| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1 | # Copyright 2017 Google Inc. |
| 2 | # |
| 3 | # Use of this source code is governed by a BSD-style license that can be |
| 4 | # found in the LICENSE file. |
| 5 | |
| 6 | # This file is generated semi-automatically with this command: |
| 7 | # $ src/jumper/build_stages.py |
| 8 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9 | #if defined(__MACH__) |
| 10 | #define HIDDEN .private_extern |
| 11 | #else |
| 12 | #define HIDDEN .hidden |
| Mike Klein | 8517b4c | 2017-03-30 20:07:52 -0400 | [diff] [blame] | 13 | .section .note.GNU-stack,"",%progbits |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 14 | #endif |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 15 | .text |
| 16 | #if defined(__aarch64__) |
| 17 | .balign 4 |
| 18 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 19 | HIDDEN _sk_start_pipeline_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 20 | .globl _sk_start_pipeline_aarch64 |
| 21 | _sk_start_pipeline_aarch64: |
| 22 | .long 0xa9bd5bf7 // stp x23, x22, [sp, #-48]! |
| 23 | .long 0xa90153f5 // stp x21, x20, [sp, #16] |
| 24 | .long 0xa9027bf3 // stp x19, x30, [sp, #32] |
| 25 | .long 0xaa0103f4 // mov x20, x1 |
| 26 | .long 0xf8408697 // ldr x23, [x20], #8 |
| 27 | .long 0xaa0003f5 // mov x21, x0 |
| 28 | .long 0xaa0303f3 // mov x19, x3 |
| 29 | .long 0x910012a8 // add x8, x21, #0x4 |
| 30 | .long 0xeb13011f // cmp x8, x19 |
| 31 | .long 0xaa0203f6 // mov x22, x2 |
| 32 | .long 0x54000069 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast |
| 33 | .long 0xaa1503e0 // mov x0, x21 |
| 34 | .long 0x14000012 // b 78 <sk_start_pipeline_aarch64+0x78> |
| 35 | .long 0x6f00e400 // movi v0.2d, #0x0 |
| 36 | .long 0x6f00e401 // movi v1.2d, #0x0 |
| 37 | .long 0x6f00e402 // movi v2.2d, #0x0 |
| 38 | .long 0x6f00e403 // movi v3.2d, #0x0 |
| 39 | .long 0x6f00e404 // movi v4.2d, #0x0 |
| 40 | .long 0x6f00e405 // movi v5.2d, #0x0 |
| 41 | .long 0x6f00e406 // movi v6.2d, #0x0 |
| 42 | .long 0x6f00e407 // movi v7.2d, #0x0 |
| 43 | .long 0xaa1503e0 // mov x0, x21 |
| 44 | .long 0xaa1403e1 // mov x1, x20 |
| 45 | .long 0xaa1603e2 // mov x2, x22 |
| 46 | .long 0xd63f02e0 // blr x23 |
| 47 | .long 0x910012a0 // add x0, x21, #0x4 |
| 48 | .long 0x910022a8 // add x8, x21, #0x8 |
| 49 | .long 0xeb13011f // cmp x8, x19 |
| 50 | .long 0xaa0003f5 // mov x21, x0 |
| 51 | .long 0x54fffe09 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast |
| 52 | .long 0xa9427bf3 // ldp x19, x30, [sp, #32] |
| 53 | .long 0xa94153f5 // ldp x21, x20, [sp, #16] |
| 54 | .long 0xa8c35bf7 // ldp x23, x22, [sp], #48 |
| 55 | .long 0xd65f03c0 // ret |
| 56 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 57 | HIDDEN _sk_just_return_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 58 | .globl _sk_just_return_aarch64 |
| 59 | _sk_just_return_aarch64: |
| 60 | .long 0xd65f03c0 // ret |
| 61 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 62 | HIDDEN _sk_seed_shader_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 63 | .globl _sk_seed_shader_aarch64 |
| 64 | _sk_seed_shader_aarch64: |
| 65 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 66 | .long 0x3dc00046 // ldr q6, [x2] |
| 67 | .long 0x4e040c00 // dup v0.4s, w0 |
| 68 | .long 0x4f0167e7 // movi v7.4s, #0x3f, lsl #24 |
| 69 | .long 0x4d40c901 // ld1r {v1.4s}, [x8] |
| 70 | .long 0x4e21d800 // scvtf v0.4s, v0.4s |
| 71 | .long 0x4e27d400 // fadd v0.4s, v0.4s, v7.4s |
| 72 | .long 0x4f03f602 // fmov v2.4s, #1.000000000000000000e+00 |
| 73 | .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| 74 | .long 0x6f00e403 // movi v3.2d, #0x0 |
| 75 | .long 0x6f00e404 // movi v4.2d, #0x0 |
| 76 | .long 0x6f00e405 // movi v5.2d, #0x0 |
| 77 | .long 0x4e26d400 // fadd v0.4s, v0.4s, v6.4s |
| 78 | .long 0x6f00e406 // movi v6.2d, #0x0 |
| 79 | .long 0x4e27d421 // fadd v1.4s, v1.4s, v7.4s |
| 80 | .long 0x6f00e407 // movi v7.2d, #0x0 |
| 81 | .long 0xd61f0060 // br x3 |
| 82 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 83 | HIDDEN _sk_constant_color_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 84 | .globl _sk_constant_color_aarch64 |
| 85 | _sk_constant_color_aarch64: |
| 86 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 87 | .long 0x3dc00103 // ldr q3, [x8] |
| 88 | .long 0x4e040460 // dup v0.4s, v3.s[0] |
| 89 | .long 0x4e0c0461 // dup v1.4s, v3.s[1] |
| 90 | .long 0x4e140462 // dup v2.4s, v3.s[2] |
| 91 | .long 0x4e1c0463 // dup v3.4s, v3.s[3] |
| 92 | .long 0xd61f0060 // br x3 |
| 93 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 94 | HIDDEN _sk_clear_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 95 | .globl _sk_clear_aarch64 |
| 96 | _sk_clear_aarch64: |
| 97 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 98 | .long 0x6f00e400 // movi v0.2d, #0x0 |
| 99 | .long 0x6f00e401 // movi v1.2d, #0x0 |
| 100 | .long 0x6f00e402 // movi v2.2d, #0x0 |
| 101 | .long 0x6f00e403 // movi v3.2d, #0x0 |
| 102 | .long 0xd61f0060 // br x3 |
| 103 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 104 | HIDDEN _sk_srcatop_aarch64 |
| 105 | .globl _sk_srcatop_aarch64 |
| 106 | _sk_srcatop_aarch64: |
| 107 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 108 | .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s |
| 109 | .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s |
| 110 | .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s |
| 111 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 112 | .long 0xf8408423 // ldr x3, [x1], #8 |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 113 | .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| 114 | .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| 115 | .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s |
| 116 | .long 0x6e27de10 // fmul v16.4s, v16.4s, v7.4s |
| 117 | .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s |
| 118 | .long 0x4eb01e03 // mov v3.16b, v16.16b |
| 119 | .long 0xd61f0060 // br x3 |
| 120 | |
| 121 | HIDDEN _sk_dstatop_aarch64 |
| 122 | .globl _sk_dstatop_aarch64 |
| 123 | _sk_dstatop_aarch64: |
| 124 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 125 | .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s |
| 126 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 127 | .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| 128 | .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| 129 | .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| 130 | .long 0x6e23de10 // fmul v16.4s, v16.4s, v3.4s |
| 131 | .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s |
| 132 | .long 0x4e23cc80 // fmla v0.4s, v4.4s, v3.4s |
| 133 | .long 0x4e23cca1 // fmla v1.4s, v5.4s, v3.4s |
| 134 | .long 0x4e23ccc2 // fmla v2.4s, v6.4s, v3.4s |
| 135 | .long 0x4eb01e03 // mov v3.16b, v16.16b |
| 136 | .long 0xd61f0060 // br x3 |
| 137 | |
| 138 | HIDDEN _sk_srcin_aarch64 |
| 139 | .globl _sk_srcin_aarch64 |
| 140 | _sk_srcin_aarch64: |
| 141 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 142 | .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s |
| 143 | .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s |
| 144 | .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s |
| 145 | .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| 146 | .long 0xd61f0060 // br x3 |
| 147 | |
| 148 | HIDDEN _sk_dstin_aarch64 |
| 149 | .globl _sk_dstin_aarch64 |
| 150 | _sk_dstin_aarch64: |
| 151 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 152 | .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s |
| 153 | .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s |
| 154 | .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s |
| 155 | .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| 156 | .long 0xd61f0060 // br x3 |
| 157 | |
| 158 | HIDDEN _sk_srcout_aarch64 |
| 159 | .globl _sk_srcout_aarch64 |
| 160 | _sk_srcout_aarch64: |
| 161 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 162 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 163 | .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s |
| 164 | .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| 165 | .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| 166 | .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| 167 | .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| 168 | .long 0xd61f0060 // br x3 |
| 169 | |
| 170 | HIDDEN _sk_dstout_aarch64 |
| 171 | .globl _sk_dstout_aarch64 |
| 172 | _sk_dstout_aarch64: |
| 173 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 174 | .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00 |
| 175 | .long 0x4ea3d403 // fsub v3.4s, v0.4s, v3.4s |
| 176 | .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s |
| 177 | .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s |
| 178 | .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s |
| 179 | .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 180 | .long 0xd61f0060 // br x3 |
| 181 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 182 | HIDDEN _sk_srcover_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 183 | .globl _sk_srcover_aarch64 |
| 184 | _sk_srcover_aarch64: |
| 185 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 186 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 187 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 188 | .long 0x4e24ce00 // fmla v0.4s, v16.4s, v4.4s |
| 189 | .long 0x4e25ce01 // fmla v1.4s, v16.4s, v5.4s |
| 190 | .long 0x4e26ce02 // fmla v2.4s, v16.4s, v6.4s |
| 191 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 192 | .long 0xd61f0060 // br x3 |
| 193 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 194 | HIDDEN _sk_dstover_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 195 | .globl _sk_dstover_aarch64 |
| 196 | _sk_dstover_aarch64: |
| 197 | .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| 198 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 199 | .long 0x4ea41c90 // mov v16.16b, v4.16b |
| 200 | .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s |
| 201 | .long 0x4ea51cb1 // mov v17.16b, v5.16b |
| 202 | .long 0x4ea61cd2 // mov v18.16b, v6.16b |
| 203 | .long 0x4ea71cf3 // mov v19.16b, v7.16b |
| 204 | .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s |
| 205 | .long 0x4e21ce91 // fmla v17.4s, v20.4s, v1.4s |
| 206 | .long 0x4e22ce92 // fmla v18.4s, v20.4s, v2.4s |
| 207 | .long 0x4e23ce93 // fmla v19.4s, v20.4s, v3.4s |
| 208 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 209 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 210 | .long 0x4eb21e42 // mov v2.16b, v18.16b |
| 211 | .long 0x4eb31e63 // mov v3.16b, v19.16b |
| 212 | .long 0xd61f0060 // br x3 |
| 213 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 214 | HIDDEN _sk_modulate_aarch64 |
| 215 | .globl _sk_modulate_aarch64 |
| 216 | _sk_modulate_aarch64: |
| 217 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 218 | .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s |
| 219 | .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s |
| 220 | .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| 221 | .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s |
| 222 | .long 0xd61f0060 // br x3 |
| 223 | |
| 224 | HIDDEN _sk_multiply_aarch64 |
| 225 | .globl _sk_multiply_aarch64 |
| 226 | _sk_multiply_aarch64: |
| 227 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 228 | .long 0x4ea7d613 // fsub v19.4s, v16.4s, v7.4s |
| 229 | .long 0x4ea3d614 // fsub v20.4s, v16.4s, v3.4s |
| 230 | .long 0x6e20de70 // fmul v16.4s, v19.4s, v0.4s |
| 231 | .long 0x6e21de71 // fmul v17.4s, v19.4s, v1.4s |
| 232 | .long 0x6e22de72 // fmul v18.4s, v19.4s, v2.4s |
| 233 | .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| 234 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 235 | .long 0x4e34cc90 // fmla v16.4s, v4.4s, v20.4s |
| 236 | .long 0x4e34ccb1 // fmla v17.4s, v5.4s, v20.4s |
| 237 | .long 0x4e34ccd2 // fmla v18.4s, v6.4s, v20.4s |
| 238 | .long 0x4e34ccf3 // fmla v19.4s, v7.4s, v20.4s |
| 239 | .long 0x4e20cc90 // fmla v16.4s, v4.4s, v0.4s |
| 240 | .long 0x4e21ccb1 // fmla v17.4s, v5.4s, v1.4s |
| 241 | .long 0x4e22ccd2 // fmla v18.4s, v6.4s, v2.4s |
| 242 | .long 0x4e23ccf3 // fmla v19.4s, v7.4s, v3.4s |
| 243 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 244 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 245 | .long 0x4eb21e42 // mov v2.16b, v18.16b |
| 246 | .long 0x4eb31e63 // mov v3.16b, v19.16b |
| 247 | .long 0xd61f0060 // br x3 |
| 248 | |
| 249 | HIDDEN _sk_plus__aarch64 |
| 250 | .globl _sk_plus__aarch64 |
| 251 | _sk_plus__aarch64: |
| 252 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 253 | .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| 254 | .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| 255 | .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| 256 | .long 0x4e27d463 // fadd v3.4s, v3.4s, v7.4s |
| 257 | .long 0xd61f0060 // br x3 |
| 258 | |
| 259 | HIDDEN _sk_screen_aarch64 |
| 260 | .globl _sk_screen_aarch64 |
| 261 | _sk_screen_aarch64: |
| 262 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 263 | .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s |
| 264 | .long 0x4e25d431 // fadd v17.4s, v1.4s, v5.4s |
| 265 | .long 0x4e26d452 // fadd v18.4s, v2.4s, v6.4s |
| 266 | .long 0x4e27d473 // fadd v19.4s, v3.4s, v7.4s |
| 267 | .long 0x4ea4cc10 // fmls v16.4s, v0.4s, v4.4s |
| 268 | .long 0x4ea5cc31 // fmls v17.4s, v1.4s, v5.4s |
| 269 | .long 0x4ea6cc52 // fmls v18.4s, v2.4s, v6.4s |
| 270 | .long 0x4ea7cc73 // fmls v19.4s, v3.4s, v7.4s |
| 271 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 272 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 273 | .long 0x4eb21e42 // mov v2.16b, v18.16b |
| 274 | .long 0x4eb31e63 // mov v3.16b, v19.16b |
| 275 | .long 0xd61f0060 // br x3 |
| 276 | |
| 277 | HIDDEN _sk_xor__aarch64 |
| 278 | .globl _sk_xor__aarch64 |
| 279 | _sk_xor__aarch64: |
| 280 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 281 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 282 | .long 0x4ea7d611 // fsub v17.4s, v16.4s, v7.4s |
| 283 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 284 | .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s |
| 285 | .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s |
| 286 | .long 0x6e22de22 // fmul v2.4s, v17.4s, v2.4s |
| 287 | .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s |
| 288 | .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| 289 | .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| 290 | .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s |
| 291 | .long 0x4e30cce3 // fmla v3.4s, v7.4s, v16.4s |
| 292 | .long 0xd61f0060 // br x3 |
| 293 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 294 | HIDDEN _sk_darken_aarch64 |
| 295 | .globl _sk_darken_aarch64 |
| 296 | _sk_darken_aarch64: |
| 297 | .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| 298 | .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| 299 | .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| 300 | .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| 301 | .long 0x4e31f610 // fmax v16.4s, v16.4s, v17.4s |
| 302 | .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| 303 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 304 | .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| 305 | .long 0x4e33f652 // fmax v18.4s, v18.4s, v19.4s |
| 306 | .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| 307 | .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| 308 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 309 | .long 0x4e33f631 // fmax v17.4s, v17.4s, v19.4s |
| 310 | .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| 311 | .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| 312 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 313 | .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s |
| 314 | .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| 315 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 316 | .long 0xd61f0060 // br x3 |
| 317 | |
| 318 | HIDDEN _sk_lighten_aarch64 |
| 319 | .globl _sk_lighten_aarch64 |
| 320 | _sk_lighten_aarch64: |
| 321 | .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| 322 | .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| 323 | .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| 324 | .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| 325 | .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s |
| 326 | .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| 327 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 328 | .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| 329 | .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s |
| 330 | .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| 331 | .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| 332 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 333 | .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s |
| 334 | .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| 335 | .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| 336 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 337 | .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s |
| 338 | .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| 339 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 340 | .long 0xd61f0060 // br x3 |
| 341 | |
| 342 | HIDDEN _sk_difference_aarch64 |
| 343 | .globl _sk_difference_aarch64 |
| 344 | _sk_difference_aarch64: |
| 345 | .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s |
| 346 | .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| 347 | .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s |
| 348 | .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s |
| 349 | .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s |
| 350 | .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s |
| 351 | .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s |
| 352 | .long 0x4e30d610 // fadd v16.4s, v16.4s, v16.4s |
| 353 | .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s |
| 354 | .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s |
| 355 | .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s |
| 356 | .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s |
| 357 | .long 0x4e32d650 // fadd v16.4s, v18.4s, v18.4s |
| 358 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 359 | .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s |
| 360 | .long 0x4eb0d421 // fsub v1.4s, v1.4s, v16.4s |
| 361 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 362 | .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s |
| 363 | .long 0x4e31d631 // fadd v17.4s, v17.4s, v17.4s |
| 364 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 365 | .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s |
| 366 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 367 | .long 0xd61f0060 // br x3 |
| 368 | |
| 369 | HIDDEN _sk_exclusion_aarch64 |
| 370 | .globl _sk_exclusion_aarch64 |
| 371 | _sk_exclusion_aarch64: |
| 372 | .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s |
| 373 | .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s |
| 374 | .long 0x4e20d400 // fadd v0.4s, v0.4s, v0.4s |
| 375 | .long 0x4ea0d600 // fsub v0.4s, v16.4s, v0.4s |
| 376 | .long 0x4e25d430 // fadd v16.4s, v1.4s, v5.4s |
| 377 | .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s |
| 378 | .long 0x4e21d421 // fadd v1.4s, v1.4s, v1.4s |
| 379 | .long 0x4ea1d601 // fsub v1.4s, v16.4s, v1.4s |
| 380 | .long 0x4e26d450 // fadd v16.4s, v2.4s, v6.4s |
| 381 | .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| 382 | .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| 383 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 384 | .long 0x4ea2d602 // fsub v2.4s, v16.4s, v2.4s |
| 385 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 386 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 387 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 388 | .long 0xd61f0060 // br x3 |
| 389 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 390 | HIDDEN _sk_colorburn_aarch64 |
| 391 | .globl _sk_colorburn_aarch64 |
| 392 | _sk_colorburn_aarch64: |
| 393 | .long 0x4ea4d4f3 // fsub v19.4s, v7.4s, v4.4s |
| 394 | .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| 395 | .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| 396 | .long 0x6e20fe73 // fdiv v19.4s, v19.4s, v0.4s |
| 397 | .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s |
| 398 | .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s |
| 399 | .long 0x6e20de95 // fmul v21.4s, v20.4s, v0.4s |
| 400 | .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s |
| 401 | .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s |
| 402 | .long 0x4e33cc75 // fmla v21.4s, v3.4s, v19.4s |
| 403 | .long 0x4ea5d4f3 // fsub v19.4s, v7.4s, v5.4s |
| 404 | .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s |
| 405 | .long 0x6e21fe73 // fdiv v19.4s, v19.4s, v1.4s |
| 406 | .long 0x4ea0d812 // fcmeq v18.4s, v0.4s, #0.0 |
| 407 | .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s |
| 408 | .long 0x6e751c12 // bsl v18.16b, v0.16b, v21.16b |
| 409 | .long 0x6e21de80 // fmul v0.4s, v20.4s, v1.4s |
| 410 | .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s |
| 411 | .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s |
| 412 | .long 0x4e33cc60 // fmla v0.4s, v3.4s, v19.4s |
| 413 | .long 0x4ea0d833 // fcmeq v19.4s, v1.4s, #0.0 |
| 414 | .long 0x6e601c33 // bsl v19.16b, v1.16b, v0.16b |
| 415 | .long 0x4ea6d4e0 // fsub v0.4s, v7.4s, v6.4s |
| 416 | .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s |
| 417 | .long 0x6e22fc00 // fdiv v0.4s, v0.4s, v2.4s |
| 418 | .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s |
| 419 | .long 0x6e22de81 // fmul v1.4s, v20.4s, v2.4s |
| 420 | .long 0x4ea0d4e0 // fsub v0.4s, v7.4s, v0.4s |
| 421 | .long 0x4e26d434 // fadd v20.4s, v1.4s, v6.4s |
| 422 | .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s |
| 423 | .long 0x4ea0d840 // fcmeq v0.4s, v2.4s, #0.0 |
| 424 | .long 0x4ea3d631 // fsub v17.4s, v17.4s, v3.4s |
| 425 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 426 | .long 0x4e27e490 // fcmeq v16.4s, v4.4s, v7.4s |
| 427 | .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b |
| 428 | .long 0x4e31cc92 // fmla v18.4s, v4.4s, v17.4s |
| 429 | .long 0x4e27e4a1 // fcmeq v1.4s, v5.4s, v7.4s |
| 430 | .long 0x4e27e4c2 // fcmeq v2.4s, v6.4s, v7.4s |
| 431 | .long 0x4e31ccb3 // fmla v19.4s, v5.4s, v17.4s |
| 432 | .long 0x4e31ccc0 // fmla v0.4s, v6.4s, v17.4s |
| 433 | .long 0x6e721ed0 // bsl v16.16b, v22.16b, v18.16b |
| 434 | .long 0x6e731ea1 // bsl v1.16b, v21.16b, v19.16b |
| 435 | .long 0x6e601e82 // bsl v2.16b, v20.16b, v0.16b |
| 436 | .long 0x4e27ce23 // fmla v3.4s, v17.4s, v7.4s |
| 437 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 438 | .long 0xd61f0060 // br x3 |
| 439 | |
| 440 | HIDDEN _sk_colordodge_aarch64 |
| 441 | .globl _sk_colordodge_aarch64 |
| 442 | _sk_colordodge_aarch64: |
| 443 | .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| 444 | .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s |
| 445 | .long 0x4ea0d474 // fsub v20.4s, v3.4s, v0.4s |
| 446 | .long 0x6e25dc75 // fmul v21.4s, v3.4s, v5.4s |
| 447 | .long 0x4ea1d476 // fsub v22.4s, v3.4s, v1.4s |
| 448 | .long 0x4ea7d657 // fsub v23.4s, v18.4s, v7.4s |
| 449 | .long 0x6e34fe31 // fdiv v17.4s, v17.4s, v20.4s |
| 450 | .long 0x6e36feb4 // fdiv v20.4s, v21.4s, v22.4s |
| 451 | .long 0x6e20def5 // fmul v21.4s, v23.4s, v0.4s |
| 452 | .long 0x4eb1f4f1 // fmin v17.4s, v7.4s, v17.4s |
| 453 | .long 0x4e23e413 // fcmeq v19.4s, v0.4s, v3.4s |
| 454 | .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s |
| 455 | .long 0x4e31cc75 // fmla v21.4s, v3.4s, v17.4s |
| 456 | .long 0x6e751c13 // bsl v19.16b, v0.16b, v21.16b |
| 457 | .long 0x6e21dee0 // fmul v0.4s, v23.4s, v1.4s |
| 458 | .long 0x4eb4f4f4 // fmin v20.4s, v7.4s, v20.4s |
| 459 | .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s |
| 460 | .long 0x4e34cc60 // fmla v0.4s, v3.4s, v20.4s |
| 461 | .long 0x4e23e434 // fcmeq v20.4s, v1.4s, v3.4s |
| 462 | .long 0x6e601c34 // bsl v20.16b, v1.16b, v0.16b |
| 463 | .long 0x6e26dc60 // fmul v0.4s, v3.4s, v6.4s |
| 464 | .long 0x4ea2d461 // fsub v1.4s, v3.4s, v2.4s |
| 465 | .long 0x6e21fc00 // fdiv v0.4s, v0.4s, v1.4s |
| 466 | .long 0x6e22dee1 // fmul v1.4s, v23.4s, v2.4s |
| 467 | .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s |
| 468 | .long 0x4e26d437 // fadd v23.4s, v1.4s, v6.4s |
| 469 | .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s |
| 470 | .long 0x4e23e440 // fcmeq v0.4s, v2.4s, v3.4s |
| 471 | .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b |
| 472 | .long 0x4ea3d641 // fsub v1.4s, v18.4s, v3.4s |
| 473 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 474 | .long 0x4ea0d890 // fcmeq v16.4s, v4.4s, #0.0 |
| 475 | .long 0x4ea0d8b1 // fcmeq v17.4s, v5.4s, #0.0 |
| 476 | .long 0x4e21cc93 // fmla v19.4s, v4.4s, v1.4s |
| 477 | .long 0x4e21ccb4 // fmla v20.4s, v5.4s, v1.4s |
| 478 | .long 0x4ea0d8c2 // fcmeq v2.4s, v6.4s, #0.0 |
| 479 | .long 0x4e21ccc0 // fmla v0.4s, v6.4s, v1.4s |
| 480 | .long 0x6e731ed0 // bsl v16.16b, v22.16b, v19.16b |
| 481 | .long 0x6e741eb1 // bsl v17.16b, v21.16b, v20.16b |
| 482 | .long 0x6e601ee2 // bsl v2.16b, v23.16b, v0.16b |
| 483 | .long 0x4e27cc23 // fmla v3.4s, v1.4s, v7.4s |
| 484 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 485 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 486 | .long 0xd61f0060 // br x3 |
| 487 | |
| 488 | HIDDEN _sk_hardlight_aarch64 |
| 489 | .globl _sk_hardlight_aarch64 |
| 490 | _sk_hardlight_aarch64: |
| 491 | .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s |
| 492 | .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s |
| 493 | .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s |
| 494 | .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s |
| 495 | .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s |
| 496 | .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s |
| 497 | .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| 498 | .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s |
| 499 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 500 | .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s |
| 501 | .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b |
| 502 | .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s |
