB  0,0                                 ; add           %al,(%eax)
   DB  61,0,0,0,50                         ; cmp           $0x32000000,%eax
@@ -57210,10 +58787,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          b129 <_sk_load_8888_dst_sse2_lowp+0x3d>
+  DB  232,0,0,0,0                         ; call          b351 <_sk_load_8888_dst_sse2_lowp+0x3d>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,32                              ; ja            b14f <_sk_load_8888_dst_sse2_lowp+0x63>
+  DB  119,32                              ; ja            b377 <_sk_load_8888_dst_sse2_lowp+0x63>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,59,1,0,0                ; mov           0x13b(%edi,%ecx,4),%ecx
@@ -57222,14 +58799,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  243,15,16,60,179                    ; movss         (%ebx,%esi,4),%xmm7
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
-  DB  235,73                              ; jmp           b198 <_sk_load_8888_dst_sse2_lowp+0xac>
+  DB  235,73                              ; jmp           b3c0 <_sk_load_8888_dst_sse2_lowp+0xac>
   DB  102,15,16,60,179                    ; movupd        (%ebx,%esi,4),%xmm7
   DB  15,16,92,179,16                     ; movups        0x10(%ebx,%esi,4),%xmm3
-  DB  235,61                              ; jmp           b198 <_sk_load_8888_dst_sse2_lowp+0xac>
+  DB  235,61                              ; jmp           b3c0 <_sk_load_8888_dst_sse2_lowp+0xac>
   DB  102,15,110,84,179,8                 ; movd          0x8(%ebx,%esi,4),%xmm2
   DB  102,15,112,250,69                   ; pshufd        $0x45,%xmm2,%xmm7
   DB  102,15,18,60,179                    ; movlpd        (%ebx,%esi,4),%xmm7
-  DB  235,43                              ; jmp           b198 <_sk_load_8888_dst_sse2_lowp+0xac>
+  DB  235,43                              ; jmp           b3c0 <_sk_load_8888_dst_sse2_lowp+0xac>
   DB  102,15,110,84,179,24                ; movd          0x18(%ebx,%esi,4),%xmm2
   DB  102,15,112,218,69                   ; pshufd        $0x45,%xmm2,%xmm3
   DB  243,15,16,84,179,20                 ; movss         0x14(%ebx,%esi,4),%xmm2
@@ -57246,7 +58823,7 @@
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,114,226,16                   ; psrad         $0x10,%xmm2
   DB  102,15,107,242                      ; packssdw      %xmm2,%xmm6
-  DB  102,15,111,151,23,38,0,0            ; movdqa        0x2617(%edi),%xmm2
+  DB  102,15,111,151,15,45,0,0            ; movdqa        0x2d0f(%edi),%xmm2
   DB  102,15,111,206                      ; movdqa        %xmm6,%xmm1
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  102,15,113,214,8                    ; psrlw         $0x8,%xmm6
@@ -57338,19 +58915,19 @@
   DB  128,227,7                           ; and           $0x7,%bl
   DB  254,203                             ; dec           %bl
   DB  128,251,6                           ; cmp           $0x6,%bl
-  DB  232,0,0,0,0                         ; call          b2f4 <_sk_store_8888_sse2_lowp+0x74>
+  DB  232,0,0,0,0                         ; call          b51c <_sk_store_8888_sse2_lowp+0x74>
   DB  89                                  ; pop           %ecx
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
   DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
-  DB  119,19                              ; ja            b311 <_sk_store_8888_sse2_lowp+0x91>
+  DB  119,19                              ; ja            b539 <_sk_store_8888_sse2_lowp+0x91>
   DB  15,182,219                          ; movzbl        %bl,%ebx
   DB  3,140,153,184,0,0,0                 ; add           0xb8(%ecx,%ebx,4),%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  102,15,126,60,183                   ; movd          %xmm7,(%edi,%esi,4)
-  DB  235,93                              ; jmp           b36e <_sk_store_8888_sse2_lowp+0xee>
+  DB  235,93                              ; jmp           b596 <_sk_store_8888_sse2_lowp+0xee>
   DB  243,15,127,60,183                   ; movdqu        %xmm7,(%edi,%esi,4)
   DB  243,15,127,116,183,16               ; movdqu        %xmm6,0x10(%edi,%esi,4)
-  DB  235,80                              ; jmp           b36e <_sk_store_8888_sse2_lowp+0xee>
+  DB  235,80                              ; jmp           b596 <_sk_store_8888_sse2_lowp+0xee>
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  102,15,112,223,78                   ; pshufd        $0x4e,%xmm7,%xmm3
@@ -57358,7 +58935,7 @@
   DB  15,40,220                           ; movaps        %xmm4,%xmm3
   DB  102,15,111,230                      ; movdqa        %xmm6,%xmm4
   DB  102,15,214,60,183                   ; movq          %xmm7,(%edi,%esi,4)
-  DB  235,48                              ; jmp           b36e <_sk_store_8888_sse2_lowp+0xee>
+  DB  235,48                              ; jmp           b596 <_sk_store_8888_sse2_lowp+0xee>
   DB  102,15,112,222,78                   ; pshufd        $0x4e,%xmm6,%xmm3
   DB  102,15,126,92,183,24                ; movd          %xmm3,0x18(%edi,%esi,4)
   DB  15,40,93,72                         ; movaps        0x48(%ebp),%xmm3
@@ -57395,7 +58972,7 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  42,0                                ; sub           (%eax),%al
   DB  0,0                                 ; add           %al,(%eax)
-  DB  117,0                               ; jne           b3ba <_sk_store_8888_sse2_lowp+0x13a>
+  DB  117,0                               ; jne           b5e2 <_sk_store_8888_sse2_lowp+0x13a>
   DB  0,0                                 ; add           %al,(%eax)
   DB  111                                 ; outsl         %ds:(%esi),(%dx)
   DB  0,0                                 ; add           %al,(%eax)
@@ -57425,10 +59002,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          b3f8 <_sk_load_bgra_sse2_lowp+0x30>
+  DB  232,0,0,0,0                         ; call          b620 <_sk_load_bgra_sse2_lowp+0x30>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,32                              ; ja            b41e <_sk_load_bgra_sse2_lowp+0x56>
+  DB  119,32                              ; ja            b646 <_sk_load_bgra_sse2_lowp+0x56>
   DB  102,15,239,246                      ; pxor          %xmm6,%xmm6
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,52,1,0,0                ; mov           0x134(%edi,%ecx,4),%ecx
@@ -57437,14 +59014,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  243,15,16,28,179                    ; movss         (%ebx,%esi,4),%xmm3
   DB  15,87,246                           ; xorps         %xmm6,%xmm6
-  DB  235,73                              ; jmp           b467 <_sk_load_bgra_sse2_lowp+0x9f>
+  DB  235,73                              ; jmp           b68f <_sk_load_bgra_sse2_lowp+0x9f>
   DB  102,15,16,28,179                    ; movupd        (%ebx,%esi,4),%xmm3
   DB  15,16,116,179,16                    ; movups        0x10(%ebx,%esi,4),%xmm6
-  DB  235,61                              ; jmp           b467 <_sk_load_bgra_sse2_lowp+0x9f>
+  DB  235,61                              ; jmp           b68f <_sk_load_bgra_sse2_lowp+0x9f>
   DB  102,15,110,68,179,8                 ; movd          0x8(%ebx,%esi,4),%xmm0
   DB  102,15,112,216,69                   ; pshufd        $0x45,%xmm0,%xmm3
   DB  102,15,18,28,179                    ; movlpd        (%ebx,%esi,4),%xmm3
-  DB  235,43                              ; jmp           b467 <_sk_load_bgra_sse2_lowp+0x9f>
+  DB  235,43                              ; jmp           b68f <_sk_load_bgra_sse2_lowp+0x9f>
   DB  102,15,110,68,179,24                ; movd          0x18(%ebx,%esi,4),%xmm0
   DB  102,15,112,240,69                   ; pshufd        $0x45,%xmm0,%xmm6
   DB  243,15,16,68,179,20                 ; movss         0x14(%ebx,%esi,4),%xmm0
@@ -57461,7 +59038,7 @@
   DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
   DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
   DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
-  DB  102,15,111,175,72,35,0,0            ; movdqa        0x2348(%edi),%xmm5
+  DB  102,15,111,175,64,42,0,0            ; movdqa        0x2a40(%edi),%xmm5
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,219,213                      ; pand          %xmm5,%xmm2
   DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
@@ -57539,10 +59116,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          b585 <_sk_load_bgra_dst_sse2_lowp+0x3d>
+  DB  232,0,0,0,0                         ; call          b7ad <_sk_load_bgra_dst_sse2_lowp+0x3d>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,32                              ; ja            b5ab <_sk_load_bgra_dst_sse2_lowp+0x63>
+  DB  119,32                              ; ja            b7d3 <_sk_load_bgra_dst_sse2_lowp+0x63>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,59,1,0,0                ; mov           0x13b(%edi,%ecx,4),%ecx
@@ -57551,14 +59128,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  243,15,16,60,179                    ; movss         (%ebx,%esi,4),%xmm7
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
-  DB  235,73                              ; jmp           b5f4 <_sk_load_bgra_dst_sse2_lowp+0xac>
+  DB  235,73                              ; jmp           b81c <_sk_load_bgra_dst_sse2_lowp+0xac>
   DB  102,15,16,60,179                    ; movupd        (%ebx,%esi,4),%xmm7
   DB  15,16,92,179,16                     ; movups        0x10(%ebx,%esi,4),%xmm3
-  DB  235,61                              ; jmp           b5f4 <_sk_load_bgra_dst_sse2_lowp+0xac>
+  DB  235,61                              ; jmp           b81c <_sk_load_bgra_dst_sse2_lowp+0xac>
   DB  102,15,110,84,179,8                 ; movd          0x8(%ebx,%esi,4),%xmm2
   DB  102,15,112,250,69                   ; pshufd        $0x45,%xmm2,%xmm7
   DB  102,15,18,60,179                    ; movlpd        (%ebx,%esi,4),%xmm7
-  DB  235,43                              ; jmp           b5f4 <_sk_load_bgra_dst_sse2_lowp+0xac>
+  DB  235,43                              ; jmp           b81c <_sk_load_bgra_dst_sse2_lowp+0xac>
   DB  102,15,110,84,179,24                ; movd          0x18(%ebx,%esi,4),%xmm2
   DB  102,15,112,218,69                   ; pshufd        $0x45,%xmm2,%xmm3
   DB  243,15,16,84,179,20                 ; movss         0x14(%ebx,%esi,4),%xmm2
@@ -57575,7 +59152,7 @@
   DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
   DB  102,15,114,226,16                   ; psrad         $0x10,%xmm2
   DB  102,15,107,242                      ; packssdw      %xmm2,%xmm6
-  DB  102,15,111,151,187,33,0,0           ; movdqa        0x21bb(%edi),%xmm2
+  DB  102,15,111,151,179,40,0,0           ; movdqa        0x28b3(%edi),%xmm2
   DB  102,15,111,206                      ; movdqa        %xmm6,%xmm1
   DB  102,15,219,202                      ; pand          %xmm2,%xmm1
   DB  102,15,113,214,8                    ; psrlw         $0x8,%xmm6
@@ -57667,19 +59244,19 @@
   DB  128,227,7                           ; and           $0x7,%bl
   DB  254,203                             ; dec           %bl
   DB  128,251,6                           ; cmp           $0x6,%bl
-  DB  232,0,0,0,0                         ; call          b750 <_sk_store_bgra_sse2_lowp+0x74>
+  DB  232,0,0,0,0                         ; call          b978 <_sk_store_bgra_sse2_lowp+0x74>
   DB  89                                  ; pop           %ecx
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
   DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
-  DB  119,19                              ; ja            b76d <_sk_store_bgra_sse2_lowp+0x91>
+  DB  119,19                              ; ja            b995 <_sk_store_bgra_sse2_lowp+0x91>
   DB  15,182,219                          ; movzbl        %bl,%ebx
   DB  3,140,153,184,0,0,0                 ; add           0xb8(%ecx,%ebx,4),%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  102,15,126,60,183                   ; movd          %xmm7,(%edi,%esi,4)
-  DB  235,93                              ; jmp           b7ca <_sk_store_bgra_sse2_lowp+0xee>
+  DB  235,93                              ; jmp           b9f2 <_sk_store_bgra_sse2_lowp+0xee>
   DB  243,15,127,60,183                   ; movdqu        %xmm7,(%edi,%esi,4)
   DB  243,15,127,116,183,16               ; movdqu        %xmm6,0x10(%edi,%esi,4)
-  DB  235,80                              ; jmp           b7ca <_sk_store_bgra_sse2_lowp+0xee>
+  DB  235,80                              ; jmp           b9f2 <_sk_store_bgra_sse2_lowp+0xee>
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  15,40,227                           ; movaps        %xmm3,%xmm4
   DB  102,15,112,223,78                   ; pshufd        $0x4e,%xmm7,%xmm3
@@ -57687,7 +59264,7 @@
   DB  15,40,220                           ; movaps        %xmm4,%xmm3
   DB  102,15,111,230                      ; movdqa        %xmm6,%xmm4
   DB  102,15,214,60,183                   ; movq          %xmm7,(%edi,%esi,4)
-  DB  235,48                              ; jmp           b7ca <_sk_store_bgra_sse2_lowp+0xee>
+  DB  235,48                              ; jmp           b9f2 <_sk_store_bgra_sse2_lowp+0xee>
   DB  102,15,112,222,78                   ; pshufd        $0x4e,%xmm6,%xmm3
   DB  102,15,126,92,183,24                ; movd          %xmm3,0x18(%edi,%esi,4)
   DB  15,40,93,72                         ; movaps        0x48(%ebp),%xmm3
@@ -57724,7 +59301,7 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  42,0                                ; sub           (%eax),%al
   DB  0,0                                 ; add           %al,(%eax)
-  DB  117,0                               ; jne           b816 <_sk_store_bgra_sse2_lowp+0x13a>
+  DB  117,0                               ; jne           ba3e <_sk_store_bgra_sse2_lowp+0x13a>
   DB  0,0                                 ; add           %al,(%eax)
   DB  111                                 ; outsl         %ds:(%esi),(%dx)
   DB  0,0                                 ; add           %al,(%eax)
@@ -57741,7 +59318,7 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,96                          ; sub           $0x60,%esp
-  DB  232,0,0,0,0                         ; call          b831 <_sk_gather_8888_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          ba59 <_sk_gather_8888_sse2_lowp+0xd>
   DB  90                                  ; pop           %edx
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
@@ -57833,7 +59410,7 @@
   DB  102,15,114,243,16                   ; pslld         $0x10,%xmm3
   DB  102,15,114,227,16                   ; psrad         $0x10,%xmm3
   DB  102,15,107,216                      ; packssdw      %xmm0,%xmm3
-  DB  102,15,111,186,15,31,0,0            ; movdqa        0x1f0f(%edx),%xmm7
+  DB  102,15,111,186,7,38,0,0             ; movdqa        0x2607(%edx),%xmm7
   DB  102,15,114,210,16                   ; psrld         $0x10,%xmm2
   DB  102,15,114,212,16                   ; psrld         $0x10,%xmm4
   DB  102,15,114,244,16                   ; pslld         $0x10,%xmm4
@@ -57863,6 +59440,134 @@
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
 