| 503 | .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s |
| 504 | .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s |
| 505 | .long 0x4e21d436 // fadd v22.4s, v1.4s, v1.4s |
| 506 | .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s |
| 507 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 508 | .long 0x6e36e476 // fcmge v22.4s, v3.4s, v22.4s |
| 509 | .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s |
| 510 | .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| 511 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 512 | .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b |
| 513 | .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s |
| 514 | .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s |
| 515 | .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| 516 | .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s |
| 517 | .long 0x4e22d454 // fadd v20.4s, v2.4s, v2.4s |
| 518 | .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s |
| 519 | .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s |
| 520 | .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s |
| 521 | .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| 522 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 523 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 524 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 525 | .long 0x6e34e474 // fcmge v20.4s, v3.4s, v20.4s |
| 526 | .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| 527 | .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| 528 | .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| 529 | .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| 530 | .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s |
| 531 | .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b |
| 532 | .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| 533 | .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s |
| 534 | .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s |
| 535 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 536 | .long 0xd61f0060 // br x3 |
| 537 | |
| 538 | HIDDEN _sk_overlay_aarch64 |
| 539 | .globl _sk_overlay_aarch64 |
| 540 | _sk_overlay_aarch64: |
| 541 | .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s |
| 542 | .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s |
| 543 | .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s |
| 544 | .long 0x4e24d491 // fadd v17.4s, v4.4s, v4.4s |
| 545 | .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s |
| 546 | .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s |
| 547 | .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| 548 | .long 0x6e31e4f1 // fcmge v17.4s, v7.4s, v17.4s |
| 549 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 550 | .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s |
| 551 | .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b |
| 552 | .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s |
| 553 | .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s |
| 554 | .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s |
| 555 | .long 0x4e25d4b6 // fadd v22.4s, v5.4s, v5.4s |
| 556 | .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s |
| 557 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 558 | .long 0x6e36e4f6 // fcmge v22.4s, v7.4s, v22.4s |
| 559 | .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s |
| 560 | .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| 561 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 562 | .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b |
| 563 | .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s |
| 564 | .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s |
| 565 | .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s |
| 566 | .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s |
| 567 | .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s |
| 568 | .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s |
| 569 | .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s |
| 570 | .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s |
| 571 | .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s |
| 572 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 573 | .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s |
| 574 | .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s |
| 575 | .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s |
| 576 | .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s |
| 577 | .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s |
| 578 | .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s |
| 579 | .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s |
| 580 | .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s |
| 581 | .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b |
| 582 | .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| 583 | .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s |
| 584 | .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s |
| 585 | .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s |
| 586 | .long 0xd61f0060 // br x3 |
| 587 | |
| 588 | HIDDEN _sk_softlight_aarch64 |
| 589 | .globl _sk_softlight_aarch64 |
| 590 | _sk_softlight_aarch64: |
| 591 | .long 0x4ea0c8f5 // fcmgt v21.4s, v7.4s, #0.0 |
| 592 | .long 0x6e27fc96 // fdiv v22.4s, v4.4s, v7.4s |
| 593 | .long 0x6e27fcb8 // fdiv v24.4s, v5.4s, v7.4s |
| 594 | .long 0x6e27fcd9 // fdiv v25.4s, v6.4s, v7.4s |
| 595 | .long 0x4e351ed6 // and v22.16b, v22.16b, v21.16b |
| 596 | .long 0x4e351f18 // and v24.16b, v24.16b, v21.16b |
| 597 | .long 0x4e351f35 // and v21.16b, v25.16b, v21.16b |
| 598 | .long 0x6ea1dad9 // frsqrte v25.4s, v22.4s |
| 599 | .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s |
| 600 | .long 0x4ebdfedd // frsqrts v29.4s, v22.4s, v29.4s |
| 601 | .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s |
| 602 | .long 0x4ea1db3d // frecpe v29.4s, v25.4s |
| 603 | .long 0x6ea0fada // fneg v26.4s, v22.4s |
| 604 | .long 0x6ea1db1b // frsqrte v27.4s, v24.4s |
| 605 | .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s |
| 606 | .long 0x4e3dcf3a // fmla v26.4s, v25.4s, v29.4s |
| 607 | .long 0x6e3bdf7d // fmul v29.4s, v27.4s, v27.4s |
| 608 | .long 0x4ebdff1d // frsqrts v29.4s, v24.4s, v29.4s |
| 609 | .long 0x6e3ddf7b // fmul v27.4s, v27.4s, v29.4s |
| 610 | .long 0x4ea1db7d // frecpe v29.4s, v27.4s |
| 611 | .long 0x6ea0fb1c // fneg v28.4s, v24.4s |
| 612 | .long 0x6ea1dab9 // frsqrte v25.4s, v21.4s |
| 613 | .long 0x4e3dff7b // frecps v27.4s, v27.4s, v29.4s |
| 614 | .long 0x4e3dcf7c // fmla v28.4s, v27.4s, v29.4s |
| 615 | .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s |
| 616 | .long 0x4ebdfebd // frsqrts v29.4s, v21.4s, v29.4s |
| 617 | .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s |
| 618 | .long 0x4ea1db3d // frecpe v29.4s, v25.4s |
| 619 | .long 0x6ea0fabb // fneg v27.4s, v21.4s |
| 620 | .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s |
| 621 | .long 0x4e3dcf3b // fmla v27.4s, v25.4s, v29.4s |
| 622 | .long 0x4e36d6d9 // fadd v25.4s, v22.4s, v22.4s |
| 623 | .long 0x4f07f613 // fmov v19.4s, #-1.000000000000000000e+00 |
| 624 | .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s |
| 625 | .long 0x4e24d497 // fadd v23.4s, v4.4s, v4.4s |
| 626 | .long 0x4e33d6dd // fadd v29.4s, v22.4s, v19.4s |
| 627 | .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s |
| 628 | .long 0x4f00f794 // fmov v20.4s, #7.000000000000000000e+00 |
| 629 | .long 0x6e39dfb9 // fmul v25.4s, v29.4s, v25.4s |
| 630 | .long 0x4e37d6f7 // fadd v23.4s, v23.4s, v23.4s |
| 631 | .long 0x6e37e4f7 // fcmge v23.4s, v7.4s, v23.4s |
| 632 | .long 0x4e36ce99 // fmla v25.4s, v20.4s, v22.4s |
| 633 | .long 0x6e7a1f37 // bsl v23.16b, v25.16b, v26.16b |
| 634 | .long 0x4e38d719 // fadd v25.4s, v24.4s, v24.4s |
| 635 | .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s |
| 636 | .long 0x4e33d71a // fadd v26.4s, v24.4s, v19.4s |
| 637 | .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s |
| 638 | .long 0x6e39df59 // fmul v25.4s, v26.4s, v25.4s |
| 639 | .long 0x4e25d4ba // fadd v26.4s, v5.4s, v5.4s |
| 640 | .long 0x4e3ad75a // fadd v26.4s, v26.4s, v26.4s |
| 641 | .long 0x6e3ae4fa // fcmge v26.4s, v7.4s, v26.4s |
| 642 | .long 0x4e38ce99 // fmla v25.4s, v20.4s, v24.4s |
| 643 | .long 0x6e7c1f3a // bsl v26.16b, v25.16b, v28.16b |
| 644 | .long 0x4e35d6bc // fadd v28.4s, v21.4s, v21.4s |
| 645 | .long 0x4e3cd79c // fadd v28.4s, v28.4s, v28.4s |
| 646 | .long 0x4e33d6b3 // fadd v19.4s, v21.4s, v19.4s |
| 647 | .long 0x4e3ccf9c // fmla v28.4s, v28.4s, v28.4s |
| 648 | .long 0x6e3cde73 // fmul v19.4s, v19.4s, v28.4s |
| 649 | .long 0x4e35ce93 // fmla v19.4s, v20.4s, v21.4s |
| 650 | .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s |
| 651 | .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s |
| 652 | .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00 |
| 653 | .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s |
| 654 | .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s |
| 655 | .long 0x6e7b1e74 // bsl v20.16b, v19.16b, v27.16b |
| 656 | .long 0x4ea7d65b // fsub v27.4s, v18.4s, v7.4s |
| 657 | .long 0x4ea31c70 // mov v16.16b, v3.16b |
| 658 | .long 0x4e21d43d // fadd v29.4s, v1.4s, v1.4s |
| 659 | .long 0x4e22d45c // fadd v28.4s, v2.4s, v2.4s |
| 660 | .long 0x6e20df60 // fmul v0.4s, v27.4s, v0.4s |
| 661 | .long 0x6e21df61 // fmul v1.4s, v27.4s, v1.4s |
| 662 | .long 0x6e22df62 // fmul v2.4s, v27.4s, v2.4s |
| 663 | .long 0x4ea3d63b // fsub v27.4s, v17.4s, v3.4s |
| 664 | .long 0x4eb6d656 // fsub v22.4s, v18.4s, v22.4s |
| 665 | .long 0x4ea31c79 // mov v25.16b, v3.16b |
| 666 | .long 0x4e3bced0 // fmla v16.4s, v22.4s, v27.4s |
| 667 | .long 0x4ea3d7b6 // fsub v22.4s, v29.4s, v3.4s |
| 668 | .long 0x4eb8d658 // fsub v24.4s, v18.4s, v24.4s |
| 669 | .long 0x4ea31c73 // mov v19.16b, v3.16b |
| 670 | .long 0x4e36cf19 // fmla v25.4s, v24.4s, v22.4s |
| 671 | .long 0x4ea3d798 // fsub v24.4s, v28.4s, v3.4s |
| 672 | .long 0x4eb5d655 // fsub v21.4s, v18.4s, v21.4s |
| 673 | .long 0x4e38ceb3 // fmla v19.4s, v21.4s, v24.4s |
| 674 | .long 0x6e27df7b // fmul v27.4s, v27.4s, v7.4s |
| 675 | .long 0x6e27ded6 // fmul v22.4s, v22.4s, v7.4s |
| 676 | .long 0x6e27df18 // fmul v24.4s, v24.4s, v7.4s |
| 677 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 678 | .long 0x6e37df77 // fmul v23.4s, v27.4s, v23.4s |
| 679 | .long 0x6e3aded6 // fmul v22.4s, v22.4s, v26.4s |
| 680 | .long 0x6e34df14 // fmul v20.4s, v24.4s, v20.4s |
| 681 | .long 0x4ea3d652 // fsub v18.4s, v18.4s, v3.4s |
| 682 | .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s |
| 683 | .long 0x6e3de475 // fcmge v21.4s, v3.4s, v29.4s |
| 684 | .long 0x6e3ce47c // fcmge v28.4s, v3.4s, v28.4s |
| 685 | .long 0x6e24de10 // fmul v16.4s, v16.4s, v4.4s |
| 686 | .long 0x6e25df39 // fmul v25.4s, v25.4s, v5.4s |
| 687 | .long 0x6e26de73 // fmul v19.4s, v19.4s, v6.4s |
| 688 | .long 0x4e23cc97 // fmla v23.4s, v4.4s, v3.4s |
| 689 | .long 0x4e23ccb6 // fmla v22.4s, v5.4s, v3.4s |
| 690 | .long 0x4e23ccd4 // fmla v20.4s, v6.4s, v3.4s |
| 691 | .long 0x4e32cc80 // fmla v0.4s, v4.4s, v18.4s |
| 692 | .long 0x4e32cca1 // fmla v1.4s, v5.4s, v18.4s |
| 693 | .long 0x4e32ccc2 // fmla v2.4s, v6.4s, v18.4s |
| 694 | .long 0x6e771e11 // bsl v17.16b, v16.16b, v23.16b |
| 695 | .long 0x6e761f35 // bsl v21.16b, v25.16b, v22.16b |
| 696 | .long 0x6e741e7c // bsl v28.16b, v19.16b, v20.16b |
| 697 | .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s |
| 698 | .long 0x4e35d421 // fadd v1.4s, v1.4s, v21.4s |
| 699 | .long 0x4e3cd442 // fadd v2.4s, v2.4s, v28.4s |
| 700 | .long 0x4e27ce43 // fmla v3.4s, v18.4s, v7.4s |
| 701 | .long 0xd61f0060 // br x3 |
| 702 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 703 | HIDDEN _sk_clamp_0_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 704 | .globl _sk_clamp_0_aarch64 |
| 705 | _sk_clamp_0_aarch64: |
| 706 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 707 | .long 0x6f00e410 // movi v16.2d, #0x0 |
| 708 | .long 0x4e30f400 // fmax v0.4s, v0.4s, v16.4s |
| 709 | .long 0x4e30f421 // fmax v1.4s, v1.4s, v16.4s |
| 710 | .long 0x4e30f442 // fmax v2.4s, v2.4s, v16.4s |
| 711 | .long 0x4e30f463 // fmax v3.4s, v3.4s, v16.4s |
| 712 | .long 0xd61f0060 // br x3 |
| 713 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 714 | HIDDEN _sk_clamp_1_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 715 | .globl _sk_clamp_1_aarch64 |
| 716 | _sk_clamp_1_aarch64: |
| 717 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 718 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 719 | .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| 720 | .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| 721 | .long 0x4eb0f442 // fmin v2.4s, v2.4s, v16.4s |
| 722 | .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s |
| 723 | .long 0xd61f0060 // br x3 |
| 724 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 725 | HIDDEN _sk_clamp_a_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 726 | .globl _sk_clamp_a_aarch64 |
| 727 | _sk_clamp_a_aarch64: |
| 728 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 729 | .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00 |
| 730 | .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s |
| 731 | .long 0x4ea3f400 // fmin v0.4s, v0.4s, v3.4s |
| 732 | .long 0x4ea3f421 // fmin v1.4s, v1.4s, v3.4s |
| 733 | .long 0x4ea3f442 // fmin v2.4s, v2.4s, v3.4s |
| 734 | .long 0xd61f0060 // br x3 |
| 735 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 736 | HIDDEN _sk_set_rgb_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 737 | .globl _sk_set_rgb_aarch64 |
| 738 | _sk_set_rgb_aarch64: |
| 739 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 740 | .long 0xaa0803e9 // mov x9, x8 |
| 741 | .long 0x4ddfc920 // ld1r {v0.4s}, [x9], #4 |
| 742 | .long 0x91002108 // add x8, x8, #0x8 |
| 743 | .long 0x4d40c902 // ld1r {v2.4s}, [x8] |
| 744 | .long 0x4d40c921 // ld1r {v1.4s}, [x9] |
| 745 | .long 0xd61f0060 // br x3 |
| 746 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 747 | HIDDEN _sk_swap_rb_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 748 | .globl _sk_swap_rb_aarch64 |
| 749 | _sk_swap_rb_aarch64: |
| 750 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 751 | .long 0x4ea01c10 // mov v16.16b, v0.16b |
| 752 | .long 0x4ea21c40 // mov v0.16b, v2.16b |
| 753 | .long 0x4eb01e02 // mov v2.16b, v16.16b |
| 754 | .long 0xd61f0060 // br x3 |
| 755 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 756 | HIDDEN _sk_swap_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 757 | .globl _sk_swap_aarch64 |
| 758 | _sk_swap_aarch64: |
| 759 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 760 | .long 0x4ea31c70 // mov v16.16b, v3.16b |
| 761 | .long 0x4ea21c51 // mov v17.16b, v2.16b |
| 762 | .long 0x4ea11c32 // mov v18.16b, v1.16b |
| 763 | .long 0x4ea01c13 // mov v19.16b, v0.16b |
| 764 | .long 0x4ea41c80 // mov v0.16b, v4.16b |
| 765 | .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| 766 | .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| 767 | .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| 768 | .long 0x4eb31e64 // mov v4.16b, v19.16b |
| 769 | .long 0x4eb21e45 // mov v5.16b, v18.16b |
| 770 | .long 0x4eb11e26 // mov v6.16b, v17.16b |
| 771 | .long 0x4eb01e07 // mov v7.16b, v16.16b |
| 772 | .long 0xd61f0060 // br x3 |
| 773 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 774 | HIDDEN _sk_move_src_dst_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 775 | .globl _sk_move_src_dst_aarch64 |
| 776 | _sk_move_src_dst_aarch64: |
| 777 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 778 | .long 0x4ea01c04 // mov v4.16b, v0.16b |
| 779 | .long 0x4ea11c25 // mov v5.16b, v1.16b |
| 780 | .long 0x4ea21c46 // mov v6.16b, v2.16b |
| 781 | .long 0x4ea31c67 // mov v7.16b, v3.16b |
| 782 | .long 0xd61f0060 // br x3 |
| 783 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 784 | HIDDEN _sk_move_dst_src_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 785 | .globl _sk_move_dst_src_aarch64 |
| 786 | _sk_move_dst_src_aarch64: |
| 787 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 788 | .long 0x4ea41c80 // mov v0.16b, v4.16b |
| 789 | .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| 790 | .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| 791 | .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| 792 | .long 0xd61f0060 // br x3 |
| 793 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 794 | HIDDEN _sk_premul_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 795 | .globl _sk_premul_aarch64 |
| 796 | _sk_premul_aarch64: |
| 797 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 798 | .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s |
| 799 | .long 0x6e23dc21 // fmul v1.4s, v1.4s, v3.4s |
| 800 | .long 0x6e23dc42 // fmul v2.4s, v2.4s, v3.4s |
| 801 | .long 0xd61f0060 // br x3 |
| 802 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 803 | HIDDEN _sk_unpremul_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 804 | .globl _sk_unpremul_aarch64 |
| 805 | _sk_unpremul_aarch64: |
| 806 | .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| 807 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 808 | .long 0x4ea0d870 // fcmeq v16.4s, v3.4s, #0.0 |
| 809 | .long 0x6e23fe31 // fdiv v17.4s, v17.4s, v3.4s |
| 810 | .long 0x4e701e30 // bic v16.16b, v17.16b, v16.16b |
| 811 | .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| 812 | .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| 813 | .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| 814 | .long 0xd61f0060 // br x3 |
| 815 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 816 | HIDDEN _sk_from_srgb_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 817 | .globl _sk_from_srgb_aarch64 |
| 818 | _sk_from_srgb_aarch64: |
| 819 | .long 0x52a7d328 // mov w8, #0x3e990000 |
| 820 | .long 0x72933348 // movk w8, #0x999a |
| 821 | .long 0x4e040d10 // dup v16.4s, w8 |
| 822 | .long 0x52a7e648 // mov w8, #0x3f320000 |
| 823 | .long 0x7291eb88 // movk w8, #0x8f5c |
| 824 | .long 0x4e040d11 // dup v17.4s, w8 |
| 825 | .long 0x52a76468 // mov w8, #0x3b230000 |
| 826 | .long 0x729ae148 // movk w8, #0xd70a |
| 827 | .long 0x4e040d12 // dup v18.4s, w8 |
| 828 | .long 0x52a7b3c8 // mov w8, #0x3d9e0000 |
| 829 | .long 0x72907228 // movk w8, #0x8391 |
| 830 | .long 0x6e22dc54 // fmul v20.4s, v2.4s, v2.4s |
| 831 | .long 0x4eb11e35 // mov v21.16b, v17.16b |
| 832 | .long 0x4eb11e37 // mov v23.16b, v17.16b |
| 833 | .long 0x4e22ce11 // fmla v17.4s, v16.4s, v2.4s |
| 834 | .long 0x4eb21e56 // mov v22.16b, v18.16b |
| 835 | .long 0x4eb21e58 // mov v24.16b, v18.16b |
| 836 | .long 0x4e34ce32 // fmla v18.4s, v17.4s, v20.4s |
| 837 | .long 0x4e040d11 // dup v17.4s, w8 |
| 838 | .long 0x52a7ac28 // mov w8, #0x3d610000 |
| 839 | .long 0x6e20dc13 // fmul v19.4s, v0.4s, v0.4s |
| 840 | .long 0x7288f5c8 // movk w8, #0x47ae |
| 841 | .long 0x4e20ce15 // fmla v21.4s, v16.4s, v0.4s |
| 842 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 843 | .long 0x6e21dc34 // fmul v20.4s, v1.4s, v1.4s |
| 844 | .long 0x4e33ceb6 // fmla v22.4s, v21.4s, v19.4s |
| 845 | .long 0x4e040d13 // dup v19.4s, w8 |
| 846 | .long 0x4e21ce17 // fmla v23.4s, v16.4s, v1.4s |
| 847 | .long 0x6e31dc15 // fmul v21.4s, v0.4s, v17.4s |
| 848 | .long 0x6ea0e660 // fcmgt v0.4s, v19.4s, v0.4s |
| 849 | .long 0x6e31dc30 // fmul v16.4s, v1.4s, v17.4s |
| 850 | .long 0x6ea1e661 // fcmgt v1.4s, v19.4s, v1.4s |
| 851 | .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s |
| 852 | .long 0x6ea2e662 // fcmgt v2.4s, v19.4s, v2.4s |
| 853 | .long 0x4e34cef8 // fmla v24.4s, v23.4s, v20.4s |
| 854 | .long 0x6e761ea0 // bsl v0.16b, v21.16b, v22.16b |
| 855 | .long 0x6e781e01 // bsl v1.16b, v16.16b, v24.16b |
| 856 | .long 0x6e721e22 // bsl v2.16b, v17.16b, v18.16b |
| 857 | .long 0xd61f0060 // br x3 |
| 858 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 859 | HIDDEN _sk_to_srgb_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 860 | .globl _sk_to_srgb_aarch64 |
| 861 | _sk_to_srgb_aarch64: |
| 862 | .long 0x52a828e8 // mov w8, #0x41470000 |
| 863 | .long 0x728b8528 // movk w8, #0x5c29 |
| 864 | .long 0x4e040d12 // dup v18.4s, w8 |
| 865 | .long 0x52a7e608 // mov w8, #0x3f300000 |
| 866 | .long 0x728df9c8 // movk w8, #0x6fce |
| 867 | .long 0x6ea1d811 // frsqrte v17.4s, v0.4s |
| 868 | .long 0x4e040d13 // dup v19.4s, w8 |
| 869 | .long 0x52b7b948 // mov w8, #0xbdca0000 |
| 870 | .long 0x728af508 // movk w8, #0x57a8 |
| 871 | .long 0x6ea1d834 // frsqrte v20.4s, v1.4s |
| 872 | .long 0x6e31de36 // fmul v22.4s, v17.4s, v17.4s |
| 873 | .long 0x4e040d10 // dup v16.4s, w8 |
| 874 | .long 0x52a77188 // mov w8, #0x3b8c0000 |
| 875 | .long 0x6ea1d855 // frsqrte v21.4s, v2.4s |
| 876 | .long 0x6e34de98 // fmul v24.4s, v20.4s, v20.4s |
| 877 | .long 0x4eb6fc16 // frsqrts v22.4s, v0.4s, v22.4s |
| 878 | .long 0x729ce088 // movk w8, #0xe704 |
| 879 | .long 0x6e35deb9 // fmul v25.4s, v21.4s, v21.4s |
| 880 | .long 0x4eb8fc38 // frsqrts v24.4s, v1.4s, v24.4s |
| 881 | .long 0x6e36de31 // fmul v17.4s, v17.4s, v22.4s |
| 882 | .long 0x4e040d17 // dup v23.4s, w8 |
| 883 | .long 0x4eb9fc59 // frsqrts v25.4s, v2.4s, v25.4s |
| 884 | .long 0x6e38de94 // fmul v20.4s, v20.4s, v24.4s |
| 885 | .long 0x4ea1da36 // frecpe v22.4s, v17.4s |
| 886 | .long 0x6e32dc1a // fmul v26.4s, v0.4s, v18.4s |
| 887 | .long 0x6ea0e6e0 // fcmgt v0.4s, v23.4s, v0.4s |
| 888 | .long 0x6e32dc3c // fmul v28.4s, v1.4s, v18.4s |
| 889 | .long 0x6ea1e6e1 // fcmgt v1.4s, v23.4s, v1.4s |
| 890 | .long 0x6e32dc52 // fmul v18.4s, v2.4s, v18.4s |
| 891 | .long 0x6ea2e6e2 // fcmgt v2.4s, v23.4s, v2.4s |
| 892 | .long 0x6e39deb5 // fmul v21.4s, v21.4s, v25.4s |
| 893 | .long 0x4ea1da97 // frecpe v23.4s, v20.4s |
| 894 | .long 0x4e36fe39 // frecps v25.4s, v17.4s, v22.4s |
| 895 | .long 0x4ea1dab8 // frecpe v24.4s, v21.4s |
| 896 | .long 0x6e39ded6 // fmul v22.4s, v22.4s, v25.4s |
| 897 | .long 0x4e37fe99 // frecps v25.4s, v20.4s, v23.4s |
| 898 | .long 0x4eb01e1b // mov v27.16b, v16.16b |
| 899 | .long 0x6e39def7 // fmul v23.4s, v23.4s, v25.4s |
| 900 | .long 0x4e38feb9 // frecps v25.4s, v21.4s, v24.4s |
| 901 | .long 0x6e39df18 // fmul v24.4s, v24.4s, v25.4s |
| 902 | .long 0x4eb01e19 // mov v25.16b, v16.16b |
| 903 | .long 0x4e36ce7b // fmla v27.4s, v19.4s, v22.4s |
| 904 | .long 0x6ea1da36 // frsqrte v22.4s, v17.4s |
| 905 | .long 0x4e37ce79 // fmla v25.4s, v19.4s, v23.4s |
| 906 | .long 0x6ea1da97 // frsqrte v23.4s, v20.4s |
| 907 | .long 0x4e38ce70 // fmla v16.4s, v19.4s, v24.4s |
| 908 | .long 0x6e36ded8 // fmul v24.4s, v22.4s, v22.4s |
| 909 | .long 0x6ea1dab3 // frsqrte v19.4s, v21.4s |
| 910 | .long 0x4eb8fe31 // frsqrts v17.4s, v17.4s, v24.4s |
| 911 | .long 0x6e37def8 // fmul v24.4s, v23.4s, v23.4s |
| 912 | .long 0x4eb8fe94 // frsqrts v20.4s, v20.4s, v24.4s |
| 913 | .long 0x6e33de78 // fmul v24.4s, v19.4s, v19.4s |
| 914 | .long 0x52a7da48 // mov w8, #0x3ed20000 |
| 915 | .long 0x4eb8feb5 // frsqrts v21.4s, v21.4s, v24.4s |
| 916 | .long 0x7290f848 // movk w8, #0x87c2 |
| 917 | .long 0x6e31ded1 // fmul v17.4s, v22.4s, v17.4s |
| 918 | .long 0x6e34def4 // fmul v20.4s, v23.4s, v20.4s |
| 919 | .long 0x6e35de73 // fmul v19.4s, v19.4s, v21.4s |
| 920 | .long 0x4e040d15 // dup v21.4s, w8 |
| 921 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 922 | .long 0x4e31cebb // fmla v27.4s, v21.4s, v17.4s |
| 923 | .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00 |
| 924 | .long 0x4e34ceb9 // fmla v25.4s, v21.4s, v20.4s |
| 925 | .long 0x4e33ceb0 // fmla v16.4s, v21.4s, v19.4s |
| 926 | .long 0x4ebbf633 // fmin v19.4s, v17.4s, v27.4s |
| 927 | .long 0x4eb9f634 // fmin v20.4s, v17.4s, v25.4s |
| 928 | .long 0x4eb0f630 // fmin v16.4s, v17.4s, v16.4s |
| 929 | .long 0x6e731f40 // bsl v0.16b, v26.16b, v19.16b |
| 930 | .long 0x6e741f81 // bsl v1.16b, v28.16b, v20.16b |
| 931 | .long 0x6e701e42 // bsl v2.16b, v18.16b, v16.16b |
| 932 | .long 0xd61f0060 // br x3 |
| 933 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 934 | HIDDEN _sk_scale_1_float_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 935 | .globl _sk_scale_1_float_aarch64 |
| 936 | _sk_scale_1_float_aarch64: |
| 937 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 938 | .long 0xbd400110 // ldr s16, [x8] |
| 939 | .long 0x4f909000 // fmul v0.4s, v0.4s, v16.s[0] |
| 940 | .long 0x4f909021 // fmul v1.4s, v1.4s, v16.s[0] |
| 941 | .long 0x4f909042 // fmul v2.4s, v2.4s, v16.s[0] |
| 942 | .long 0x4f909063 // fmul v3.4s, v3.4s, v16.s[0] |
| 943 | .long 0xd61f0060 // br x3 |
| 944 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 945 | HIDDEN _sk_scale_u8_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 946 | .globl _sk_scale_u8_aarch64 |
| 947 | _sk_scale_u8_aarch64: |
| 948 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 949 | .long 0x52a77009 // mov w9, #0x3b800000 |
| 950 | .long 0x72901029 // movk w9, #0x8081 |
| 951 | .long 0x4e040d30 // dup v16.4s, w9 |
| 952 | .long 0xf9400108 // ldr x8, [x8] |
| 953 | .long 0x8b000108 // add x8, x8, x0 |
| 954 | .long 0x39400109 // ldrb w9, [x8] |
| 955 | .long 0x3940050a // ldrb w10, [x8, #1] |
| 956 | .long 0x3940090b // ldrb w11, [x8, #2] |
| 957 | .long 0x39400d08 // ldrb w8, [x8, #3] |
| 958 | .long 0x4e021d31 // mov v17.h[0], w9 |
| 959 | .long 0x4e061d51 // mov v17.h[1], w10 |
| 960 | .long 0x4e0a1d71 // mov v17.h[2], w11 |
| 961 | .long 0x4e0e1d11 // mov v17.h[3], w8 |
| 962 | .long 0x2f10a631 // uxtl v17.4s, v17.4h |
| 963 | .long 0x6e21da31 // ucvtf v17.4s, v17.4s |
| 964 | .long 0x6e30de30 // fmul v16.4s, v17.4s, v16.4s |
| 965 | .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s |
| 966 | .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s |
| 967 | .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s |
| 968 | .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| 969 | .long 0xd61f0060 // br x3 |
| 970 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 971 | HIDDEN _sk_lerp_1_float_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 972 | .globl _sk_lerp_1_float_aarch64 |
| 973 | _sk_lerp_1_float_aarch64: |
| 974 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 975 | .long 0x4ea4d411 // fsub v17.4s, v0.4s, v4.4s |
| 976 | .long 0x4ea41c80 // mov v0.16b, v4.16b |
| 977 | .long 0x4ea5d432 // fsub v18.4s, v1.4s, v5.4s |
| 978 | .long 0xbd400110 // ldr s16, [x8] |