+PUBLIC _sk_gather_bgra_sse2_lowp
+_sk_gather_bgra_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,80                          ; sub           $0x50,%esp
+  DB  232,0,0,0,0                         ; call          bc54 <_sk_gather_bgra_sse2_lowp+0xd>
+  DB  90                                  ; pop           %edx
+  DB  139,69,8                            ; mov           0x8(%ebp),%eax
+  DB  139,77,12                           ; mov           0xc(%ebp),%ecx
+  DB  139,57                              ; mov           (%ecx),%edi
+  DB  243,15,16,103,8                     ; movss         0x8(%edi),%xmm4
+  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
+  DB  102,15,118,237                      ; pcmpeqd       %xmm5,%xmm5
+  DB  102,15,254,229                      ; paddd         %xmm5,%xmm4
+  DB  15,87,246                           ; xorps         %xmm6,%xmm6
+  DB  15,95,198                           ; maxps         %xmm6,%xmm0
+  DB  15,95,206                           ; maxps         %xmm6,%xmm1
+  DB  15,93,204                           ; minps         %xmm4,%xmm1
+  DB  15,93,196                           ; minps         %xmm4,%xmm0
+  DB  243,15,16,103,12                    ; movss         0xc(%edi),%xmm4
+  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
+  DB  102,15,254,229                      ; paddd         %xmm5,%xmm4
+  DB  15,95,222                           ; maxps         %xmm6,%xmm3
+  DB  15,95,214                           ; maxps         %xmm6,%xmm2
+  DB  15,93,212                           ; minps         %xmm4,%xmm2
+  DB  15,93,220                           ; minps         %xmm4,%xmm3
+  DB  139,55                              ; mov           (%edi),%esi
+  DB  243,15,91,219                       ; cvttps2dq     %xmm3,%xmm3
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,103,4                    ; movd          0x4(%edi),%xmm4
+  DB  102,15,112,228,0                    ; pshufd        $0x0,%xmm4,%xmm4
+  DB  102,15,112,234,245                  ; pshufd        $0xf5,%xmm2,%xmm5
+  DB  102,15,244,236                      ; pmuludq       %xmm4,%xmm5
+  DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
+  DB  102,15,112,251,245                  ; pshufd        $0xf5,%xmm3,%xmm7
+  DB  102,15,244,252                      ; pmuludq       %xmm4,%xmm7
+  DB  102,15,244,226                      ; pmuludq       %xmm2,%xmm4
+  DB  102,15,112,212,232                  ; pshufd        $0xe8,%xmm4,%xmm2
+  DB  102,15,112,229,232                  ; pshufd        $0xe8,%xmm5,%xmm4
+  DB  102,15,98,212                       ; punpckldq     %xmm4,%xmm2
+  DB  102,15,244,243                      ; pmuludq       %xmm3,%xmm6
+  DB  102,15,112,222,232                  ; pshufd        $0xe8,%xmm6,%xmm3
+  DB  102,15,112,231,232                  ; pshufd        $0xe8,%xmm7,%xmm4
+  DB  102,15,98,220                       ; punpckldq     %xmm4,%xmm3
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
+  DB  102,15,112,208,78                   ; pshufd        $0x4e,%xmm0,%xmm2
+  DB  102,15,112,216,231                  ; pshufd        $0xe7,%xmm0,%xmm3
+  DB  102,15,126,223                      ; movd          %xmm3,%edi
+  DB  102,15,110,36,190                   ; movd          (%esi,%edi,4),%xmm4
+  DB  102,15,126,215                      ; movd          %xmm2,%edi
+  DB  102,15,110,52,190                   ; movd          (%esi,%edi,4),%xmm6
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  102,15,112,192,229                  ; pshufd        $0xe5,%xmm0,%xmm0
+  DB  102,15,110,28,190                   ; movd          (%esi,%edi,4),%xmm3
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  102,15,112,193,78                   ; pshufd        $0x4e,%xmm1,%xmm0
+  DB  102,15,112,209,231                  ; pshufd        $0xe7,%xmm1,%xmm2
+  DB  102,15,110,44,190                   ; movd          (%esi,%edi,4),%xmm5
+  DB  102,15,126,215                      ; movd          %xmm2,%edi
+  DB  102,15,110,60,190                   ; movd          (%esi,%edi,4),%xmm7
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  102,15,110,4,190                    ; movd          (%esi,%edi,4),%xmm0
+  DB  102,15,126,207                      ; movd          %xmm1,%edi
+  DB  102,15,112,201,229                  ; pshufd        $0xe5,%xmm1,%xmm1
+  DB  102,15,110,20,190                   ; movd          (%esi,%edi,4),%xmm2
+  DB  102,15,126,207                      ; movd          %xmm1,%edi
+  DB  102,15,110,12,190                   ; movd          (%esi,%edi,4),%xmm1
+  DB  139,117,16                          ; mov           0x10(%ebp),%esi
+  DB  139,125,20                          ; mov           0x14(%ebp),%edi
+  DB  102,15,98,244                       ; punpckldq     %xmm4,%xmm6
+  DB  102,15,98,221                       ; punpckldq     %xmm5,%xmm3
+  DB  15,40,109,40                        ; movaps        0x28(%ebp),%xmm5
+  DB  102,15,108,222                      ; punpcklqdq    %xmm6,%xmm3
+  DB  15,40,117,56                        ; movaps        0x38(%ebp),%xmm6
+  DB  102,15,98,199                       ; punpckldq     %xmm7,%xmm0
+  DB  102,15,98,209                       ; punpckldq     %xmm1,%xmm2
+  DB  102,15,108,208                      ; punpcklqdq    %xmm0,%xmm2
+  DB  102,15,111,194                      ; movdqa        %xmm2,%xmm0
+  DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
+  DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
+  DB  102,15,111,203                      ; movdqa        %xmm3,%xmm1
+  DB  102,15,114,241,16                   ; pslld         $0x10,%xmm1
+  DB  102,15,114,225,16                   ; psrad         $0x10,%xmm1
+  DB  102,15,107,200                      ; packssdw      %xmm0,%xmm1
+  DB  102,15,111,195                      ; movdqa        %xmm3,%xmm0
+  DB  102,15,111,250                      ; movdqa        %xmm2,%xmm7
+  DB  102,15,114,210,24                   ; psrld         $0x18,%xmm2
+  DB  102,15,114,242,16                   ; pslld         $0x10,%xmm2
+  DB  102,15,114,226,16                   ; psrad         $0x10,%xmm2
+  DB  102,15,114,211,24                   ; psrld         $0x18,%xmm3
+  DB  102,15,114,243,16                   ; pslld         $0x10,%xmm3
+  DB  102,15,114,227,16                   ; psrad         $0x10,%xmm3
+  DB  102,15,107,218                      ; packssdw      %xmm2,%xmm3
+  DB  102,15,111,162,12,36,0,0            ; movdqa        0x240c(%edx),%xmm4
+  DB  102,15,114,208,16                   ; psrld         $0x10,%xmm0
+  DB  102,15,114,215,16                   ; psrld         $0x10,%xmm7
+  DB  102,15,114,247,16                   ; pslld         $0x10,%xmm7
+  DB  102,15,114,231,16                   ; psrad         $0x10,%xmm7
+  DB  102,15,114,240,16                   ; pslld         $0x10,%xmm0
+  DB  102,15,114,224,16                   ; psrad         $0x10,%xmm0
+  DB  102,15,107,199                      ; packssdw      %xmm7,%xmm0
+  DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
+  DB  102,15,219,212                      ; pand          %xmm4,%xmm2
+  DB  102,15,219,196                      ; pand          %xmm4,%xmm0
+  DB  15,40,101,72                        ; movaps        0x48(%ebp),%xmm4
+  DB  102,15,113,209,8                    ; psrlw         $0x8,%xmm1
+  DB  141,81,8                            ; lea           0x8(%ecx),%edx
+  DB  15,41,100,36,64                     ; movaps        %xmm4,0x40(%esp)
+  DB  15,41,116,36,48                     ; movaps        %xmm6,0x30(%esp)
+  DB  15,41,108,36,32                     ; movaps        %xmm5,0x20(%esp)
+  DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
+  DB  15,41,100,36,16                     ; movaps        %xmm4,0x10(%esp)
+  DB  137,124,36,12                       ; mov           %edi,0xc(%esp)
+  DB  137,116,36,8                        ; mov           %esi,0x8(%esp)
+  DB  137,84,36,4                         ; mov           %edx,0x4(%esp)
+  DB  137,4,36                            ; mov           %eax,(%esp)
+  DB  255,81,4                            ; call          *0x4(%ecx)
+  DB  131,196,80                          ; add           $0x50,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_load_565_sse2_lowp
 _sk_load_565_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -57886,25 +59591,25 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          ba5a <_sk_load_565_sse2_lowp+0x3b>
+  DB  232,0,0,0,0                         ; call          be78 <_sk_load_565_sse2_lowp+0x3b>
   DB  94                                  ; pop           %esi
   DB  139,125,16                          ; mov           0x10(%ebp),%edi
-  DB  119,28                              ; ja            ba7c <_sk_load_565_sse2_lowp+0x5d>
+  DB  119,28                              ; ja            be9a <_sk_load_565_sse2_lowp+0x5d>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,201                          ; movzbl        %cl,%ecx
-  DB  139,140,142,246,0,0,0               ; mov           0xf6(%esi,%ecx,4),%ecx
+  DB  139,140,142,244,0,0,0               ; mov           0xf4(%esi,%ecx,4),%ecx
   DB  1,241                               ; add           %esi,%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  15,183,12,123                       ; movzwl        (%ebx,%edi,2),%ecx
   DB  102,15,110,193                      ; movd          %ecx,%xmm0
-  DB  235,59                              ; jmp           bab7 <_sk_load_565_sse2_lowp+0x98>
+  DB  235,59                              ; jmp           bed5 <_sk_load_565_sse2_lowp+0x98>
   DB  243,15,111,4,123                    ; movdqu        (%ebx,%edi,2),%xmm0
-  DB  235,52                              ; jmp           bab7 <_sk_load_565_sse2_lowp+0x98>
+  DB  235,52                              ; jmp           bed5 <_sk_load_565_sse2_lowp+0x98>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,68,123,4,2               ; pinsrw        $0x2,0x4(%ebx,%edi,2),%xmm0
   DB  243,15,16,12,123                    ; movss         (%ebx,%edi,2),%xmm1
   DB  243,15,16,193                       ; movss         %xmm1,%xmm0
-  DB  235,30                              ; jmp           bab7 <_sk_load_565_sse2_lowp+0x98>
+  DB  235,30                              ; jmp           bed5 <_sk_load_565_sse2_lowp+0x98>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,68,123,12,6              ; pinsrw        $0x6,0xc(%ebx,%edi,2),%xmm0
   DB  102,15,196,68,123,10,5              ; pinsrw        $0x5,0xa(%ebx,%edi,2),%xmm0
@@ -57912,11 +59617,11 @@
   DB  102,15,18,4,123                     ; movlpd        (%ebx,%edi,2),%xmm0
   DB  102,15,111,248                      ; movdqa        %xmm0,%xmm7
   DB  102,15,113,215,8                    ; psrlw         $0x8,%xmm7
-  DB  102,15,219,190,38,29,0,0            ; pand          0x1d26(%esi),%xmm7
+  DB  102,15,219,190,40,34,0,0            ; pand          0x2228(%esi),%xmm7
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
-  DB  102,15,219,142,54,29,0,0            ; pand          0x1d36(%esi),%xmm1
-  DB  102,15,111,150,70,29,0,0            ; movdqa        0x1d46(%esi),%xmm2
+  DB  102,15,219,142,56,34,0,0            ; pand          0x2238(%esi),%xmm1
+  DB  102,15,111,150,72,34,0,0            ; movdqa        0x2248(%esi),%xmm2
   DB  102,15,219,208                      ; pand          %xmm0,%xmm2
   DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
   DB  102,15,235,199                      ; por           %xmm7,%xmm0
@@ -57938,7 +59643,7 @@
   DB  137,124,36,8                        ; mov           %edi,0x8(%esp)
   DB  137,76,36,4                         ; mov           %ecx,0x4(%esp)
   DB  137,20,36                           ; mov           %edx,(%esp)
-  DB  15,40,158,230,28,0,0                ; movaps        0x1ce6(%esi),%xmm3
+  DB  15,40,158,232,33,0,0                ; movaps        0x21e8(%esi),%xmm3
   DB  255,80,4                            ; call          *0x4(%eax)
   DB  131,196,92                          ; add           $0x5c,%esp
   DB  94                                  ; pop           %esi
@@ -57946,7 +59651,7 @@
   DB  91                                  ; pop           %ebx
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
-  DB  15,31,0                             ; nopl          (%eax)
+  DB  144                                 ; nop
   DB  24,0                                ; sbb           %al,(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  52,0                                ; xor           $0x0,%al
@@ -57982,10 +59687,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          bb97 <_sk_load_565_dst_sse2_lowp+0x2b>
+  DB  232,0,0,0,0                         ; call          bfb3 <_sk_load_565_dst_sse2_lowp+0x2b>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            bbb9 <_sk_load_565_dst_sse2_lowp+0x4d>
+  DB  119,28                              ; ja            bfd5 <_sk_load_565_dst_sse2_lowp+0x4d>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,249,0,0,0               ; mov           0xf9(%edi,%ecx,4),%ecx
@@ -57993,14 +59698,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,183,12,115                       ; movzwl        (%ebx,%esi,2),%ecx
   DB  102,15,110,225                      ; movd          %ecx,%xmm4
-  DB  235,59                              ; jmp           bbf4 <_sk_load_565_dst_sse2_lowp+0x88>
+  DB  235,59                              ; jmp           c010 <_sk_load_565_dst_sse2_lowp+0x88>
   DB  243,15,111,36,115                   ; movdqu        (%ebx,%esi,2),%xmm4
-  DB  235,52                              ; jmp           bbf4 <_sk_load_565_dst_sse2_lowp+0x88>
+  DB  235,52                              ; jmp           c010 <_sk_load_565_dst_sse2_lowp+0x88>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,100,115,4,2              ; pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm4
   DB  243,15,16,44,115                    ; movss         (%ebx,%esi,2),%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,30                              ; jmp           bbf4 <_sk_load_565_dst_sse2_lowp+0x88>
+  DB  235,30                              ; jmp           c010 <_sk_load_565_dst_sse2_lowp+0x88>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,100,115,12,6             ; pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm4
   DB  102,15,196,100,115,10,5             ; pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm4
@@ -58008,11 +59713,11 @@
   DB  102,15,18,36,115                    ; movlpd        (%ebx,%esi,2),%xmm4
   DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
   DB  102,15,113,214,8                    ; psrlw         $0x8,%xmm6
-  DB  102,15,219,183,233,27,0,0           ; pand          0x1be9(%edi),%xmm6
+  DB  102,15,219,183,237,32,0,0           ; pand          0x20ed(%edi),%xmm6
   DB  102,15,111,236                      ; movdqa        %xmm4,%xmm5
   DB  102,15,113,213,5                    ; psrlw         $0x5,%xmm5
-  DB  102,15,219,175,249,27,0,0           ; pand          0x1bf9(%edi),%xmm5
-  DB  102,15,111,191,9,28,0,0             ; movdqa        0x1c09(%edi),%xmm7
+  DB  102,15,219,175,253,32,0,0           ; pand          0x20fd(%edi),%xmm5
+  DB  102,15,111,191,13,33,0,0            ; movdqa        0x210d(%edi),%xmm7
   DB  102,15,219,252                      ; pand          %xmm4,%xmm7
   DB  102,15,113,212,13                   ; psrlw         $0xd,%xmm4
   DB  102,15,235,230                      ; por           %xmm6,%xmm4
@@ -58025,7 +59730,7 @@
   DB  102,15,113,215,2                    ; psrlw         $0x2,%xmm7
   DB  102,15,235,254                      ; por           %xmm6,%xmm7
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
-  DB  15,40,183,169,27,0,0                ; movaps        0x1ba9(%edi),%xmm6
+  DB  15,40,183,173,32,0,0                ; movaps        0x20ad(%edi),%xmm6
   DB  15,41,116,36,64                     ; movaps        %xmm6,0x40(%esp)
   DB  139,125,20                          ; mov           0x14(%ebp),%edi
   DB  137,124,36,12                       ; mov           %edi,0xc(%esp)
@@ -58068,15 +59773,15 @@
   DB  86                                  ; push          %esi
   DB  131,236,92                          ; sub           $0x5c,%esp
   DB  102,15,111,226                      ; movdqa        %xmm2,%xmm4
-  DB  232,0,0,0,0                         ; call          bcbe <_sk_store_565_sse2_lowp+0x12>
+  DB  232,0,0,0,0                         ; call          c0da <_sk_store_565_sse2_lowp+0x12>
   DB  95                                  ; pop           %edi
   DB  15,40,109,72                        ; movaps        0x48(%ebp),%xmm5
   DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
   DB  102,15,113,242,8                    ; psllw         $0x8,%xmm2
-  DB  102,15,219,151,242,26,0,0           ; pand          0x1af2(%edi),%xmm2
+  DB  102,15,219,151,246,31,0,0           ; pand          0x1ff6(%edi),%xmm2
   DB  102,15,111,241                      ; movdqa        %xmm1,%xmm6
   DB  102,15,113,246,3                    ; psllw         $0x3,%xmm6
-  DB  102,15,219,183,2,27,0,0             ; pand          0x1b02(%edi),%xmm6
+  DB  102,15,219,183,6,32,0,0             ; pand          0x2006(%edi),%xmm6
   DB  102,15,235,242                      ; por           %xmm2,%xmm6
   DB  102,15,111,212                      ; movdqa        %xmm4,%xmm2
   DB  102,15,113,210,3                    ; psrlw         $0x3,%xmm2
@@ -58095,19 +59800,19 @@
   DB  128,249,6                           ; cmp           $0x6,%cl
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
   DB  15,40,125,24                        ; movaps        0x18(%ebp),%xmm7
-  DB  119,22                              ; ja            bd36 <_sk_store_565_sse2_lowp+0x8a>
+  DB  119,22                              ; ja            c152 <_sk_store_565_sse2_lowp+0x8a>
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  3,188,143,242,0,0,0                 ; add           0xf2(%edi,%ecx,4),%edi
   DB  255,231                             ; jmp           *%edi
   DB  102,15,126,209                      ; movd          %xmm2,%ecx
   DB  102,137,12,115                      ; mov           %cx,(%ebx,%esi,2)
-  DB  235,59                              ; jmp           bd71 <_sk_store_565_sse2_lowp+0xc5>
+  DB  235,59                              ; jmp           c18d <_sk_store_565_sse2_lowp+0xc5>
   DB  243,15,127,20,115                   ; movdqu        %xmm2,(%ebx,%esi,2)
-  DB  235,52                              ; jmp           bd71 <_sk_store_565_sse2_lowp+0xc5>
+  DB  235,52                              ; jmp           c18d <_sk_store_565_sse2_lowp+0xc5>
   DB  102,15,197,202,2                    ; pextrw        $0x2,%xmm2,%ecx
   DB  102,137,76,115,4                    ; mov           %cx,0x4(%ebx,%esi,2)
   DB  102,15,126,20,115                   ; movd          %xmm2,(%ebx,%esi,2)
-  DB  235,35                              ; jmp           bd71 <_sk_store_565_sse2_lowp+0xc5>
+  DB  235,35                              ; jmp           c18d <_sk_store_565_sse2_lowp+0xc5>
   DB  102,15,197,202,6                    ; pextrw        $0x6,%xmm2,%ecx
   DB  102,137,76,115,12                   ; mov           %cx,0xc(%ebx,%esi,2)
   DB  102,15,197,202,5                    ; pextrw        $0x5,%xmm2,%ecx
@@ -58145,6 +59850,138 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  0                                   ; .byte         0x0
 
+PUBLIC _sk_gather_565_sse2_lowp
+_sk_gather_565_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  83                                  ; push          %ebx
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,108                         ; sub           $0x6c,%esp
+  DB  139,69,12                           ; mov           0xc(%ebp),%eax
+  DB  139,16                              ; mov           (%eax),%edx
+  DB  243,15,16,98,8                      ; movss         0x8(%edx),%xmm4
+  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
+  DB  102,15,118,237                      ; pcmpeqd       %xmm5,%xmm5
+  DB  102,15,254,229                      ; paddd         %xmm5,%xmm4
+  DB  15,87,246                           ; xorps         %xmm6,%xmm6
+  DB  15,95,198                           ; maxps         %xmm6,%xmm0
+  DB  15,95,206                           ; maxps         %xmm6,%xmm1
+  DB  15,93,204                           ; minps         %xmm4,%xmm1
+  DB  15,93,196                           ; minps         %xmm4,%xmm0
+  DB  243,15,16,98,12                     ; movss         0xc(%edx),%xmm4
+  DB  15,198,228,0                        ; shufps        $0x0,%xmm4,%xmm4
+  DB  102,15,254,229                      ; paddd         %xmm5,%xmm4
+  DB  15,95,222                           ; maxps         %xmm6,%xmm3
+  DB  15,95,214                           ; maxps         %xmm6,%xmm2
+  DB  15,93,212                           ; minps         %xmm4,%xmm2
+  DB  15,93,220                           ; minps         %xmm4,%xmm3
+  DB  243,15,91,219                       ; cvttps2dq     %xmm3,%xmm3
+  DB  243,15,91,210                       ; cvttps2dq     %xmm2,%xmm2
+  DB  102,15,110,98,4                     ; movd          0x4(%edx),%xmm4
+  DB  102,15,112,228,0                    ; pshufd        $0x0,%xmm4,%xmm4
+  DB  102,15,112,234,245                  ; pshufd        $0xf5,%xmm2,%xmm5
+  DB  102,15,244,236                      ; pmuludq       %xmm4,%xmm5
+  DB  102,15,111,244                      ; movdqa        %xmm4,%xmm6
+  DB  102,15,244,242                      ; pmuludq       %xmm2,%xmm6
+  DB  102,15,112,211,245                  ; pshufd        $0xf5,%xmm3,%xmm2
+  DB  102,15,244,212                      ; pmuludq       %xmm4,%xmm2
+  DB  232,0,0,0,0                         ; call          c260 <_sk_gather_565_sse2_lowp+0x78>
+  DB  89                                  ; pop           %ecx
+  DB  139,50                              ; mov           (%edx),%esi
+  DB  102,15,112,246,232                  ; pshufd        $0xe8,%xmm6,%xmm6
+  DB  102,15,112,237,232                  ; pshufd        $0xe8,%xmm5,%xmm5
+  DB  102,15,98,245                       ; punpckldq     %xmm5,%xmm6
+  DB  102,15,244,227                      ; pmuludq       %xmm3,%xmm4
+  DB  102,15,112,220,232                  ; pshufd        $0xe8,%xmm4,%xmm3
+  DB  102,15,112,210,232                  ; pshufd        $0xe8,%xmm2,%xmm2
+  DB  102,15,98,218                       ; punpckldq     %xmm2,%xmm3
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,198                      ; paddd         %xmm6,%xmm0
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,15,254,203                      ; paddd         %xmm3,%xmm1
+  DB  102,15,126,194                      ; movd          %xmm0,%edx
+  DB  102,15,112,208,229                  ; pshufd        $0xe5,%xmm0,%xmm2
+  DB  102,15,127,85,216                   ; movdqa        %xmm2,-0x28(%ebp)
+  DB  102,15,112,216,78                   ; pshufd        $0x4e,%xmm0,%xmm3
+  DB  102,15,112,192,231                  ; pshufd        $0xe7,%xmm0,%xmm0
+  DB  102,15,112,225,78                   ; pshufd        $0x4e,%xmm1,%xmm4
+  DB  102,15,112,233,231                  ; pshufd        $0xe7,%xmm1,%xmm5
+  DB  102,15,126,239                      ; movd          %xmm5,%edi
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,239                      ; movd          %edi,%xmm5
+  DB  102,15,126,231                      ; movd          %xmm4,%edi
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,231                      ; movd          %edi,%xmm4
+  DB  102,15,126,207                      ; movd          %xmm1,%edi
+  DB  102,15,112,201,229                  ; pshufd        $0xe5,%xmm1,%xmm1
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,247                      ; movd          %edi,%xmm6
+  DB  102,15,126,207                      ; movd          %xmm1,%edi
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,215                      ; movd          %edi,%xmm2
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,255                      ; movd          %edi,%xmm7
+  DB  102,15,126,223                      ; movd          %xmm3,%edi
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  102,15,110,207                      ; movd          %edi,%xmm1
+  DB  102,15,111,69,216                   ; movdqa        -0x28(%ebp),%xmm0
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  15,183,20,86                        ; movzwl        (%esi,%edx,2),%edx
+  DB  102,15,110,194                      ; movd          %edx,%xmm0
+  DB  139,85,8                            ; mov           0x8(%ebp),%edx
+  DB  15,183,60,126                       ; movzwl        (%esi,%edi,2),%edi
+  DB  139,117,16                          ; mov           0x10(%ebp),%esi
+  DB  102,15,97,229                       ; punpcklwd     %xmm5,%xmm4
+  DB  102,15,110,239                      ; movd          %edi,%xmm5
+  DB  139,125,20                          ; mov           0x14(%ebp),%edi
+  DB  102,15,97,242                       ; punpcklwd     %xmm2,%xmm6
+  DB  15,40,93,24                         ; movaps        0x18(%ebp),%xmm3
+  DB  102,15,98,244                       ; punpckldq     %xmm4,%xmm6
+  DB  15,40,101,40                        ; movaps        0x28(%ebp),%xmm4
+  DB  102,15,97,207                       ; punpcklwd     %xmm7,%xmm1
+  DB  15,40,125,56                        ; movaps        0x38(%ebp),%xmm7
+  DB  102,15,97,197                       ; punpcklwd     %xmm5,%xmm0
+  DB  102,15,98,193                       ; punpckldq     %xmm1,%xmm0
+  DB  102,15,108,198                      ; punpcklqdq    %xmm6,%xmm0
+  DB  102,15,111,232                      ; movdqa        %xmm0,%xmm5
+  DB  102,15,113,213,8                    ; psrlw         $0x8,%xmm5
+  DB  102,15,219,169,64,30,0,0            ; pand          0x1e40(%ecx),%xmm5
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,113,209,5                    ; psrlw         $0x5,%xmm1
+  DB  102,15,219,137,80,30,0,0            ; pand          0x1e50(%ecx),%xmm1
+  DB  102,15,111,145,96,30,0,0            ; movdqa        0x1e60(%ecx),%xmm2
+  DB  102,15,219,208                      ; pand          %xmm0,%xmm2
+  DB  102,15,113,208,13                   ; psrlw         $0xd,%xmm0
+  DB  102,15,235,197                      ; por           %xmm5,%xmm0
+  DB  102,15,111,233                      ; movdqa        %xmm1,%xmm5
+  DB  102,15,113,245,2                    ; psllw         $0x2,%xmm5
+  DB  102,15,113,209,4                    ; psrlw         $0x4,%xmm1
+  DB  102,15,235,205                      ; por           %xmm5,%xmm1
+  DB  102,15,111,234                      ; movdqa        %xmm2,%xmm5
+  DB  102,15,113,245,3                    ; psllw         $0x3,%xmm5
+  DB  102,15,113,210,2                    ; psrlw         $0x2,%xmm2
+  DB  102,15,235,213                      ; por           %xmm5,%xmm2
+  DB  15,40,109,72                        ; movaps        0x48(%ebp),%xmm5
+  DB  141,88,8                            ; lea           0x8(%eax),%ebx
+  DB  15,41,108,36,64                     ; movaps        %xmm5,0x40(%esp)
+  DB  15,41,124,36,48                     ; movaps        %xmm7,0x30(%esp)
+  DB  15,41,100,36,32                     ; movaps        %xmm4,0x20(%esp)
+  DB  15,41,92,36,16                      ; movaps        %xmm3,0x10(%esp)
+  DB  137,124,36,12                       ; mov           %edi,0xc(%esp)
+  DB  137,116,36,8                        ; mov           %esi,0x8(%esp)
+  DB  137,92,36,4                         ; mov           %ebx,0x4(%esp)
+  DB  137,20,36                           ; mov           %edx,(%esp)
+  DB  15,40,153,0,30,0,0                  ; movaps        0x1e00(%ecx),%xmm3
+  DB  255,80,4                            ; call          *0x4(%eax)
+  DB  131,196,108                         ; add           $0x6c,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  91                                  ; pop           %ebx
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_load_a8_sse2_lowp
 _sk_load_a8_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -58167,21 +60004,21 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          be06 <_sk_load_a8_sse2_lowp+0x3a>
+  DB  232,0,0,0,0                         ; call          c41d <_sk_load_a8_sse2_lowp+0x3a>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            be28 <_sk_load_a8_sse2_lowp+0x5c>
+  DB  119,28                              ; ja            c43f <_sk_load_a8_sse2_lowp+0x5c>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
-  DB  139,140,143,198,0,0,0               ; mov           0xc6(%edi,%ecx,4),%ecx
+  DB  139,140,143,195,0,0,0               ; mov           0xc3(%edi,%ecx,4),%ecx
   DB  1,249                               ; add           %edi,%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,51                        ; movzbl        (%ebx,%esi,1),%ecx
   DB  102,15,110,217                      ; movd          %ecx,%xmm3
-  DB  235,90                              ; jmp           be82 <_sk_load_a8_sse2_lowp+0xb6>
+  DB  235,90                              ; jmp           c499 <_sk_load_a8_sse2_lowp+0xb6>
   DB  243,15,126,28,51                    ; movq          (%ebx,%esi,1),%xmm3
   DB  102,15,96,216                       ; punpcklbw     %xmm0,%xmm3
-  DB  235,79                              ; jmp           be82 <_sk_load_a8_sse2_lowp+0xb6>
+  DB  235,79                              ; jmp           c499 <_sk_load_a8_sse2_lowp+0xb6>
   DB  15,182,76,51,2                      ; movzbl        0x2(%ebx,%esi,1),%ecx
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,217,2                    ; pinsrw        $0x2,%ecx,%xmm3
@@ -58189,7 +60026,7 @@
   DB  102,15,110,233                      ; movd          %ecx,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,221                       ; movss         %xmm5,%xmm3
-  DB  235,47                              ; jmp           be82 <_sk_load_a8_sse2_lowp+0xb6>
+  DB  235,47                              ; jmp           c499 <_sk_load_a8_sse2_lowp+0xb6>
   DB  15,182,76,51,6                      ; movzbl        0x6(%ebx,%esi,1),%ecx
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,217,6                    ; pinsrw        $0x6,%ecx,%xmm3
@@ -58200,7 +60037,7 @@
   DB  102,15,110,44,51                    ; movd          (%ebx,%esi,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,221                       ; movsd         %xmm5,%xmm3
-  DB  102,15,219,159,58,25,0,0            ; pand          0x193a(%edi),%xmm3
+  DB  102,15,219,159,67,28,0,0            ; pand          0x1c43(%edi),%xmm3
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  15,41,100,36,64                     ; movaps        %xmm4,0x40(%esp)
   DB  15,41,84,36,48                      ; movaps        %xmm2,0x30(%esp)
@@ -58221,7 +60058,6 @@
   DB  91                                  ; pop           %ebx
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
-  DB  15,31,0                             ; nopl          (%eax)
   DB  24,0                                ; sbb           %al,(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  59,0                                ; cmp           (%eax),%eax
@@ -58253,10 +60089,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          bf11 <_sk_load_a8_dst_sse2_lowp+0x29>
+  DB  232,0,0,0,0                         ; call          c525 <_sk_load_a8_dst_sse2_lowp+0x29>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            bf33 <_sk_load_a8_dst_sse2_lowp+0x4b>
+  DB  119,28                              ; ja            c547 <_sk_load_a8_dst_sse2_lowp+0x4b>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,195,0,0,0               ; mov           0xc3(%edi,%ecx,4),%ecx
@@ -58264,10 +60100,10 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,51                        ; movzbl        (%ebx,%esi,1),%ecx
   DB  102,15,110,225                      ; movd          %ecx,%xmm4
-  DB  235,90                              ; jmp           bf8d <_sk_load_a8_dst_sse2_lowp+0xa5>
+  DB  235,90                              ; jmp           c5a1 <_sk_load_a8_dst_sse2_lowp+0xa5>
   DB  243,15,126,36,51                    ; movq          (%ebx,%esi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
-  DB  235,79                              ; jmp           bf8d <_sk_load_a8_dst_sse2_lowp+0xa5>
+  DB  235,79                              ; jmp           c5a1 <_sk_load_a8_dst_sse2_lowp+0xa5>
   DB  15,182,76,51,2                      ; movzbl        0x2(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,2                    ; pinsrw        $0x2,%ecx,%xmm4
@@ -58275,7 +60111,7 @@
   DB  102,15,110,233                      ; movd          %ecx,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,47                              ; jmp           bf8d <_sk_load_a8_dst_sse2_lowp+0xa5>
+  DB  235,47                              ; jmp           c5a1 <_sk_load_a8_dst_sse2_lowp+0xa5>
   DB  15,182,76,51,6                      ; movzbl        0x6(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,6                    ; pinsrw        $0x6,%ecx,%xmm4
@@ -58286,7 +60122,7 @@
   DB  102,15,110,44,51                    ; movd          (%ebx,%esi,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,229                       ; movsd         %xmm5,%xmm4
-  DB  102,15,219,167,47,24,0,0            ; pand          0x182f(%edi),%xmm4
+  DB  102,15,219,167,59,27,0,0            ; pand          0x1b3b(%edi),%xmm4
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  102,15,127,100,36,64                ; movdqa        %xmm4,0x40(%esp)
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
@@ -58341,32 +60177,32 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c029 <_sk_store_a8_sse2_lowp+0x39>
+  DB  232,0,0,0,0                         ; call          c63d <_sk_store_a8_sse2_lowp+0x39>
   DB  91                                  ; pop           %ebx
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,26                              ; ja            c049 <_sk_store_a8_sse2_lowp+0x59>
+  DB  119,26                              ; ja            c65d <_sk_store_a8_sse2_lowp+0x59>
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,139,223,0,0,0               ; mov           0xdf(%ebx,%ecx,4),%ecx
   DB  1,217                               ; add           %ebx,%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  102,15,126,217                      ; movd          %xmm3,%ecx
   DB  136,12,55                           ; mov           %cl,(%edi,%esi,1)
-  DB  233,136,0,0,0                       ; jmp           c0d1 <_sk_store_a8_sse2_lowp+0xe1>
-  DB  102,15,111,163,23,23,0,0            ; movdqa        0x1717(%ebx),%xmm4
+  DB  233,136,0,0,0                       ; jmp           c6e5 <_sk_store_a8_sse2_lowp+0xe1>
+  DB  102,15,111,163,35,26,0,0            ; movdqa        0x1a23(%ebx),%xmm4
   DB  102,15,219,227                      ; pand          %xmm3,%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
   DB  102,15,214,36,55                    ; movq          %xmm4,(%edi,%esi,1)
   DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
-  DB  235,109                             ; jmp           c0d1 <_sk_store_a8_sse2_lowp+0xe1>
+  DB  235,109                             ; jmp           c6e5 <_sk_store_a8_sse2_lowp+0xe1>
   DB  102,15,197,203,2                    ; pextrw        $0x2,%xmm3,%ecx
   DB  136,76,55,2                         ; mov           %cl,0x2(%edi,%esi,1)
-  DB  102,15,111,163,23,23,0,0            ; movdqa        0x1717(%ebx),%xmm4
+  DB  102,15,111,163,35,26,0,0            ; movdqa        0x1a23(%ebx),%xmm4
   DB  102,15,219,227                      ; pand          %xmm3,%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
   DB  102,15,126,225                      ; movd          %xmm4,%ecx
   DB  15,40,101,24                        ; movaps        0x18(%ebp),%xmm4
   DB  102,137,12,55                       ; mov           %cx,(%edi,%esi,1)
-  DB  235,70                              ; jmp           c0d1 <_sk_store_a8_sse2_lowp+0xe1>
+  DB  235,70                              ; jmp           c6e5 <_sk_store_a8_sse2_lowp+0xe1>
   DB  102,15,197,203,6                    ; pextrw        $0x6,%xmm3,%ecx
   DB  136,76,55,6                         ; mov           %cl,0x6(%edi,%esi,1)
   DB  102,15,197,203,5                    ; pextrw        $0x5,%xmm3,%ecx
@@ -58376,7 +60212,7 @@
   DB  15,40,254                           ; movaps        %xmm6,%xmm7
   DB  15,40,245                           ; movaps        %xmm5,%xmm6
   DB  15,40,236                           ; movaps        %xmm4,%xmm5
-  DB  102,15,111,163,23,23,0,0            ; movdqa        0x1717(%ebx),%xmm4
+  DB  102,15,111,163,35,26,0,0            ; movdqa        0x1a23(%ebx),%xmm4
   DB  102,15,219,227                      ; pand          %xmm3,%xmm4
   DB  102,15,103,228                      ; packuswb      %xmm4,%xmm4
   DB  102,15,126,36,55                    ; movd          %xmm4,(%edi,%esi,1)
@@ -58410,12 +60246,120 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  0,125,0                             ; add           %bh,0x0(%ebp)
   DB  0,0                                 ; add           %al,(%eax)
-  DB  116,0                               ; je            c11a <_sk_store_a8_sse2_lowp+0x12a>
+  DB  116,0                               ; je            c72e <_sk_store_a8_sse2_lowp+0x12a>
   DB  0,0                                 ; add           %al,(%eax)
   DB  107,0,0                             ; imul          $0x0,(%eax),%eax
   DB  0,98,0                              ; add           %ah,0x0(%edx)
   DB  0,0                                 ; add           %al,(%eax)
 