| 979 | .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| 980 | .long 0x4f901220 // fmla v0.4s, v17.4s, v16.s[0] |
| 981 | .long 0x4ea6d451 // fsub v17.4s, v2.4s, v6.4s |
| 982 | .long 0x4f901241 // fmla v1.4s, v18.4s, v16.s[0] |
| 983 | .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| 984 | .long 0x4ea7d472 // fsub v18.4s, v3.4s, v7.4s |
| 985 | .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| 986 | .long 0x4f901222 // fmla v2.4s, v17.4s, v16.s[0] |
| 987 | .long 0x4f901243 // fmla v3.4s, v18.4s, v16.s[0] |
| 988 | .long 0xd61f0060 // br x3 |
| 989 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 990 | HIDDEN _sk_lerp_u8_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 991 | .globl _sk_lerp_u8_aarch64 |
| 992 | _sk_lerp_u8_aarch64: |
| 993 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 994 | .long 0x52a77009 // mov w9, #0x3b800000 |
| 995 | .long 0x72901029 // movk w9, #0x8081 |
| 996 | .long 0x4e040d30 // dup v16.4s, w9 |
| 997 | .long 0xf9400108 // ldr x8, [x8] |
| 998 | .long 0x4ea4d412 // fsub v18.4s, v0.4s, v4.4s |
| 999 | .long 0x8b000108 // add x8, x8, x0 |
| 1000 | .long 0x3940010a // ldrb w10, [x8] |
| 1001 | .long 0x39400509 // ldrb w9, [x8, #1] |
| 1002 | .long 0x3940090b // ldrb w11, [x8, #2] |
| 1003 | .long 0x39400d08 // ldrb w8, [x8, #3] |
| 1004 | .long 0x4e021d51 // mov v17.h[0], w10 |
| 1005 | .long 0x4e061d31 // mov v17.h[1], w9 |
| 1006 | .long 0x4e0a1d71 // mov v17.h[2], w11 |
| 1007 | .long 0x4e0e1d11 // mov v17.h[3], w8 |
| 1008 | .long 0x2f10a620 // uxtl v0.4s, v17.4h |
| 1009 | .long 0x6e21d800 // ucvtf v0.4s, v0.4s |
| 1010 | .long 0x6e30dc10 // fmul v16.4s, v0.4s, v16.4s |
| 1011 | .long 0x4ea41c80 // mov v0.16b, v4.16b |
| 1012 | .long 0x4ea5d431 // fsub v17.4s, v1.4s, v5.4s |
| 1013 | .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| 1014 | .long 0x4e32ce00 // fmla v0.4s, v16.4s, v18.4s |
| 1015 | .long 0x4ea6d452 // fsub v18.4s, v2.4s, v6.4s |
| 1016 | .long 0x4e31ce01 // fmla v1.4s, v16.4s, v17.4s |
| 1017 | .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| 1018 | .long 0x4ea7d471 // fsub v17.4s, v3.4s, v7.4s |
| 1019 | .long 0x4ea71ce3 // mov v3.16b, v7.16b |
| 1020 | .long 0x4e32ce02 // fmla v2.4s, v16.4s, v18.4s |
| 1021 | .long 0x4e31ce03 // fmla v3.4s, v16.4s, v17.4s |
| 1022 | .long 0xd61f0060 // br x3 |
| 1023 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1024 | HIDDEN _sk_lerp_565_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1025 | .globl _sk_lerp_565_aarch64 |
| 1026 | _sk_lerp_565_aarch64: |
| 1027 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1028 | .long 0xd37ff809 // lsl x9, x0, #1 |
| 1029 | .long 0x4f072710 // movi v16.4s, #0xf8, lsl #8 |
| 1030 | .long 0x4ea4d413 // fsub v19.4s, v0.4s, v4.4s |
| 1031 | .long 0xf9400108 // ldr x8, [x8] |
| 1032 | .long 0xfc696903 // ldr d3, [x8, x9] |
| 1033 | .long 0x52a6f088 // mov w8, #0x37840000 |
| 1034 | .long 0x72842108 // movk w8, #0x2108 |
| 1035 | .long 0x4e040d11 // dup v17.4s, w8 |
| 1036 | .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| 1037 | .long 0x321b17e8 // orr w8, wzr, #0x7e0 |
| 1038 | .long 0x4e301c60 // and v0.16b, v3.16b, v16.16b |
| 1039 | .long 0x4e040d12 // dup v18.4s, w8 |
| 1040 | .long 0x52a74048 // mov w8, #0x3a020000 |
| 1041 | .long 0x4e21d800 // scvtf v0.4s, v0.4s |
| 1042 | .long 0x72810428 // movk w8, #0x821 |
| 1043 | .long 0x6e31dc10 // fmul v16.4s, v0.4s, v17.4s |
| 1044 | .long 0x4ea41c80 // mov v0.16b, v4.16b |
| 1045 | .long 0x4e33ce00 // fmla v0.4s, v16.4s, v19.4s |
| 1046 | .long 0x4f0007f0 // movi v16.4s, #0x1f |
| 1047 | .long 0x4e040d11 // dup v17.4s, w8 |
| 1048 | .long 0x52a7a088 // mov w8, #0x3d040000 |
| 1049 | .long 0x4e321c72 // and v18.16b, v3.16b, v18.16b |
| 1050 | .long 0x72842108 // movk w8, #0x2108 |
| 1051 | .long 0x4e301c63 // and v3.16b, v3.16b, v16.16b |
| 1052 | .long 0x4ea6d450 // fsub v16.4s, v2.4s, v6.4s |
| 1053 | .long 0x4e21da42 // scvtf v2.4s, v18.4s |
| 1054 | .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s |
| 1055 | .long 0x4e040d02 // dup v2.4s, w8 |
| 1056 | .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| 1057 | .long 0x4ea5d433 // fsub v19.4s, v1.4s, v5.4s |
| 1058 | .long 0x4ea51ca1 // mov v1.16b, v5.16b |
| 1059 | .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s |
| 1060 | .long 0x4ea61cc2 // mov v2.16b, v6.16b |
| 1061 | .long 0x4e33ce21 // fmla v1.4s, v17.4s, v19.4s |
| 1062 | .long 0x4e30cc62 // fmla v2.4s, v3.4s, v16.4s |
| 1063 | .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| 1064 | .long 0xd61f0060 // br x3 |
| 1065 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1066 | HIDDEN _sk_load_tables_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1067 | .globl _sk_load_tables_aarch64 |
| 1068 | _sk_load_tables_aarch64: |
| 1069 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1070 | .long 0xd37ef409 // lsl x9, x0, #2 |
| 1071 | .long 0x6f00e620 // movi v0.2d, #0xff000000ff |
| 1072 | .long 0x52a7700b // mov w11, #0x3b800000 |
| 1073 | .long 0xa940310a // ldp x10, x12, [x8] |
| 1074 | .long 0x7290102b // movk w11, #0x8081 |
| 1075 | .long 0x4e040d63 // dup v3.4s, w11 |
| 1076 | .long 0x3ce96942 // ldr q2, [x10, x9] |
| 1077 | .long 0xa9412109 // ldp x9, x8, [x8, #16] |
| 1078 | .long 0x4e201c41 // and v1.16b, v2.16b, v0.16b |
| 1079 | .long 0x1e26002e // fmov w14, s1 |
| 1080 | .long 0x6f380450 // ushr v16.4s, v2.4s, #8 |
| 1081 | .long 0x6f300451 // ushr v17.4s, v2.4s, #16 |
| 1082 | .long 0x8b2e498e // add x14, x12, w14, uxtw #2 |
| 1083 | .long 0x0e0c3c2a // mov w10, v1.s[1] |
| 1084 | .long 0x0e143c2b // mov w11, v1.s[2] |
| 1085 | .long 0x0e1c3c2d // mov w13, v1.s[3] |
| 1086 | .long 0x4e201e01 // and v1.16b, v16.16b, v0.16b |
| 1087 | .long 0x4e201e30 // and v16.16b, v17.16b, v0.16b |
| 1088 | .long 0x0d4081c0 // ld1 {v0.s}[0], [x14] |
| 1089 | .long 0x8b2a498a // add x10, x12, w10, uxtw #2 |
| 1090 | .long 0xbc6b5991 // ldr s17, [x12, w11, uxtw #2] |
| 1091 | .long 0xbc6d5992 // ldr s18, [x12, w13, uxtw #2] |
| 1092 | .long 0x0e0c3c2b // mov w11, v1.s[1] |
| 1093 | .long 0x0e143c2c // mov w12, v1.s[2] |
| 1094 | .long 0x0e1c3c2d // mov w13, v1.s[3] |
| 1095 | .long 0x1e26002e // fmov w14, s1 |
| 1096 | .long 0x8b2e492e // add x14, x9, w14, uxtw #2 |
| 1097 | .long 0xbc6c5933 // ldr s19, [x9, w12, uxtw #2] |
| 1098 | .long 0xbc6d5934 // ldr s20, [x9, w13, uxtw #2] |
| 1099 | .long 0x8b2b4929 // add x9, x9, w11, uxtw #2 |
| 1100 | .long 0x1e26020b // fmov w11, s16 |
| 1101 | .long 0x6f280442 // ushr v2.4s, v2.4s, #24 |
| 1102 | .long 0x0d409140 // ld1 {v0.s}[1], [x10] |
| 1103 | .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| 1104 | .long 0x8b2b490a // add x10, x8, w11, uxtw #2 |
| 1105 | .long 0x0d4081c1 // ld1 {v1.s}[0], [x14] |
| 1106 | .long 0x6e23dc43 // fmul v3.4s, v2.4s, v3.4s |
| 1107 | .long 0x0d408142 // ld1 {v2.s}[0], [x10] |
| 1108 | .long 0x0e0c3e0f // mov w15, v16.s[1] |
| 1109 | .long 0x0e143e0c // mov w12, v16.s[2] |
| 1110 | .long 0x8b2f490a // add x10, x8, w15, uxtw #2 |
| 1111 | .long 0x0e1c3e0d // mov w13, v16.s[3] |
| 1112 | .long 0xbc6c5910 // ldr s16, [x8, w12, uxtw #2] |
| 1113 | .long 0x0d409121 // ld1 {v1.s}[1], [x9] |
| 1114 | .long 0x0d409142 // ld1 {v2.s}[1], [x10] |
| 1115 | .long 0x6e140620 // mov v0.s[2], v17.s[0] |
| 1116 | .long 0xbc6d5911 // ldr s17, [x8, w13, uxtw #2] |
| 1117 | .long 0x6e140661 // mov v1.s[2], v19.s[0] |
| 1118 | .long 0x6e140602 // mov v2.s[2], v16.s[0] |
| 1119 | .long 0x6e1c0640 // mov v0.s[3], v18.s[0] |
| 1120 | .long 0x6e1c0681 // mov v1.s[3], v20.s[0] |
| 1121 | .long 0x6e1c0622 // mov v2.s[3], v17.s[0] |
| 1122 | .long 0xd61f0060 // br x3 |
| 1123 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1124 | HIDDEN _sk_load_a8_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1125 | .globl _sk_load_a8_aarch64 |
| 1126 | _sk_load_a8_aarch64: |
| 1127 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1128 | .long 0x52a77009 // mov w9, #0x3b800000 |
| 1129 | .long 0x72901029 // movk w9, #0x8081 |
| 1130 | .long 0x4e040d22 // dup v2.4s, w9 |
| 1131 | .long 0xf9400108 // ldr x8, [x8] |
| 1132 | .long 0x6f00e400 // movi v0.2d, #0x0 |
| 1133 | .long 0x6f00e401 // movi v1.2d, #0x0 |
| 1134 | .long 0x8b000108 // add x8, x8, x0 |
| 1135 | .long 0x3940010a // ldrb w10, [x8] |
| 1136 | .long 0x3940050b // ldrb w11, [x8, #1] |
| 1137 | .long 0x3940090c // ldrb w12, [x8, #2] |
| 1138 | .long 0x39400d08 // ldrb w8, [x8, #3] |
| 1139 | .long 0x4e021d43 // mov v3.h[0], w10 |
| 1140 | .long 0x4e061d63 // mov v3.h[1], w11 |
| 1141 | .long 0x4e0a1d83 // mov v3.h[2], w12 |
| 1142 | .long 0x4e0e1d03 // mov v3.h[3], w8 |
| 1143 | .long 0x2f10a463 // uxtl v3.4s, v3.4h |
| 1144 | .long 0x6e21d863 // ucvtf v3.4s, v3.4s |
| 1145 | .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s |
| 1146 | .long 0x6f00e402 // movi v2.2d, #0x0 |
| 1147 | .long 0xd61f0060 // br x3 |
| 1148 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1149 | HIDDEN _sk_store_a8_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1150 | .globl _sk_store_a8_aarch64 |
| 1151 | _sk_store_a8_aarch64: |
| 1152 | .long 0xf9400028 // ldr x8, [x1] |
| 1153 | .long 0x52a86fe9 // mov w9, #0x437f0000 |
| 1154 | .long 0x4e040d30 // dup v16.4s, w9 |
| 1155 | .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| 1156 | .long 0xf9400108 // ldr x8, [x8] |
| 1157 | .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| 1158 | .long 0x0e612a10 // xtn v16.4h, v16.4s |
| 1159 | .long 0x0e0e3e09 // umov w9, v16.h[3] |
| 1160 | .long 0x8b000108 // add x8, x8, x0 |
| 1161 | .long 0x39000d09 // strb w9, [x8, #3] |
| 1162 | .long 0x0e0a3e09 // umov w9, v16.h[2] |
| 1163 | .long 0x39000909 // strb w9, [x8, #2] |
| 1164 | .long 0x0e063e09 // umov w9, v16.h[1] |
| 1165 | .long 0x39000509 // strb w9, [x8, #1] |
| 1166 | .long 0x0e023e09 // umov w9, v16.h[0] |
| 1167 | .long 0x39000109 // strb w9, [x8] |
| 1168 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1169 | .long 0x91004021 // add x1, x1, #0x10 |
| 1170 | .long 0xd61f0060 // br x3 |
| 1171 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 1172 | HIDDEN _sk_load_g8_aarch64 |
| 1173 | .globl _sk_load_g8_aarch64 |
| 1174 | _sk_load_g8_aarch64: |
| 1175 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1176 | .long 0x52a77009 // mov w9, #0x3b800000 |
| 1177 | .long 0x72901029 // movk w9, #0x8081 |
| 1178 | .long 0x4e040d20 // dup v0.4s, w9 |
| 1179 | .long 0xf9400108 // ldr x8, [x8] |
| 1180 | .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| 1181 | .long 0x8b000108 // add x8, x8, x0 |
| 1182 | .long 0x3940010a // ldrb w10, [x8] |
| 1183 | .long 0x39400509 // ldrb w9, [x8, #1] |
| 1184 | .long 0x3940090b // ldrb w11, [x8, #2] |
| 1185 | .long 0x39400d08 // ldrb w8, [x8, #3] |
| 1186 | .long 0x4e021d41 // mov v1.h[0], w10 |
| 1187 | .long 0x4e061d21 // mov v1.h[1], w9 |
| 1188 | .long 0x4e0a1d61 // mov v1.h[2], w11 |
| 1189 | .long 0x4e0e1d01 // mov v1.h[3], w8 |
| 1190 | .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| 1191 | .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| 1192 | .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| 1193 | .long 0x4ea01c01 // mov v1.16b, v0.16b |
| 1194 | .long 0x4ea01c02 // mov v2.16b, v0.16b |
| 1195 | .long 0xd61f0060 // br x3 |
| 1196 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1197 | HIDDEN _sk_load_565_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1198 | .globl _sk_load_565_aarch64 |
| 1199 | _sk_load_565_aarch64: |
| 1200 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1201 | .long 0xd37ff809 // lsl x9, x0, #1 |
| 1202 | .long 0x4f072701 // movi v1.4s, #0xf8, lsl #8 |
| 1203 | .long 0x4f0007e3 // movi v3.4s, #0x1f |
| 1204 | .long 0xf9400108 // ldr x8, [x8] |
| 1205 | .long 0xfc696900 // ldr d0, [x8, x9] |
| 1206 | .long 0x321b17e8 // orr w8, wzr, #0x7e0 |
| 1207 | .long 0x4e040d02 // dup v2.4s, w8 |
| 1208 | .long 0x52a6f088 // mov w8, #0x37840000 |
| 1209 | .long 0x72842108 // movk w8, #0x2108 |
| 1210 | .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| 1211 | .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b |
| 1212 | .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b |
| 1213 | .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b |
| 1214 | .long 0x4e040d00 // dup v0.4s, w8 |
| 1215 | .long 0x52a74048 // mov w8, #0x3a020000 |
| 1216 | .long 0x72810428 // movk w8, #0x821 |
| 1217 | .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| 1218 | .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| 1219 | .long 0x4e040d01 // dup v1.4s, w8 |
| 1220 | .long 0x52a7a088 // mov w8, #0x3d040000 |
| 1221 | .long 0x72842108 // movk w8, #0x2108 |
| 1222 | .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| 1223 | .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s |
| 1224 | .long 0x4e040d02 // dup v2.4s, w8 |
| 1225 | .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| 1226 | .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s |
| 1227 | .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00 |
| 1228 | .long 0xd61f0060 // br x3 |
| 1229 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1230 | HIDDEN _sk_store_565_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1231 | .globl _sk_store_565_aarch64 |
| 1232 | _sk_store_565_aarch64: |
| 1233 | .long 0xf9400028 // ldr x8, [x1] |
| 1234 | .long 0x52a84f8a // mov w10, #0x427c0000 |
| 1235 | .long 0x4f01f7f0 // fmov v16.4s, #3.100000000000000000e+01 |
| 1236 | .long 0x4e040d52 // dup v18.4s, w10 |
| 1237 | .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| 1238 | .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s |
| 1239 | .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| 1240 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1241 | .long 0x6e30dc50 // fmul v16.4s, v2.4s, v16.4s |
| 1242 | .long 0x4f2b5631 // shl v17.4s, v17.4s, #11 |
| 1243 | .long 0xf9400108 // ldr x8, [x8] |
| 1244 | .long 0x4f255652 // shl v18.4s, v18.4s, #5 |
| 1245 | .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b |
| 1246 | .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| 1247 | .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b |
| 1248 | .long 0xd37ff809 // lsl x9, x0, #1 |
| 1249 | .long 0x0e612a10 // xtn v16.4h, v16.4s |
| 1250 | .long 0xfc296910 // str d16, [x8, x9] |
| 1251 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1252 | .long 0x91004021 // add x1, x1, #0x10 |
| 1253 | .long 0xd61f0060 // br x3 |
| 1254 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 1255 | HIDDEN _sk_load_4444_aarch64 |
| 1256 | .globl _sk_load_4444_aarch64 |
| 1257 | _sk_load_4444_aarch64: |
| 1258 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1259 | .long 0xd37ff809 // lsl x9, x0, #1 |
| 1260 | .long 0x4f072601 // movi v1.4s, #0xf0, lsl #8 |
| 1261 | .long 0x4f0025e2 // movi v2.4s, #0xf, lsl #8 |
| 1262 | .long 0xf9400108 // ldr x8, [x8] |
| 1263 | .long 0x4f070603 // movi v3.4s, #0xf0 |
| 1264 | .long 0x4f0005f0 // movi v16.4s, #0xf |
| 1265 | .long 0xfc696900 // ldr d0, [x8, x9] |
| 1266 | .long 0x52a6f108 // mov w8, #0x37880000 |
| 1267 | .long 0x72911128 // movk w8, #0x8889 |
| 1268 | .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| 1269 | .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b |
| 1270 | .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b |
| 1271 | .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b |
| 1272 | .long 0x4e301c10 // and v16.16b, v0.16b, v16.16b |
| 1273 | .long 0x4e040d00 // dup v0.4s, w8 |
| 1274 | .long 0x52a73108 // mov w8, #0x39880000 |
| 1275 | .long 0x72911128 // movk w8, #0x8889 |
| 1276 | .long 0x4e21d821 // scvtf v1.4s, v1.4s |
| 1277 | .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s |
| 1278 | .long 0x4e040d01 // dup v1.4s, w8 |
| 1279 | .long 0x52a77108 // mov w8, #0x3b880000 |
| 1280 | .long 0x72911128 // movk w8, #0x8889 |
| 1281 | .long 0x4e21d842 // scvtf v2.4s, v2.4s |
| 1282 | .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s |
| 1283 | .long 0x4e040d02 // dup v2.4s, w8 |
| 1284 | .long 0x52a7b108 // mov w8, #0x3d880000 |
| 1285 | .long 0x72911128 // movk w8, #0x8889 |
| 1286 | .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| 1287 | .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s |
| 1288 | .long 0x4e040d03 // dup v3.4s, w8 |
| 1289 | .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| 1290 | .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s |
| 1291 | .long 0xd61f0060 // br x3 |
| 1292 | |
| 1293 | HIDDEN _sk_store_4444_aarch64 |
| 1294 | .globl _sk_store_4444_aarch64 |
| 1295 | _sk_store_4444_aarch64: |
| 1296 | .long 0x4f01f5d0 // fmov v16.4s, #1.500000000000000000e+01 |
| 1297 | .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| 1298 | .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s |
| 1299 | .long 0xf9400028 // ldr x8, [x1] |
| 1300 | .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| 1301 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1302 | .long 0x4f2c5631 // shl v17.4s, v17.4s, #12 |
| 1303 | .long 0x4f285652 // shl v18.4s, v18.4s, #8 |
| 1304 | .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b |
| 1305 | .long 0x6e30dc52 // fmul v18.4s, v2.4s, v16.4s |
| 1306 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1307 | .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| 1308 | .long 0x4f245652 // shl v18.4s, v18.4s, #4 |
| 1309 | .long 0xf9400108 // ldr x8, [x8] |
| 1310 | .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b |
| 1311 | .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| 1312 | .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b |
| 1313 | .long 0xd37ff809 // lsl x9, x0, #1 |
| 1314 | .long 0x0e612a10 // xtn v16.4h, v16.4s |
| 1315 | .long 0xfc296910 // str d16, [x8, x9] |
| 1316 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1317 | .long 0x91004021 // add x1, x1, #0x10 |
| 1318 | .long 0xd61f0060 // br x3 |
| 1319 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1320 | HIDDEN _sk_load_8888_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1321 | .globl _sk_load_8888_aarch64 |
| 1322 | _sk_load_8888_aarch64: |
| 1323 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1324 | .long 0xd37ef409 // lsl x9, x0, #2 |
| 1325 | .long 0x6f00e621 // movi v1.2d, #0xff000000ff |
| 1326 | .long 0xf9400108 // ldr x8, [x8] |
| 1327 | .long 0x3ce96900 // ldr q0, [x8, x9] |
| 1328 | .long 0x52a77008 // mov w8, #0x3b800000 |
| 1329 | .long 0x72901028 // movk w8, #0x8081 |
| 1330 | .long 0x4e040d02 // dup v2.4s, w8 |
| 1331 | .long 0x6f380410 // ushr v16.4s, v0.4s, #8 |
| 1332 | .long 0x6f300411 // ushr v17.4s, v0.4s, #16 |
| 1333 | .long 0x4e211c03 // and v3.16b, v0.16b, v1.16b |
| 1334 | .long 0x6f280400 // ushr v0.4s, v0.4s, #24 |
| 1335 | .long 0x4e211e10 // and v16.16b, v16.16b, v1.16b |
| 1336 | .long 0x4e211e21 // and v1.16b, v17.16b, v1.16b |
| 1337 | .long 0x4e21d863 // scvtf v3.4s, v3.4s |
| 1338 | .long 0x4e21d811 // scvtf v17.4s, v0.4s |
| 1339 | .long 0x4e21da10 // scvtf v16.4s, v16.4s |
| 1340 | .long 0x4e21d832 // scvtf v18.4s, v1.4s |
| 1341 | .long 0x6e22dc60 // fmul v0.4s, v3.4s, v2.4s |
| 1342 | .long 0x6e22de23 // fmul v3.4s, v17.4s, v2.4s |
| 1343 | .long 0x6e22de01 // fmul v1.4s, v16.4s, v2.4s |
| 1344 | .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s |
| 1345 | .long 0xd61f0060 // br x3 |
| 1346 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1347 | HIDDEN _sk_store_8888_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1348 | .globl _sk_store_8888_aarch64 |
| 1349 | _sk_store_8888_aarch64: |
| 1350 | .long 0x52a86fea // mov w10, #0x437f0000 |
| 1351 | .long 0x4e040d50 // dup v16.4s, w10 |
| 1352 | .long 0xf9400028 // ldr x8, [x1] |
| 1353 | .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s |
| 1354 | .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| 1355 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1356 | .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| 1357 | .long 0x4f285652 // shl v18.4s, v18.4s, #8 |
| 1358 | .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b |
| 1359 | .long 0x6e30dc52 // fmul v18.4s, v2.4s, v16.4s |
| 1360 | .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| 1361 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1362 | .long 0xf9400108 // ldr x8, [x8] |
| 1363 | .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| 1364 | .long 0x4f305652 // shl v18.4s, v18.4s, #16 |
| 1365 | .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b |
| 1366 | .long 0x4f385610 // shl v16.4s, v16.4s, #24 |
| 1367 | .long 0xd37ef409 // lsl x9, x0, #2 |
| 1368 | .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b |
| 1369 | .long 0x3ca96910 // str q16, [x8, x9] |
| 1370 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1371 | .long 0x91004021 // add x1, x1, #0x10 |
| 1372 | .long 0xd61f0060 // br x3 |
| 1373 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1374 | HIDDEN _sk_load_f16_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1375 | .globl _sk_load_f16_aarch64 |
| 1376 | _sk_load_f16_aarch64: |
| 1377 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1378 | .long 0xf9400108 // ldr x8, [x8] |
| 1379 | .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| 1380 | .long 0x0c400510 // ld4 {v16.4h-v19.4h}, [x8] |
| 1381 | .long 0x0e217a00 // fcvtl v0.4s, v16.4h |
| 1382 | .long 0x0e217a21 // fcvtl v1.4s, v17.4h |
| 1383 | .long 0x0e217a42 // fcvtl v2.4s, v18.4h |
| 1384 | .long 0x0e217a63 // fcvtl v3.4s, v19.4h |
| 1385 | .long 0xd61f0060 // br x3 |
| 1386 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1387 | HIDDEN _sk_store_f16_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1388 | .globl _sk_store_f16_aarch64 |
| 1389 | _sk_store_f16_aarch64: |
| 1390 | .long 0xf9400028 // ldr x8, [x1] |
| 1391 | .long 0x0e216810 // fcvtn v16.4h, v0.4s |
| 1392 | .long 0x0e216831 // fcvtn v17.4h, v1.4s |
| 1393 | .long 0x0e216852 // fcvtn v18.4h, v2.4s |
| 1394 | .long 0xf9400108 // ldr x8, [x8] |
| 1395 | .long 0x0e216873 // fcvtn v19.4h, v3.4s |
| 1396 | .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| 1397 | .long 0x0c000510 // st4 {v16.4h-v19.4h}, [x8] |
| 1398 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1399 | .long 0x91004021 // add x1, x1, #0x10 |
| 1400 | .long 0xd61f0060 // br x3 |
| 1401 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 1402 | HIDDEN _sk_load_u16_be_aarch64 |
| 1403 | .globl _sk_load_u16_be_aarch64 |
| 1404 | _sk_load_u16_be_aarch64: |
| 1405 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1406 | .long 0xf9400108 // ldr x8, [x8] |
| 1407 | .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| 1408 | .long 0x0c400500 // ld4 {v0.4h-v3.4h}, [x8] |
| 1409 | .long 0x52a6f008 // mov w8, #0x37800000 |
| 1410 | .long 0x72801008 // movk w8, #0x80 |
| 1411 | .long 0x0f185410 // shl v16.4h, v0.4h, #8 |
| 1412 | .long 0x2f180411 // ushr v17.4h, v0.4h, #8 |
| 1413 | .long 0x0f185432 // shl v18.4h, v1.4h, #8 |
| 1414 | .long 0x2f180433 // ushr v19.4h, v1.4h, #8 |
| 1415 | .long 0x0f185454 // shl v20.4h, v2.4h, #8 |
| 1416 | .long 0x2f180455 // ushr v21.4h, v2.4h, #8 |
| 1417 | .long 0x0f185476 // shl v22.4h, v3.4h, #8 |
| 1418 | .long 0x2f180460 // ushr v0.4h, v3.4h, #8 |
| 1419 | .long 0x0eb11e01 // orr v1.8b, v16.8b, v17.8b |
| 1420 | .long 0x0eb31e42 // orr v2.8b, v18.8b, v19.8b |
| 1421 | .long 0x0eb51e90 // orr v16.8b, v20.8b, v21.8b |
| 1422 | .long 0x0ea01ec0 // orr v0.8b, v22.8b, v0.8b |
| 1423 | .long 0x2f10a421 // uxtl v1.4s, v1.4h |
| 1424 | .long 0x2f10a442 // uxtl v2.4s, v2.4h |
| 1425 | .long 0x2f10a610 // uxtl v16.4s, v16.4h |
| 1426 | .long 0x2f10a400 // uxtl v0.4s, v0.4h |
| 1427 | .long 0x4e040d03 // dup v3.4s, w8 |
| 1428 | .long 0x6e21d821 // ucvtf v1.4s, v1.4s |
| 1429 | .long 0x6e21d842 // ucvtf v2.4s, v2.4s |
| 1430 | .long 0x6e21da10 // ucvtf v16.4s, v16.4s |
| 1431 | .long 0x6e21d811 // ucvtf v17.4s, v0.4s |
| 1432 | .long 0x6e23dc20 // fmul v0.4s, v1.4s, v3.4s |
| 1433 | .long 0x6e23dc41 // fmul v1.4s, v2.4s, v3.4s |
| 1434 | .long 0x6e23de02 // fmul v2.4s, v16.4s, v3.4s |
| 1435 | .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s |
| 1436 | .long 0xd61f0060 // br x3 |
| 1437 | |
| 1438 | HIDDEN _sk_store_u16_be_aarch64 |
| 1439 | .globl _sk_store_u16_be_aarch64 |
| 1440 | _sk_store_u16_be_aarch64: |
| 1441 | .long 0x52a8efe9 // mov w9, #0x477f0000 |
| 1442 | .long 0x729fe009 // movk w9, #0xff00 |
| 1443 | .long 0x4e040d30 // dup v16.4s, w9 |
| 1444 | .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s |
| 1445 | .long 0xf9400028 // ldr x8, [x1] |
| 1446 | .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| 1447 | .long 0x0e612a31 // xtn v17.4h, v17.4s |
| 1448 | .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s |
| 1449 | .long 0x0f185633 // shl v19.4h, v17.4h, #8 |
| 1450 | .long 0x2f180631 // ushr v17.4h, v17.4h, #8 |
| 1451 | .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s |
| 1452 | .long 0x0eb11e75 // orr v21.8b, v19.8b, v17.8b |
| 1453 | .long 0x6e30dc51 // fmul v17.4s, v2.4s, v16.4s |
| 1454 | .long 0x0e612a52 // xtn v18.4h, v18.4s |
| 1455 | .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s |
| 1456 | .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s |
| 1457 | .long 0xf9400108 // ldr x8, [x8] |
| 1458 | .long 0x0f185654 // shl v20.4h, v18.4h, #8 |
| 1459 | .long 0x2f180652 // ushr v18.4h, v18.4h, #8 |
| 1460 | .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s |
| 1461 | .long 0x0e612a31 // xtn v17.4h, v17.4s |
| 1462 | .long 0x0eb21e96 // orr v22.8b, v20.8b, v18.8b |
| 1463 | .long 0x0e612a10 // xtn v16.4h, v16.4s |
| 1464 | .long 0x0f185632 // shl v18.4h, v17.4h, #8 |
| 1465 | .long 0x2f180631 // ushr v17.4h, v17.4h, #8 |
| 1466 | .long 0x0eb11e57 // orr v23.8b, v18.8b, v17.8b |
| 1467 | .long 0x0f185611 // shl v17.4h, v16.4h, #8 |
| 1468 | .long 0x2f180610 // ushr v16.4h, v16.4h, #8 |
| 1469 | .long 0x8b000d08 // add x8, x8, x0, lsl #3 |
| 1470 | .long 0x0eb01e38 // orr v24.8b, v17.8b, v16.8b |
| 1471 | .long 0x0c000515 // st4 {v21.4h-v24.4h}, [x8] |
| 1472 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1473 | .long 0x91004021 // add x1, x1, #0x10 |
| 1474 | .long 0xd61f0060 // br x3 |
| 1475 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1476 | HIDDEN _sk_store_f32_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1477 | .globl _sk_store_f32_aarch64 |
| 1478 | _sk_store_f32_aarch64: |
| 1479 | .long 0xf9400028 // ldr x8, [x1] |
| 1480 | .long 0xf9400108 // ldr x8, [x8] |
| 1481 | .long 0x8b001108 // add x8, x8, x0, lsl #4 |
| 1482 | .long 0x4c000900 // st4 {v0.4s-v3.4s}, [x8] |
| 1483 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1484 | .long 0x91004021 // add x1, x1, #0x10 |
| 1485 | .long 0xd61f0060 // br x3 |
| 1486 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1487 | HIDDEN _sk_clamp_x_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1488 | .globl _sk_clamp_x_aarch64 |
| 1489 | _sk_clamp_x_aarch64: |
| 1490 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1491 | .long 0x6f00e411 // movi v17.2d, #0x0 |
| 1492 | .long 0x4e20f620 // fmax v0.4s, v17.4s, v0.4s |
| 1493 | .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| 1494 | .long 0x4d40c910 // ld1r {v16.4s}, [x8] |
| 1495 | .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s |
| 1496 | .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| 1497 | .long 0xd61f0060 // br x3 |
| 1498 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1499 | HIDDEN _sk_clamp_y_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1500 | .globl _sk_clamp_y_aarch64 |
| 1501 | _sk_clamp_y_aarch64: |
| 1502 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1503 | .long 0x6f00e411 // movi v17.2d, #0x0 |
| 1504 | .long 0x4e21f621 // fmax v1.4s, v17.4s, v1.4s |
| 1505 | .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| 1506 | .long 0x4d40c910 // ld1r {v16.4s}, [x8] |
| 1507 | .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s |
| 1508 | .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| 1509 | .long 0xd61f0060 // br x3 |
| 1510 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1511 | HIDDEN _sk_repeat_x_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1512 | .globl _sk_repeat_x_aarch64 |
| 1513 | _sk_repeat_x_aarch64: |
| 1514 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1515 | .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| 1516 | .long 0xbd400110 // ldr s16, [x8] |
| 1517 | .long 0x4e040612 // dup v18.4s, v16.s[0] |
| 1518 | .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s |
| 1519 | .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s |
| 1520 | .long 0x4e219a52 // frintm v18.4s, v18.4s |
| 1521 | .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] |
| 1522 | .long 0x4eb1f400 // fmin v0.4s, v0.4s, v17.4s |
| 1523 | .long 0xd61f0060 // br x3 |
| 1524 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1525 | HIDDEN _sk_repeat_y_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1526 | .globl _sk_repeat_y_aarch64 |
| 1527 | _sk_repeat_y_aarch64: |
| 1528 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1529 | .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff |
| 1530 | .long 0xbd400110 // ldr s16, [x8] |
| 1531 | .long 0x4e040612 // dup v18.4s, v16.s[0] |
| 1532 | .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s |
| 1533 | .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s |
| 1534 | .long 0x4e219a52 // frintm v18.4s, v18.4s |
| 1535 | .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] |
| 1536 | .long 0x4eb1f421 // fmin v1.4s, v1.4s, v17.4s |
| 1537 | .long 0xd61f0060 // br x3 |
| 1538 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1539 | HIDDEN _sk_mirror_x_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1540 | .globl _sk_mirror_x_aarch64 |
| 1541 | _sk_mirror_x_aarch64: |
| 1542 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1543 | .long 0xbd400110 // ldr s16, [x8] |
| 1544 | .long 0x4e040611 // dup v17.4s, v16.s[0] |
| 1545 | .long 0x1e302a10 // fadd s16, s16, s16 |
| 1546 | .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s |
| 1547 | .long 0x4e040612 // dup v18.4s, v16.s[0] |
| 1548 | .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s |
| 1549 | .long 0x4e219a52 // frintm v18.4s, v18.4s |
| 1550 | .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0] |
| 1551 | .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff |
| 1552 | .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s |
| 1553 | .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s |
| 1554 | .long 0x4ea0f800 // fabs v0.4s, v0.4s |
| 1555 | .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s |
| 1556 | .long 0xd61f0060 // br x3 |
| 1557 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1558 | HIDDEN _sk_mirror_y_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1559 | .globl _sk_mirror_y_aarch64 |
| 1560 | _sk_mirror_y_aarch64: |
| 1561 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1562 | .long 0xbd400110 // ldr s16, [x8] |
| 1563 | .long 0x4e040611 // dup v17.4s, v16.s[0] |
| 1564 | .long 0x1e302a10 // fadd s16, s16, s16 |
| 1565 | .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s |
| 1566 | .long 0x4e040612 // dup v18.4s, v16.s[0] |
| 1567 | .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s |
| 1568 | .long 0x4e219a52 // frintm v18.4s, v18.4s |
| 1569 | .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0] |
| 1570 | .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff |
| 1571 | .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s |
| 1572 | .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s |
| 1573 | .long 0x4ea0f821 // fabs v1.4s, v1.4s |
| 1574 | .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s |
| 1575 | .long 0xd61f0060 // br x3 |
| 1576 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1577 | HIDDEN _sk_luminance_to_alpha_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1578 | .globl _sk_luminance_to_alpha_aarch64 |
| 1579 | _sk_luminance_to_alpha_aarch64: |
| 1580 | .long 0x52a7cb28 // mov w8, #0x3e590000 |
| 1581 | .long 0x72967a08 // movk w8, #0xb3d0 |
| 1582 | .long 0x4e040d11 // dup v17.4s, w8 |
| 1583 | .long 0x52a7e6e8 // mov w8, #0x3f370000 |
| 1584 | .long 0x7282eb28 // movk w8, #0x1759 |
| 1585 | .long 0x4ea01c10 // mov v16.16b, v0.16b |
| 1586 | .long 0x4e040d00 // dup v0.4s, w8 |
| 1587 | .long 0x52a7b268 // mov w8, #0x3d930000 |
| 1588 | .long 0xf8408423 // ldr x3, [x1], #8 |
| 1589 | .long 0x729bb308 // movk w8, #0xdd98 |
| 1590 | .long 0x6e20dc23 // fmul v3.4s, v1.4s, v0.4s |
| 1591 | .long 0x4e30ce23 // fmla v3.4s, v17.4s, v16.4s |
| 1592 | .long 0x4e040d10 // dup v16.4s, w8 |
| 1593 | .long 0x6f00e400 // movi v0.2d, #0x0 |
| 1594 | .long 0x6f00e401 // movi v1.2d, #0x0 |
| 1595 | .long 0x4e22ce03 // fmla v3.4s, v16.4s, v2.4s |
| 1596 | .long 0x6f00e402 // movi v2.2d, #0x0 |
| 1597 | .long 0xd61f0060 // br x3 |
| 1598 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1599 | HIDDEN _sk_matrix_2x3_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1600 | .globl _sk_matrix_2x3_aarch64 |
| 1601 | _sk_matrix_2x3_aarch64: |
| 1602 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1603 | .long 0xaa0803e9 // mov x9, x8 |
| 1604 | .long 0x9100410a // add x10, x8, #0x10 |
| 1605 | .long 0x4ddfc932 // ld1r {v18.4s}, [x9], #4 |
| 1606 | .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| 1607 | .long 0x2d415113 // ldp s19, s20, [x8, #8] |
| 1608 | .long 0x9100510a // add x10, x8, #0x14 |
| 1609 | .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| 1610 | .long 0x4f931030 // fmla v16.4s, v1.4s, v19.s[0] |
| 1611 | .long 0xbd400133 // ldr s19, [x9] |
| 1612 | .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0] |
| 1613 | .long 0x4e20ce50 // fmla v16.4s, v18.4s, v0.4s |
| 1614 | .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0] |
| 1615 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 1616 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 1617 | .long 0xd61f0060 // br x3 |
| 1618 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1619 | HIDDEN _sk_matrix_3x4_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1620 | .globl _sk_matrix_3x4_aarch64 |
| 1621 | _sk_matrix_3x4_aarch64: |
| 1622 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1623 | .long 0xaa0803e9 // mov x9, x8 |
| 1624 | .long 0x9100910a // add x10, x8, #0x24 |
| 1625 | .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4 |
| 1626 | .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| 1627 | .long 0x9100a10a // add x10, x8, #0x28 |
| 1628 | .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| 1629 | .long 0x9100b10a // add x10, x8, #0x2c |
| 1630 | .long 0x2d435514 // ldp s20, s21, [x8, #24] |
| 1631 | .long 0xbd402116 // ldr s22, [x8, #32] |
| 1632 | .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| 1633 | .long 0x4f941050 // fmla v16.4s, v2.4s, v20.s[0] |
| 1634 | .long 0x4f951051 // fmla v17.4s, v2.4s, v21.s[0] |
| 1635 | .long 0x4f961052 // fmla v18.4s, v2.4s, v22.s[0] |
| 1636 | .long 0x2d425502 // ldp s2, s21, [x8, #16] |
| 1637 | .long 0x2d415d14 // ldp s20, s23, [x8, #8] |
| 1638 | .long 0x4f821031 // fmla v17.4s, v1.4s, v2.s[0] |
| 1639 | .long 0xbd400122 // ldr s2, [x9] |
| 1640 | .long 0x4f971030 // fmla v16.4s, v1.4s, v23.s[0] |
| 1641 | .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0] |
| 1642 | .long 0x4e20ce70 // fmla v16.4s, v19.4s, v0.4s |
| 1643 | .long 0x4f941012 // fmla v18.4s, v0.4s, v20.s[0] |
| 1644 | .long 0x4f821011 // fmla v17.4s, v0.4s, v2.s[0] |
| 1645 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 1646 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 1647 | .long 0x4eb21e42 // mov v2.16b, v18.16b |
| 1648 | .long 0xd61f0060 // br x3 |
| 1649 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1650 | HIDDEN _sk_matrix_4x5_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1651 | .globl _sk_matrix_4x5_aarch64 |
| 1652 | _sk_matrix_4x5_aarch64: |
| 1653 | .long 0xf9400029 // ldr x9, [x1] |
| 1654 | .long 0xaa0903e8 // mov x8, x9 |
| 1655 | .long 0x9101012a // add x10, x9, #0x40 |
| 1656 | .long 0x4ddfc914 // ld1r {v20.4s}, [x8], #4 |
| 1657 | .long 0x4d40c950 // ld1r {v16.4s}, [x10] |
| 1658 | .long 0x9101112a // add x10, x9, #0x44 |
| 1659 | .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| 1660 | .long 0x9101212a // add x10, x9, #0x48 |
| 1661 | .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| 1662 | .long 0x2d465533 // ldp s19, s21, [x9, #48] |
| 1663 | .long 0x2d475d36 // ldp s22, s23, [x9, #56] |
| 1664 | .long 0x9101312a // add x10, x9, #0x4c |
| 1665 | .long 0xf9400423 // ldr x3, [x1, #8] |
| 1666 | .long 0x4f931070 // fmla v16.4s, v3.4s, v19.s[0] |
| 1667 | .long 0x4d40c953 // ld1r {v19.4s}, [x10] |
| 1668 | .long 0x4f951071 // fmla v17.4s, v3.4s, v21.s[0] |
| 1669 | .long 0x4f961072 // fmla v18.4s, v3.4s, v22.s[0] |
| 1670 | .long 0x2d445935 // ldp s21, s22, [x9, #32] |
| 1671 | .long 0x4f971073 // fmla v19.4s, v3.4s, v23.s[0] |
| 1672 | .long 0x2d455d23 // ldp s3, s23, [x9, #40] |
| 1673 | .long 0x91004021 // add x1, x1, #0x10 |
| 1674 | .long 0x4f951050 // fmla v16.4s, v2.4s, v21.s[0] |
| 1675 | .long 0x4f961051 // fmla v17.4s, v2.4s, v22.s[0] |
| 1676 | .long 0x2d425935 // ldp s21, s22, [x9, #16] |
| 1677 | .long 0x4f971053 // fmla v19.4s, v2.4s, v23.s[0] |
| 1678 | .long 0x4f831052 // fmla v18.4s, v2.4s, v3.s[0] |
| 1679 | .long 0x2d410d22 // ldp s2, s3, [x9, #8] |
| 1680 | .long 0x4f951030 // fmla v16.4s, v1.4s, v21.s[0] |
| 1681 | .long 0x2d435d35 // ldp s21, s23, [x9, #24] |
| 1682 | .long 0x4f961031 // fmla v17.4s, v1.4s, v22.s[0] |
| 1683 | .long 0xbd400116 // ldr s22, [x8] |
| 1684 | .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s |
| 1685 | .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0] |
| 1686 | .long 0x4f971033 // fmla v19.4s, v1.4s, v23.s[0] |
| 1687 | .long 0x4f821012 // fmla v18.4s, v0.4s, v2.s[0] |
| 1688 | .long 0x4f831013 // fmla v19.4s, v0.4s, v3.s[0] |
| 1689 | .long 0x4f961011 // fmla v17.4s, v0.4s, v22.s[0] |
| 1690 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 1691 | .long 0x4eb11e21 // mov v1.16b, v17.16b |
| 1692 | .long 0x4eb21e42 // mov v2.16b, v18.16b |
| 1693 | .long 0x4eb31e63 // mov v3.16b, v19.16b |
| 1694 | .long 0xd61f0060 // br x3 |
| 1695 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1696 | HIDDEN _sk_matrix_perspective_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1697 | .globl _sk_matrix_perspective_aarch64 |
| 1698 | _sk_matrix_perspective_aarch64: |
| 1699 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1700 | .long 0xaa0803e9 // mov x9, x8 |
| 1701 | .long 0x9100510a // add x10, x8, #0x14 |
| 1702 | .long 0x4ddfc930 // ld1r {v16.4s}, [x9], #4 |
| 1703 | .long 0x4d40c951 // ld1r {v17.4s}, [x10] |
| 1704 | .long 0x9100810a // add x10, x8, #0x20 |
| 1705 | .long 0x4d40c952 // ld1r {v18.4s}, [x10] |
| 1706 | .long 0x2d41d113 // ldp s19, s20, [x8, #12] |
| 1707 | .long 0x2d435915 // ldp s21, s22, [x8, #24] |
| 1708 | .long 0x91002108 // add x8, x8, #0x8 |
| 1709 | .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0] |
| 1710 | .long 0x4d40c914 // ld1r {v20.4s}, [x8] |
| 1711 | .long 0x4f961032 // fmla v18.4s, v1.4s, v22.s[0] |
| 1712 | .long 0xbd400136 // ldr s22, [x9] |
| 1713 | .long 0x4f951012 // fmla v18.4s, v0.4s, v21.s[0] |
| 1714 | .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0] |
| 1715 | .long 0x4f961034 // fmla v20.4s, v1.4s, v22.s[0] |
| 1716 | .long 0x4ea1da41 // frecpe v1.4s, v18.4s |
| 1717 | .long 0x4e21fe52 // frecps v18.4s, v18.4s, v1.4s |
| 1718 | .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s |
| 1719 | .long 0x4e20ce14 // fmla v20.4s, v16.4s, v0.4s |
| 1720 | .long 0x6e32de21 // fmul v1.4s, v17.4s, v18.4s |
| 1721 | .long 0x6e32de80 // fmul v0.4s, v20.4s, v18.4s |
| 1722 | .long 0xd61f0060 // br x3 |
| 1723 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1724 | HIDDEN _sk_linear_gradient_2stops_aarch64 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1725 | .globl _sk_linear_gradient_2stops_aarch64 |
| 1726 | _sk_linear_gradient_2stops_aarch64: |
| 1727 | .long 0xa8c10c28 // ldp x8, x3, [x1], #16 |
| 1728 | .long 0xad404503 // ldp q3, q17, [x8] |
| 1729 | .long 0x4e040470 // dup v16.4s, v3.s[0] |
| 1730 | .long 0x4e0c0461 // dup v1.4s, v3.s[1] |
| 1731 | .long 0x4e140462 // dup v2.4s, v3.s[2] |
| 1732 | .long 0x4e1c0463 // dup v3.4s, v3.s[3] |
| 1733 | .long 0x4f911010 // fmla v16.4s, v0.4s, v17.s[0] |
| 1734 | .long 0x4fb11001 // fmla v1.4s, v0.4s, v17.s[1] |
| 1735 | .long 0x4f911802 // fmla v2.4s, v0.4s, v17.s[2] |
| 1736 | .long 0x4fb11803 // fmla v3.4s, v0.4s, v17.s[3] |
| 1737 | .long 0x4eb01e00 // mov v0.16b, v16.16b |
| 1738 | .long 0xd61f0060 // br x3 |
| 1739 | #elif defined(__arm__) |
| 1740 | .balign 4 |
| 1741 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1742 | HIDDEN _sk_start_pipeline_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1743 | .globl _sk_start_pipeline_vfp4 |
| 1744 | _sk_start_pipeline_vfp4: |
| 1745 | .long 0xe92d41f0 // push {r4, r5, r6, r7, r8, lr} |
| 1746 | .long 0xe1a04000 // mov r4, r0 |
| 1747 | .long 0xe2840002 // add r0, r4, #2 |
| 1748 | .long 0xe1a05003 // mov r5, r3 |
| 1749 | .long 0xe1a08002 // mov r8, r2 |
| 1750 | .long 0xe1a07001 // mov r7, r1 |
| 1751 | .long 0xe1500005 // cmp r0, r5 |
| 1752 | .long 0x8a000010 // bhi 64 <sk_start_pipeline_vfp4+0x64> |
| 1753 | .long 0xe4976004 // ldr r6, [r7], #4 |
| 1754 | .long 0xf2800010 // vmov.i32 d0, #0 |
| 1755 | .long 0xe1a00004 // mov r0, r4 |
| 1756 | .long 0xf2801010 // vmov.i32 d1, #0 |
| 1757 | .long 0xe1a01007 // mov r1, r7 |
| 1758 | .long 0xf2802010 // vmov.i32 d2, #0 |
| 1759 | .long 0xe1a02008 // mov r2, r8 |
| 1760 | .long 0xf2803010 // vmov.i32 d3, #0 |
| 1761 | .long 0xf2804010 // vmov.i32 d4, #0 |
| 1762 | .long 0xf2805010 // vmov.i32 d5, #0 |
| 1763 | .long 0xf2806010 // vmov.i32 d6, #0 |
| 1764 | .long 0xf2807010 // vmov.i32 d7, #0 |
| 1765 | .long 0xe12fff36 // blx r6 |
| 1766 | .long 0xe2840004 // add r0, r4, #4 |
| 1767 | .long 0xe2844002 // add r4, r4, #2 |
| 1768 | .long 0xe1500005 // cmp r0, r5 |
| 1769 | .long 0x9affffef // bls 24 <sk_start_pipeline_vfp4+0x24> |
| 1770 | .long 0xe1a00004 // mov r0, r4 |
| 1771 | .long 0xe8bd81f0 // pop {r4, r5, r6, r7, r8, pc} |
| 1772 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1773 | HIDDEN _sk_just_return_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1774 | .globl _sk_just_return_vfp4 |
| 1775 | _sk_just_return_vfp4: |
| 1776 | .long 0xe12fff1e // bx lr |
| 1777 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1778 | HIDDEN _sk_seed_shader_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1779 | .globl _sk_seed_shader_vfp4 |
| 1780 | _sk_seed_shader_vfp4: |
| 1781 | .long 0xee800b90 // vdup.32 d16, r0 |
| 1782 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 1783 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 1784 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 1785 | .long 0xedd23b00 // vldr d19, [r2] |
| 1786 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 1787 | .long 0xf2872f10 // vmov.f32 d2, #1 |
| 1788 | .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| 1789 | .long 0xe2811008 // add r1, r1, #8 |
| 1790 | .long 0xf2400da1 // vadd.f32 d16, d16, d17 |
| 1791 | .long 0xf2803010 // vmov.i32 d3, #0 |
| 1792 | .long 0xf2804010 // vmov.i32 d4, #0 |
| 1793 | .long 0xf2021da1 // vadd.f32 d1, d18, d17 |
| 1794 | .long 0xf2000da3 // vadd.f32 d0, d16, d19 |
| 1795 | .long 0xf2805010 // vmov.i32 d5, #0 |
| 1796 | .long 0xf2806010 // vmov.i32 d6, #0 |
| 1797 | .long 0xf2807010 // vmov.i32 d7, #0 |
| 1798 | .long 0xe12fff1c // bx ip |
| 1799 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1800 | HIDDEN _sk_constant_color_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1801 | .globl _sk_constant_color_vfp4 |
| 1802 | _sk_constant_color_vfp4: |
| 1803 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 1804 | .long 0xe2811008 // add r1, r1, #8 |
| 1805 | .long 0xf4630a0f // vld1.8 {d16-d17}, [r3] |
| 1806 | .long 0xf3b40c20 // vdup.32 d0, d16[0] |
| 1807 | .long 0xf3bc1c20 // vdup.32 d1, d16[1] |
| 1808 | .long 0xf3b42c21 // vdup.32 d2, d17[0] |
| 1809 | .long 0xf3bc3c21 // vdup.32 d3, d17[1] |
| 1810 | .long 0xe12fff1c // bx ip |
| 1811 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1812 | HIDDEN _sk_clear_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1813 | .globl _sk_clear_vfp4 |
| 1814 | _sk_clear_vfp4: |
| 1815 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1816 | .long 0xf2800010 // vmov.i32 d0, #0 |
| 1817 | .long 0xf2801010 // vmov.i32 d1, #0 |
| 1818 | .long 0xf2802010 // vmov.i32 d2, #0 |
| 1819 | .long 0xf2803010 // vmov.i32 d3, #0 |
| 1820 | .long 0xe12fff13 // bx r3 |
| 1821 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 1822 | HIDDEN _sk_srcatop_vfp4 |
| 1823 | .globl _sk_srcatop_vfp4 |
| 1824 | _sk_srcatop_vfp4: |
| 1825 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1826 | .long 0xe4913004 // ldr r3, [r1], #4 |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 1827 | .long 0xf2603d83 // vsub.f32 d19, d16, d3 |
| 1828 | .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| 1829 | .long 0xf3430d94 // vmul.f32 d16, d19, d4 |
| 1830 | .long 0xf3431d95 // vmul.f32 d17, d19, d5 |
| 1831 | .long 0xf3432d96 // vmul.f32 d18, d19, d6 |
| 1832 | .long 0xf2400c17 // vfma.f32 d16, d0, d7 |
| 1833 | .long 0xf2411c17 // vfma.f32 d17, d1, d7 |
| 1834 | .long 0xf2422c17 // vfma.f32 d18, d2, d7 |
| 1835 | .long 0xf2033c97 // vfma.f32 d3, d19, d7 |
| 1836 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 1837 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 1838 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 1839 | .long 0xe12fff13 // bx r3 |
| 1840 | |
| 1841 | HIDDEN _sk_dstatop_vfp4 |
| 1842 | .globl _sk_dstatop_vfp4 |
| 1843 | _sk_dstatop_vfp4: |
| 1844 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1845 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1846 | .long 0xf3431d15 // vmul.f32 d17, d3, d5 |
| 1847 | .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| 1848 | .long 0xf3430d14 // vmul.f32 d16, d3, d4 |
| 1849 | .long 0xf3432d16 // vmul.f32 d18, d3, d6 |
| 1850 | .long 0xf3433d17 // vmul.f32 d19, d3, d7 |
| 1851 | .long 0xf2440c90 // vfma.f32 d16, d20, d0 |
| 1852 | .long 0xf2441c91 // vfma.f32 d17, d20, d1 |
| 1853 | .long 0xf2442c92 // vfma.f32 d18, d20, d2 |
| 1854 | .long 0xf2443c93 // vfma.f32 d19, d20, d3 |
| 1855 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 1856 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 1857 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 1858 | .long 0xf22331b3 // vorr d3, d19, d19 |
| 1859 | .long 0xe12fff13 // bx r3 |
| 1860 | |
| 1861 | HIDDEN _sk_srcin_vfp4 |
| 1862 | .globl _sk_srcin_vfp4 |
| 1863 | _sk_srcin_vfp4: |
| 1864 | .long 0xf3000d17 // vmul.f32 d0, d0, d7 |
| 1865 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1866 | .long 0xf3011d17 // vmul.f32 d1, d1, d7 |
| 1867 | .long 0xf3022d17 // vmul.f32 d2, d2, d7 |
| 1868 | .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| 1869 | .long 0xe12fff13 // bx r3 |
| 1870 | |
| 1871 | HIDDEN _sk_dstin_vfp4 |
| 1872 | .globl _sk_dstin_vfp4 |
| 1873 | _sk_dstin_vfp4: |
| 1874 | .long 0xf3030d14 // vmul.f32 d0, d3, d4 |
| 1875 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1876 | .long 0xf3031d15 // vmul.f32 d1, d3, d5 |
| 1877 | .long 0xf3032d16 // vmul.f32 d2, d3, d6 |
| 1878 | .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| 1879 | .long 0xe12fff13 // bx r3 |
| 1880 | |
| 1881 | HIDDEN _sk_srcout_vfp4 |
| 1882 | .globl _sk_srcout_vfp4 |
| 1883 | _sk_srcout_vfp4: |
| 1884 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1885 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1886 | .long 0xf2600d87 // vsub.f32 d16, d16, d7 |
| 1887 | .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| 1888 | .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| 1889 | .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| 1890 | .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| 1891 | .long 0xe12fff13 // bx r3 |
| 1892 | |
| 1893 | HIDDEN _sk_dstout_vfp4 |
| 1894 | .globl _sk_dstout_vfp4 |
| 1895 | _sk_dstout_vfp4: |
| 1896 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1897 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1898 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 1899 | .long 0xf3000d94 // vmul.f32 d0, d16, d4 |
| 1900 | .long 0xf3001d95 // vmul.f32 d1, d16, d5 |
| 1901 | .long 0xf3002d96 // vmul.f32 d2, d16, d6 |
| 1902 | .long 0xf3003d97 // vmul.f32 d3, d16, d7 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1903 | .long 0xe12fff13 // bx r3 |
| 1904 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1905 | HIDDEN _sk_srcover_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1906 | .globl _sk_srcover_vfp4 |
| 1907 | _sk_srcover_vfp4: |
| 1908 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1909 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1910 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 1911 | .long 0xf2040c30 // vfma.f32 d0, d4, d16 |
| 1912 | .long 0xf2051c30 // vfma.f32 d1, d5, d16 |
| 1913 | .long 0xf2062c30 // vfma.f32 d2, d6, d16 |
| 1914 | .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| 1915 | .long 0xe12fff13 // bx r3 |
| 1916 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 1917 | HIDDEN _sk_dstover_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 1918 | .globl _sk_dstover_vfp4 |
| 1919 | _sk_dstover_vfp4: |
| 1920 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1921 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1922 | .long 0xf2651115 // vorr d17, d5, d5 |
| 1923 | .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| 1924 | .long 0xf2640114 // vorr d16, d4, d4 |
| 1925 | .long 0xf2662116 // vorr d18, d6, d6 |
| 1926 | .long 0xf2673117 // vorr d19, d7, d7 |
| 1927 | .long 0xf2400c34 // vfma.f32 d16, d0, d20 |
| 1928 | .long 0xf2411c34 // vfma.f32 d17, d1, d20 |
| 1929 | .long 0xf2422c34 // vfma.f32 d18, d2, d20 |
| 1930 | .long 0xf2433c34 // vfma.f32 d19, d3, d20 |
| 1931 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 1932 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 1933 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 1934 | .long 0xf22331b3 // vorr d3, d19, d19 |
| 1935 | .long 0xe12fff13 // bx r3 |
| 1936 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 1937 | HIDDEN _sk_modulate_vfp4 |
| 1938 | .globl _sk_modulate_vfp4 |
| 1939 | _sk_modulate_vfp4: |
| 1940 | .long 0xf3000d14 // vmul.f32 d0, d0, d4 |
| 1941 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1942 | .long 0xf3011d15 // vmul.f32 d1, d1, d5 |
| 1943 | .long 0xf3022d16 // vmul.f32 d2, d2, d6 |
| 1944 | .long 0xf3033d17 // vmul.f32 d3, d3, d7 |
| 1945 | .long 0xe12fff13 // bx r3 |
| 1946 | |
| 1947 | HIDDEN _sk_multiply_vfp4 |
| 1948 | .globl _sk_multiply_vfp4 |
| 1949 | _sk_multiply_vfp4: |
| 1950 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 1951 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1952 | .long 0xf2601d87 // vsub.f32 d17, d16, d7 |
| 1953 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 1954 | .long 0xf3412d90 // vmul.f32 d18, d17, d0 |
| 1955 | .long 0xf3403d94 // vmul.f32 d19, d16, d4 |
| 1956 | .long 0xf3414d91 // vmul.f32 d20, d17, d1 |
| 1957 | .long 0xf3405d95 // vmul.f32 d21, d16, d5 |
| 1958 | .long 0xf3416d92 // vmul.f32 d22, d17, d2 |
| 1959 | .long 0xf3418d93 // vmul.f32 d24, d17, d3 |
| 1960 | .long 0xf3407d96 // vmul.f32 d23, d16, d6 |
| 1961 | .long 0xf3409d97 // vmul.f32 d25, d16, d7 |
| 1962 | .long 0xf2430da2 // vadd.f32 d16, d19, d18 |
| 1963 | .long 0xf2451da4 // vadd.f32 d17, d21, d20 |
| 1964 | .long 0xf2472da6 // vadd.f32 d18, d23, d22 |
| 1965 | .long 0xf2493da8 // vadd.f32 d19, d25, d24 |
| 1966 | .long 0xf2400c14 // vfma.f32 d16, d0, d4 |
| 1967 | .long 0xf2411c15 // vfma.f32 d17, d1, d5 |
| 1968 | .long 0xf2422c16 // vfma.f32 d18, d2, d6 |
| 1969 | .long 0xf2433c17 // vfma.f32 d19, d3, d7 |
| 1970 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 1971 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 1972 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 1973 | .long 0xf22331b3 // vorr d3, d19, d19 |
| 1974 | .long 0xe12fff13 // bx r3 |
| 1975 | |
| 1976 | HIDDEN _sk_plus__vfp4 |
| 1977 | .globl _sk_plus__vfp4 |
| 1978 | _sk_plus__vfp4: |
| 1979 | .long 0xf2000d04 // vadd.f32 d0, d0, d4 |
| 1980 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1981 | .long 0xf2011d05 // vadd.f32 d1, d1, d5 |
| 1982 | .long 0xf2022d06 // vadd.f32 d2, d2, d6 |
| 1983 | .long 0xf2033d07 // vadd.f32 d3, d3, d7 |
| 1984 | .long 0xe12fff13 // bx r3 |
| 1985 | |
| 1986 | HIDDEN _sk_screen_vfp4 |
| 1987 | .globl _sk_screen_vfp4 |
| 1988 | _sk_screen_vfp4: |
| 1989 | .long 0xf2400d04 // vadd.f32 d16, d0, d4 |
| 1990 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 1991 | .long 0xf2411d05 // vadd.f32 d17, d1, d5 |