+PUBLIC _sk_gather_a8_sse2_lowp
+_sk_gather_a8_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,80                          ; sub           $0x50,%esp
+  DB  139,69,12                           ; mov           0xc(%ebp),%eax
+  DB  139,8                               ; mov           (%eax),%ecx
+  DB  243,15,16,105,8                     ; movss         0x8(%ecx),%xmm5
+  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
+  DB  102,15,118,246                      ; pcmpeqd       %xmm6,%xmm6
+  DB  102,15,254,238                      ; paddd         %xmm6,%xmm5
+  DB  15,87,228                           ; xorps         %xmm4,%xmm4
+  DB  15,95,196                           ; maxps         %xmm4,%xmm0
+  DB  15,95,204                           ; maxps         %xmm4,%xmm1
+  DB  15,93,205                           ; minps         %xmm5,%xmm1
+  DB  15,93,197                           ; minps         %xmm5,%xmm0
+  DB  243,15,16,105,12                    ; movss         0xc(%ecx),%xmm5
+  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
+  DB  102,15,254,238                      ; paddd         %xmm6,%xmm5
+  DB  15,95,220                           ; maxps         %xmm4,%xmm3
+  DB  15,95,212                           ; maxps         %xmm4,%xmm2
+  DB  15,93,213                           ; minps         %xmm5,%xmm2
+  DB  15,93,221                           ; minps         %xmm5,%xmm3
+  DB  243,15,91,234                       ; cvttps2dq     %xmm2,%xmm5
+  DB  102,15,110,81,4                     ; movd          0x4(%ecx),%xmm2
+  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
+  DB  102,15,111,242                      ; movdqa        %xmm2,%xmm6
+  DB  102,15,244,245                      ; pmuludq       %xmm5,%xmm6
+  DB  102,15,112,237,245                  ; pshufd        $0xf5,%xmm5,%xmm5
+  DB  102,15,244,234                      ; pmuludq       %xmm2,%xmm5
+  DB  102,15,112,246,232                  ; pshufd        $0xe8,%xmm6,%xmm6
+  DB  102,15,112,237,232                  ; pshufd        $0xe8,%xmm5,%xmm5
+  DB  102,15,98,245                       ; punpckldq     %xmm5,%xmm6
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,198                      ; paddd         %xmm6,%xmm0
+  DB  102,15,112,232,78                   ; pshufd        $0x4e,%xmm0,%xmm5
+  DB  102,15,126,234                      ; movd          %xmm5,%edx
+  DB  102,15,112,232,231                  ; pshufd        $0xe7,%xmm0,%xmm5
+  DB  102,15,126,238                      ; movd          %xmm5,%esi
+  DB  139,9                               ; mov           (%ecx),%ecx
+  DB  243,15,91,219                       ; cvttps2dq     %xmm3,%xmm3
+  DB  102,15,126,199                      ; movd          %xmm0,%edi
+  DB  102,15,112,192,229                  ; pshufd        $0xe5,%xmm0,%xmm0
+  DB  15,182,20,17                        ; movzbl        (%ecx,%edx,1),%edx
+  DB  15,182,52,49                        ; movzbl        (%ecx,%esi,1),%esi
+  DB  193,230,8                           ; shl           $0x8,%esi
+  DB  9,214                               ; or            %edx,%esi
+  DB  102,15,126,194                      ; movd          %xmm0,%edx
+  DB  102,15,112,195,245                  ; pshufd        $0xf5,%xmm3,%xmm0
+  DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
+  DB  15,182,60,57                        ; movzbl        (%ecx,%edi,1),%edi
+  DB  15,182,20,17                        ; movzbl        (%ecx,%edx,1),%edx
+  DB  193,226,8                           ; shl           $0x8,%edx
+  DB  9,250                               ; or            %edi,%edx
+  DB  139,125,8                           ; mov           0x8(%ebp),%edi
+  DB  102,15,244,211                      ; pmuludq       %xmm3,%xmm2
+  DB  102,15,112,210,232                  ; pshufd        $0xe8,%xmm2,%xmm2
+  DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
+  DB  102,15,98,208                       ; punpckldq     %xmm0,%xmm2
+  DB  243,15,91,193                       ; cvttps2dq     %xmm1,%xmm0
+  DB  102,15,254,194                      ; paddd         %xmm2,%xmm0
+  DB  102,15,110,218                      ; movd          %edx,%xmm3
+  DB  102,15,126,194                      ; movd          %xmm0,%edx
+  DB  102,15,112,200,229                  ; pshufd        $0xe5,%xmm0,%xmm1
+  DB  102,15,196,222,1                    ; pinsrw        $0x1,%esi,%xmm3
+  DB  102,15,126,206                      ; movd          %xmm1,%esi
+  DB  102,15,112,200,78                   ; pshufd        $0x4e,%xmm0,%xmm1
+  DB  15,182,20,17                        ; movzbl        (%ecx,%edx,1),%edx
+  DB  15,182,52,49                        ; movzbl        (%ecx,%esi,1),%esi
+  DB  193,230,8                           ; shl           $0x8,%esi
+  DB  9,214                               ; or            %edx,%esi
+  DB  102,15,126,202                      ; movd          %xmm1,%edx
+  DB  102,15,112,192,231                  ; pshufd        $0xe7,%xmm0,%xmm0
+  DB  102,15,196,222,2                    ; pinsrw        $0x2,%esi,%xmm3
+  DB  102,15,126,198                      ; movd          %xmm0,%esi
+  DB  15,182,20,17                        ; movzbl        (%ecx,%edx,1),%edx
+  DB  15,182,12,49                        ; movzbl        (%ecx,%esi,1),%ecx
+  DB  139,117,16                          ; mov           0x10(%ebp),%esi
+  DB  193,225,8                           ; shl           $0x8,%ecx
+  DB  9,209                               ; or            %edx,%ecx
+  DB  139,85,20                           ; mov           0x14(%ebp),%edx
+  DB  15,40,69,24                         ; movaps        0x18(%ebp),%xmm0
+  DB  15,40,77,40                         ; movaps        0x28(%ebp),%xmm1
+  DB  15,40,85,56                         ; movaps        0x38(%ebp),%xmm2
+  DB  102,15,196,217,3                    ; pinsrw        $0x3,%ecx,%xmm3
+  DB  102,15,96,220                       ; punpcklbw     %xmm4,%xmm3
+  DB  15,40,101,72                        ; movaps        0x48(%ebp),%xmm4
+  DB  141,72,8                            ; lea           0x8(%eax),%ecx
+  DB  15,41,100,36,64                     ; movaps        %xmm4,0x40(%esp)
+  DB  15,41,84,36,48                      ; movaps        %xmm2,0x30(%esp)
+  DB  15,41,76,36,32                      ; movaps        %xmm1,0x20(%esp)
+  DB  15,41,68,36,16                      ; movaps        %xmm0,0x10(%esp)
+  DB  137,84,36,12                        ; mov           %edx,0xc(%esp)
+  DB  137,116,36,8                        ; mov           %esi,0x8(%esp)
+  DB  137,76,36,4                         ; mov           %ecx,0x4(%esp)
+  DB  137,60,36                           ; mov           %edi,(%esp)
+  DB  15,87,192                           ; xorps         %xmm0,%xmm0
+  DB  15,87,201                           ; xorps         %xmm1,%xmm1
+  DB  15,87,210                           ; xorps         %xmm2,%xmm2
+  DB  255,80,4                            ; call          *0x4(%eax)
+  DB  131,196,80                          ; add           $0x50,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_load_g8_sse2_lowp
 _sk_load_g8_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -58438,21 +60382,21 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c15d <_sk_load_g8_sse2_lowp+0x39>
+  DB  232,0,0,0,0                         ; call          c8f0 <_sk_load_g8_sse2_lowp+0x39>
   DB  94                                  ; pop           %esi
   DB  139,125,16                          ; mov           0x10(%ebp),%edi
-  DB  119,28                              ; ja            c17f <_sk_load_g8_sse2_lowp+0x5b>
+  DB  119,28                              ; ja            c912 <_sk_load_g8_sse2_lowp+0x5b>
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  15,182,201                          ; movzbl        %cl,%ecx
-  DB  139,140,142,199,0,0,0               ; mov           0xc7(%esi,%ecx,4),%ecx
+  DB  139,140,142,200,0,0,0               ; mov           0xc8(%esi,%ecx,4),%ecx
   DB  1,241                               ; add           %esi,%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,59                        ; movzbl        (%ebx,%edi,1),%ecx
   DB  102,15,110,193                      ; movd          %ecx,%xmm0
-  DB  235,90                              ; jmp           c1d9 <_sk_load_g8_sse2_lowp+0xb5>
+  DB  235,90                              ; jmp           c96c <_sk_load_g8_sse2_lowp+0xb5>
   DB  243,15,126,4,59                     ; movq          (%ebx,%edi,1),%xmm0
   DB  102,15,96,192                       ; punpcklbw     %xmm0,%xmm0
-  DB  235,79                              ; jmp           c1d9 <_sk_load_g8_sse2_lowp+0xb5>
+  DB  235,79                              ; jmp           c96c <_sk_load_g8_sse2_lowp+0xb5>
   DB  15,182,76,59,2                      ; movzbl        0x2(%ebx,%edi,1),%ecx
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,193,2                    ; pinsrw        $0x2,%ecx,%xmm0
@@ -58460,7 +60404,7 @@
   DB  102,15,110,233                      ; movd          %ecx,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,197                       ; movss         %xmm5,%xmm0
-  DB  235,47                              ; jmp           c1d9 <_sk_load_g8_sse2_lowp+0xb5>
+  DB  235,47                              ; jmp           c96c <_sk_load_g8_sse2_lowp+0xb5>
   DB  15,182,76,59,6                      ; movzbl        0x6(%ebx,%edi,1),%ecx
   DB  102,15,239,192                      ; pxor          %xmm0,%xmm0
   DB  102,15,196,193,6                    ; pinsrw        $0x6,%ecx,%xmm0
@@ -58471,7 +60415,7 @@
   DB  102,15,110,44,59                    ; movd          (%ebx,%edi,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,197                       ; movsd         %xmm5,%xmm0
-  DB  102,15,219,134,227,21,0,0           ; pand          0x15e3(%esi),%xmm0
+  DB  102,15,219,134,112,23,0,0           ; pand          0x1770(%esi),%xmm0
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
   DB  15,41,100,36,64                     ; movaps        %xmm4,0x40(%esp)
   DB  15,41,92,36,48                      ; movaps        %xmm3,0x30(%esp)
@@ -58482,7 +60426,7 @@
   DB  137,124,36,8                        ; mov           %edi,0x8(%esp)
   DB  137,76,36,4                         ; mov           %ecx,0x4(%esp)
   DB  137,20,36                           ; mov           %edx,(%esp)
-  DB  15,40,158,227,21,0,0                ; movaps        0x15e3(%esi),%xmm3
+  DB  15,40,158,112,23,0,0                ; movaps        0x1770(%esi),%xmm3
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
   DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
   DB  255,80,4                            ; call          *0x4(%eax)
@@ -58492,6 +60436,7 @@
   DB  91                                  ; pop           %ebx
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
+  DB  144                                 ; nop
   DB  24,0                                ; sbb           %al,(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  59,0                                ; cmp           (%eax),%eax
@@ -58523,10 +60468,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c269 <_sk_load_g8_dst_sse2_lowp+0x29>
+  DB  232,0,0,0,0                         ; call          c9fd <_sk_load_g8_dst_sse2_lowp+0x29>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            c28b <_sk_load_g8_dst_sse2_lowp+0x4b>
+  DB  119,28                              ; ja            ca1f <_sk_load_g8_dst_sse2_lowp+0x4b>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,195,0,0,0               ; mov           0xc3(%edi,%ecx,4),%ecx
@@ -58534,10 +60479,10 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,51                        ; movzbl        (%ebx,%esi,1),%ecx
   DB  102,15,110,225                      ; movd          %ecx,%xmm4
-  DB  235,90                              ; jmp           c2e5 <_sk_load_g8_dst_sse2_lowp+0xa5>
+  DB  235,90                              ; jmp           ca79 <_sk_load_g8_dst_sse2_lowp+0xa5>
   DB  243,15,126,36,51                    ; movq          (%ebx,%esi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
-  DB  235,79                              ; jmp           c2e5 <_sk_load_g8_dst_sse2_lowp+0xa5>
+  DB  235,79                              ; jmp           ca79 <_sk_load_g8_dst_sse2_lowp+0xa5>
   DB  15,182,76,51,2                      ; movzbl        0x2(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,2                    ; pinsrw        $0x2,%ecx,%xmm4
@@ -58545,7 +60490,7 @@
   DB  102,15,110,233                      ; movd          %ecx,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,47                              ; jmp           c2e5 <_sk_load_g8_dst_sse2_lowp+0xa5>
+  DB  235,47                              ; jmp           ca79 <_sk_load_g8_dst_sse2_lowp+0xa5>
   DB  15,182,76,51,6                      ; movzbl        0x6(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,6                    ; pinsrw        $0x6,%ecx,%xmm4
@@ -58556,9 +60501,9 @@
   DB  102,15,110,44,51                    ; movd          (%ebx,%esi,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,229                       ; movsd         %xmm5,%xmm4
-  DB  102,15,219,167,215,20,0,0           ; pand          0x14d7(%edi),%xmm4
+  DB  102,15,219,167,99,22,0,0            ; pand          0x1663(%edi),%xmm4
   DB  141,72,8                            ; lea           0x8(%eax),%ecx
-  DB  15,40,175,215,20,0,0                ; movaps        0x14d7(%edi),%xmm5
+  DB  15,40,175,99,22,0,0                 ; movaps        0x1663(%edi),%xmm5
   DB  15,41,108,36,64                     ; movaps        %xmm5,0x40(%esp)
   DB  102,15,127,100,36,48                ; movdqa        %xmm4,0x30(%esp)
   DB  102,15,127,100,36,32                ; movdqa        %xmm4,0x20(%esp)
@@ -58597,7 +60542,7 @@
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
-  DB  232,0,0,0,0                         ; call          c359 <_sk_luminance_to_alpha_sse2_lowp+0x11>
+  DB  232,0,0,0,0                         ; call          caed <_sk_luminance_to_alpha_sse2_lowp+0x11>
   DB  88                                  ; pop           %eax
   DB  139,77,8                            ; mov           0x8(%ebp),%ecx
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
@@ -58607,10 +60552,10 @@
   DB  15,40,109,56                        ; movaps        0x38(%ebp),%xmm5
   DB  15,40,117,72                        ; movaps        0x48(%ebp),%xmm6
   DB  139,125,12                          ; mov           0xc(%ebp),%edi
-  DB  102,15,213,128,119,20,0,0           ; pmullw        0x1477(%eax),%xmm0
-  DB  102,15,213,136,135,20,0,0           ; pmullw        0x1487(%eax),%xmm1
+  DB  102,15,213,128,3,22,0,0             ; pmullw        0x1603(%eax),%xmm0
+  DB  102,15,213,136,19,22,0,0            ; pmullw        0x1613(%eax),%xmm1
   DB  102,15,253,200                      ; paddw         %xmm0,%xmm1
-  DB  102,15,213,152,151,20,0,0           ; pmullw        0x1497(%eax),%xmm3
+  DB  102,15,213,152,35,22,0,0            ; pmullw        0x1623(%eax),%xmm3
   DB  102,15,253,217                      ; paddw         %xmm1,%xmm3
   DB  102,15,113,211,8                    ; psrlw         $0x8,%xmm3
   DB  141,71,4                            ; lea           0x4(%edi),%eax
@@ -58632,6 +60577,118 @@
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
 