| 1992 | .long 0xf2422d06 // vadd.f32 d18, d2, d6 |
| 1993 | .long 0xf2433d07 // vadd.f32 d19, d3, d7 |
| 1994 | .long 0xf2600c14 // vfms.f32 d16, d0, d4 |
| 1995 | .long 0xf2611c15 // vfms.f32 d17, d1, d5 |
| 1996 | .long 0xf2622c16 // vfms.f32 d18, d2, d6 |
| 1997 | .long 0xf2633c17 // vfms.f32 d19, d3, d7 |
| 1998 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 1999 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 2000 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 2001 | .long 0xf22331b3 // vorr d3, d19, d19 |
| 2002 | .long 0xe12fff13 // bx r3 |
| 2003 | |
| 2004 | HIDDEN _sk_xor__vfp4 |
| 2005 | .globl _sk_xor__vfp4 |
| 2006 | _sk_xor__vfp4: |
| 2007 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2008 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2009 | .long 0xf2603d83 // vsub.f32 d19, d16, d3 |
| 2010 | .long 0xf2604d87 // vsub.f32 d20, d16, d7 |
| 2011 | .long 0xf3430d94 // vmul.f32 d16, d19, d4 |
| 2012 | .long 0xf3431d95 // vmul.f32 d17, d19, d5 |
| 2013 | .long 0xf3432d96 // vmul.f32 d18, d19, d6 |
| 2014 | .long 0xf3433d97 // vmul.f32 d19, d19, d7 |
| 2015 | .long 0xf2440c90 // vfma.f32 d16, d20, d0 |
| 2016 | .long 0xf2441c91 // vfma.f32 d17, d20, d1 |
| 2017 | .long 0xf2442c92 // vfma.f32 d18, d20, d2 |
| 2018 | .long 0xf2443c93 // vfma.f32 d19, d20, d3 |
| 2019 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 2020 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 2021 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 2022 | .long 0xf22331b3 // vorr d3, d19, d19 |
| 2023 | .long 0xe12fff13 // bx r3 |
| 2024 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 2025 | HIDDEN _sk_darken_vfp4 |
| 2026 | .globl _sk_darken_vfp4 |
| 2027 | _sk_darken_vfp4: |
| 2028 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2029 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2030 | .long 0xf3431d14 // vmul.f32 d17, d3, d4 |
| 2031 | .long 0xf3402d17 // vmul.f32 d18, d0, d7 |
| 2032 | .long 0xf3433d15 // vmul.f32 d19, d3, d5 |
| 2033 | .long 0xf3414d17 // vmul.f32 d20, d1, d7 |
| 2034 | .long 0xf3435d16 // vmul.f32 d21, d3, d6 |
| 2035 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 2036 | .long 0xf3426d17 // vmul.f32 d22, d2, d7 |
| 2037 | .long 0xf2421fa1 // vmax.f32 d17, d18, d17 |
| 2038 | .long 0xf2407d04 // vadd.f32 d23, d0, d4 |
| 2039 | .long 0xf2443fa3 // vmax.f32 d19, d20, d19 |
| 2040 | .long 0xf2412d05 // vadd.f32 d18, d1, d5 |
| 2041 | .long 0xf2424d06 // vadd.f32 d20, d2, d6 |
| 2042 | .long 0xf2465fa5 // vmax.f32 d21, d22, d21 |
| 2043 | .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| 2044 | .long 0xf2270da1 // vsub.f32 d0, d23, d17 |
| 2045 | .long 0xf2221da3 // vsub.f32 d1, d18, d19 |
| 2046 | .long 0xf2242da5 // vsub.f32 d2, d20, d21 |
| 2047 | .long 0xe12fff13 // bx r3 |
| 2048 | |
| 2049 | HIDDEN _sk_lighten_vfp4 |
| 2050 | .globl _sk_lighten_vfp4 |
| 2051 | _sk_lighten_vfp4: |
| 2052 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2053 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2054 | .long 0xf3431d14 // vmul.f32 d17, d3, d4 |
| 2055 | .long 0xf3402d17 // vmul.f32 d18, d0, d7 |
| 2056 | .long 0xf3433d15 // vmul.f32 d19, d3, d5 |
| 2057 | .long 0xf3414d17 // vmul.f32 d20, d1, d7 |
| 2058 | .long 0xf3435d16 // vmul.f32 d21, d3, d6 |
| 2059 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 2060 | .long 0xf3426d17 // vmul.f32 d22, d2, d7 |
| 2061 | .long 0xf2621fa1 // vmin.f32 d17, d18, d17 |
| 2062 | .long 0xf2407d04 // vadd.f32 d23, d0, d4 |
| 2063 | .long 0xf2643fa3 // vmin.f32 d19, d20, d19 |
| 2064 | .long 0xf2412d05 // vadd.f32 d18, d1, d5 |
| 2065 | .long 0xf2424d06 // vadd.f32 d20, d2, d6 |
| 2066 | .long 0xf2665fa5 // vmin.f32 d21, d22, d21 |
| 2067 | .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| 2068 | .long 0xf2270da1 // vsub.f32 d0, d23, d17 |
| 2069 | .long 0xf2221da3 // vsub.f32 d1, d18, d19 |
| 2070 | .long 0xf2242da5 // vsub.f32 d2, d20, d21 |
| 2071 | .long 0xe12fff13 // bx r3 |
| 2072 | |
| 2073 | HIDDEN _sk_difference_vfp4 |
| 2074 | .globl _sk_difference_vfp4 |
| 2075 | _sk_difference_vfp4: |
| 2076 | .long 0xf3430d14 // vmul.f32 d16, d3, d4 |
| 2077 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2078 | .long 0xf3401d17 // vmul.f32 d17, d0, d7 |
| 2079 | .long 0xf3432d15 // vmul.f32 d18, d3, d5 |
| 2080 | .long 0xf3413d17 // vmul.f32 d19, d1, d7 |
| 2081 | .long 0xf3434d16 // vmul.f32 d20, d3, d6 |
| 2082 | .long 0xf3425d17 // vmul.f32 d21, d2, d7 |
| 2083 | .long 0xf2c76f10 // vmov.f32 d22, #1 |
| 2084 | .long 0xf2610fa0 // vmin.f32 d16, d17, d16 |
| 2085 | .long 0xf2631fa2 // vmin.f32 d17, d19, d18 |
| 2086 | .long 0xf2662d83 // vsub.f32 d18, d22, d3 |
| 2087 | .long 0xf2653fa4 // vmin.f32 d19, d21, d20 |
| 2088 | .long 0xf2404d04 // vadd.f32 d20, d0, d4 |
| 2089 | .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| 2090 | .long 0xf2073c32 // vfma.f32 d3, d7, d18 |
| 2091 | .long 0xf2415d05 // vadd.f32 d21, d1, d5 |
| 2092 | .long 0xf2411da1 // vadd.f32 d17, d17, d17 |
| 2093 | .long 0xf2426d06 // vadd.f32 d22, d2, d6 |
| 2094 | .long 0xf2432da3 // vadd.f32 d18, d19, d19 |
| 2095 | .long 0xf2240da0 // vsub.f32 d0, d20, d16 |
| 2096 | .long 0xf2251da1 // vsub.f32 d1, d21, d17 |
| 2097 | .long 0xf2262da2 // vsub.f32 d2, d22, d18 |
| 2098 | .long 0xe12fff13 // bx r3 |
| 2099 | |
| 2100 | HIDDEN _sk_exclusion_vfp4 |
| 2101 | .globl _sk_exclusion_vfp4 |
| 2102 | _sk_exclusion_vfp4: |
| 2103 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2104 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2105 | .long 0xf3401d14 // vmul.f32 d17, d0, d4 |
| 2106 | .long 0xf3412d15 // vmul.f32 d18, d1, d5 |
| 2107 | .long 0xf3423d16 // vmul.f32 d19, d2, d6 |
| 2108 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 2109 | .long 0xf2404d04 // vadd.f32 d20, d0, d4 |
| 2110 | .long 0xf2411da1 // vadd.f32 d17, d17, d17 |
| 2111 | .long 0xf2415d05 // vadd.f32 d21, d1, d5 |
| 2112 | .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| 2113 | .long 0xf2426d06 // vadd.f32 d22, d2, d6 |
| 2114 | .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| 2115 | .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| 2116 | .long 0xf2240da1 // vsub.f32 d0, d20, d17 |
| 2117 | .long 0xf2251da2 // vsub.f32 d1, d21, d18 |
| 2118 | .long 0xf2262da3 // vsub.f32 d2, d22, d19 |
| 2119 | .long 0xe12fff13 // bx r3 |
| 2120 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 2121 | HIDDEN _sk_colorburn_vfp4 |
| 2122 | .globl _sk_colorburn_vfp4 |
| 2123 | _sk_colorburn_vfp4: |
| 2124 | .long 0xed2d8b08 // vpush {d8-d11} |
| 2125 | .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| 2126 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2127 | .long 0xf2671d06 // vsub.f32 d17, d7, d6 |
| 2128 | .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| 2129 | .long 0xf3008d93 // vmul.f32 d8, d16, d3 |
| 2130 | .long 0xf3019d93 // vmul.f32 d9, d17, d3 |
| 2131 | .long 0xf302ad93 // vmul.f32 d10, d18, d3 |
| 2132 | .long 0xf2c71f10 // vmov.f32 d17, #1 |
| 2133 | .long 0xeec8baa0 // vdiv.f32 s23, s17, s1 |
| 2134 | .long 0xee88ba00 // vdiv.f32 s22, s16, s0 |
| 2135 | .long 0xeec98aa2 // vdiv.f32 s17, s19, s5 |
| 2136 | .long 0xee898a02 // vdiv.f32 s16, s18, s4 |
| 2137 | .long 0xeeca9aa1 // vdiv.f32 s19, s21, s3 |
| 2138 | .long 0xee8a9a01 // vdiv.f32 s18, s20, s2 |
| 2139 | .long 0xf2672f08 // vmin.f32 d18, d7, d8 |
| 2140 | .long 0xf2673f09 // vmin.f32 d19, d7, d9 |
| 2141 | .long 0xf2670f0b // vmin.f32 d16, d7, d11 |
| 2142 | .long 0xf2614d87 // vsub.f32 d20, d17, d7 |
| 2143 | .long 0xf2672d22 // vsub.f32 d18, d7, d18 |
| 2144 | .long 0xf2673d23 // vsub.f32 d19, d7, d19 |
| 2145 | .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| 2146 | .long 0xf2670d20 // vsub.f32 d16, d7, d16 |
| 2147 | .long 0xf3445d90 // vmul.f32 d21, d20, d0 |
| 2148 | .long 0xf3446d92 // vmul.f32 d22, d20, d2 |
| 2149 | .long 0xf3422d93 // vmul.f32 d18, d18, d3 |
| 2150 | .long 0xf3444d91 // vmul.f32 d20, d20, d1 |
| 2151 | .long 0xf3433d93 // vmul.f32 d19, d19, d3 |
| 2152 | .long 0xf3400d93 // vmul.f32 d16, d16, d3 |
| 2153 | .long 0xf3417d95 // vmul.f32 d23, d17, d5 |
| 2154 | .long 0xf3418d94 // vmul.f32 d24, d17, d4 |
| 2155 | .long 0xf3419d96 // vmul.f32 d25, d17, d6 |
| 2156 | .long 0xf2443da3 // vadd.f32 d19, d20, d19 |
| 2157 | .long 0xf2462da2 // vadd.f32 d18, d22, d18 |
| 2158 | .long 0xf245ada0 // vadd.f32 d26, d21, d16 |
| 2159 | .long 0xf247bd81 // vadd.f32 d27, d23, d1 |
| 2160 | .long 0xf248cd80 // vadd.f32 d28, d24, d0 |
| 2161 | .long 0xf249dd82 // vadd.f32 d29, d25, d2 |
| 2162 | .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| 2163 | .long 0xf2499da2 // vadd.f32 d25, d25, d18 |
| 2164 | .long 0xf2473da3 // vadd.f32 d19, d23, d19 |
| 2165 | .long 0xf3f97501 // vceq.f32 d23, d1, #0 |
| 2166 | .long 0xf2455d84 // vadd.f32 d21, d21, d4 |
| 2167 | .long 0xf2444d85 // vadd.f32 d20, d20, d5 |
| 2168 | .long 0xf2440e07 // vceq.f32 d16, d4, d7 |
| 2169 | .long 0xf2466d86 // vadd.f32 d22, d22, d6 |
| 2170 | .long 0xf2451e07 // vceq.f32 d17, d5, d7 |
| 2171 | .long 0xf2462e07 // vceq.f32 d18, d6, d7 |
| 2172 | .long 0xf35b71b3 // vbsl d23, d27, d19 |
| 2173 | .long 0xf3f93500 // vceq.f32 d19, d0, #0 |
| 2174 | .long 0xf2488daa // vadd.f32 d24, d24, d26 |
| 2175 | .long 0xf35c31b8 // vbsl d19, d28, d24 |
| 2176 | .long 0xf3f98502 // vceq.f32 d24, d2, #0 |
| 2177 | .long 0xf35d81b9 // vbsl d24, d29, d25 |
| 2178 | .long 0xf35501b3 // vbsl d16, d21, d19 |
| 2179 | .long 0xf35411b7 // vbsl d17, d20, d23 |
| 2180 | .long 0xf35621b8 // vbsl d18, d22, d24 |
| 2181 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 2182 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 2183 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 2184 | .long 0xecbd8b08 // vpop {d8-d11} |
| 2185 | .long 0xe12fff13 // bx r3 |
| 2186 | |
| 2187 | HIDDEN _sk_colordodge_vfp4 |
| 2188 | .globl _sk_colordodge_vfp4 |
| 2189 | _sk_colordodge_vfp4: |
| 2190 | .long 0xed2d8b0e // vpush {d8-d14} |
| 2191 | .long 0xf2238d02 // vsub.f32 d8, d3, d2 |
| 2192 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2193 | .long 0xf3039d16 // vmul.f32 d9, d3, d6 |
| 2194 | .long 0xf223ad01 // vsub.f32 d10, d3, d1 |
| 2195 | .long 0xf303bd15 // vmul.f32 d11, d3, d5 |
| 2196 | .long 0xf223cd00 // vsub.f32 d12, d3, d0 |
| 2197 | .long 0xf303dd14 // vmul.f32 d13, d3, d4 |
| 2198 | .long 0xeec9eaa8 // vdiv.f32 s29, s19, s17 |
| 2199 | .long 0xee89ea08 // vdiv.f32 s28, s18, s16 |
| 2200 | .long 0xeecb8aaa // vdiv.f32 s17, s23, s21 |
| 2201 | .long 0xeecd9aac // vdiv.f32 s19, s27, s25 |
| 2202 | .long 0xee8b8a0a // vdiv.f32 s16, s22, s20 |
| 2203 | .long 0xee8d9a0c // vdiv.f32 s18, s26, s24 |
| 2204 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2205 | .long 0xf2672f0e // vmin.f32 d18, d7, d14 |
| 2206 | .long 0xf2601d87 // vsub.f32 d17, d16, d7 |
| 2207 | .long 0xf2673f08 // vmin.f32 d19, d7, d8 |
| 2208 | .long 0xf2674f09 // vmin.f32 d20, d7, d9 |
| 2209 | .long 0xf2600d83 // vsub.f32 d16, d16, d3 |
| 2210 | .long 0xf3415d92 // vmul.f32 d21, d17, d2 |
| 2211 | .long 0xf3422d93 // vmul.f32 d18, d18, d3 |
| 2212 | .long 0xf3416d91 // vmul.f32 d22, d17, d1 |
| 2213 | .long 0xf3433d93 // vmul.f32 d19, d19, d3 |
| 2214 | .long 0xf3411d90 // vmul.f32 d17, d17, d0 |
| 2215 | .long 0xf3444d93 // vmul.f32 d20, d20, d3 |
| 2216 | .long 0xf3407d95 // vmul.f32 d23, d16, d5 |
| 2217 | .long 0xf3408d94 // vmul.f32 d24, d16, d4 |
| 2218 | .long 0xf3409d96 // vmul.f32 d25, d16, d6 |
| 2219 | .long 0xf2452da2 // vadd.f32 d18, d21, d18 |
| 2220 | .long 0xf2463da3 // vadd.f32 d19, d22, d19 |
| 2221 | .long 0xf2414da4 // vadd.f32 d20, d17, d20 |
| 2222 | .long 0xf241ae03 // vceq.f32 d26, d1, d3 |
| 2223 | .long 0xf247bd81 // vadd.f32 d27, d23, d1 |
| 2224 | .long 0xf3b91505 // vceq.f32 d1, d5, #0 |
| 2225 | .long 0xf240ce03 // vceq.f32 d28, d0, d3 |
| 2226 | .long 0xf248dd80 // vadd.f32 d29, d24, d0 |
| 2227 | .long 0xf3b90504 // vceq.f32 d0, d4, #0 |
| 2228 | .long 0xf242ee03 // vceq.f32 d30, d2, d3 |
| 2229 | .long 0xf249fd82 // vadd.f32 d31, d25, d2 |
| 2230 | .long 0xf3b92506 // vceq.f32 d2, d6, #0 |
| 2231 | .long 0xf2073c30 // vfma.f32 d3, d7, d16 |
| 2232 | .long 0xf2410d84 // vadd.f32 d16, d17, d4 |
| 2233 | .long 0xf2491da2 // vadd.f32 d17, d25, d18 |
| 2234 | .long 0xf2462d85 // vadd.f32 d18, d22, d5 |
| 2235 | .long 0xf2455d86 // vadd.f32 d21, d21, d6 |
| 2236 | .long 0xf2473da3 // vadd.f32 d19, d23, d19 |
| 2237 | .long 0xf2484da4 // vadd.f32 d20, d24, d20 |
| 2238 | .long 0xf35fe1b1 // vbsl d30, d31, d17 |
| 2239 | .long 0xf35ba1b3 // vbsl d26, d27, d19 |
| 2240 | .long 0xf35dc1b4 // vbsl d28, d29, d20 |
| 2241 | .long 0xf31001bc // vbsl d0, d16, d28 |
| 2242 | .long 0xf31211ba // vbsl d1, d18, d26 |
| 2243 | .long 0xf31521be // vbsl d2, d21, d30 |
| 2244 | .long 0xecbd8b0e // vpop {d8-d14} |
| 2245 | .long 0xe12fff13 // bx r3 |
| 2246 | |
| 2247 | HIDDEN _sk_hardlight_vfp4 |
| 2248 | .globl _sk_hardlight_vfp4 |
| 2249 | _sk_hardlight_vfp4: |
| 2250 | .long 0xf2c71f10 // vmov.f32 d17, #1 |
| 2251 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2252 | .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| 2253 | .long 0xf2617d87 // vsub.f32 d23, d17, d7 |
| 2254 | .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| 2255 | .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| 2256 | .long 0xf2674d06 // vsub.f32 d20, d7, d6 |
| 2257 | .long 0xf2633d00 // vsub.f32 d19, d3, d0 |
| 2258 | .long 0xf2635d01 // vsub.f32 d21, d3, d1 |
| 2259 | .long 0xf2636d02 // vsub.f32 d22, d3, d2 |
| 2260 | .long 0xf347bd90 // vmul.f32 d27, d23, d0 |
| 2261 | .long 0xf341cd94 // vmul.f32 d28, d17, d4 |
| 2262 | .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| 2263 | .long 0xf3463db4 // vmul.f32 d19, d22, d20 |
| 2264 | .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| 2265 | .long 0xf2404d00 // vadd.f32 d20, d0, d0 |
| 2266 | .long 0xf3405d14 // vmul.f32 d21, d0, d4 |
| 2267 | .long 0xf2416d01 // vadd.f32 d22, d1, d1 |
| 2268 | .long 0xf3418d15 // vmul.f32 d24, d1, d5 |
| 2269 | .long 0xf2429d02 // vadd.f32 d25, d2, d2 |
| 2270 | .long 0xf342ad16 // vmul.f32 d26, d2, d6 |
| 2271 | .long 0xf347dd91 // vmul.f32 d29, d23, d1 |
| 2272 | .long 0xf341fd95 // vmul.f32 d31, d17, d5 |
| 2273 | .long 0xf24cbdab // vadd.f32 d27, d28, d27 |
| 2274 | .long 0xf3477d92 // vmul.f32 d23, d23, d2 |
| 2275 | .long 0xf341cd96 // vmul.f32 d28, d17, d6 |
| 2276 | .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| 2277 | .long 0xf343ed17 // vmul.f32 d30, d3, d7 |
| 2278 | .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| 2279 | .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| 2280 | .long 0xf3434e24 // vcge.f32 d20, d3, d20 |
| 2281 | .long 0xf2455da5 // vadd.f32 d21, d21, d21 |
| 2282 | .long 0xf3436e26 // vcge.f32 d22, d3, d22 |
| 2283 | .long 0xf3439e29 // vcge.f32 d25, d3, d25 |
| 2284 | .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| 2285 | .long 0xf24aadaa // vadd.f32 d26, d26, d26 |
| 2286 | .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| 2287 | .long 0xf24fddad // vadd.f32 d29, d31, d29 |
| 2288 | .long 0xf24c1da7 // vadd.f32 d17, d28, d23 |
| 2289 | .long 0xf26e0da0 // vsub.f32 d16, d30, d16 |
| 2290 | .long 0xf26e2da2 // vsub.f32 d18, d30, d18 |
| 2291 | .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| 2292 | .long 0xf35541b0 // vbsl d20, d21, d16 |
| 2293 | .long 0xf35861b2 // vbsl d22, d24, d18 |
| 2294 | .long 0xf35a91b3 // vbsl d25, d26, d19 |
| 2295 | .long 0xf20b0da4 // vadd.f32 d0, d27, d20 |
| 2296 | .long 0xf20d1da6 // vadd.f32 d1, d29, d22 |
| 2297 | .long 0xf2012da9 // vadd.f32 d2, d17, d25 |
| 2298 | .long 0xe12fff13 // bx r3 |
| 2299 | |
| 2300 | HIDDEN _sk_overlay_vfp4 |
| 2301 | .globl _sk_overlay_vfp4 |
| 2302 | _sk_overlay_vfp4: |
| 2303 | .long 0xf2c71f10 // vmov.f32 d17, #1 |
| 2304 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2305 | .long 0xf2670d04 // vsub.f32 d16, d7, d4 |
| 2306 | .long 0xf2617d87 // vsub.f32 d23, d17, d7 |
| 2307 | .long 0xf2611d83 // vsub.f32 d17, d17, d3 |
| 2308 | .long 0xf2672d05 // vsub.f32 d18, d7, d5 |
| 2309 | .long 0xf2674d06 // vsub.f32 d20, d7, d6 |
| 2310 | .long 0xf2633d00 // vsub.f32 d19, d3, d0 |
| 2311 | .long 0xf2635d01 // vsub.f32 d21, d3, d1 |
| 2312 | .long 0xf2636d02 // vsub.f32 d22, d3, d2 |
| 2313 | .long 0xf347bd90 // vmul.f32 d27, d23, d0 |
| 2314 | .long 0xf341cd94 // vmul.f32 d28, d17, d4 |
| 2315 | .long 0xf3430db0 // vmul.f32 d16, d19, d16 |
| 2316 | .long 0xf3463db4 // vmul.f32 d19, d22, d20 |
| 2317 | .long 0xf3452db2 // vmul.f32 d18, d21, d18 |
| 2318 | .long 0xf2444d04 // vadd.f32 d20, d4, d4 |
| 2319 | .long 0xf3405d14 // vmul.f32 d21, d0, d4 |
| 2320 | .long 0xf2456d05 // vadd.f32 d22, d5, d5 |
| 2321 | .long 0xf3418d15 // vmul.f32 d24, d1, d5 |
| 2322 | .long 0xf2469d06 // vadd.f32 d25, d6, d6 |
| 2323 | .long 0xf342ad16 // vmul.f32 d26, d2, d6 |
| 2324 | .long 0xf347dd91 // vmul.f32 d29, d23, d1 |
| 2325 | .long 0xf341fd95 // vmul.f32 d31, d17, d5 |
| 2326 | .long 0xf24cbdab // vadd.f32 d27, d28, d27 |
| 2327 | .long 0xf3477d92 // vmul.f32 d23, d23, d2 |
| 2328 | .long 0xf341cd96 // vmul.f32 d28, d17, d6 |
| 2329 | .long 0xf343ed17 // vmul.f32 d30, d3, d7 |
| 2330 | .long 0xf2400da0 // vadd.f32 d16, d16, d16 |
| 2331 | .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| 2332 | .long 0xf2433da3 // vadd.f32 d19, d19, d19 |
| 2333 | .long 0xf3474e24 // vcge.f32 d20, d7, d20 |
| 2334 | .long 0xf2455da5 // vadd.f32 d21, d21, d21 |
| 2335 | .long 0xf3476e26 // vcge.f32 d22, d7, d22 |
| 2336 | .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| 2337 | .long 0xf3479e29 // vcge.f32 d25, d7, d25 |
| 2338 | .long 0xf24aadaa // vadd.f32 d26, d26, d26 |
| 2339 | .long 0xf2073c31 // vfma.f32 d3, d7, d17 |
| 2340 | .long 0xf24fddad // vadd.f32 d29, d31, d29 |
| 2341 | .long 0xf24c1da7 // vadd.f32 d17, d28, d23 |
| 2342 | .long 0xf26e0da0 // vsub.f32 d16, d30, d16 |
| 2343 | .long 0xf26e2da2 // vsub.f32 d18, d30, d18 |
| 2344 | .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| 2345 | .long 0xf35541b0 // vbsl d20, d21, d16 |
| 2346 | .long 0xf35861b2 // vbsl d22, d24, d18 |
| 2347 | .long 0xf35a91b3 // vbsl d25, d26, d19 |
| 2348 | .long 0xf20b0da4 // vadd.f32 d0, d27, d20 |
| 2349 | .long 0xf20d1da6 // vadd.f32 d1, d29, d22 |
| 2350 | .long 0xf2012da9 // vadd.f32 d2, d17, d25 |
| 2351 | .long 0xe12fff13 // bx r3 |
| 2352 | |
| 2353 | HIDDEN _sk_softlight_vfp4 |
| 2354 | .globl _sk_softlight_vfp4 |
| 2355 | _sk_softlight_vfp4: |
| 2356 | .long 0xed2d8b06 // vpush {d8-d10} |
| 2357 | .long 0xeec58aa7 // vdiv.f32 s17, s11, s15 |
| 2358 | .long 0xf3f90407 // vcgt.f32 d16, d7, #0 |
| 2359 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2360 | .long 0xeec49aa7 // vdiv.f32 s19, s9, s15 |
| 2361 | .long 0xeec6aaa7 // vdiv.f32 s21, s13, s15 |
| 2362 | .long 0xee858a07 // vdiv.f32 s16, s10, s14 |
| 2363 | .long 0xee849a07 // vdiv.f32 s18, s8, s14 |
| 2364 | .long 0xee86aa07 // vdiv.f32 s20, s12, s14 |
| 2365 | .long 0xf26021b0 // vorr d18, d16, d16 |
| 2366 | .long 0xf2c01010 // vmov.i32 d17, #0 |
| 2367 | .long 0xf3582131 // vbsl d18, d8, d17 |
| 2368 | .long 0xf26031b0 // vorr d19, d16, d16 |
| 2369 | .long 0xf3fb45a2 // vrsqrte.f32 d20, d18 |
| 2370 | .long 0xf3593131 // vbsl d19, d9, d17 |
| 2371 | .long 0xf35a0131 // vbsl d16, d10, d17 |
| 2372 | .long 0xf3fb15a3 // vrsqrte.f32 d17, d19 |
| 2373 | .long 0xf3fb55a0 // vrsqrte.f32 d21, d16 |
| 2374 | .long 0xf3446db4 // vmul.f32 d22, d20, d20 |
| 2375 | .long 0xf243ada3 // vadd.f32 d26, d19, d19 |
| 2376 | .long 0xf240bda0 // vadd.f32 d27, d16, d16 |
| 2377 | .long 0xf3417db1 // vmul.f32 d23, d17, d17 |
| 2378 | .long 0xf3458db5 // vmul.f32 d24, d21, d21 |
| 2379 | .long 0xf2626fb6 // vrsqrts.f32 d22, d18, d22 |
| 2380 | .long 0xf2429da2 // vadd.f32 d25, d18, d18 |
| 2381 | .long 0xf2637fb7 // vrsqrts.f32 d23, d19, d23 |
| 2382 | .long 0xf2608fb8 // vrsqrts.f32 d24, d16, d24 |
| 2383 | .long 0xf2818f1c // vmov.f32 d8, #7 |
| 2384 | .long 0xf2499da9 // vadd.f32 d25, d25, d25 |
| 2385 | .long 0xf3444db6 // vmul.f32 d20, d20, d22 |
| 2386 | .long 0xf24a6daa // vadd.f32 d22, d26, d26 |
| 2387 | .long 0xf24badab // vadd.f32 d26, d27, d27 |
| 2388 | .long 0xf3411db7 // vmul.f32 d17, d17, d23 |
| 2389 | .long 0xf3455db8 // vmul.f32 d21, d21, d24 |
| 2390 | .long 0xf3fb7524 // vrecpe.f32 d23, d20 |
| 2391 | .long 0xf3498db9 // vmul.f32 d24, d25, d25 |
| 2392 | .long 0xf3fbd521 // vrecpe.f32 d29, d17 |
| 2393 | .long 0xf34aedba // vmul.f32 d30, d26, d26 |
| 2394 | .long 0xf3fbf525 // vrecpe.f32 d31, d21 |
| 2395 | .long 0xf2444fb7 // vrecps.f32 d20, d20, d23 |
| 2396 | .long 0xf346cdb6 // vmul.f32 d28, d22, d22 |
| 2397 | .long 0xf2411fbd // vrecps.f32 d17, d17, d29 |
| 2398 | .long 0xf3c7bf10 // vmov.f32 d27, #-1 |
| 2399 | .long 0xf2455fbf // vrecps.f32 d21, d21, d31 |
| 2400 | .long 0xf24aadae // vadd.f32 d26, d26, d30 |
| 2401 | .long 0xf2498da8 // vadd.f32 d24, d25, d24 |
| 2402 | .long 0xf2429dab // vadd.f32 d25, d18, d27 |
| 2403 | .long 0xf2466dac // vadd.f32 d22, d22, d28 |
| 2404 | .long 0xf243cdab // vadd.f32 d28, d19, d27 |
| 2405 | .long 0xf240bdab // vadd.f32 d27, d16, d27 |
| 2406 | .long 0xf3474db4 // vmul.f32 d20, d23, d20 |
| 2407 | .long 0xf2c7ef10 // vmov.f32 d30, #1 |
| 2408 | .long 0xf34d1db1 // vmul.f32 d17, d29, d17 |
| 2409 | .long 0xf34badba // vmul.f32 d26, d27, d26 |
| 2410 | .long 0xf242bd02 // vadd.f32 d27, d2, d2 |
| 2411 | .long 0xf26edda0 // vsub.f32 d29, d30, d16 |
| 2412 | .long 0xf3498db8 // vmul.f32 d24, d25, d24 |
| 2413 | .long 0xf3429d98 // vmul.f32 d25, d18, d8 |
| 2414 | .long 0xf34f5db5 // vmul.f32 d21, d31, d21 |
| 2415 | .long 0xf26efda2 // vsub.f32 d31, d30, d18 |
| 2416 | .long 0xf2642da2 // vsub.f32 d18, d20, d18 |
| 2417 | .long 0xf26b4d83 // vsub.f32 d20, d27, d3 |
| 2418 | .long 0xf2498da8 // vadd.f32 d24, d25, d24 |
| 2419 | .long 0xf34c6db6 // vmul.f32 d22, d28, d22 |
| 2420 | .long 0xf3437d98 // vmul.f32 d23, d19, d8 |
| 2421 | .long 0xf3449dbd // vmul.f32 d25, d20, d29 |
| 2422 | .long 0xf245dd05 // vadd.f32 d29, d5, d5 |
| 2423 | .long 0xf340cd98 // vmul.f32 d28, d16, d8 |
| 2424 | .long 0xf2476da6 // vadd.f32 d22, d23, d22 |
| 2425 | .long 0xf2611da3 // vsub.f32 d17, d17, d19 |
| 2426 | .long 0xf24dddad // vadd.f32 d29, d29, d29 |
| 2427 | .long 0xf24c7daa // vadd.f32 d23, d28, d26 |
| 2428 | .long 0xf2650da0 // vsub.f32 d16, d21, d16 |
| 2429 | .long 0xf26e3da3 // vsub.f32 d19, d30, d19 |
| 2430 | .long 0xf347de2d // vcge.f32 d29, d7, d29 |
| 2431 | .long 0xf241ad01 // vadd.f32 d26, d1, d1 |
| 2432 | .long 0xf3444d97 // vmul.f32 d20, d20, d7 |
| 2433 | .long 0xf358d1b2 // vbsl d29, d24, d18 |
| 2434 | .long 0xf2448d04 // vadd.f32 d24, d4, d4 |
| 2435 | .long 0xf2462d06 // vadd.f32 d18, d6, d6 |
| 2436 | .long 0xf26a5d83 // vsub.f32 d21, d26, d3 |
| 2437 | .long 0xf2488da8 // vadd.f32 d24, d24, d24 |
| 2438 | .long 0xf2422da2 // vadd.f32 d18, d18, d18 |
| 2439 | .long 0xf345cdbf // vmul.f32 d28, d21, d31 |
| 2440 | .long 0xf3455d97 // vmul.f32 d21, d21, d7 |
| 2441 | .long 0xf3478e28 // vcge.f32 d24, d7, d24 |
| 2442 | .long 0xf3472e22 // vcge.f32 d18, d7, d18 |
| 2443 | .long 0xf343fd14 // vmul.f32 d31, d3, d4 |
| 2444 | .long 0xf3455dbd // vmul.f32 d21, d21, d29 |
| 2445 | .long 0xf35681b1 // vbsl d24, d22, d17 |
| 2446 | .long 0xf2401d00 // vadd.f32 d17, d0, d0 |
| 2447 | .long 0xf35721b0 // vbsl d18, d23, d16 |
| 2448 | .long 0xf24c0d83 // vadd.f32 d16, d28, d3 |
| 2449 | .long 0xf2496d83 // vadd.f32 d22, d25, d3 |
| 2450 | .long 0xf2617d83 // vsub.f32 d23, d17, d3 |
| 2451 | .long 0xf3442db2 // vmul.f32 d18, d20, d18 |
| 2452 | .long 0xf3434e2a // vcge.f32 d20, d3, d26 |
| 2453 | .long 0xf343ae2b // vcge.f32 d26, d3, d27 |
| 2454 | .long 0xf3473db3 // vmul.f32 d19, d23, d19 |
| 2455 | .long 0xf3477d97 // vmul.f32 d23, d23, d7 |
| 2456 | .long 0xf3431e21 // vcge.f32 d17, d3, d17 |
| 2457 | .long 0xf3400d95 // vmul.f32 d16, d16, d5 |
| 2458 | .long 0xf2433d83 // vadd.f32 d19, d19, d3 |
| 2459 | .long 0xf3477db8 // vmul.f32 d23, d23, d24 |
| 2460 | .long 0xf26e8d87 // vsub.f32 d24, d30, d7 |
| 2461 | .long 0xf26eed83 // vsub.f32 d30, d30, d3 |
| 2462 | .long 0xf3433d94 // vmul.f32 d19, d19, d4 |
| 2463 | .long 0xf24f7da7 // vadd.f32 d23, d31, d23 |
| 2464 | .long 0xf3489d91 // vmul.f32 d25, d24, d1 |
| 2465 | .long 0xf348cd90 // vmul.f32 d28, d24, d0 |
| 2466 | .long 0xf34edd94 // vmul.f32 d29, d30, d4 |
| 2467 | .long 0xf34ebd95 // vmul.f32 d27, d30, d5 |
| 2468 | .long 0xf3488d92 // vmul.f32 d24, d24, d2 |
| 2469 | .long 0xf34efd96 // vmul.f32 d31, d30, d6 |
| 2470 | .long 0xf24dcdac // vadd.f32 d28, d29, d28 |
| 2471 | .long 0xf343dd15 // vmul.f32 d29, d3, d5 |
| 2472 | .long 0xf24b9da9 // vadd.f32 d25, d27, d25 |
| 2473 | .long 0xf343bd16 // vmul.f32 d27, d3, d6 |
| 2474 | .long 0xf3466d96 // vmul.f32 d22, d22, d6 |
| 2475 | .long 0xf24f8da8 // vadd.f32 d24, d31, d24 |
| 2476 | .long 0xf24d5da5 // vadd.f32 d21, d29, d21 |
| 2477 | .long 0xf24b2da2 // vadd.f32 d18, d27, d18 |
| 2478 | .long 0xf35311b7 // vbsl d17, d19, d23 |
| 2479 | .long 0xf35041b5 // vbsl d20, d16, d21 |
| 2480 | .long 0xf356a1b2 // vbsl d26, d22, d18 |
| 2481 | .long 0xf2073c3e // vfma.f32 d3, d7, d30 |
| 2482 | .long 0xf20c0da1 // vadd.f32 d0, d28, d17 |
| 2483 | .long 0xf2091da4 // vadd.f32 d1, d25, d20 |
| 2484 | .long 0xf2082daa // vadd.f32 d2, d24, d26 |
| 2485 | .long 0xecbd8b06 // vpop {d8-d10} |
| 2486 | .long 0xe12fff13 // bx r3 |
| 2487 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2488 | HIDDEN _sk_clamp_0_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2489 | .globl _sk_clamp_0_vfp4 |
| 2490 | _sk_clamp_0_vfp4: |
| 2491 | .long 0xf2c00010 // vmov.i32 d16, #0 |
| 2492 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2493 | .long 0xf2000f20 // vmax.f32 d0, d0, d16 |
| 2494 | .long 0xf2011f20 // vmax.f32 d1, d1, d16 |
| 2495 | .long 0xf2022f20 // vmax.f32 d2, d2, d16 |
| 2496 | .long 0xf2033f20 // vmax.f32 d3, d3, d16 |
| 2497 | .long 0xe12fff13 // bx r3 |
| 2498 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2499 | HIDDEN _sk_clamp_1_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2500 | .globl _sk_clamp_1_vfp4 |
| 2501 | _sk_clamp_1_vfp4: |
| 2502 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2503 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2504 | .long 0xf2200f20 // vmin.f32 d0, d0, d16 |
| 2505 | .long 0xf2211f20 // vmin.f32 d1, d1, d16 |
| 2506 | .long 0xf2222f20 // vmin.f32 d2, d2, d16 |
| 2507 | .long 0xf2233f20 // vmin.f32 d3, d3, d16 |
| 2508 | .long 0xe12fff13 // bx r3 |
| 2509 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2510 | HIDDEN _sk_clamp_a_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2511 | .globl _sk_clamp_a_vfp4 |
| 2512 | _sk_clamp_a_vfp4: |
| 2513 | .long 0xf2c70f10 // vmov.f32 d16, #1 |
| 2514 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2515 | .long 0xf2233f20 // vmin.f32 d3, d3, d16 |
| 2516 | .long 0xf2200f03 // vmin.f32 d0, d0, d3 |
| 2517 | .long 0xf2211f03 // vmin.f32 d1, d1, d3 |
| 2518 | .long 0xf2222f03 // vmin.f32 d2, d2, d3 |
| 2519 | .long 0xe12fff13 // bx r3 |
| 2520 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2521 | HIDDEN _sk_set_rgb_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2522 | .globl _sk_set_rgb_vfp4 |
| 2523 | _sk_set_rgb_vfp4: |
| 2524 | .long 0xe92d4800 // push {fp, lr} |
| 2525 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2526 | .long 0xe2811008 // add r1, r1, #8 |
| 2527 | .long 0xe283e008 // add lr, r3, #8 |
| 2528 | .long 0xf4a30c9d // vld1.32 {d0[]}, [r3 :32]! |
| 2529 | .long 0xf4ae2c9f // vld1.32 {d2[]}, [lr :32] |
| 2530 | .long 0xf4a31c9f // vld1.32 {d1[]}, [r3 :32] |
| 2531 | .long 0xe8bd4800 // pop {fp, lr} |
| 2532 | .long 0xe12fff1c // bx ip |
| 2533 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2534 | HIDDEN _sk_swap_rb_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2535 | .globl _sk_swap_rb_vfp4 |
| 2536 | _sk_swap_rb_vfp4: |
| 2537 | .long 0xeef00b40 // vmov.f64 d16, d0 |
| 2538 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2539 | .long 0xeeb00b42 // vmov.f64 d0, d2 |
| 2540 | .long 0xeeb02b60 // vmov.f64 d2, d16 |
| 2541 | .long 0xe12fff13 // bx r3 |
| 2542 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2543 | HIDDEN _sk_swap_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2544 | .globl _sk_swap_vfp4 |
| 2545 | _sk_swap_vfp4: |
| 2546 | .long 0xeef00b43 // vmov.f64 d16, d3 |
| 2547 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2548 | .long 0xeef01b42 // vmov.f64 d17, d2 |
| 2549 | .long 0xeef02b41 // vmov.f64 d18, d1 |
| 2550 | .long 0xeef03b40 // vmov.f64 d19, d0 |
| 2551 | .long 0xeeb00b44 // vmov.f64 d0, d4 |
| 2552 | .long 0xeeb01b45 // vmov.f64 d1, d5 |
| 2553 | .long 0xeeb02b46 // vmov.f64 d2, d6 |
| 2554 | .long 0xeeb03b47 // vmov.f64 d3, d7 |
| 2555 | .long 0xeeb04b63 // vmov.f64 d4, d19 |
| 2556 | .long 0xeeb05b62 // vmov.f64 d5, d18 |
| 2557 | .long 0xeeb06b61 // vmov.f64 d6, d17 |
| 2558 | .long 0xeeb07b60 // vmov.f64 d7, d16 |
| 2559 | .long 0xe12fff13 // bx r3 |
| 2560 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2561 | HIDDEN _sk_move_src_dst_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2562 | .globl _sk_move_src_dst_vfp4 |
| 2563 | _sk_move_src_dst_vfp4: |
| 2564 | .long 0xeeb04b40 // vmov.f64 d4, d0 |
| 2565 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2566 | .long 0xeeb05b41 // vmov.f64 d5, d1 |
| 2567 | .long 0xeeb06b42 // vmov.f64 d6, d2 |
| 2568 | .long 0xeeb07b43 // vmov.f64 d7, d3 |
| 2569 | .long 0xe12fff13 // bx r3 |
| 2570 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2571 | HIDDEN _sk_move_dst_src_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2572 | .globl _sk_move_dst_src_vfp4 |
| 2573 | _sk_move_dst_src_vfp4: |
| 2574 | .long 0xeeb00b44 // vmov.f64 d0, d4 |
| 2575 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2576 | .long 0xeeb01b45 // vmov.f64 d1, d5 |
| 2577 | .long 0xeeb02b46 // vmov.f64 d2, d6 |
| 2578 | .long 0xeeb03b47 // vmov.f64 d3, d7 |
| 2579 | .long 0xe12fff13 // bx r3 |
| 2580 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2581 | HIDDEN _sk_premul_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2582 | .globl _sk_premul_vfp4 |
| 2583 | _sk_premul_vfp4: |
| 2584 | .long 0xf3000d13 // vmul.f32 d0, d0, d3 |
| 2585 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2586 | .long 0xf3011d13 // vmul.f32 d1, d1, d3 |
| 2587 | .long 0xf3022d13 // vmul.f32 d2, d2, d3 |
| 2588 | .long 0xe12fff13 // bx r3 |
| 2589 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2590 | HIDDEN _sk_unpremul_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2591 | .globl _sk_unpremul_vfp4 |
| 2592 | _sk_unpremul_vfp4: |
| 2593 | .long 0xed2d8b04 // vpush {d8-d9} |
| 2594 | .long 0xeeb78a00 // vmov.f32 s16, #112 |
| 2595 | .long 0xf3f91503 // vceq.f32 d17, d3, #0 |
| 2596 | .long 0xf2c00010 // vmov.i32 d16, #0 |
| 2597 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2598 | .long 0xeec89a23 // vdiv.f32 s19, s16, s7 |
| 2599 | .long 0xee889a03 // vdiv.f32 s18, s16, s6 |
| 2600 | .long 0xf3501199 // vbsl d17, d16, d9 |
| 2601 | .long 0xf3010d90 // vmul.f32 d0, d17, d0 |
| 2602 | .long 0xf3011d91 // vmul.f32 d1, d17, d1 |
| 2603 | .long 0xf3012d92 // vmul.f32 d2, d17, d2 |
| 2604 | .long 0xecbd8b04 // vpop {d8-d9} |
| 2605 | .long 0xe12fff13 // bx r3 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 2606 | .long 0xe320f000 // nop {0} |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2607 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2608 | HIDDEN _sk_from_srgb_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2609 | .globl _sk_from_srgb_vfp4 |
| 2610 | _sk_from_srgb_vfp4: |
| 2611 | .long 0xeddf3b20 // vldr d19, [pc, #128] |
| 2612 | .long 0xf3408d10 // vmul.f32 d24, d0, d0 |
| 2613 | .long 0xeddf0b1c // vldr d16, [pc, #112] |
| 2614 | .long 0xf26341b3 // vorr d20, d19, d19 |
| 2615 | .long 0xf26351b3 // vorr d21, d19, d19 |
| 2616 | .long 0xeddf9b1f // vldr d25, [pc, #124] |
| 2617 | .long 0xf2404c30 // vfma.f32 d20, d0, d16 |
| 2618 | .long 0xeddf2b1b // vldr d18, [pc, #108] |
| 2619 | .long 0xf2415c30 // vfma.f32 d21, d1, d16 |
| 2620 | .long 0xeddfcb1d // vldr d28, [pc, #116] |
| 2621 | .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| 2622 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2623 | .long 0xf3426d12 // vmul.f32 d22, d2, d2 |
| 2624 | .long 0xf3417d11 // vmul.f32 d23, d1, d1 |
| 2625 | .long 0xf3620e80 // vcgt.f32 d16, d18, d0 |
| 2626 | .long 0xf3621e81 // vcgt.f32 d17, d18, d1 |
| 2627 | .long 0xf341ad39 // vmul.f32 d26, d1, d25 |
| 2628 | .long 0xf342bd39 // vmul.f32 d27, d2, d25 |
| 2629 | .long 0xf3622e82 // vcgt.f32 d18, d18, d2 |
| 2630 | .long 0xf3409d39 // vmul.f32 d25, d0, d25 |
| 2631 | .long 0xf26cd1bc // vorr d29, d28, d28 |
| 2632 | .long 0xf248dcb4 // vfma.f32 d29, d24, d20 |
| 2633 | .long 0xf26c41bc // vorr d20, d28, d28 |
| 2634 | .long 0xf2474cb5 // vfma.f32 d20, d23, d21 |
| 2635 | .long 0xf246ccb3 // vfma.f32 d28, d22, d19 |
| 2636 | .long 0xf35901bd // vbsl d16, d25, d29 |
| 2637 | .long 0xf35a11b4 // vbsl d17, d26, d20 |
| 2638 | .long 0xf35b21bc // vbsl d18, d27, d28 |
| 2639 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 2640 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 2641 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 2642 | .long 0xe12fff13 // bx r3 |
| 2643 | .long 0x3e99999a // .word 0x3e99999a |
| 2644 | .long 0x3e99999a // .word 0x3e99999a |
| 2645 | .long 0x3f328f5c // .word 0x3f328f5c |
| 2646 | .long 0x3f328f5c // .word 0x3f328f5c |
| 2647 | .long 0x3d6147ae // .word 0x3d6147ae |
| 2648 | .long 0x3d6147ae // .word 0x3d6147ae |
| 2649 | .long 0x3d9e8391 // .word 0x3d9e8391 |
| 2650 | .long 0x3d9e8391 // .word 0x3d9e8391 |
| 2651 | .long 0x3b23d70a // .word 0x3b23d70a |
| 2652 | .long 0x3b23d70a // .word 0x3b23d70a |
| 2653 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2654 | HIDDEN _sk_to_srgb_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2655 | .globl _sk_to_srgb_vfp4 |
| 2656 | _sk_to_srgb_vfp4: |
| 2657 | .long 0xf3fb0582 // vrsqrte.f32 d16, d2 |
| 2658 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 2659 | .long 0xf3fb1581 // vrsqrte.f32 d17, d1 |
| 2660 | .long 0xf3fb2580 // vrsqrte.f32 d18, d0 |
| 2661 | .long 0xf3403db0 // vmul.f32 d19, d16, d16 |
| 2662 | .long 0xf3414db1 // vmul.f32 d20, d17, d17 |
| 2663 | .long 0xf3425db2 // vmul.f32 d21, d18, d18 |
| 2664 | .long 0xf2623f33 // vrsqrts.f32 d19, d2, d19 |
| 2665 | .long 0xf2614f34 // vrsqrts.f32 d20, d1, d20 |
| 2666 | .long 0xf2605f35 // vrsqrts.f32 d21, d0, d21 |
| 2667 | .long 0xf3400db3 // vmul.f32 d16, d16, d19 |
| 2668 | .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| 2669 | .long 0xf3422db5 // vmul.f32 d18, d18, d21 |
| 2670 | .long 0xf3fb3520 // vrecpe.f32 d19, d16 |
| 2671 | .long 0xf3fb4521 // vrecpe.f32 d20, d17 |
| 2672 | .long 0xf3fb6522 // vrecpe.f32 d22, d18 |
| 2673 | .long 0xf3fb55a0 // vrsqrte.f32 d21, d16 |
| 2674 | .long 0xf3fb75a1 // vrsqrte.f32 d23, d17 |
| 2675 | .long 0xf3fb85a2 // vrsqrte.f32 d24, d18 |
| 2676 | .long 0xf2409fb3 // vrecps.f32 d25, d16, d19 |
| 2677 | .long 0xf241afb4 // vrecps.f32 d26, d17, d20 |
| 2678 | .long 0xf242bfb6 // vrecps.f32 d27, d18, d22 |
| 2679 | .long 0xf345cdb5 // vmul.f32 d28, d21, d21 |
| 2680 | .long 0xf347ddb7 // vmul.f32 d29, d23, d23 |
| 2681 | .long 0xf348edb8 // vmul.f32 d30, d24, d24 |
| 2682 | .long 0xf2600fbc // vrsqrts.f32 d16, d16, d28 |
| 2683 | .long 0xf2611fbd // vrsqrts.f32 d17, d17, d29 |
| 2684 | .long 0xf2622fbe // vrsqrts.f32 d18, d18, d30 |
| 2685 | .long 0xf3433db9 // vmul.f32 d19, d19, d25 |
| 2686 | .long 0xeddf9b21 // vldr d25, [pc, #132] |
| 2687 | .long 0xf3444dba // vmul.f32 d20, d20, d26 |
| 2688 | .long 0xeddfab21 // vldr d26, [pc, #132] |
| 2689 | .long 0xf3466dbb // vmul.f32 d22, d22, d27 |
| 2690 | .long 0xf26ab1ba // vorr d27, d26, d26 |
| 2691 | .long 0xf243bcb9 // vfma.f32 d27, d19, d25 |
| 2692 | .long 0xf26a31ba // vorr d19, d26, d26 |
| 2693 | .long 0xf2443cb9 // vfma.f32 d19, d20, d25 |
| 2694 | .long 0xeddf4b1d // vldr d20, [pc, #116] |
| 2695 | .long 0xf246acb9 // vfma.f32 d26, d22, d25 |
| 2696 | .long 0xf3450db0 // vmul.f32 d16, d21, d16 |
| 2697 | .long 0xeddf5b1c // vldr d21, [pc, #112] |
| 2698 | .long 0xf3471db1 // vmul.f32 d17, d23, d17 |
| 2699 | .long 0xf3482db2 // vmul.f32 d18, d24, d18 |
| 2700 | .long 0xf3406d35 // vmul.f32 d22, d0, d21 |
| 2701 | .long 0xf240bcb4 // vfma.f32 d27, d16, d20 |
| 2702 | .long 0xf2413cb4 // vfma.f32 d19, d17, d20 |
| 2703 | .long 0xf242acb4 // vfma.f32 d26, d18, d20 |
| 2704 | .long 0xeddf2b17 // vldr d18, [pc, #92] |
| 2705 | .long 0xf3417d35 // vmul.f32 d23, d1, d21 |
| 2706 | .long 0xf3620e80 // vcgt.f32 d16, d18, d0 |
| 2707 | .long 0xf3621e81 // vcgt.f32 d17, d18, d1 |
| 2708 | .long 0xf3622e82 // vcgt.f32 d18, d18, d2 |
| 2709 | .long 0xf3425d35 // vmul.f32 d21, d2, d21 |
| 2710 | .long 0xf2c74f10 // vmov.f32 d20, #1 |
| 2711 | .long 0xf2648faa // vmin.f32 d24, d20, d26 |
| 2712 | .long 0xf2643fa3 // vmin.f32 d19, d20, d19 |
| 2713 | .long 0xf2644fab // vmin.f32 d20, d20, d27 |
| 2714 | .long 0xf35601b8 // vbsl d16, d22, d24 |
| 2715 | .long 0xf35711b3 // vbsl d17, d23, d19 |
| 2716 | .long 0xf35521b4 // vbsl d18, d21, d20 |
| 2717 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 2718 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 2719 | .long 0xf22221b2 // vorr d2, d18, d18 |
| 2720 | .long 0xe12fff13 // bx r3 |
| 2721 | .long 0x3f306fce // .word 0x3f306fce |
| 2722 | .long 0x3f306fce // .word 0x3f306fce |
| 2723 | .long 0xbdca57a8 // .word 0xbdca57a8 |
| 2724 | .long 0xbdca57a8 // .word 0xbdca57a8 |
| 2725 | .long 0x3ed287c2 // .word 0x3ed287c2 |
| 2726 | .long 0x3ed287c2 // .word 0x3ed287c2 |
| 2727 | .long 0x41475c29 // .word 0x41475c29 |
| 2728 | .long 0x41475c29 // .word 0x41475c29 |
| 2729 | .long 0x3b8ce704 // .word 0x3b8ce704 |
| 2730 | .long 0x3b8ce704 // .word 0x3b8ce704 |
| 2731 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2732 | HIDDEN _sk_scale_1_float_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2733 | .globl _sk_scale_1_float_vfp4 |
| 2734 | _sk_scale_1_float_vfp4: |
| 2735 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2736 | .long 0xe2811008 // add r1, r1, #8 |
| 2737 | .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| 2738 | .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| 2739 | .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| 2740 | .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| 2741 | .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| 2742 | .long 0xe12fff1c // bx ip |
| 2743 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2744 | HIDDEN _sk_scale_u8_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2745 | .globl _sk_scale_u8_vfp4 |
| 2746 | _sk_scale_u8_vfp4: |
| 2747 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2748 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2749 | .long 0xe2811008 // add r1, r1, #8 |
| 2750 | .long 0xe5933000 // ldr r3, [r3] |
| 2751 | .long 0xe0833000 // add r3, r3, r0 |
| 2752 | .long 0xe1d330b0 // ldrh r3, [r3] |
| 2753 | .long 0xe1cd30b0 // strh r3, [sp] |
| 2754 | .long 0xe1a0300d // mov r3, sp |
| 2755 | .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16] |
| 2756 | .long 0xf3c80a30 // vmovl.u8 q8, d16 |
| 2757 | .long 0xf3d00a30 // vmovl.u16 q8, d16 |
| 2758 | .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| 2759 | .long 0xeddf1b06 // vldr d17, [pc, #24] |
| 2760 | .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| 2761 | .long 0xf3000d90 // vmul.f32 d0, d16, d0 |
| 2762 | .long 0xf3001d91 // vmul.f32 d1, d16, d1 |
| 2763 | .long 0xf3002d92 // vmul.f32 d2, d16, d2 |
| 2764 | .long 0xf3003d93 // vmul.f32 d3, d16, d3 |
| 2765 | .long 0xe28dd004 // add sp, sp, #4 |
| 2766 | .long 0xe12fff1c // bx ip |
| 2767 | .long 0x3b808081 // .word 0x3b808081 |
| 2768 | .long 0x3b808081 // .word 0x3b808081 |
| 2769 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2770 | HIDDEN _sk_lerp_1_float_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2771 | .globl _sk_lerp_1_float_vfp4 |
| 2772 | _sk_lerp_1_float_vfp4: |
| 2773 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2774 | .long 0xf2600d04 // vsub.f32 d16, d0, d4 |
| 2775 | .long 0xf2611d05 // vsub.f32 d17, d1, d5 |
| 2776 | .long 0xf2622d06 // vsub.f32 d18, d2, d6 |
| 2777 | .long 0xe2811008 // add r1, r1, #8 |
| 2778 | .long 0xf2633d07 // vsub.f32 d19, d3, d7 |
| 2779 | .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| 2780 | .long 0xf2240114 // vorr d0, d4, d4 |
| 2781 | .long 0xf2251115 // vorr d1, d5, d5 |
| 2782 | .long 0xf2262116 // vorr d2, d6, d6 |
| 2783 | .long 0xf2273117 // vorr d3, d7, d7 |
| 2784 | .long 0xf2000cb4 // vfma.f32 d0, d16, d20 |
| 2785 | .long 0xf2011cb4 // vfma.f32 d1, d17, d20 |
| 2786 | .long 0xf2022cb4 // vfma.f32 d2, d18, d20 |
| 2787 | .long 0xf2033cb4 // vfma.f32 d3, d19, d20 |
| 2788 | .long 0xe12fff1c // bx ip |
| 2789 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2790 | HIDDEN _sk_lerp_u8_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2791 | .globl _sk_lerp_u8_vfp4 |
| 2792 | _sk_lerp_u8_vfp4: |
| 2793 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2794 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2795 | .long 0xf2602d04 // vsub.f32 d18, d0, d4 |
| 2796 | .long 0xf2623d06 // vsub.f32 d19, d2, d6 |
| 2797 | .long 0xf2634d07 // vsub.f32 d20, d3, d7 |
| 2798 | .long 0xe2811008 // add r1, r1, #8 |
| 2799 | .long 0xe5933000 // ldr r3, [r3] |
| 2800 | .long 0xf2240114 // vorr d0, d4, d4 |
| 2801 | .long 0xf2262116 // vorr d2, d6, d6 |
| 2802 | .long 0xe0833000 // add r3, r3, r0 |
| 2803 | .long 0xf2273117 // vorr d3, d7, d7 |
| 2804 | .long 0xe1d330b0 // ldrh r3, [r3] |
| 2805 | .long 0xe1cd30b0 // strh r3, [sp] |
| 2806 | .long 0xe1a0300d // mov r3, sp |
| 2807 | .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16] |
| 2808 | .long 0xf3c80a30 // vmovl.u8 q8, d16 |
| 2809 | .long 0xf3d00a30 // vmovl.u16 q8, d16 |
| 2810 | .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| 2811 | .long 0xeddf1b08 // vldr d17, [pc, #32] |
| 2812 | .long 0xf3400db1 // vmul.f32 d16, d16, d17 |
| 2813 | .long 0xf2611d05 // vsub.f32 d17, d1, d5 |
| 2814 | .long 0xf2251115 // vorr d1, d5, d5 |
| 2815 | .long 0xf2020cb0 // vfma.f32 d0, d18, d16 |
| 2816 | .long 0xf2011cb0 // vfma.f32 d1, d17, d16 |
| 2817 | .long 0xf2032cb0 // vfma.f32 d2, d19, d16 |
| 2818 | .long 0xf2043cb0 // vfma.f32 d3, d20, d16 |
| 2819 | .long 0xe28dd004 // add sp, sp, #4 |
| 2820 | .long 0xe12fff1c // bx ip |
| 2821 | .long 0x3b808081 // .word 0x3b808081 |
| 2822 | .long 0x3b808081 // .word 0x3b808081 |
| 2823 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2824 | HIDDEN _sk_lerp_565_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2825 | .globl _sk_lerp_565_vfp4 |
| 2826 | _sk_lerp_565_vfp4: |
| 2827 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2828 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2829 | .long 0xf3c72218 // vmov.i32 d18, #63488 |
| 2830 | .long 0xf2c1101f // vmov.i32 d17, #31 |
| 2831 | .long 0xf2603d04 // vsub.f32 d19, d0, d4 |
| 2832 | .long 0xe2811008 // add r1, r1, #8 |
| 2833 | .long 0xe5933000 // ldr r3, [r3] |
| 2834 | .long 0xf2616d05 // vsub.f32 d22, d1, d5 |
| 2835 | .long 0xf2240114 // vorr d0, d4, d4 |
| 2836 | .long 0xf2251115 // vorr d1, d5, d5 |
| 2837 | .long 0xe7933080 // ldr r3, [r3, r0, lsl #1] |
| 2838 | .long 0xf2873f10 // vmov.f32 d3, #1 |
| 2839 | .long 0xe58d3000 // str r3, [sp] |
| 2840 | .long 0xe1a0300d // mov r3, sp |
| 2841 | .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| 2842 | .long 0xe3a03e7e // mov r3, #2016 |
| 2843 | .long 0xf3d04a30 // vmovl.u16 q10, d16 |
| 2844 | .long 0xee803b90 // vdup.32 d16, r3 |
| 2845 | .long 0xf24421b2 // vand d18, d20, d18 |
| 2846 | .long 0xf24411b1 // vand d17, d20, d17 |
| 2847 | .long 0xeddf5b12 // vldr d21, [pc, #72] |
| 2848 | .long 0xf24401b0 // vand d16, d20, d16 |
| 2849 | .long 0xeddf4b0e // vldr d20, [pc, #56] |
| 2850 | .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| 2851 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 2852 | .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| 2853 | .long 0xf3422db4 // vmul.f32 d18, d18, d20 |
| 2854 | .long 0xeddf4b0d // vldr d20, [pc, #52] |
| 2855 | .long 0xf3400db5 // vmul.f32 d16, d16, d21 |
| 2856 | .long 0xf2625d06 // vsub.f32 d21, d2, d6 |
| 2857 | .long 0xf3411db4 // vmul.f32 d17, d17, d20 |
| 2858 | .long 0xf2262116 // vorr d2, d6, d6 |
| 2859 | .long 0xf2030cb2 // vfma.f32 d0, d19, d18 |
| 2860 | .long 0xf2061cb0 // vfma.f32 d1, d22, d16 |
| 2861 | .long 0xf2052cb1 // vfma.f32 d2, d21, d17 |
| 2862 | .long 0xe28dd004 // add sp, sp, #4 |
| 2863 | .long 0xe12fff1c // bx ip |
| 2864 | .long 0xe320f000 // nop {0} |
| 2865 | .long 0x37842108 // .word 0x37842108 |
| 2866 | .long 0x37842108 // .word 0x37842108 |
| 2867 | .long 0x3a020821 // .word 0x3a020821 |
| 2868 | .long 0x3a020821 // .word 0x3a020821 |
| 2869 | .long 0x3d042108 // .word 0x3d042108 |
| 2870 | .long 0x3d042108 // .word 0x3d042108 |
| 2871 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2872 | HIDDEN _sk_load_tables_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2873 | .globl _sk_load_tables_vfp4 |
| 2874 | _sk_load_tables_vfp4: |
| 2875 | .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| 2876 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2877 | .long 0xf3c7001f // vmov.i32 d16, #255 |
| 2878 | .long 0xe2811008 // add r1, r1, #8 |
| 2879 | .long 0xe593e000 // ldr lr, [r3] |
| 2880 | .long 0xe99300b0 // ldmib r3, {r4, r5, r7} |
| 2881 | .long 0xe08e3100 // add r3, lr, r0, lsl #2 |
| 2882 | .long 0xedd31b00 // vldr d17, [r3] |
| 2883 | .long 0xf24121b0 // vand d18, d17, d16 |
| 2884 | .long 0xf3f83031 // vshr.u32 d19, d17, #8 |
| 2885 | .long 0xee323b90 // vmov.32 r3, d18[1] |
| 2886 | .long 0xee126b90 // vmov.32 r6, d18[0] |
| 2887 | .long 0xf3f02031 // vshr.u32 d18, d17, #16 |
| 2888 | .long 0xf24221b0 // vand d18, d18, d16 |
| 2889 | .long 0xf24301b0 // vand d16, d19, d16 |
| 2890 | .long 0xe0843103 // add r3, r4, r3, lsl #2 |
| 2891 | .long 0xedd30a00 // vldr s1, [r3] |
| 2892 | .long 0xe0843106 // add r3, r4, r6, lsl #2 |
| 2893 | .long 0xee326b90 // vmov.32 r6, d18[1] |
| 2894 | .long 0xed930a00 // vldr s0, [r3] |
| 2895 | .long 0xee303b90 // vmov.32 r3, d16[1] |
| 2896 | .long 0xee104b90 // vmov.32 r4, d16[0] |
| 2897 | .long 0xf3e80031 // vshr.u32 d16, d17, #24 |
| 2898 | .long 0xeddf1b0d // vldr d17, [pc, #52] |
| 2899 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 2900 | .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| 2901 | .long 0xe087e106 // add lr, r7, r6, lsl #2 |
| 2902 | .long 0xee126b90 // vmov.32 r6, d18[0] |
| 2903 | .long 0xe0853103 // add r3, r5, r3, lsl #2 |
| 2904 | .long 0xedde2a00 // vldr s5, [lr] |
| 2905 | .long 0xedd31a00 // vldr s3, [r3] |
| 2906 | .long 0xe0853104 // add r3, r5, r4, lsl #2 |
| 2907 | .long 0xed931a00 // vldr s2, [r3] |
| 2908 | .long 0xe0873106 // add r3, r7, r6, lsl #2 |
| 2909 | .long 0xed932a00 // vldr s4, [r3] |
| 2910 | .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| 2911 | .long 0xe12fff1c // bx ip |
| 2912 | .long 0xe320f000 // nop {0} |
| 2913 | .long 0x3b808081 // .word 0x3b808081 |
| 2914 | .long 0x3b808081 // .word 0x3b808081 |
| 2915 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2916 | HIDDEN _sk_load_a8_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2917 | .globl _sk_load_a8_vfp4 |
| 2918 | _sk_load_a8_vfp4: |
| 2919 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2920 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2921 | .long 0xe2811008 // add r1, r1, #8 |
| 2922 | .long 0xf2800010 // vmov.i32 d0, #0 |
| 2923 | .long 0xf2801010 // vmov.i32 d1, #0 |
| 2924 | .long 0xe5933000 // ldr r3, [r3] |
| 2925 | .long 0xf2802010 // vmov.i32 d2, #0 |
| 2926 | .long 0xe0833000 // add r3, r3, r0 |
| 2927 | .long 0xe1d330b0 // ldrh r3, [r3] |
| 2928 | .long 0xe1cd30b0 // strh r3, [sp] |
| 2929 | .long 0xe1a0300d // mov r3, sp |
| 2930 | .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16] |
| 2931 | .long 0xf3c80a30 // vmovl.u8 q8, d16 |
| 2932 | .long 0xf3d00a30 // vmovl.u16 q8, d16 |
| 2933 | .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| 2934 | .long 0xeddf1b03 // vldr d17, [pc, #12] |
| 2935 | .long 0xf3003db1 // vmul.f32 d3, d16, d17 |
| 2936 | .long 0xe28dd004 // add sp, sp, #4 |
| 2937 | .long 0xe12fff1c // bx ip |
| 2938 | .long 0xe320f000 // nop {0} |
| 2939 | .long 0x3b808081 // .word 0x3b808081 |
| 2940 | .long 0x3b808081 // .word 0x3b808081 |
| 2941 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2942 | HIDDEN _sk_store_a8_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2943 | .globl _sk_store_a8_vfp4 |
| 2944 | _sk_store_a8_vfp4: |
| 2945 | .long 0xe92d4800 // push {fp, lr} |
| 2946 | .long 0xeddf0b0d // vldr d16, [pc, #52] |
| 2947 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 2948 | .long 0xf2431c30 // vfma.f32 d17, d3, d16 |
| 2949 | .long 0xe5913000 // ldr r3, [r1] |
| 2950 | .long 0xe5933000 // ldr r3, [r3] |
| 2951 | .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17 |
| 2952 | .long 0xee10eb90 // vmov.32 lr, d16[0] |
| 2953 | .long 0xee30cb90 // vmov.32 ip, d16[1] |
| 2954 | .long 0xe7e3e000 // strb lr, [r3, r0]! |
| 2955 | .long 0xe5c3c001 // strb ip, [r3, #1] |
| 2956 | .long 0xe2813008 // add r3, r1, #8 |
| 2957 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 2958 | .long 0xe1a01003 // mov r1, r3 |
| 2959 | .long 0xe8bd4800 // pop {fp, lr} |
| 2960 | .long 0xe12fff1c // bx ip |
| 2961 | .long 0x437f0000 // .word 0x437f0000 |
| 2962 | .long 0x437f0000 // .word 0x437f0000 |
| 2963 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 2964 | HIDDEN _sk_load_g8_vfp4 |
| 2965 | .globl _sk_load_g8_vfp4 |
| 2966 | _sk_load_g8_vfp4: |
| 2967 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2968 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2969 | .long 0xe2811008 // add r1, r1, #8 |
| 2970 | .long 0xf2873f10 // vmov.f32 d3, #1 |
| 2971 | .long 0xe5933000 // ldr r3, [r3] |
| 2972 | .long 0xe0833000 // add r3, r3, r0 |
| 2973 | .long 0xe1d330b0 // ldrh r3, [r3] |
| 2974 | .long 0xe1cd30b0 // strh r3, [sp] |
| 2975 | .long 0xe1a0300d // mov r3, sp |
| 2976 | .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16] |
| 2977 | .long 0xf3c80a30 // vmovl.u8 q8, d16 |
| 2978 | .long 0xf3d00a30 // vmovl.u16 q8, d16 |
| 2979 | .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| 2980 | .long 0xeddf1b05 // vldr d17, [pc, #20] |
| 2981 | .long 0xf3000db1 // vmul.f32 d0, d16, d17 |
| 2982 | .long 0xf2201110 // vorr d1, d0, d0 |
| 2983 | .long 0xf2202110 // vorr d2, d0, d0 |
| 2984 | .long 0xe28dd004 // add sp, sp, #4 |
| 2985 | .long 0xe12fff1c // bx ip |
| 2986 | .long 0xe320f000 // nop {0} |
| 2987 | .long 0x3b808081 // .word 0x3b808081 |
| 2988 | .long 0x3b808081 // .word 0x3b808081 |
| 2989 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 2990 | HIDDEN _sk_load_565_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 2991 | .globl _sk_load_565_vfp4 |
| 2992 | _sk_load_565_vfp4: |
| 2993 | .long 0xe24dd004 // sub sp, sp, #4 |
| 2994 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 2995 | .long 0xf2c1101f // vmov.i32 d17, #31 |
| 2996 | .long 0xf3c72218 // vmov.i32 d18, #63488 |
| 2997 | .long 0xeddf3b16 // vldr d19, [pc, #88] |
| 2998 | .long 0xe2811008 // add r1, r1, #8 |
| 2999 | .long 0xe5933000 // ldr r3, [r3] |
| 3000 | .long 0xf2873f10 // vmov.f32 d3, #1 |
| 3001 | .long 0xe7933080 // ldr r3, [r3, r0, lsl #1] |
| 3002 | .long 0xe58d3000 // str r3, [sp] |
| 3003 | .long 0xe1a0300d // mov r3, sp |
| 3004 | .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| 3005 | .long 0xe3a03e7e // mov r3, #2016 |
| 3006 | .long 0xf3d04a30 // vmovl.u16 q10, d16 |
| 3007 | .long 0xee803b90 // vdup.32 d16, r3 |
| 3008 | .long 0xf24411b1 // vand d17, d20, d17 |
| 3009 | .long 0xeddf5b0e // vldr d21, [pc, #56] |
| 3010 | .long 0xf24421b2 // vand d18, d20, d18 |
| 3011 | .long 0xf24401b0 // vand d16, d20, d16 |
| 3012 | .long 0xeddf4b09 // vldr d20, [pc, #36] |
| 3013 | .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| 3014 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 3015 | .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| 3016 | .long 0xf3020db3 // vmul.f32 d0, d18, d19 |
| 3017 | .long 0xf3001db4 // vmul.f32 d1, d16, d20 |
| 3018 | .long 0xf3012db5 // vmul.f32 d2, d17, d21 |
| 3019 | .long 0xe28dd004 // add sp, sp, #4 |
| 3020 | .long 0xe12fff1c // bx ip |
| 3021 | .long 0x37842108 // .word 0x37842108 |
| 3022 | .long 0x37842108 // .word 0x37842108 |
| 3023 | .long 0x3a020821 // .word 0x3a020821 |
| 3024 | .long 0x3a020821 // .word 0x3a020821 |
| 3025 | .long 0x3d042108 // .word 0x3d042108 |
| 3026 | .long 0x3d042108 // .word 0x3d042108 |