+PUBLIC _sk_gather_g8_sse2_lowp
+_sk_gather_g8_sse2_lowp LABEL PROC
+  DB  85                                  ; push          %ebp
+  DB  137,229                             ; mov           %esp,%ebp
+  DB  83                                  ; push          %ebx
+  DB  87                                  ; push          %edi
+  DB  86                                  ; push          %esi
+  DB  131,236,92                          ; sub           $0x5c,%esp
+  DB  139,69,12                           ; mov           0xc(%ebp),%eax
+  DB  139,8                               ; mov           (%eax),%ecx
+  DB  243,15,16,105,8                     ; movss         0x8(%ecx),%xmm5
+  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
+  DB  102,15,118,246                      ; pcmpeqd       %xmm6,%xmm6
+  DB  102,15,254,238                      ; paddd         %xmm6,%xmm5
+  DB  15,87,228                           ; xorps         %xmm4,%xmm4
+  DB  15,95,196                           ; maxps         %xmm4,%xmm0
+  DB  15,95,204                           ; maxps         %xmm4,%xmm1
+  DB  15,93,205                           ; minps         %xmm5,%xmm1
+  DB  15,93,197                           ; minps         %xmm5,%xmm0
+  DB  243,15,16,105,12                    ; movss         0xc(%ecx),%xmm5
+  DB  15,198,237,0                        ; shufps        $0x0,%xmm5,%xmm5
+  DB  102,15,254,238                      ; paddd         %xmm6,%xmm5
+  DB  15,95,220                           ; maxps         %xmm4,%xmm3
+  DB  15,95,212                           ; maxps         %xmm4,%xmm2
+  DB  15,93,213                           ; minps         %xmm5,%xmm2
+  DB  15,93,221                           ; minps         %xmm5,%xmm3
+  DB  243,15,91,234                       ; cvttps2dq     %xmm2,%xmm5
+  DB  102,15,110,81,4                     ; movd          0x4(%ecx),%xmm2
+  DB  102,15,112,210,0                    ; pshufd        $0x0,%xmm2,%xmm2
+  DB  102,15,111,242                      ; movdqa        %xmm2,%xmm6
+  DB  102,15,244,245                      ; pmuludq       %xmm5,%xmm6
+  DB  102,15,112,237,245                  ; pshufd        $0xf5,%xmm5,%xmm5
+  DB  102,15,244,234                      ; pmuludq       %xmm2,%xmm5
+  DB  102,15,112,246,232                  ; pshufd        $0xe8,%xmm6,%xmm6
+  DB  102,15,112,237,232                  ; pshufd        $0xe8,%xmm5,%xmm5
+  DB  102,15,98,245                       ; punpckldq     %xmm5,%xmm6
+  DB  243,15,91,192                       ; cvttps2dq     %xmm0,%xmm0
+  DB  102,15,254,198                      ; paddd         %xmm6,%xmm0
+  DB  102,15,112,232,78                   ; pshufd        $0x4e,%xmm0,%xmm5
+  DB  102,15,126,234                      ; movd          %xmm5,%edx
+  DB  102,15,112,232,231                  ; pshufd        $0xe7,%xmm0,%xmm5
+  DB  102,15,126,238                      ; movd          %xmm5,%esi
+  DB  139,57                              ; mov           (%ecx),%edi
+  DB  243,15,91,219                       ; cvttps2dq     %xmm3,%xmm3
+  DB  102,15,126,193                      ; movd          %xmm0,%ecx
+  DB  102,15,112,192,229                  ; pshufd        $0xe5,%xmm0,%xmm0
+  DB  15,182,20,23                        ; movzbl        (%edi,%edx,1),%edx
+  DB  15,182,52,55                        ; movzbl        (%edi,%esi,1),%esi
+  DB  193,230,8                           ; shl           $0x8,%esi
+  DB  9,214                               ; or            %edx,%esi
+  DB  102,15,126,194                      ; movd          %xmm0,%edx
+  DB  102,15,112,195,245                  ; pshufd        $0xf5,%xmm3,%xmm0
+  DB  102,15,244,194                      ; pmuludq       %xmm2,%xmm0
+  DB  15,182,12,15                        ; movzbl        (%edi,%ecx,1),%ecx
+  DB  15,182,20,23                        ; movzbl        (%edi,%edx,1),%edx
+  DB  193,226,8                           ; shl           $0x8,%edx
+  DB  9,202                               ; or            %ecx,%edx
+  DB  232,0,0,0,0                         ; call          cc32 <_sk_gather_g8_sse2_lowp+0xc9>
+  DB  89                                  ; pop           %ecx
+  DB  102,15,244,211                      ; pmuludq       %xmm3,%xmm2
+  DB  102,15,112,210,232                  ; pshufd        $0xe8,%xmm2,%xmm2
+  DB  102,15,112,192,232                  ; pshufd        $0xe8,%xmm0,%xmm0
+  DB  102,15,98,208                       ; punpckldq     %xmm0,%xmm2
+  DB  243,15,91,201                       ; cvttps2dq     %xmm1,%xmm1
+  DB  102,15,254,202                      ; paddd         %xmm2,%xmm1
+  DB  102,15,110,194                      ; movd          %edx,%xmm0
+  DB  102,15,126,202                      ; movd          %xmm1,%edx
+  DB  102,15,112,209,229                  ; pshufd        $0xe5,%xmm1,%xmm2
+  DB  102,15,196,198,1                    ; pinsrw        $0x1,%esi,%xmm0
+  DB  102,15,126,214                      ; movd          %xmm2,%esi
+  DB  102,15,112,209,78                   ; pshufd        $0x4e,%xmm1,%xmm2
+  DB  15,182,20,23                        ; movzbl        (%edi,%edx,1),%edx
+  DB  15,182,52,55                        ; movzbl        (%edi,%esi,1),%esi
+  DB  193,230,8                           ; shl           $0x8,%esi
+  DB  9,214                               ; or            %edx,%esi
+  DB  102,15,126,210                      ; movd          %xmm2,%edx
+  DB  102,15,112,201,231                  ; pshufd        $0xe7,%xmm1,%xmm1
+  DB  102,15,196,198,2                    ; pinsrw        $0x2,%esi,%xmm0
+  DB  102,15,126,206                      ; movd          %xmm1,%esi
+  DB  15,182,20,23                        ; movzbl        (%edi,%edx,1),%edx
+  DB  15,182,52,55                        ; movzbl        (%edi,%esi,1),%esi
+  DB  139,125,8                           ; mov           0x8(%ebp),%edi
+  DB  193,230,8                           ; shl           $0x8,%esi
+  DB  9,214                               ; or            %edx,%esi
+  DB  139,85,16                           ; mov           0x10(%ebp),%edx
+  DB  102,15,196,198,3                    ; pinsrw        $0x3,%esi,%xmm0
+  DB  139,117,20                          ; mov           0x14(%ebp),%esi
+  DB  15,40,77,24                         ; movaps        0x18(%ebp),%xmm1
+  DB  15,40,85,40                         ; movaps        0x28(%ebp),%xmm2
+  DB  15,40,93,56                         ; movaps        0x38(%ebp),%xmm3
+  DB  102,15,96,196                       ; punpcklbw     %xmm4,%xmm0
+  DB  15,40,101,72                        ; movaps        0x48(%ebp),%xmm4
+  DB  141,88,8                            ; lea           0x8(%eax),%ebx
+  DB  15,41,100,36,64                     ; movaps        %xmm4,0x40(%esp)
+  DB  15,41,92,36,48                      ; movaps        %xmm3,0x30(%esp)
+  DB  15,41,84,36,32                      ; movaps        %xmm2,0x20(%esp)
+  DB  15,41,76,36,16                      ; movaps        %xmm1,0x10(%esp)
+  DB  137,116,36,12                       ; mov           %esi,0xc(%esp)
+  DB  137,84,36,8                         ; mov           %edx,0x8(%esp)
+  DB  137,92,36,4                         ; mov           %ebx,0x4(%esp)
+  DB  137,60,36                           ; mov           %edi,(%esp)
+  DB  15,40,153,46,20,0,0                 ; movaps        0x142e(%ecx),%xmm3
+  DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
+  DB  102,15,111,208                      ; movdqa        %xmm0,%xmm2
+  DB  255,80,4                            ; call          *0x4(%eax)
+  DB  131,196,92                          ; add           $0x5c,%esp
+  DB  94                                  ; pop           %esi
+  DB  95                                  ; pop           %edi
+  DB  91                                  ; pop           %ebx
+  DB  93                                  ; pop           %ebp
+  DB  195                                 ; ret
+
 PUBLIC _sk_scale_1_float_sse2_lowp
 _sk_scale_1_float_sse2_lowp LABEL PROC
   DB  85                                  ; push          %ebp
@@ -58639,15 +60696,15 @@
   DB  87                                  ; push          %edi
   DB  86                                  ; push          %esi
   DB  131,236,80                          ; sub           $0x50,%esp
-  DB  232,0,0,0,0                         ; call          c3e2 <_sk_scale_1_float_sse2_lowp+0xd>
+  DB  232,0,0,0,0                         ; call          cd03 <_sk_scale_1_float_sse2_lowp+0xd>
   DB  94                                  ; pop           %esi
   DB  139,69,8                            ; mov           0x8(%ebp),%eax
   DB  139,85,16                           ; mov           0x10(%ebp),%edx
   DB  139,77,12                           ; mov           0xc(%ebp),%ecx
   DB  139,57                              ; mov           (%ecx),%edi
   DB  243,15,16,39                        ; movss         (%edi),%xmm4
-  DB  243,15,89,166,34,20,0,0             ; mulss         0x1422(%esi),%xmm4
-  DB  243,15,88,166,30,20,0,0             ; addss         0x141e(%esi),%xmm4
+  DB  243,15,89,166,33,20,0,0             ; mulss         0x1421(%esi),%xmm4
+  DB  243,15,88,166,29,20,0,0             ; addss         0x141d(%esi),%xmm4
   DB  243,15,44,252                       ; cvttss2si     %xmm4,%edi
   DB  102,15,110,239                      ; movd          %edi,%xmm5
   DB  139,125,20                          ; mov           0x14(%ebp),%edi
@@ -58660,7 +60717,7 @@
   DB  102,15,213,205                      ; pmullw        %xmm5,%xmm1
   DB  102,15,213,213                      ; pmullw        %xmm5,%xmm2
   DB  102,15,213,221                      ; pmullw        %xmm5,%xmm3
-  DB  102,15,111,174,94,19,0,0            ; movdqa        0x135e(%esi),%xmm5
+  DB  102,15,111,174,93,19,0,0            ; movdqa        0x135d(%esi),%xmm5
   DB  102,15,253,197                      ; paddw         %xmm5,%xmm0
   DB  102,15,253,205                      ; paddw         %xmm5,%xmm1
   DB  102,15,253,213                      ; paddw         %xmm5,%xmm2
@@ -58697,19 +60754,19 @@
   DB  102,15,111,218                      ; movdqa        %xmm2,%xmm3
   DB  102,15,111,209                      ; movdqa        %xmm1,%xmm2
   DB  102,15,111,200                      ; movdqa        %xmm0,%xmm1
-  DB  232,0,0,0,0                         ; call          c4b0 <_sk_lerp_1_float_sse2_lowp+0x1d>
+  DB  232,0,0,0,0                         ; call          cdd1 <_sk_lerp_1_float_sse2_lowp+0x1d>
   DB  89                                  ; pop           %ecx
   DB  102,15,111,109,24                   ; movdqa        0x18(%ebp),%xmm5
   DB  139,69,12                           ; mov           0xc(%ebp),%eax
   DB  139,16                              ; mov           (%eax),%edx
   DB  243,15,16,50                        ; movss         (%edx),%xmm6
-  DB  243,15,89,177,84,19,0,0             ; mulss         0x1354(%ecx),%xmm6
-  DB  243,15,88,177,80,19,0,0             ; addss         0x1350(%ecx),%xmm6
+  DB  243,15,89,177,83,19,0,0             ; mulss         0x1353(%ecx),%xmm6
+  DB  243,15,88,177,79,19,0,0             ; addss         0x134f(%ecx),%xmm6
   DB  243,15,44,214                       ; cvttss2si     %xmm6,%edx
   DB  102,15,110,242                      ; movd          %edx,%xmm6
   DB  242,15,112,246,0                    ; pshuflw       $0x0,%xmm6,%xmm6
   DB  102,15,112,246,80                   ; pshufd        $0x50,%xmm6,%xmm6
-  DB  102,15,111,129,144,18,0,0           ; movdqa        0x1290(%ecx),%xmm0
+  DB  102,15,111,129,143,18,0,0           ; movdqa        0x128f(%ecx),%xmm0
   DB  102,15,213,206                      ; pmullw        %xmm6,%xmm1
   DB  102,15,253,200                      ; paddw         %xmm0,%xmm1
   DB  102,15,213,214                      ; pmullw        %xmm6,%xmm2
@@ -58780,21 +60837,21 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c5d9 <_sk_scale_u8_sse2_lowp+0x31>
+  DB  232,0,0,0,0                         ; call          cefa <_sk_scale_u8_sse2_lowp+0x31>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            c5fb <_sk_scale_u8_sse2_lowp+0x53>
+  DB  119,28                              ; ja            cf1c <_sk_scale_u8_sse2_lowp+0x53>
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  15,182,201                          ; movzbl        %cl,%ecx
-  DB  139,140,143,3,1,0,0                 ; mov           0x103(%edi,%ecx,4),%ecx
+  DB  139,140,143,2,1,0,0                 ; mov           0x102(%edi,%ecx,4),%ecx
   DB  1,249                               ; add           %edi,%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,51                        ; movzbl        (%ebx,%esi,1),%ecx
   DB  102,15,110,225                      ; movd          %ecx,%xmm4
-  DB  235,90                              ; jmp           c655 <_sk_scale_u8_sse2_lowp+0xad>
+  DB  235,90                              ; jmp           cf76 <_sk_scale_u8_sse2_lowp+0xad>
   DB  243,15,126,36,51                    ; movq          (%ebx,%esi,1),%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
-  DB  235,79                              ; jmp           c655 <_sk_scale_u8_sse2_lowp+0xad>
+  DB  235,79                              ; jmp           cf76 <_sk_scale_u8_sse2_lowp+0xad>
   DB  15,182,76,51,2                      ; movzbl        0x2(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,2                    ; pinsrw        $0x2,%ecx,%xmm4
@@ -58802,7 +60859,7 @@
   DB  102,15,110,233                      ; movd          %ecx,%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  243,15,16,229                       ; movss         %xmm5,%xmm4
-  DB  235,47                              ; jmp           c655 <_sk_scale_u8_sse2_lowp+0xad>
+  DB  235,47                              ; jmp           cf76 <_sk_scale_u8_sse2_lowp+0xad>
   DB  15,182,76,51,6                      ; movzbl        0x6(%ebx,%esi,1),%ecx
   DB  102,15,239,228                      ; pxor          %xmm4,%xmm4
   DB  102,15,196,225,6                    ; pinsrw        $0x6,%ecx,%xmm4
@@ -58813,9 +60870,9 @@
   DB  102,15,110,44,51                    ; movd          (%ebx,%esi,1),%xmm5
   DB  102,15,96,232                       ; punpcklbw     %xmm0,%xmm5
   DB  242,15,16,229                       ; movsd         %xmm5,%xmm4
-  DB  102,15,219,167,103,17,0,0           ; pand          0x1167(%edi),%xmm4
+  DB  102,15,219,167,102,17,0,0           ; pand          0x1166(%edi),%xmm4
   DB  102,15,213,196                      ; pmullw        %xmm4,%xmm0
-  DB  102,15,111,175,103,17,0,0           ; movdqa        0x1167(%edi),%xmm5
+  DB  102,15,111,175,102,17,0,0           ; movdqa        0x1166(%edi),%xmm5
   DB  102,15,253,197                      ; paddw         %xmm5,%xmm0
   DB  102,15,113,208,8                    ; psrlw         $0x8,%xmm0
   DB  102,15,213,204                      ; pmullw        %xmm4,%xmm1
@@ -58847,7 +60904,7 @@
   DB  91                                  ; pop           %ebx
   DB  93                                  ; pop           %ebp
   DB  195                                 ; ret
-  DB  15,31,0                             ; nopl          (%eax)
+  DB  102,144                             ; xchg          %ax,%ax
   DB  24,0                                ; sbb           %al,(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  59,0                                ; cmp           (%eax),%eax
@@ -58880,10 +60937,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c726 <_sk_lerp_u8_sse2_lowp+0x2e>