| 3027 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3028 | HIDDEN _sk_store_565_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3029 | .globl _sk_store_565_vfp4 |
| 3030 | _sk_store_565_vfp4: |
| 3031 | .long 0xf2c30f1f // vmov.f32 d16, #31 |
| 3032 | .long 0xeddf1b15 // vldr d17, [pc, #84] |
| 3033 | .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| 3034 | .long 0xe5913000 // ldr r3, [r1] |
| 3035 | .long 0xf2413c31 // vfma.f32 d19, d1, d17 |
| 3036 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 3037 | .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| 3038 | .long 0xe5933000 // ldr r3, [r3] |
| 3039 | .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| 3040 | .long 0xf2422c30 // vfma.f32 d18, d2, d16 |
| 3041 | .long 0xe0833080 // add r3, r3, r0, lsl #1 |
| 3042 | .long 0xf3fb07a3 // vcvt.u32.f32 d16, d19 |
| 3043 | .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| 3044 | .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18 |
| 3045 | .long 0xf2e50530 // vshl.s32 d16, d16, #5 |
| 3046 | .long 0xf2eb1531 // vshl.s32 d17, d17, #11 |
| 3047 | .long 0xf26001b1 // vorr d16, d16, d17 |
| 3048 | .long 0xf26001b2 // vorr d16, d16, d18 |
| 3049 | .long 0xf3f60121 // vuzp.16 d16, d17 |
| 3050 | .long 0xf4c3080f // vst1.32 {d16[0]}, [r3] |
| 3051 | .long 0xe2813008 // add r3, r1, #8 |
| 3052 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3053 | .long 0xe1a01003 // mov r1, r3 |
| 3054 | .long 0xe12fff1c // bx ip |
| 3055 | .long 0x427c0000 // .word 0x427c0000 |
| 3056 | .long 0x427c0000 // .word 0x427c0000 |
| 3057 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 3058 | HIDDEN _sk_load_4444_vfp4 |
| 3059 | .globl _sk_load_4444_vfp4 |
| 3060 | _sk_load_4444_vfp4: |
| 3061 | .long 0xe24dd004 // sub sp, sp, #4 |
| 3062 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3063 | .long 0xf3c71210 // vmov.i32 d17, #61440 |
| 3064 | .long 0xf3c74010 // vmov.i32 d20, #240 |
| 3065 | .long 0xf2c0501f // vmov.i32 d21, #15 |
| 3066 | .long 0xeddf6b1d // vldr d22, [pc, #116] |
| 3067 | .long 0xe5933000 // ldr r3, [r3] |
| 3068 | .long 0xe2811008 // add r1, r1, #8 |
| 3069 | .long 0xe7933080 // ldr r3, [r3, r0, lsl #1] |
| 3070 | .long 0xe58d3000 // str r3, [sp] |
| 3071 | .long 0xe1a0300d // mov r3, sp |
| 3072 | .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32] |
| 3073 | .long 0xf3d02a30 // vmovl.u16 q9, d16 |
| 3074 | .long 0xf2c0021f // vmov.i32 d16, #3840 |
| 3075 | .long 0xf24211b1 // vand d17, d18, d17 |
| 3076 | .long 0xf24201b0 // vand d16, d18, d16 |
| 3077 | .long 0xf24241b4 // vand d20, d18, d20 |
| 3078 | .long 0xf24221b5 // vand d18, d18, d21 |
| 3079 | .long 0xeddf3b0c // vldr d19, [pc, #48] |
| 3080 | .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| 3081 | .long 0xeddf5b0c // vldr d21, [pc, #48] |
| 3082 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 3083 | .long 0xf3fb4624 // vcvt.f32.s32 d20, d20 |
| 3084 | .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| 3085 | .long 0xf3010db3 // vmul.f32 d0, d17, d19 |
| 3086 | .long 0xeddf1b0b // vldr d17, [pc, #44] |
| 3087 | .long 0xf3001db5 // vmul.f32 d1, d16, d21 |
| 3088 | .long 0xf3042db6 // vmul.f32 d2, d20, d22 |
| 3089 | .long 0xf3023db1 // vmul.f32 d3, d18, d17 |
| 3090 | .long 0xe28dd004 // add sp, sp, #4 |
| 3091 | .long 0xe12fff1c // bx ip |
| 3092 | .long 0xe320f000 // nop {0} |
| 3093 | .long 0x37888889 // .word 0x37888889 |
| 3094 | .long 0x37888889 // .word 0x37888889 |
| 3095 | .long 0x39888889 // .word 0x39888889 |
| 3096 | .long 0x39888889 // .word 0x39888889 |
| 3097 | .long 0x3b888889 // .word 0x3b888889 |
| 3098 | .long 0x3b888889 // .word 0x3b888889 |
| 3099 | .long 0x3d888889 // .word 0x3d888889 |
| 3100 | .long 0x3d888889 // .word 0x3d888889 |
| 3101 | |
| 3102 | HIDDEN _sk_store_4444_vfp4 |
| 3103 | .globl _sk_store_4444_vfp4 |
| 3104 | _sk_store_4444_vfp4: |
| 3105 | .long 0xf2c20f1e // vmov.f32 d16, #15 |
| 3106 | .long 0xe5913000 // ldr r3, [r1] |
| 3107 | .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| 3108 | .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| 3109 | .long 0xf2402c30 // vfma.f32 d18, d0, d16 |
| 3110 | .long 0xe5933000 // ldr r3, [r3] |
| 3111 | .long 0xf2413c30 // vfma.f32 d19, d1, d16 |
| 3112 | .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| 3113 | .long 0xe0833080 // add r3, r3, r0, lsl #1 |
| 3114 | .long 0xf2424c30 // vfma.f32 d20, d2, d16 |
| 3115 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 3116 | .long 0xf2431c30 // vfma.f32 d17, d3, d16 |
| 3117 | .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| 3118 | .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19 |
| 3119 | .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20 |
| 3120 | .long 0xf2ec0530 // vshl.s32 d16, d16, #12 |
| 3121 | .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| 3122 | .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| 3123 | .long 0xf2e43533 // vshl.s32 d19, d19, #4 |
| 3124 | .long 0xf26201b0 // vorr d16, d18, d16 |
| 3125 | .long 0xf26001b3 // vorr d16, d16, d19 |
| 3126 | .long 0xf26001b1 // vorr d16, d16, d17 |
| 3127 | .long 0xf3f60121 // vuzp.16 d16, d17 |
| 3128 | .long 0xf4c3080f // vst1.32 {d16[0]}, [r3] |
| 3129 | .long 0xe2813008 // add r3, r1, #8 |
| 3130 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3131 | .long 0xe1a01003 // mov r1, r3 |
| 3132 | .long 0xe12fff1c // bx ip |
| 3133 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3134 | HIDDEN _sk_load_8888_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3135 | .globl _sk_load_8888_vfp4 |
| 3136 | _sk_load_8888_vfp4: |
| 3137 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3138 | .long 0xf3c7001f // vmov.i32 d16, #255 |
| 3139 | .long 0xe2811008 // add r1, r1, #8 |
| 3140 | .long 0xe5933000 // ldr r3, [r3] |
| 3141 | .long 0xe0833100 // add r3, r3, r0, lsl #2 |
| 3142 | .long 0xedd31b00 // vldr d17, [r3] |
| 3143 | .long 0xf24121b0 // vand d18, d17, d16 |
| 3144 | .long 0xf3f83031 // vshr.u32 d19, d17, #8 |
| 3145 | .long 0xf3e84031 // vshr.u32 d20, d17, #24 |
| 3146 | .long 0xf3f01031 // vshr.u32 d17, d17, #16 |
| 3147 | .long 0xf24331b0 // vand d19, d19, d16 |
| 3148 | .long 0xf24101b0 // vand d16, d17, d16 |
| 3149 | .long 0xeddf1b08 // vldr d17, [pc, #32] |
| 3150 | .long 0xf3fb2622 // vcvt.f32.s32 d18, d18 |
| 3151 | .long 0xf3fb4624 // vcvt.f32.s32 d20, d20 |
| 3152 | .long 0xf3fb3623 // vcvt.f32.s32 d19, d19 |
| 3153 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 3154 | .long 0xf3020db1 // vmul.f32 d0, d18, d17 |
| 3155 | .long 0xf3043db1 // vmul.f32 d3, d20, d17 |
| 3156 | .long 0xf3031db1 // vmul.f32 d1, d19, d17 |
| 3157 | .long 0xf3002db1 // vmul.f32 d2, d16, d17 |
| 3158 | .long 0xe12fff1c // bx ip |
| 3159 | .long 0x3b808081 // .word 0x3b808081 |
| 3160 | .long 0x3b808081 // .word 0x3b808081 |
| 3161 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3162 | HIDDEN _sk_store_8888_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3163 | .globl _sk_store_8888_vfp4 |
| 3164 | _sk_store_8888_vfp4: |
| 3165 | .long 0xeddf0b1a // vldr d16, [pc, #104] |
| 3166 | .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| 3167 | .long 0xf2412c30 // vfma.f32 d18, d1, d16 |
| 3168 | .long 0xe5913000 // ldr r3, [r1] |
| 3169 | .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| 3170 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 3171 | .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| 3172 | .long 0xe5933000 // ldr r3, [r3] |
| 3173 | .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| 3174 | .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| 3175 | .long 0xe0833100 // add r3, r3, r0, lsl #2 |
| 3176 | .long 0xf2434c30 // vfma.f32 d20, d3, d16 |
| 3177 | .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| 3178 | .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19 |
| 3179 | .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| 3180 | .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20 |
| 3181 | .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| 3182 | .long 0xf2f02532 // vshl.s32 d18, d18, #16 |
| 3183 | .long 0xf26001b1 // vorr d16, d16, d17 |
| 3184 | .long 0xf2f81533 // vshl.s32 d17, d19, #24 |
| 3185 | .long 0xf26001b2 // vorr d16, d16, d18 |
| 3186 | .long 0xf26001b1 // vorr d16, d16, d17 |
| 3187 | .long 0xedc30b00 // vstr d16, [r3] |
| 3188 | .long 0xe2813008 // add r3, r1, #8 |
| 3189 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3190 | .long 0xe1a01003 // mov r1, r3 |
| 3191 | .long 0xe12fff1c // bx ip |
| 3192 | .long 0xe320f000 // nop {0} |
| 3193 | .long 0x437f0000 // .word 0x437f0000 |
| 3194 | .long 0x437f0000 // .word 0x437f0000 |
| 3195 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3196 | HIDDEN _sk_load_f16_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3197 | .globl _sk_load_f16_vfp4 |
| 3198 | _sk_load_f16_vfp4: |
| 3199 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3200 | .long 0xe2811008 // add r1, r1, #8 |
| 3201 | .long 0xe5933000 // ldr r3, [r3] |
| 3202 | .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 3203 | .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]! |
| 3204 | .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3] |
| 3205 | .long 0xf3b60720 // vcvt.f32.f16 q0, d16 |
| 3206 | .long 0xf3b62722 // vcvt.f32.f16 q1, d18 |
| 3207 | .long 0xf3f64721 // vcvt.f32.f16 q10, d17 |
| 3208 | .long 0xf3f60723 // vcvt.f32.f16 q8, d19 |
| 3209 | .long 0xf22411b4 // vorr d1, d20, d20 |
| 3210 | .long 0xf22031b0 // vorr d3, d16, d16 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3211 | .long 0xe12fff1c // bx ip |
| 3212 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3213 | HIDDEN _sk_store_f16_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3214 | .globl _sk_store_f16_vfp4 |
| 3215 | _sk_store_f16_vfp4: |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 3216 | .long 0xf2630113 // vorr d16, d3, d3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3217 | .long 0xe5913000 // ldr r3, [r1] |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 3218 | .long 0xf2612111 // vorr d18, d1, d1 |
| 3219 | .long 0xf3f67620 // vcvt.f16.f32 d23, q8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3220 | .long 0xe5933000 // ldr r3, [r3] |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 3221 | .long 0xf3f66602 // vcvt.f16.f32 d22, q1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3222 | .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 3223 | .long 0xf3f65622 // vcvt.f16.f32 d21, q9 |
| 3224 | .long 0xf3f64600 // vcvt.f16.f32 d20, q0 |
| 3225 | .long 0xf22211b2 // vorr d1, d18, d18 |
| 3226 | .long 0xf22031b0 // vorr d3, d16, d16 |
| 3227 | .long 0xf4c3470d // vst4.16 {d20[0],d21[0],d22[0],d23[0]}, [r3]! |
| 3228 | .long 0xf4c3474f // vst4.16 {d20[1],d21[1],d22[1],d23[1]}, [r3] |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3229 | .long 0xe2813008 // add r3, r1, #8 |
| 3230 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3231 | .long 0xe1a01003 // mov r1, r3 |
| 3232 | .long 0xe12fff1c // bx ip |
| 3233 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 3234 | HIDDEN _sk_load_u16_be_vfp4 |
| 3235 | .globl _sk_load_u16_be_vfp4 |
| 3236 | _sk_load_u16_be_vfp4: |
| 3237 | .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr} |
| 3238 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3239 | .long 0xe2811008 // add r1, r1, #8 |
| 3240 | .long 0xe5933000 // ldr r3, [r3] |
| 3241 | .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| 3242 | .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]! |
| 3243 | .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3] |
| 3244 | .long 0xee903bb0 // vmov.u16 r3, d16[0] |
| 3245 | .long 0xee92ebb0 // vmov.u16 lr, d18[0] |
| 3246 | .long 0xee914bb0 // vmov.u16 r4, d17[0] |
| 3247 | .long 0xee937bb0 // vmov.u16 r7, d19[0] |
| 3248 | .long 0xee905bf0 // vmov.u16 r5, d16[1] |
| 3249 | .long 0xee926bf0 // vmov.u16 r6, d18[1] |
| 3250 | .long 0xee043b90 // vmov.32 d20[0], r3 |
| 3251 | .long 0xee05eb90 // vmov.32 d21[0], lr |
| 3252 | .long 0xee93ebf0 // vmov.u16 lr, d19[1] |
| 3253 | .long 0xee913bf0 // vmov.u16 r3, d17[1] |
| 3254 | .long 0xf3c71c1f // vmov.i32 d17, #65535 |
| 3255 | .long 0xee004b90 // vmov.32 d16[0], r4 |
| 3256 | .long 0xee027b90 // vmov.32 d18[0], r7 |
| 3257 | .long 0xee245b90 // vmov.32 d20[1], r5 |
| 3258 | .long 0xf24431b1 // vand d19, d20, d17 |
| 3259 | .long 0xee256b90 // vmov.32 d21[1], r6 |
| 3260 | .long 0xf2e84534 // vshl.s32 d20, d20, #8 |
| 3261 | .long 0xf24561b1 // vand d22, d21, d17 |
| 3262 | .long 0xf3f83033 // vshr.u32 d19, d19, #8 |
| 3263 | .long 0xf2e85535 // vshl.s32 d21, d21, #8 |
| 3264 | .long 0xf26431b3 // vorr d19, d20, d19 |
| 3265 | .long 0xf3f86036 // vshr.u32 d22, d22, #8 |
| 3266 | .long 0xf24331b1 // vand d19, d19, d17 |
| 3267 | .long 0xf26551b6 // vorr d21, d21, d22 |
| 3268 | .long 0xf3fb36a3 // vcvt.f32.u32 d19, d19 |
| 3269 | .long 0xee22eb90 // vmov.32 d18[1], lr |
| 3270 | .long 0xee203b90 // vmov.32 d16[1], r3 |
| 3271 | .long 0xf24281b1 // vand d24, d18, d17 |
| 3272 | .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| 3273 | .long 0xf24071b1 // vand d23, d16, d17 |
| 3274 | .long 0xf3f84038 // vshr.u32 d20, d24, #8 |
| 3275 | .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| 3276 | .long 0xf3f87037 // vshr.u32 d23, d23, #8 |
| 3277 | .long 0xf26221b4 // vorr d18, d18, d20 |
| 3278 | .long 0xf26001b7 // vorr d16, d16, d23 |
| 3279 | .long 0xf24541b1 // vand d20, d21, d17 |
| 3280 | .long 0xf24001b1 // vand d16, d16, d17 |
| 3281 | .long 0xf24211b1 // vand d17, d18, d17 |
| 3282 | .long 0xeddf2b09 // vldr d18, [pc, #36] |
| 3283 | .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16 |
| 3284 | .long 0xf3fb46a4 // vcvt.f32.u32 d20, d20 |
| 3285 | .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17 |
| 3286 | .long 0xf3030db2 // vmul.f32 d0, d19, d18 |
| 3287 | .long 0xf3001db2 // vmul.f32 d1, d16, d18 |
| 3288 | .long 0xf3042db2 // vmul.f32 d2, d20, d18 |
| 3289 | .long 0xf3013db2 // vmul.f32 d3, d17, d18 |
| 3290 | .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr} |
| 3291 | .long 0xe12fff1c // bx ip |
| 3292 | .long 0xe320f000 // nop {0} |
| 3293 | .long 0x37800080 // .word 0x37800080 |
| 3294 | .long 0x37800080 // .word 0x37800080 |
| 3295 | |
| 3296 | HIDDEN _sk_store_u16_be_vfp4 |
| 3297 | .globl _sk_store_u16_be_vfp4 |
| 3298 | _sk_store_u16_be_vfp4: |
| 3299 | .long 0xeddf0b2a // vldr d16, [pc, #168] |
| 3300 | .long 0xf2c3261f // vmov.i32 d18, #1056964608 |
| 3301 | .long 0xf2c3361f // vmov.i32 d19, #1056964608 |
| 3302 | .long 0xe5913000 // ldr r3, [r1] |
| 3303 | .long 0xf2432c30 // vfma.f32 d18, d3, d16 |
| 3304 | .long 0xf2c3461f // vmov.i32 d20, #1056964608 |
| 3305 | .long 0xf2423c30 // vfma.f32 d19, d2, d16 |
| 3306 | .long 0xf2c3161f // vmov.i32 d17, #1056964608 |
| 3307 | .long 0xf2414c30 // vfma.f32 d20, d1, d16 |
| 3308 | .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| 3309 | .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18 |
| 3310 | .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19 |
| 3311 | .long 0xf3c73c1f // vmov.i32 d19, #65535 |
| 3312 | .long 0xf3fb47a4 // vcvt.u32.f32 d20, d20 |
| 3313 | .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17 |
| 3314 | .long 0xf24051b3 // vand d21, d16, d19 |
| 3315 | .long 0xf24261b3 // vand d22, d18, d19 |
| 3316 | .long 0xf24471b3 // vand d23, d20, d19 |
| 3317 | .long 0xf24131b3 // vand d19, d17, d19 |
| 3318 | .long 0xf2e80530 // vshl.s32 d16, d16, #8 |
| 3319 | .long 0xf3f85035 // vshr.u32 d21, d21, #8 |
| 3320 | .long 0xf2e82532 // vshl.s32 d18, d18, #8 |
| 3321 | .long 0xf3f86036 // vshr.u32 d22, d22, #8 |
| 3322 | .long 0xf260b1b5 // vorr d27, d16, d21 |
| 3323 | .long 0xf2e84534 // vshl.s32 d20, d20, #8 |
| 3324 | .long 0xf3f87037 // vshr.u32 d23, d23, #8 |
| 3325 | .long 0xf262a1b6 // vorr d26, d18, d22 |
| 3326 | .long 0xf2e81531 // vshl.s32 d17, d17, #8 |
| 3327 | .long 0xf3f83033 // vshr.u32 d19, d19, #8 |
| 3328 | .long 0xf26491b7 // vorr d25, d20, d23 |
| 3329 | .long 0xf26181b3 // vorr d24, d17, d19 |
| 3330 | .long 0xf3f6b120 // vuzp.16 d27, d16 |
| 3331 | .long 0xe5933000 // ldr r3, [r3] |
| 3332 | .long 0xf3f6a120 // vuzp.16 d26, d16 |
| 3333 | .long 0xe0833180 // add r3, r3, r0, lsl #3 |
| 3334 | .long 0xf3f69120 // vuzp.16 d25, d16 |
| 3335 | .long 0xf3f68120 // vuzp.16 d24, d16 |
| 3336 | .long 0xf4c3870d // vst4.16 {d24[0],d25[0],d26[0],d27[0]}, [r3]! |
| 3337 | .long 0xf4c3874f // vst4.16 {d24[1],d25[1],d26[1],d27[1]}, [r3] |
| 3338 | .long 0xe2813008 // add r3, r1, #8 |
| 3339 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3340 | .long 0xe1a01003 // mov r1, r3 |
| 3341 | .long 0xe12fff1c // bx ip |
| 3342 | .long 0xe320f000 // nop {0} |
| 3343 | .long 0x477fff00 // .word 0x477fff00 |
| 3344 | .long 0x477fff00 // .word 0x477fff00 |
| 3345 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3346 | HIDDEN _sk_store_f32_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3347 | .globl _sk_store_f32_vfp4 |
| 3348 | _sk_store_f32_vfp4: |
| 3349 | .long 0xe5913000 // ldr r3, [r1] |
| 3350 | .long 0xe5933000 // ldr r3, [r3] |
| 3351 | .long 0xe0833200 // add r3, r3, r0, lsl #4 |
| 3352 | .long 0xf403008f // vst4.32 {d0-d3}, [r3] |
| 3353 | .long 0xe2813008 // add r3, r1, #8 |
| 3354 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3355 | .long 0xe1a01003 // mov r1, r3 |
| 3356 | .long 0xe12fff1c // bx ip |
| 3357 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3358 | HIDDEN _sk_clamp_x_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3359 | .globl _sk_clamp_x_vfp4 |
| 3360 | _sk_clamp_x_vfp4: |
| 3361 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3362 | .long 0xf2c00010 // vmov.i32 d16, #0 |
| 3363 | .long 0xf3c71e1f // vmov.i8 d17, #255 |
| 3364 | .long 0xf2400f80 // vmax.f32 d16, d16, d0 |
| 3365 | .long 0xe2811008 // add r1, r1, #8 |
| 3366 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3367 | .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| 3368 | .long 0xf2200fa1 // vmin.f32 d0, d16, d17 |
| 3369 | .long 0xe12fff1c // bx ip |
| 3370 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3371 | HIDDEN _sk_clamp_y_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3372 | .globl _sk_clamp_y_vfp4 |
| 3373 | _sk_clamp_y_vfp4: |
| 3374 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3375 | .long 0xf2c00010 // vmov.i32 d16, #0 |
| 3376 | .long 0xf3c71e1f // vmov.i8 d17, #255 |
| 3377 | .long 0xf2400f81 // vmax.f32 d16, d16, d1 |
| 3378 | .long 0xe2811008 // add r1, r1, #8 |
| 3379 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3380 | .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| 3381 | .long 0xf2201fa1 // vmin.f32 d1, d16, d17 |
| 3382 | .long 0xe12fff1c // bx ip |
| 3383 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3384 | HIDDEN _sk_repeat_x_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3385 | .globl _sk_repeat_x_vfp4 |
| 3386 | _sk_repeat_x_vfp4: |
| 3387 | .long 0xed2d8b04 // vpush {d8-d9} |
| 3388 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3389 | .long 0xf2c02010 // vmov.i32 d18, #0 |
| 3390 | .long 0xe2811008 // add r1, r1, #8 |
| 3391 | .long 0xeddf3b10 // vldr d19, [pc, #64] |
| 3392 | .long 0xed938a00 // vldr s16, [r3] |
| 3393 | .long 0xeec09a88 // vdiv.f32 s19, s1, s16 |
| 3394 | .long 0xee809a08 // vdiv.f32 s18, s0, s16 |
| 3395 | .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 |
| 3396 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 3397 | .long 0xf3601e89 // vcgt.f32 d17, d16, d9 |
| 3398 | .long 0xf35311b2 // vbsl d17, d19, d18 |
| 3399 | .long 0xf3f42c08 // vdup.32 d18, d8[0] |
| 3400 | .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| 3401 | .long 0xf3c71e1f // vmov.i8 d17, #255 |
| 3402 | .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| 3403 | .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] |
| 3404 | .long 0xf2600d20 // vsub.f32 d16, d0, d16 |
| 3405 | .long 0xf2200fa1 // vmin.f32 d0, d16, d17 |
| 3406 | .long 0xecbd8b04 // vpop {d8-d9} |
| 3407 | .long 0xe12fff1c // bx ip |
| 3408 | .long 0xe320f000 // nop {0} |
| 3409 | .long 0x3f800000 // .word 0x3f800000 |
| 3410 | .long 0x3f800000 // .word 0x3f800000 |
| 3411 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3412 | HIDDEN _sk_repeat_y_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3413 | .globl _sk_repeat_y_vfp4 |
| 3414 | _sk_repeat_y_vfp4: |
| 3415 | .long 0xed2d8b04 // vpush {d8-d9} |
| 3416 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3417 | .long 0xf2c02010 // vmov.i32 d18, #0 |
| 3418 | .long 0xe2811008 // add r1, r1, #8 |
| 3419 | .long 0xeddf3b10 // vldr d19, [pc, #64] |
| 3420 | .long 0xed938a00 // vldr s16, [r3] |
| 3421 | .long 0xeec19a88 // vdiv.f32 s19, s3, s16 |
| 3422 | .long 0xee819a08 // vdiv.f32 s18, s2, s16 |
| 3423 | .long 0xf3fb0709 // vcvt.s32.f32 d16, d9 |
| 3424 | .long 0xf3fb0620 // vcvt.f32.s32 d16, d16 |
| 3425 | .long 0xf3601e89 // vcgt.f32 d17, d16, d9 |
| 3426 | .long 0xf35311b2 // vbsl d17, d19, d18 |
| 3427 | .long 0xf3f42c08 // vdup.32 d18, d8[0] |
| 3428 | .long 0xf2600da1 // vsub.f32 d16, d16, d17 |
| 3429 | .long 0xf3c71e1f // vmov.i8 d17, #255 |
| 3430 | .long 0xf26218a1 // vadd.i32 d17, d18, d17 |
| 3431 | .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0] |
| 3432 | .long 0xf2610d20 // vsub.f32 d16, d1, d16 |
| 3433 | .long 0xf2201fa1 // vmin.f32 d1, d16, d17 |
| 3434 | .long 0xecbd8b04 // vpop {d8-d9} |
| 3435 | .long 0xe12fff1c // bx ip |
| 3436 | .long 0xe320f000 // nop {0} |
| 3437 | .long 0x3f800000 // .word 0x3f800000 |
| 3438 | .long 0x3f800000 // .word 0x3f800000 |
| 3439 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3440 | HIDDEN _sk_mirror_x_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3441 | .globl _sk_mirror_x_vfp4 |
| 3442 | _sk_mirror_x_vfp4: |
| 3443 | .long 0xed2d8b04 // vpush {d8-d9} |
| 3444 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3445 | .long 0xf2c03010 // vmov.i32 d19, #0 |
| 3446 | .long 0xe2811008 // add r1, r1, #8 |
| 3447 | .long 0xeddf4b14 // vldr d20, [pc, #80] |
| 3448 | .long 0xed938a00 // vldr s16, [r3] |
| 3449 | .long 0xee389a08 // vadd.f32 s18, s16, s16 |
| 3450 | .long 0xf3f40c08 // vdup.32 d16, d8[0] |
| 3451 | .long 0xf2200d20 // vsub.f32 d0, d0, d16 |
| 3452 | .long 0xeec08a89 // vdiv.f32 s17, s1, s18 |
| 3453 | .long 0xee808a09 // vdiv.f32 s16, s0, s18 |
| 3454 | .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 |
| 3455 | .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| 3456 | .long 0xf3612e88 // vcgt.f32 d18, d17, d8 |
| 3457 | .long 0xf35421b3 // vbsl d18, d20, d19 |
| 3458 | .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| 3459 | .long 0xf3c72e1f // vmov.i8 d18, #255 |
| 3460 | .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] |
| 3461 | .long 0xf2601d21 // vsub.f32 d17, d0, d17 |
| 3462 | .long 0xf2611da0 // vsub.f32 d17, d17, d16 |
| 3463 | .long 0xf26008a2 // vadd.i32 d16, d16, d18 |
| 3464 | .long 0xf3f91721 // vabs.f32 d17, d17 |
| 3465 | .long 0xf2210fa0 // vmin.f32 d0, d17, d16 |
| 3466 | .long 0xecbd8b04 // vpop {d8-d9} |
| 3467 | .long 0xe12fff1c // bx ip |
| 3468 | .long 0xe320f000 // nop {0} |
| 3469 | .long 0x3f800000 // .word 0x3f800000 |
| 3470 | .long 0x3f800000 // .word 0x3f800000 |
| 3471 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3472 | HIDDEN _sk_mirror_y_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3473 | .globl _sk_mirror_y_vfp4 |
| 3474 | _sk_mirror_y_vfp4: |
| 3475 | .long 0xed2d8b04 // vpush {d8-d9} |
| 3476 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3477 | .long 0xf2c03010 // vmov.i32 d19, #0 |
| 3478 | .long 0xe2811008 // add r1, r1, #8 |
| 3479 | .long 0xeddf4b14 // vldr d20, [pc, #80] |
| 3480 | .long 0xed938a00 // vldr s16, [r3] |
| 3481 | .long 0xee389a08 // vadd.f32 s18, s16, s16 |
| 3482 | .long 0xf3f40c08 // vdup.32 d16, d8[0] |
| 3483 | .long 0xf2211d20 // vsub.f32 d1, d1, d16 |
| 3484 | .long 0xeec18a89 // vdiv.f32 s17, s3, s18 |
| 3485 | .long 0xee818a09 // vdiv.f32 s16, s2, s18 |
| 3486 | .long 0xf3fb1708 // vcvt.s32.f32 d17, d8 |
| 3487 | .long 0xf3fb1621 // vcvt.f32.s32 d17, d17 |
| 3488 | .long 0xf3612e88 // vcgt.f32 d18, d17, d8 |
| 3489 | .long 0xf35421b3 // vbsl d18, d20, d19 |
| 3490 | .long 0xf2611da2 // vsub.f32 d17, d17, d18 |
| 3491 | .long 0xf3c72e1f // vmov.i8 d18, #255 |
| 3492 | .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0] |
| 3493 | .long 0xf2611d21 // vsub.f32 d17, d1, d17 |
| 3494 | .long 0xf2611da0 // vsub.f32 d17, d17, d16 |
| 3495 | .long 0xf26008a2 // vadd.i32 d16, d16, d18 |
| 3496 | .long 0xf3f91721 // vabs.f32 d17, d17 |
| 3497 | .long 0xf2211fa0 // vmin.f32 d1, d17, d16 |
| 3498 | .long 0xecbd8b04 // vpop {d8-d9} |
| 3499 | .long 0xe12fff1c // bx ip |
| 3500 | .long 0xe320f000 // nop {0} |
| 3501 | .long 0x3f800000 // .word 0x3f800000 |
| 3502 | .long 0x3f800000 // .word 0x3f800000 |
| 3503 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3504 | HIDDEN _sk_luminance_to_alpha_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3505 | .globl _sk_luminance_to_alpha_vfp4 |
| 3506 | _sk_luminance_to_alpha_vfp4: |
| 3507 | .long 0xeddf0b0a // vldr d16, [pc, #40] |
| 3508 | .long 0xeddf1b0b // vldr d17, [pc, #44] |
| 3509 | .long 0xf3410d30 // vmul.f32 d16, d1, d16 |
| 3510 | .long 0xe4913004 // ldr r3, [r1], #4 |
| 3511 | .long 0xf3401d31 // vmul.f32 d17, d0, d17 |
| 3512 | .long 0xf2800010 // vmov.i32 d0, #0 |
| 3513 | .long 0xf2801010 // vmov.i32 d1, #0 |
| 3514 | .long 0xf2013da0 // vadd.f32 d3, d17, d16 |
| 3515 | .long 0xeddf0b06 // vldr d16, [pc, #24] |
| 3516 | .long 0xf2023c30 // vfma.f32 d3, d2, d16 |
| 3517 | .long 0xf2802010 // vmov.i32 d2, #0 |
| 3518 | .long 0xe12fff13 // bx r3 |
| 3519 | .long 0x3f371759 // .word 0x3f371759 |
| 3520 | .long 0x3f371759 // .word 0x3f371759 |
| 3521 | .long 0x3e59b3d0 // .word 0x3e59b3d0 |
| 3522 | .long 0x3e59b3d0 // .word 0x3e59b3d0 |
| 3523 | .long 0x3d93dd98 // .word 0x3d93dd98 |
| 3524 | .long 0x3d93dd98 // .word 0x3d93dd98 |
| 3525 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3526 | HIDDEN _sk_matrix_2x3_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3527 | .globl _sk_matrix_2x3_vfp4 |
| 3528 | _sk_matrix_2x3_vfp4: |
| 3529 | .long 0xe92d4800 // push {fp, lr} |
| 3530 | .long 0xe591e000 // ldr lr, [r1] |
| 3531 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3532 | .long 0xe2811008 // add r1, r1, #8 |
| 3533 | .long 0xe28e300c // add r3, lr, #12 |