+  DB  232,0,0,0,0                         ; call          d046 <_sk_lerp_u8_sse2_lowp+0x2e>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            c748 <_sk_lerp_u8_sse2_lowp+0x50>
+  DB  119,28                              ; ja            d068 <_sk_lerp_u8_sse2_lowp+0x50>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,74,1,0,0                ; mov           0x14a(%edi,%ecx,4),%ecx
@@ -58891,10 +60948,10 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,182,12,51                        ; movzbl        (%ebx,%esi,1),%ecx
   DB  102,15,110,217                      ; movd          %ecx,%xmm3
-  DB  235,90                              ; jmp           c7a2 <_sk_lerp_u8_sse2_lowp+0xaa>
+  DB  235,90                              ; jmp           d0c2 <_sk_lerp_u8_sse2_lowp+0xaa>
   DB  243,15,126,28,51                    ; movq          (%ebx,%esi,1),%xmm3
   DB  102,15,96,216                       ; punpcklbw     %xmm0,%xmm3
-  DB  235,79                              ; jmp           c7a2 <_sk_lerp_u8_sse2_lowp+0xaa>
+  DB  235,79                              ; jmp           d0c2 <_sk_lerp_u8_sse2_lowp+0xaa>
   DB  15,182,76,51,2                      ; movzbl        0x2(%ebx,%esi,1),%ecx
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,217,2                    ; pinsrw        $0x2,%ecx,%xmm3
@@ -58902,7 +60959,7 @@
   DB  102,15,110,225                      ; movd          %ecx,%xmm4
   DB  102,15,96,224                       ; punpcklbw     %xmm0,%xmm4
   DB  243,15,16,220                       ; movss         %xmm4,%xmm3
-  DB  235,47                              ; jmp           c7a2 <_sk_lerp_u8_sse2_lowp+0xaa>
+  DB  235,47                              ; jmp           d0c2 <_sk_lerp_u8_sse2_lowp+0xaa>
   DB  15,182,76,51,6                      ; movzbl        0x6(%ebx,%esi,1),%ecx
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,217,6                    ; pinsrw        $0x6,%ecx,%xmm3
@@ -59001,10 +61058,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          c8cf <_sk_scale_565_sse2_lowp+0x43>
+  DB  232,0,0,0,0                         ; call          d1ef <_sk_scale_565_sse2_lowp+0x43>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            c8f1 <_sk_scale_565_sse2_lowp+0x65>
+  DB  119,28                              ; ja            d211 <_sk_scale_565_sse2_lowp+0x65>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,249,1,0,0               ; mov           0x1f9(%edi,%ecx,4),%ecx
@@ -59012,14 +61069,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,183,12,115                       ; movzwl        (%ebx,%esi,2),%ecx
   DB  102,15,110,217                      ; movd          %ecx,%xmm3
-  DB  235,59                              ; jmp           c92c <_sk_scale_565_sse2_lowp+0xa0>
+  DB  235,59                              ; jmp           d24c <_sk_scale_565_sse2_lowp+0xa0>
   DB  243,15,111,28,115                   ; movdqu        (%ebx,%esi,2),%xmm3
-  DB  235,52                              ; jmp           c92c <_sk_scale_565_sse2_lowp+0xa0>
+  DB  235,52                              ; jmp           d24c <_sk_scale_565_sse2_lowp+0xa0>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,92,115,4,2               ; pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm3
   DB  243,15,16,12,115                    ; movss         (%ebx,%esi,2),%xmm1
   DB  243,15,16,217                       ; movss         %xmm1,%xmm3
-  DB  235,30                              ; jmp           c92c <_sk_scale_565_sse2_lowp+0xa0>
+  DB  235,30                              ; jmp           d24c <_sk_scale_565_sse2_lowp+0xa0>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,92,115,12,6              ; pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm3
   DB  102,15,196,92,115,10,5              ; pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm3
@@ -59162,10 +61219,10 @@
   DB  128,225,7                           ; and           $0x7,%cl
   DB  254,201                             ; dec           %cl
   DB  128,249,6                           ; cmp           $0x6,%cl
-  DB  232,0,0,0,0                         ; call          cb27 <_sk_lerp_565_sse2_lowp+0x43>
+  DB  232,0,0,0,0                         ; call          d447 <_sk_lerp_565_sse2_lowp+0x43>
   DB  95                                  ; pop           %edi
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
-  DB  119,28                              ; ja            cb49 <_sk_lerp_565_sse2_lowp+0x65>
+  DB  119,28                              ; ja            d469 <_sk_lerp_565_sse2_lowp+0x65>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  15,182,201                          ; movzbl        %cl,%ecx
   DB  139,140,143,73,2,0,0                ; mov           0x249(%edi,%ecx,4),%ecx
@@ -59173,14 +61230,14 @@
   DB  255,225                             ; jmp           *%ecx
   DB  15,183,12,115                       ; movzwl        (%ebx,%esi,2),%ecx
   DB  102,15,110,217                      ; movd          %ecx,%xmm3
-  DB  235,59                              ; jmp           cb84 <_sk_lerp_565_sse2_lowp+0xa0>
+  DB  235,59                              ; jmp           d4a4 <_sk_lerp_565_sse2_lowp+0xa0>
   DB  243,15,111,28,115                   ; movdqu        (%ebx,%esi,2),%xmm3
-  DB  235,52                              ; jmp           cb84 <_sk_lerp_565_sse2_lowp+0xa0>
+  DB  235,52                              ; jmp           d4a4 <_sk_lerp_565_sse2_lowp+0xa0>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,92,115,4,2               ; pinsrw        $0x2,0x4(%ebx,%esi,2),%xmm3
   DB  243,15,16,12,115                    ; movss         (%ebx,%esi,2),%xmm1
   DB  243,15,16,217                       ; movss         %xmm1,%xmm3
-  DB  235,30                              ; jmp           cb84 <_sk_lerp_565_sse2_lowp+0xa0>
+  DB  235,30                              ; jmp           d4a4 <_sk_lerp_565_sse2_lowp+0xa0>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  102,15,196,92,115,12,6              ; pinsrw        $0x6,0xc(%ebx,%esi,2),%xmm3
   DB  102,15,196,92,115,10,5              ; pinsrw        $0x5,0xa(%ebx,%esi,2),%xmm3
@@ -59340,11 +61397,11 @@
   DB  128,227,7                           ; and           $0x7,%bl
   DB  254,203                             ; dec           %bl
   DB  128,251,6                           ; cmp           $0x6,%bl
-  DB  232,0,0,0,0                         ; call          cdcb <_sk_srcover_rgba_8888_sse2_lowp+0x3f>
+  DB  232,0,0,0,0                         ; call          d6eb <_sk_srcover_rgba_8888_sse2_lowp+0x3f>
   DB  89                                  ; pop           %ecx
   DB  139,117,16                          ; mov           0x10(%ebp),%esi
   DB  15,182,211                          ; movzbl        %bl,%edx
-  DB  119,29                              ; ja            cdf1 <_sk_srcover_rgba_8888_sse2_lowp+0x65>
+  DB  119,29                              ; ja            d711 <_sk_srcover_rgba_8888_sse2_lowp+0x65>
   DB  102,15,239,219                      ; pxor          %xmm3,%xmm3
   DB  139,132,145,77,2,0,0                ; mov           0x24d(%ecx,%edx,4),%eax
   DB  1,200                               ; add           %ecx,%eax
@@ -59352,14 +61409,14 @@
   DB  255,224                             ; jmp           *%eax
   DB  243,15,16,52,183                    ; movss         (%edi,%esi,4),%xmm6
   DB  15,87,219                           ; xorps         %xmm3,%xmm3
-  DB  235,73                              ; jmp           ce3a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  DB  235,73                              ; jmp           d75a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   DB  102,15,16,52,183                    ; movupd        (%edi,%esi,4),%xmm6
   DB  15,16,92,183,16                     ; movups        0x10(%edi,%esi,4),%xmm3
-  DB  235,61                              ; jmp           ce3a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  DB  235,61                              ; jmp           d75a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   DB  102,15,110,68,183,8                 ; movd          0x8(%edi,%esi,4),%xmm0
   DB  102,15,112,240,69                   ; pshufd        $0x45,%xmm0,%xmm6
   DB  102,15,18,52,183                    ; movlpd        (%edi,%esi,4),%xmm6
-  DB  235,43                              ; jmp           ce3a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
+  DB  235,43                              ; jmp           d75a <_sk_srcover_rgba_8888_sse2_lowp+0xae>
   DB  102,15,110,68,183,24                ; movd          0x18(%edi,%esi,4),%xmm0
   DB  102,15,112,216,69                   ; pshufd        $0x45,%xmm0,%xmm3
   DB  243,15,16,68,183,20                 ; movss         0x14(%edi,%esi,4),%xmm0
@@ -59437,18 +61494,18 @@
   DB  102,15,105,220                      ; punpckhwd     %xmm4,%xmm3
   DB  102,15,235,222                      ; por           %xmm6,%xmm3
   DB  128,251,6                           ; cmp           $0x6,%bl
-  DB  119,16                              ; ja            cf7b <_sk_srcover_rgba_8888_sse2_lowp+0x1ef>
+  DB  119,16                              ; ja            d89b <_sk_srcover_rgba_8888_sse2_lowp+0x1ef>
   DB  3,140,145,105,2,0,0                 ; add           0x269(%ecx,%edx,4),%ecx
   DB  255,225                             ; jmp           *%ecx
   DB  102,15,126,4,183                    ; movd          %xmm0,(%edi,%esi,4)
-  DB  235,64                              ; jmp           cfbb <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  DB  235,64                              ; jmp           d8db <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   DB  243,15,127,4,183                    ; movdqu        %xmm0,(%edi,%esi,4)
   DB  243,15,127,92,183,16                ; movdqu        %xmm3,0x10(%edi,%esi,4)
-  DB  235,51                              ; jmp           cfbb <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  DB  235,51                              ; jmp           d8db <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   DB  102,15,112,208,78                   ; pshufd        $0x4e,%xmm0,%xmm2
   DB  102,15,126,84,183,8                 ; movd          %xmm2,0x8(%edi,%esi,4)
   DB  102,15,214,4,183                    ; movq          %xmm0,(%edi,%esi,4)
-  DB  235,33                              ; jmp           cfbb <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
+  DB  235,33                              ; jmp           d8db <_sk_srcover_rgba_8888_sse2_lowp+0x22f>
   DB  102,15,112,211,78                   ; pshufd        $0x4e,%xmm3,%xmm2
   DB  102,15,126,84,183,24                ; movd          %xmm2,0x18(%edi,%esi,4)
   DB  102,15,112,211,229                  ; pshufd        $0xe5,%xmm3,%xmm2
@@ -59568,11 +61625,11 @@
   DB  0,128,191,0,0,128                   ; add           %al,-0x7fffff41(%eax)
   DB  191,0,0,224,64                      ; mov           $0x40e00000,%edi
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,64                              ; loopne        d128 <.literal16+0xd8>
+  DB  224,64                              ; loopne        da48 <.literal16+0xd8>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,64                              ; loopne        d12c <.literal16+0xdc>
+  DB  224,64                              ; loopne        da4c <.literal16+0xdc>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,64                              ; loopne        d130 <.literal16+0xe0>
+  DB  224,64                              ; loopne        da50 <.literal16+0xe0>
   DB  154,153,153,62,154,153,153          ; lcall         $0x9999,$0x9a3e9999
   DB  62,154,153,153,62,154,153,153       ; ds            lcall $0x9999,$0x9a3e9999
   DB  62,61,10,23,63,61                   ; ds            cmp $0x3d3f170a,%eax
@@ -59583,16 +61640,16 @@
   DB  63                                  ; aas
   DB  174                                 ; scas          %es:(%edi),%al
   DB  71                                  ; inc           %edi
-  DB  225,61                              ; loope         d151 <.literal16+0x101>
+  DB  225,61                              ; loope         da71 <.literal16+0x101>
   DB  174                                 ; scas          %es:(%edi),%al
   DB  71                                  ; inc           %edi
-  DB  225,61                              ; loope         d155 <.literal16+0x105>
+  DB  225,61                              ; loope         da75 <.literal16+0x105>
   DB  174                                 ; scas          %es:(%edi),%al
   DB  71                                  ; inc           %edi
-  DB  225,61                              ; loope         d159 <.literal16+0x109>
+  DB  225,61                              ; loope         da79 <.literal16+0x109>
   DB  174                                 ; scas          %es:(%edi),%al
   DB  71                                  ; inc           %edi
-  DB  225,61                              ; loope         d15d <.literal16+0x10d>
+  DB  225,61                              ; loope         da7d <.literal16+0x10d>
   DB  255,0                               ; incl          (%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  255,0                               ; incl          (%eax)
@@ -59602,13 +61659,13 @@
   DB  255,0                               ; incl          (%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  0,0                                 ; add           %al,(%eax)
-  DB  127,67                              ; jg            d177 <.literal16+0x127>
+  DB  127,67                              ; jg            da97 <.literal16+0x127>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  127,67                              ; jg            d17b <.literal16+0x12b>
+  DB  127,67                              ; jg            da9b <.literal16+0x12b>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  127,67                              ; jg            d17f <.literal16+0x12f>
+  DB  127,67                              ; jg            da9f <.literal16+0x12f>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  127,67                              ; jg            d183 <.literal16+0x133>
+  DB  127,67                              ; jg            daa3 <.literal16+0x133>
   DB  0,0                                 ; add           %al,(%eax)
   DB  128,127,0,0                         ; cmpb          $0x0,0x0(%edi)
   DB  128,127,0,0                         ; cmpb          $0x0,0x0(%edi)
@@ -59764,13 +61821,13 @@
   DB  132,55                              ; test          %dh,(%edi)
   DB  8,33                                ; or            %ah,(%ecx)
   DB  132,55                              ; test          %dh,(%edi)
-  DB  224,7                               ; loopne        d299 <.literal16+0x249>
+  DB  224,7                               ; loopne        dbb9 <.literal16+0x249>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,7                               ; loopne        d29d <.literal16+0x24d>
+  DB  224,7                               ; loopne        dbbd <.literal16+0x24d>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,7                               ; loopne        d2a1 <.literal16+0x251>
+  DB  224,7                               ; loopne        dbc1 <.literal16+0x251>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  224,7                               ; loopne        d2a5 <.literal16+0x255>
+  DB  224,7                               ; loopne        dbc5 <.literal16+0x255>
   DB  0,0                                 ; add           %al,(%eax)
   DB  33,8                                ; and           %ecx,(%eax)
   DB  2,58                                ; add           (%edx),%bh
@@ -59819,17 +61876,17 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  52,255                              ; xor           $0xff,%al
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d304 <.literal16+0x2b4>
+  DB  127,0                               ; jg            dc24 <.literal16+0x2b4>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d308 <.literal16+0x2b8>
+  DB  127,0                               ; jg            dc28 <.literal16+0x2b8>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d30c <.literal16+0x2bc>
+  DB  127,0                               ; jg            dc2c <.literal16+0x2bc>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d310 <.literal16+0x2c0>
-  DB  119,115                             ; ja            d385 <.literal16+0x335>
+  DB  127,0                               ; jg            dc30 <.literal16+0x2c0>
+  DB  119,115                             ; ja            dca5 <.literal16+0x335>
   DB  248                                 ; clc
   DB  194,119,115                         ; ret           $0x7377
   DB  248                                 ; clc
@@ -59840,7 +61897,7 @@
   DB  194,117,191                         ; ret           $0xbf75
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; aas
-  DB  117,191                             ; jne           d2e9 <.literal16+0x299>
+  DB  117,191                             ; jne           dc09 <.literal16+0x299>
   DB  191,63,117,191,191                  ; mov           $0xbfbf753f,%edi
   DB  63                                  ; aas
   DB  249                                 ; stc
@@ -59856,7 +61913,7 @@
   DB  68                                  ; inc           %esp
   DB  180,62                              ; mov           $0x3e,%ah
   DB  163,233,220,63,163                  ; mov           %eax,0xa33fdce9
-  DB  233,220,63,163,233                  ; jmp           e9a41326 <_sk_srcover_rgba_8888_sse2_lowp+0xe9a3459a>
+  DB  233,220,63,163,233                  ; jmp           e9a41c46 <_sk_srcover_rgba_8888_sse2_lowp+0xe9a3459a>
   DB  220,63                              ; fdivrl        (%edi)
   DB  163,233,220,63,81                   ; mov           %eax,0x513fdce9
   DB  140,242                             ; mov           %?,%edx
@@ -59952,16 +62009,16 @@
   DB  128,3,62                            ; addb          $0x3e,(%ebx)
   DB  31                                  ; pop           %ds
   DB  215                                 ; xlat          %ds:(%ebx)
-  DB  118,63                              ; jbe           d473 <.literal16+0x423>
+  DB  118,63                              ; jbe           dd93 <.literal16+0x423>
   DB  31                                  ; pop           %ds
   DB  215                                 ; xlat          %ds:(%ebx)
-  DB  118,63                              ; jbe           d477 <.literal16+0x427>
+  DB  118,63                              ; jbe           dd97 <.literal16+0x427>
   DB  31                                  ; pop           %ds
   DB  215                                 ; xlat          %ds:(%ebx)
-  DB  118,63                              ; jbe           d47b <.literal16+0x42b>
+  DB  118,63                              ; jbe           dd9b <.literal16+0x42b>
   DB  31                                  ; pop           %ds
   DB  215                                 ; xlat          %ds:(%ebx)
-  DB  118,63                              ; jbe           d47f <.literal16+0x42f>
+  DB  118,63                              ; jbe           dd9f <.literal16+0x42f>
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%eax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%eax)
   DB  246,64,83,63                        ; testb         $0x3f,0x53(%eax)
@@ -59987,13 +62044,13 @@
   DB  248                                 ; clc
   DB  65                                  ; inc           %ecx
   DB  0,0                                 ; add           %al,(%eax)
-  DB  124,66                              ; jl            d4b6 <.literal16+0x466>
+  DB  124,66                              ; jl            ddd6 <.literal16+0x466>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  124,66                              ; jl            d4ba <.literal16+0x46a>
+  DB  124,66                              ; jl            ddda <.literal16+0x46a>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  124,66                              ; jl            d4be <.literal16+0x46e>
+  DB  124,66                              ; jl            ddde <.literal16+0x46e>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  124,66                              ; jl            d4c2 <.literal16+0x472>
+  DB  124,66                              ; jl            dde2 <.literal16+0x472>
   DB  0,240                               ; add           %dh,%al
   DB  0,0                                 ; add           %al,(%eax)
   DB  0,240                               ; add           %dh,%al
@@ -60039,25 +62096,25 @@
   DB  136,136,61,137,136,136              ; mov           %cl,-0x777776c3(%eax)
   DB  61,137,136,136,61                   ; cmp           $0x3d888889,%eax
   DB  0,0                                 ; add           %al,(%eax)
-  DB  112,65                              ; jo            d545 <.literal16+0x4f5>
+  DB  112,65                              ; jo            de65 <.literal16+0x4f5>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  112,65                              ; jo            d549 <.literal16+0x4f9>
+  DB  112,65                              ; jo            de69 <.literal16+0x4f9>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  112,65                              ; jo            d54d <.literal16+0x4fd>
+  DB  112,65                              ; jo            de6d <.literal16+0x4fd>
   DB  0,0                                 ; add           %al,(%eax)
-  DB  112,65                              ; jo            d551 <.literal16+0x501>
+  DB  112,65                              ; jo            de71 <.literal16+0x501>
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  0,128,0,0,0,128                     ; add           %al,-0x80000000(%eax)
   DB  0,0                                 ; add           %al,(%eax)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d523 <.literal16+0x4d3>
+  DB  127,0                               ; jg            de43 <.literal16+0x4d3>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            d527 <.literal16+0x4d7>
+  DB  127,0                               ; jg            de47 <.literal16+0x4d7>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            d52b <.literal16+0x4db>
+  DB  127,0                               ; jg            de4b <.literal16+0x4db>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,0                               ; jg            d52f <.literal16+0x4df>
+  DB  127,0                               ; jg            de4f <.literal16+0x4df>
   DB  0,0                                 ; add           %al,(%eax)
   DB  4,0                                 ; add           $0x0,%al
   DB  0,0                                 ; add           %al,(%eax)
@@ -60083,13 +62140,13 @@
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,255                             ; jg            d568 <.literal16+0x518>
+  DB  127,255                             ; jg            de88 <.literal16+0x518>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,255                             ; jg            d56c <.literal16+0x51c>
+  DB  127,255                             ; jg            de8c <.literal16+0x51c>
   DB  255                                 ; (bad)
   DB  255                                 ; (bad)
-  DB  127,0                               ; jg            d571 <.literal16+0x521>
+  DB  127,0                               ; jg            de91 <.literal16+0x521>
   DB  0,128,56,0,0,128                    ; add           %al,-0x7fffffc8(%eax)
   DB  56,0                                ; cmp           %al,(%eax)
   DB  0,128,56,0,0,128                    ; add           %al,-0x7fffffc8(%eax)
@@ -60103,13 +62160,13 @@
   DB  0,255                               ; add           %bh,%bh
   DB  255,0                               ; incl          (%eax)
   DB  255                                 ; (bad)
-  DB  127,71                              ; jg            d5db <.literal16+0x58b>
+  DB  127,71                              ; jg            defb <.literal16+0x58b>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            d5df <.literal16+0x58f>
+  DB  127,71                              ; jg            deff <.literal16+0x58f>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            d5e3 <.literal16+0x593>
+  DB  127,71                              ; jg            df03 <.literal16+0x593>
   DB  0,255                               ; add           %bh,%bh
-  DB  127,71                              ; jg            d5e7 <.literal16+0x597>
+  DB  127,71                              ; jg            df07 <.literal16+0x597>
   DB  208                                 ; (bad)
   DB  179,89                              ; mov           $0x59,%bl
   DB  62,208                              ; ds            (bad)
@@ -60162,7 +62219,7 @@
   DB  192,45,16,17,192,45,16              ; shrb          $0x10,0x2dc01110
   DB  17,192                              ; adc           %eax,%eax
   DB  45,16,17,192,18                     ; sub           $0x12c01110,%eax
-  DB  120,57                              ; js            d64c <.literal16+0x5fc>
+  DB  120,57                              ; js            df6c <.literal16+0x5fc>
   DB  64                                  ; inc           %eax
   DB  18,120,57                           ; adc           0x39(%eax),%bh
   DB  64                                  ; inc           %eax
@@ -60233,13 +62290,13 @@
   DB  192,191,0,0,192,191,0               ; sarb          $0x0,-0x40400000(%edi)
   DB  0,192                               ; add           %al,%al
   DB  191,0,0,192,191                     ; mov           $0xbfc00000,%edi
-  DB  114,28                              ; jb            d6fe <.literal16+0x6ae>
+  DB  114,28                              ; jb            e01e <.literal16+0x6ae>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         d702 <.literal16+0x6b2>
+  DB  62,114,28                           ; jb,pt         e022 <.literal16+0x6b2>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         d706 <.literal16+0x6b6>
+  DB  62,114,28                           ; jb,pt         e026 <.literal16+0x6b6>
   DB  199                                 ; (bad)
-  DB  62,114,28                           ; jb,pt         d70a <.literal16+0x6ba>
+  DB  62,114,28                           ; jb,pt         e02a <.literal16+0x6ba>
   DB  199                                 ; (bad)
   DB  62,85                               ; ds            push %ebp
   DB  85                                  ; push          %ebp
@@ -60260,13 +62317,13 @@
   DB  57,142,99,61,57,142                 ; cmp           %ecx,-0x71c6c29d(%esi)
   DB  99,61,57,142,99,61                  ; arpl          %di,0x3d638e39
   DB  57,142,99,61,114,249                ; cmp           %ecx,-0x68dc29d(%esi)
-  DB  127,63                              ; jg            d763 <.literal16+0x713>
-  DB  114,249                             ; jb            d71f <.literal16+0x6cf>
-  DB  127,63                              ; jg            d767 <.literal16+0x717>
-  DB  114,249                             ; jb            d723 <.literal16+0x6d3>
-  DB  127,63                              ; jg            d76b <.literal16+0x71b>
-  DB  114,249                             ; jb            d727 <.literal16+0x6d7>
-  DB  127,63                              ; jg            d76f <.literal16+0x71f>
+  DB  127,63                              ; jg            e083 <.literal16+0x713>
+  DB  114,249                             ; jb            e03f <.literal16+0x6cf>
+  DB  127,63                              ; jg            e087 <.literal16+0x717>
+  DB  114,249                             ; jb            e043 <.literal16+0x6d3>
+  DB  127,63                              ; jg            e08b <.literal16+0x71b>
+  DB  114,249                             ; jb            e047 <.literal16+0x6d7>
+  DB  127,63                              ; jg            e08f <.literal16+0x71f>
   DB  3,0                                 ; add           (%eax),%eax
   DB  0,0                                 ; add           %al,(%eax)
   DB  3,0                                 ; add           (%eax),%eax
@@ -60334,14 +62391,14 @@
   DB  0,248                               ; add           %bh,%al
   DB  0,248                               ; add           %bh,%al
   DB  0,248                               ; add           %bh,%al
-  DB  224,255                             ; loopne        d7c1 <.literal16+0x771>
-  DB  224,255                             ; loopne        d7c3 <.literal16+0x773>
-  DB  224,255                             ; loopne        d7c5 <.literal16+0x775>
-  DB  224,255                             ; loopne        d7c7 <.literal16+0x777>
-  DB  224,255                             ; loopne        d7c9 <.literal16+0x779>
-  DB  224,255                             ; loopne        d7cb <.literal16+0x77b>
-  DB  224,255                             ; loopne        d7cd <.literal16+0x77d>
-  DB  224,255                             ; loopne        d7cf <.literal16+0x77f>
+  DB  224,255                             ; loopne        e0e1 <.literal16+0x771>
+  DB  224,255                             ; loopne        e0e3 <.literal16+0x773>
+  DB  224,255                             ; loopne        e0e5 <.literal16+0x775>
+  DB  224,255                             ; loopne        e0e7 <.literal16+0x777>
+  DB  224,255                             ; loopne        e0e9 <.literal16+0x779>
+  DB  224,255                             ; loopne        e0eb <.literal16+0x77b>
+  DB  224,255                             ; loopne        e0ed <.literal16+0x77d>
+  DB  224,255                             ; loopne        e0ef <.literal16+0x77f>
   DB  54,0,54                             ; add           %dh,%ss:(%esi)
   DB  0,54                                ; add           %dh,(%esi)
   DB  0,54                                ; add           %dh,(%esi)
@@ -60365,6 +62422,6 @@
   DB  0,0                                 ; add           %al,(%eax)
   DB  0,63                                ; add           %bh,(%edi)
   DB  0,0                                 ; add           %al,(%eax)
-  DB  127,67                              ; jg            d84b <_sk_srcover_rgba_8888_sse2_lowp+0xabf>
+  DB  127,67                              ; jg            e16b <_sk_srcover_rgba_8888_sse2_lowp+0xabf>
 ENDIF
 END
diff --git a/src/jumper/SkJumper_stages_lowp.cpp b/src/jumper/SkJumper_stages_lowp.cpp
index 247bdf6..aa04693 100644
--- a/src/jumper/SkJumper_stages_lowp.cpp
+++ b/src/jumper/SkJumper_stages_lowp.cpp
@@ -181,6 +181,14 @@
     memcpy((char*)&v + 1*sizeof(S), &hi, sizeof(S));
     return v;
 }
+template <typename V, typename H>
+SI V map(V v, H (*fn)(H)) {
+    H lo,hi;
+    split(v, &lo,&hi);
+    lo = fn(lo);
+    hi = fn(hi);
+    return join<V>(lo,hi);
+}
 
 // TODO: do we need platform-specific intrinsics for any of these?
 SI F if_then_else(I32 c, F t, F e) {
@@ -192,6 +200,21 @@
 SI F mad(F f, F m, F a) { return f*m+a; }
 SI U32 trunc_(F x) { return (U32)cast<I32>(x); }
 
+SI F rcp(F x) {
+#if defined(__AVX2__)
+    return map(x, _mm256_rcp_ps);
+#elif defined(__SSE__)
+    return map(x, _mm_rcp_ps);
+#elif defined(__ARM_NEON)
+    return map(x, +[](float32x4_t v) {
+        auto est = vrecpeq_f32(v);
+        return vrecpsq_f32(v,est)*est;
+    });
+#else
+    return 1.0f / x;
+#endif
+}
+
 // ~~~~~~ Basic / misc. stages ~~~~~~ //
 
 STAGE_GG(seed_shader, const float* iota) {
@@ -199,12 +222,28 @@
     y = cast<F>(I32(dy)) + 0.5f;
 }
 
+STAGE_GG(matrix_translate, const float* m) {
+    x += m[0];
+    y += m[1];
+}
+STAGE_GG(matrix_scale_translate, const float* m) {
+    x = mad(x,m[0], m[2]);
+    y = mad(y,m[1], m[3]);
+}
 STAGE_GG(matrix_2x3, const float* m) {
     auto X = mad(x,m[0], mad(y,m[2], m[4])),
          Y = mad(x,m[1], mad(y,m[3], m[5]));
     x = X;
     y = Y;
 }
+STAGE_GG(matrix_perspective, const float* m) {
+    // N.B. Unlike the other matrix_ stages, this matrix is row-major.
+    auto X = mad(x,m[0], mad(y,m[1], m[2])),
+         Y = mad(x,m[3], mad(y,m[4], m[5])),
+         Z = mad(x,m[6], mad(y,m[7], m[8]));
+    x = X * rcp(Z);
+    y = Y * rcp(Z);
+}
 
 STAGE_PP(uniform_color, const SkJumper_UniformColorCtx* c) {
     r = c->rgba[0];
@@ -490,12 +529,16 @@
     U32 ix = ix_and_ptr(&ptr, ctx, x,y);
     from_8888(gather<U32>(ptr, ix), &r, &g, &b, &a);
 }
+STAGE_GP(gather_bgra, const SkJumper_GatherCtx* ctx) {
+    const uint32_t* ptr;
+    U32 ix = ix_and_ptr(&ptr, ctx, x,y);
+    from_8888(gather<U32>(ptr, ix), &b, &g, &r, &a);
+}
 
 // ~~~~~~ 16-bit memory loads and stores ~~~~~~ //
 
-SI void load_565(const uint16_t* ptr, size_t tail, U16* r, U16* g, U16* b) {
+SI void from_565(U16 rgb, U16* r, U16* g, U16* b) {
     // Format for 565 buffers: 15|rrrrr gggggg bbbbb|0
-    U16 rgb = load<U16>(ptr, tail);
     U16 R = (rgb >> 11) & 31,
         G = (rgb >>  5) & 63,
         B = (rgb >>  0) & 31;
@@ -505,6 +548,9 @@
     *g = (G << 2) | (G >> 4);
     *b = (B << 3) | (B >> 2);
 }
+SI void load_565(const uint16_t* ptr, size_t tail, U16* r, U16* g, U16* b) {
+    from_565(load<U16>(ptr, tail), r,g,b);
+}
 SI void store_565(uint16_t* ptr, size_t tail, U16 r, U16 g, U16 b) {
     // Select the top 5,6,5 bits.
     U16 R = r >> 3,
@@ -527,6 +573,12 @@
 STAGE_PP(store_565, const SkJumper_MemoryCtx* ctx) {
     store_565(ptr_at_xy<uint16_t>(ctx, dx,dy), tail, r,g,b);
 }
+STAGE_GP(gather_565, const SkJumper_GatherCtx* ctx) {
+    const uint16_t* ptr;
+    U32 ix = ix_and_ptr(&ptr, ctx, x,y);
+    from_565(gather<U16>(ptr, ix), &r, &g, &b);
+    a = 255;
+}
 
 // ~~~~~~ 8-bit memory loads and stores ~~~~~~ //
 
@@ -548,6 +600,12 @@
 STAGE_PP(store_a8, const SkJumper_MemoryCtx* ctx) {
     store_8(ptr_at_xy<uint8_t>(ctx, dx,dy), tail, a);
 }
+STAGE_GP(gather_a8, const SkJumper_GatherCtx* ctx) {
+    const uint8_t* ptr;
+    U32 ix = ix_and_ptr(&ptr, ctx, x,y);
+    r = g = b = 0;
+    a = cast<U16>(gather<U8>(ptr, ix));
+}
 
 STAGE_PP(load_g8, const SkJumper_MemoryCtx* ctx) {
     r = g = b = load_8(ptr_at_xy<const uint8_t>(ctx, dx,dy), tail);
@@ -561,6 +619,12 @@
     a = (r*54 + g*183 + b*19)/256;  // 0.2126, 0.7152, 0.0722 with 256 denominator.
     r = g = b = 0;
 }
+STAGE_GP(gather_g8, const SkJumper_GatherCtx* ctx) {
+    const uint8_t* ptr;
+    U32 ix = ix_and_ptr(&ptr, ctx, x,y);
+    r = g = b = cast<U16>(gather<U8>(ptr, ix));
+    a = 255;
+}
 
 // ~~~~~~ Coverage scales / lerps ~~~~~~ //
 
diff --git a/tests/SkRasterPipelineTest.cpp b/tests/SkRasterPipelineTest.cpp
index 773c0d5..1923f1c 100644
--- a/tests/SkRasterPipelineTest.cpp
+++ b/tests/SkRasterPipelineTest.cpp
@@ -232,33 +232,3 @@
         }
     }
 }
-
-DEF_TEST(SkRasterPipeline_2d, r) {
-    uint32_t rgba[2*2] = {0,0,0,0};
-
-    SkSTArenaAlloc<256> alloc;
-    SkRasterPipeline p(&alloc);
-
-    // Splat out the (2d) dst coordinates: (0.5,0.5), (1.5,0.5), (0.5,1.5), (1.5,1.5).
-    p.append_seed_shader();
-
-    // Scale down to [0,1] range to write out as bytes.
-    p.append_matrix(&alloc, SkMatrix::Concat(SkMatrix::MakeScale(0.5f),
-                                             SkMatrix::MakeTrans(-0.5f, -0.5f)));
-
-    // Write out to rgba, with row stride = 2 pixels.
-    SkJumper_MemoryCtx ctx = { rgba, 2 };
-    p.append(SkRasterPipeline::store_8888, &ctx);
-
-    p.run(0,0, 2,2);
-
-    REPORTER_ASSERT(r, ((rgba[0] >> 0) & 0xff) ==   0);
-    REPORTER_ASSERT(r, ((rgba[1] >> 0) & 0xff) == 128);
-    REPORTER_ASSERT(r, ((rgba[2] >> 0) & 0xff) ==   0);
-    REPORTER_ASSERT(r, ((rgba[3] >> 0) & 0xff) == 128);
-
-    REPORTER_ASSERT(r, ((rgba[0] >> 8) & 0xff) ==   0);
-    REPORTER_ASSERT(r, ((rgba[1] >> 8) & 0xff) ==   0);
-    REPORTER_ASSERT(r, ((rgba[2] >> 8) & 0xff) == 128);
-    REPORTER_ASSERT(r, ((rgba[3] >> 8) & 0xff) == 128);
-}