| 3534 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3535 | .long 0xe28e3008 // add r3, lr, #8 |
| 3536 | .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| 3537 | .long 0xe28e3010 // add r3, lr, #16 |
| 3538 | .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| 3539 | .long 0xe28e3014 // add r3, lr, #20 |
| 3540 | .long 0xf2410c31 // vfma.f32 d16, d1, d17 |
| 3541 | .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| 3542 | .long 0xf2411c32 // vfma.f32 d17, d1, d18 |
| 3543 | .long 0xf4ee2c9d // vld1.32 {d18[]}, [lr :32]! |
| 3544 | .long 0xf4ee3c9f // vld1.32 {d19[]}, [lr :32] |
| 3545 | .long 0xf2400c32 // vfma.f32 d16, d0, d18 |
| 3546 | .long 0xf2401c33 // vfma.f32 d17, d0, d19 |
| 3547 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 3548 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 3549 | .long 0xe8bd4800 // pop {fp, lr} |
| 3550 | .long 0xe12fff1c // bx ip |
| 3551 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3552 | HIDDEN _sk_matrix_3x4_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3553 | .globl _sk_matrix_3x4_vfp4 |
| 3554 | _sk_matrix_3x4_vfp4: |
| 3555 | .long 0xe92d4800 // push {fp, lr} |
| 3556 | .long 0xe591e000 // ldr lr, [r1] |
| 3557 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3558 | .long 0xe2811008 // add r1, r1, #8 |
| 3559 | .long 0xe28e3020 // add r3, lr, #32 |
| 3560 | .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| 3561 | .long 0xe28e302c // add r3, lr, #44 |
| 3562 | .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| 3563 | .long 0xe28e301c // add r3, lr, #28 |
| 3564 | .long 0xf2420c33 // vfma.f32 d16, d2, d19 |
| 3565 | .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| 3566 | .long 0xe28e3018 // add r3, lr, #24 |
| 3567 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3568 | .long 0xe28e3024 // add r3, lr, #36 |
| 3569 | .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| 3570 | .long 0xe28e3028 // add r3, lr, #40 |
| 3571 | .long 0xf2421c32 // vfma.f32 d17, d2, d18 |
| 3572 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3573 | .long 0xe28e3010 // add r3, lr, #16 |
| 3574 | .long 0xf2422c34 // vfma.f32 d18, d2, d20 |
| 3575 | .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| 3576 | .long 0xe28e300c // add r3, lr, #12 |
| 3577 | .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| 3578 | .long 0xe28e3014 // add r3, lr, #20 |
| 3579 | .long 0xf2411c34 // vfma.f32 d17, d1, d20 |
| 3580 | .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| 3581 | .long 0xf2410c34 // vfma.f32 d16, d1, d20 |
| 3582 | .long 0xe28e3008 // add r3, lr, #8 |
| 3583 | .long 0xf2412c33 // vfma.f32 d18, d1, d19 |
| 3584 | .long 0xf4ee3c9d // vld1.32 {d19[]}, [lr :32]! |
| 3585 | .long 0xf4ee4c9f // vld1.32 {d20[]}, [lr :32] |
| 3586 | .long 0xf2401c33 // vfma.f32 d17, d0, d19 |
| 3587 | .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32] |
| 3588 | .long 0xf2400c33 // vfma.f32 d16, d0, d19 |
| 3589 | .long 0xf2402c34 // vfma.f32 d18, d0, d20 |
| 3590 | .long 0xf22101b1 // vorr d0, d17, d17 |
| 3591 | .long 0xf22021b0 // vorr d2, d16, d16 |
| 3592 | .long 0xf22211b2 // vorr d1, d18, d18 |
| 3593 | .long 0xe8bd4800 // pop {fp, lr} |
| 3594 | .long 0xe12fff1c // bx ip |
| 3595 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3596 | HIDDEN _sk_matrix_4x5_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3597 | .globl _sk_matrix_4x5_vfp4 |
| 3598 | _sk_matrix_4x5_vfp4: |
| 3599 | .long 0xe92d4010 // push {r4, lr} |
| 3600 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3601 | .long 0xf2620112 // vorr d16, d2, d2 |
| 3602 | .long 0xe2811008 // add r1, r1, #8 |
| 3603 | .long 0xe2834014 // add r4, r3, #20 |
| 3604 | .long 0xe1a0e003 // mov lr, r3 |
| 3605 | .long 0xf4e45c9f // vld1.32 {d21[]}, [r4 :32] |
| 3606 | .long 0xe2834028 // add r4, r3, #40 |
| 3607 | .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32] |
| 3608 | .long 0xe2834038 // add r4, r3, #56 |
| 3609 | .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32] |
| 3610 | .long 0xe2834048 // add r4, r3, #72 |
| 3611 | .long 0xf4a42c9f // vld1.32 {d2[]}, [r4 :32] |
| 3612 | .long 0xe2834034 // add r4, r3, #52 |
| 3613 | .long 0xf2032c37 // vfma.f32 d2, d3, d23 |
| 3614 | .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| 3615 | .long 0xe2834044 // add r4, r3, #68 |
| 3616 | .long 0xf4e41c9f // vld1.32 {d17[]}, [r4 :32] |
| 3617 | .long 0xe2834030 // add r4, r3, #48 |
| 3618 | .long 0xf2431c38 // vfma.f32 d17, d3, d24 |
| 3619 | .long 0xf4e49c9f // vld1.32 {d25[]}, [r4 :32] |
| 3620 | .long 0xe283403c // add r4, r3, #60 |
| 3621 | .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| 3622 | .long 0xe283404c // add r4, r3, #76 |
| 3623 | .long 0xf2002cb6 // vfma.f32 d2, d16, d22 |
| 3624 | .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| 3625 | .long 0xe2834040 // add r4, r3, #64 |
| 3626 | .long 0xf2432c33 // vfma.f32 d18, d3, d19 |
| 3627 | .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| 3628 | .long 0xe2834020 // add r4, r3, #32 |
| 3629 | .long 0xf2433c39 // vfma.f32 d19, d3, d25 |
| 3630 | .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32] |
| 3631 | .long 0xe283402c // add r4, r3, #44 |
| 3632 | .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| 3633 | .long 0xe2834024 // add r4, r3, #36 |
| 3634 | .long 0xf2402cb8 // vfma.f32 d18, d16, d24 |
| 3635 | .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32] |
| 3636 | .long 0xf2401cb8 // vfma.f32 d17, d16, d24 |
| 3637 | .long 0xe2834010 // add r4, r3, #16 |
| 3638 | .long 0xf2403cb7 // vfma.f32 d19, d16, d23 |
| 3639 | .long 0xf4ee4c9d // vld1.32 {d20[]}, [lr :32]! |
| 3640 | .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| 3641 | .long 0xe283401c // add r4, r3, #28 |
| 3642 | .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32] |
| 3643 | .long 0xe2834018 // add r4, r3, #24 |
| 3644 | .long 0xf2412c36 // vfma.f32 d18, d1, d22 |
| 3645 | .long 0xf2411c35 // vfma.f32 d17, d1, d21 |
| 3646 | .long 0xf4ee5c9f // vld1.32 {d21[]}, [lr :32] |
| 3647 | .long 0xf2413c30 // vfma.f32 d19, d1, d16 |
| 3648 | .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| 3649 | .long 0xe2834008 // add r4, r3, #8 |
| 3650 | .long 0xe283300c // add r3, r3, #12 |
| 3651 | .long 0xf2012c30 // vfma.f32 d2, d1, d16 |
| 3652 | .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32] |
| 3653 | .long 0xf2401c35 // vfma.f32 d17, d0, d21 |
| 3654 | .long 0xf2403c34 // vfma.f32 d19, d0, d20 |
| 3655 | .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32] |
| 3656 | .long 0xf2402c34 // vfma.f32 d18, d0, d20 |
| 3657 | .long 0xf2002c30 // vfma.f32 d2, d0, d16 |
| 3658 | .long 0xf22111b1 // vorr d1, d17, d17 |
| 3659 | .long 0xf22301b3 // vorr d0, d19, d19 |
| 3660 | .long 0xf22231b2 // vorr d3, d18, d18 |
| 3661 | .long 0xe8bd4010 // pop {r4, lr} |
| 3662 | .long 0xe12fff1c // bx ip |
| 3663 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3664 | HIDDEN _sk_matrix_perspective_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3665 | .globl _sk_matrix_perspective_vfp4 |
| 3666 | _sk_matrix_perspective_vfp4: |
| 3667 | .long 0xe92d4010 // push {r4, lr} |
| 3668 | .long 0xe591e000 // ldr lr, [r1] |
| 3669 | .long 0xe591c004 // ldr ip, [r1, #4] |
| 3670 | .long 0xe2811008 // add r1, r1, #8 |
| 3671 | .long 0xe28e301c // add r3, lr, #28 |
| 3672 | .long 0xe28e4010 // add r4, lr, #16 |
| 3673 | .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| 3674 | .long 0xe28e3020 // add r3, lr, #32 |
| 3675 | .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32] |
| 3676 | .long 0xe28e3018 // add r3, lr, #24 |
| 3677 | .long 0xf2411c30 // vfma.f32 d17, d1, d16 |
| 3678 | .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32] |
| 3679 | .long 0xe1a0300e // mov r3, lr |
| 3680 | .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32] |
| 3681 | .long 0xe28e4008 // add r4, lr, #8 |
| 3682 | .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32] |
| 3683 | .long 0xf2401c30 // vfma.f32 d17, d0, d16 |
| 3684 | .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]! |
| 3685 | .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32] |
| 3686 | .long 0xe28e3014 // add r3, lr, #20 |
| 3687 | .long 0xf2413c35 // vfma.f32 d19, d1, d21 |
| 3688 | .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32] |
| 3689 | .long 0xe28e300c // add r3, lr, #12 |
| 3690 | .long 0xf2415c32 // vfma.f32 d21, d1, d18 |
| 3691 | .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32] |
| 3692 | .long 0xf3fb4521 // vrecpe.f32 d20, d17 |
| 3693 | .long 0xf2403c30 // vfma.f32 d19, d0, d16 |
| 3694 | .long 0xf2411fb4 // vrecps.f32 d17, d17, d20 |
| 3695 | .long 0xf2405c32 // vfma.f32 d21, d0, d18 |
| 3696 | .long 0xf3440db1 // vmul.f32 d16, d20, d17 |
| 3697 | .long 0xf3030db0 // vmul.f32 d0, d19, d16 |
| 3698 | .long 0xf3051db0 // vmul.f32 d1, d21, d16 |
| 3699 | .long 0xe8bd4010 // pop {r4, lr} |
| 3700 | .long 0xe12fff1c // bx ip |
| 3701 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3702 | HIDDEN _sk_linear_gradient_2stops_vfp4 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3703 | .globl _sk_linear_gradient_2stops_vfp4 |
| 3704 | _sk_linear_gradient_2stops_vfp4: |
| 3705 | .long 0xe8911008 // ldm r1, {r3, ip} |
| 3706 | .long 0xe2811008 // add r1, r1, #8 |
| 3707 | .long 0xf4632a0d // vld1.8 {d18-d19}, [r3]! |
| 3708 | .long 0xf4634a0f // vld1.8 {d20-d21}, [r3] |
| 3709 | .long 0xf3f40c22 // vdup.32 d16, d18[0] |
| 3710 | .long 0xf3f41c24 // vdup.32 d17, d20[0] |
| 3711 | .long 0xf2400c31 // vfma.f32 d16, d0, d17 |
| 3712 | .long 0xf3fc6c24 // vdup.32 d22, d20[1] |
| 3713 | .long 0xf3bc1c22 // vdup.32 d1, d18[1] |
| 3714 | .long 0xf3b42c23 // vdup.32 d2, d19[0] |
| 3715 | .long 0xf2001c36 // vfma.f32 d1, d0, d22 |
| 3716 | .long 0xf3f41c25 // vdup.32 d17, d21[0] |
| 3717 | .long 0xf3fc4c25 // vdup.32 d20, d21[1] |
| 3718 | .long 0xf2002c31 // vfma.f32 d2, d0, d17 |
| 3719 | .long 0xf3bc3c23 // vdup.32 d3, d19[1] |
| 3720 | .long 0xf2003c34 // vfma.f32 d3, d0, d20 |
| 3721 | .long 0xf22001b0 // vorr d0, d16, d16 |
| 3722 | .long 0xe12fff1c // bx ip |
| 3723 | #elif defined(__x86_64__) |
| 3724 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3725 | HIDDEN _sk_start_pipeline_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3726 | .globl _sk_start_pipeline_hsw |
| 3727 | _sk_start_pipeline_hsw: |
| 3728 | .byte 65,87 // push %r15 |
| 3729 | .byte 65,86 // push %r14 |
| 3730 | .byte 65,85 // push %r13 |
| 3731 | .byte 65,84 // push %r12 |
| 3732 | .byte 83 // push %rbx |
| 3733 | .byte 73,137,205 // mov %rcx,%r13 |
| 3734 | .byte 73,137,214 // mov %rdx,%r14 |
| 3735 | .byte 72,137,251 // mov %rdi,%rbx |
| 3736 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3737 | .byte 73,137,199 // mov %rax,%r15 |
| 3738 | .byte 73,137,244 // mov %rsi,%r12 |
| 3739 | .byte 72,141,67,8 // lea 0x8(%rbx),%rax |
| 3740 | .byte 76,57,232 // cmp %r13,%rax |
| 3741 | .byte 118,5 // jbe 28 <_sk_start_pipeline_hsw+0x28> |
| 3742 | .byte 72,137,223 // mov %rbx,%rdi |
| 3743 | .byte 235,65 // jmp 69 <_sk_start_pipeline_hsw+0x69> |
| 3744 | .byte 185,0,0,0,0 // mov $0x0,%ecx |
| 3745 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 3746 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 3747 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 3748 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 3749 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 3750 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 3751 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 3752 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 3753 | .byte 72,137,223 // mov %rbx,%rdi |
| 3754 | .byte 76,137,230 // mov %r12,%rsi |
| 3755 | .byte 76,137,242 // mov %r14,%rdx |
| 3756 | .byte 65,255,215 // callq *%r15 |
| 3757 | .byte 72,141,123,8 // lea 0x8(%rbx),%rdi |
| 3758 | .byte 72,131,195,16 // add $0x10,%rbx |
| 3759 | .byte 76,57,235 // cmp %r13,%rbx |
| 3760 | .byte 72,137,251 // mov %rdi,%rbx |
| 3761 | .byte 118,191 // jbe 28 <_sk_start_pipeline_hsw+0x28> |
| 3762 | .byte 76,137,233 // mov %r13,%rcx |
| 3763 | .byte 72,41,249 // sub %rdi,%rcx |
| 3764 | .byte 116,41 // je 9a <_sk_start_pipeline_hsw+0x9a> |
| 3765 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 3766 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 3767 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 3768 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 3769 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 3770 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 3771 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 3772 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 3773 | .byte 76,137,230 // mov %r12,%rsi |
| 3774 | .byte 76,137,242 // mov %r14,%rdx |
| 3775 | .byte 65,255,215 // callq *%r15 |
| 3776 | .byte 76,137,232 // mov %r13,%rax |
| 3777 | .byte 91 // pop %rbx |
| 3778 | .byte 65,92 // pop %r12 |
| 3779 | .byte 65,93 // pop %r13 |
| 3780 | .byte 65,94 // pop %r14 |
| 3781 | .byte 65,95 // pop %r15 |
| 3782 | .byte 197,248,119 // vzeroupper |
| 3783 | .byte 195 // retq |
| 3784 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3785 | HIDDEN _sk_just_return_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3786 | .globl _sk_just_return_hsw |
| 3787 | _sk_just_return_hsw: |
| 3788 | .byte 195 // retq |
| 3789 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3790 | HIDDEN _sk_seed_shader_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3791 | .globl _sk_seed_shader_hsw |
| 3792 | _sk_seed_shader_hsw: |
| 3793 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3794 | .byte 197,249,110,199 // vmovd %edi,%xmm0 |
| 3795 | .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| 3796 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 3797 | .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| 3798 | .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| 3799 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 3800 | .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| 3801 | .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 |
| 3802 | .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 |
| 3803 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 3804 | .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| 3805 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3806 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 3807 | .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| 3808 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3809 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 3810 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 3811 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 3812 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 3813 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 3814 | .byte 255,224 // jmpq *%rax |
| 3815 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3816 | HIDDEN _sk_constant_color_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3817 | .globl _sk_constant_color_hsw |
| 3818 | _sk_constant_color_hsw: |
| 3819 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3820 | .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| 3821 | .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| 3822 | .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| 3823 | .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| 3824 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3825 | .byte 255,224 // jmpq *%rax |
| 3826 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3827 | HIDDEN _sk_clear_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3828 | .globl _sk_clear_hsw |
| 3829 | _sk_clear_hsw: |
| 3830 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3831 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 3832 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 3833 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 3834 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 3835 | .byte 255,224 // jmpq *%rax |
| 3836 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 3837 | HIDDEN _sk_srcatop_hsw |
| 3838 | .globl _sk_srcatop_hsw |
| 3839 | _sk_srcatop_hsw: |
| 3840 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3841 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3842 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3843 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 3844 | .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 |
| 3845 | .byte 196,194,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm0 |
| 3846 | .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9 |
| 3847 | .byte 196,194,69,168,201 // vfmadd213ps %ymm9,%ymm7,%ymm1 |
| 3848 | .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9 |
| 3849 | .byte 196,194,69,168,209 // vfmadd213ps %ymm9,%ymm7,%ymm2 |
| 3850 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 3851 | .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| 3852 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3853 | .byte 255,224 // jmpq *%rax |
| 3854 | |
| 3855 | HIDDEN _sk_dstatop_hsw |
| 3856 | .globl _sk_dstatop_hsw |
| 3857 | _sk_dstatop_hsw: |
| 3858 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3859 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3860 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3861 | .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| 3862 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 3863 | .byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0 |
| 3864 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 3865 | .byte 196,226,101,184,205 // vfmadd231ps %ymm5,%ymm3,%ymm1 |
| 3866 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 3867 | .byte 196,226,101,184,214 // vfmadd231ps %ymm6,%ymm3,%ymm2 |
| 3868 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 3869 | .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| 3870 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3871 | .byte 255,224 // jmpq *%rax |
| 3872 | |
| 3873 | HIDDEN _sk_srcin_hsw |
| 3874 | .globl _sk_srcin_hsw |
| 3875 | _sk_srcin_hsw: |
| 3876 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 3877 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 3878 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 3879 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 3880 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3881 | .byte 255,224 // jmpq *%rax |
| 3882 | |
| 3883 | HIDDEN _sk_dstin_hsw |
| 3884 | .globl _sk_dstin_hsw |
| 3885 | _sk_dstin_hsw: |
| 3886 | .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| 3887 | .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| 3888 | .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| 3889 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 3890 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3891 | .byte 255,224 // jmpq *%rax |
| 3892 | |
| 3893 | HIDDEN _sk_srcout_hsw |
| 3894 | .globl _sk_srcout_hsw |
| 3895 | _sk_srcout_hsw: |
| 3896 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3897 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3898 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3899 | .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| 3900 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 3901 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 3902 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 3903 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 3904 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3905 | .byte 255,224 // jmpq *%rax |
| 3906 | |
| 3907 | HIDDEN _sk_dstout_hsw |
| 3908 | .globl _sk_dstout_hsw |
| 3909 | _sk_dstout_hsw: |
| 3910 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3911 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 3912 | .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| 3913 | .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 |
| 3914 | .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| 3915 | .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| 3916 | .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| 3917 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3918 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3919 | .byte 255,224 // jmpq *%rax |
| 3920 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3921 | HIDDEN _sk_srcover_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3922 | .globl _sk_srcover_hsw |
| 3923 | _sk_srcover_hsw: |
| 3924 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3925 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3926 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3927 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 3928 | .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0 |
| 3929 | .byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1 |
| 3930 | .byte 196,194,77,184,208 // vfmadd231ps %ymm8,%ymm6,%ymm2 |
| 3931 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 3932 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3933 | .byte 255,224 // jmpq *%rax |
| 3934 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 3935 | HIDDEN _sk_dstover_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 3936 | .globl _sk_dstover_hsw |
| 3937 | _sk_dstover_hsw: |
| 3938 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3939 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3940 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3941 | .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| 3942 | .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| 3943 | .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| 3944 | .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| 3945 | .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| 3946 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3947 | .byte 255,224 // jmpq *%rax |
| 3948 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 3949 | HIDDEN _sk_modulate_hsw |
| 3950 | .globl _sk_modulate_hsw |
| 3951 | _sk_modulate_hsw: |
| 3952 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 3953 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 3954 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 3955 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 3956 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3957 | .byte 255,224 // jmpq *%rax |
| 3958 | |
| 3959 | HIDDEN _sk_multiply_hsw |
| 3960 | .globl _sk_multiply_hsw |
| 3961 | _sk_multiply_hsw: |
| 3962 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 3963 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 3964 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 3965 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 3966 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 3967 | .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| 3968 | .byte 196,98,53,184,208 // vfmadd231ps %ymm0,%ymm9,%ymm10 |
| 3969 | .byte 196,194,93,168,194 // vfmadd213ps %ymm10,%ymm4,%ymm0 |
| 3970 | .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10 |
| 3971 | .byte 196,98,61,184,213 // vfmadd231ps %ymm5,%ymm8,%ymm10 |
| 3972 | .byte 196,194,85,168,202 // vfmadd213ps %ymm10,%ymm5,%ymm1 |
| 3973 | .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10 |
| 3974 | .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| 3975 | .byte 196,194,77,168,210 // vfmadd213ps %ymm10,%ymm6,%ymm2 |
| 3976 | .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| 3977 | .byte 196,66,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm8 |
| 3978 | .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3 |
| 3979 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3980 | .byte 255,224 // jmpq *%rax |
| 3981 | |
| 3982 | HIDDEN _sk_plus__hsw |
| 3983 | .globl _sk_plus__hsw |
| 3984 | _sk_plus__hsw: |
| 3985 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 3986 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 3987 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 3988 | .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| 3989 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 3990 | .byte 255,224 // jmpq *%rax |
| 3991 | |
| 3992 | HIDDEN _sk_screen_hsw |
| 3993 | .globl _sk_screen_hsw |
| 3994 | _sk_screen_hsw: |
| 3995 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 3996 | .byte 196,194,93,172,192 // vfnmadd213ps %ymm8,%ymm4,%ymm0 |
| 3997 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 3998 | .byte 196,194,85,172,200 // vfnmadd213ps %ymm8,%ymm5,%ymm1 |
| 3999 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 4000 | .byte 196,194,77,172,208 // vfnmadd213ps %ymm8,%ymm6,%ymm2 |
| 4001 | .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8 |
| 4002 | .byte 196,194,69,172,216 // vfnmadd213ps %ymm8,%ymm7,%ymm3 |
| 4003 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4004 | .byte 255,224 // jmpq *%rax |
| 4005 | |
| 4006 | HIDDEN _sk_xor__hsw |
| 4007 | .globl _sk_xor__hsw |
| 4008 | _sk_xor__hsw: |
| 4009 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4010 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4011 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4012 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 4013 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4014 | .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| 4015 | .byte 196,194,53,168,194 // vfmadd213ps %ymm10,%ymm9,%ymm0 |
| 4016 | .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| 4017 | .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1 |
| 4018 | .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| 4019 | .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2 |
| 4020 | .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| 4021 | .byte 196,98,69,168,195 // vfmadd213ps %ymm3,%ymm7,%ymm8 |
| 4022 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4023 | .byte 197,124,41,195 // vmovaps %ymm8,%ymm3 |
| 4024 | .byte 255,224 // jmpq *%rax |
| 4025 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 4026 | HIDDEN _sk_darken_hsw |
| 4027 | .globl _sk_darken_hsw |
| 4028 | _sk_darken_hsw: |
| 4029 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 4030 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 4031 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 4032 | .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0 |
| 4033 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 4034 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 4035 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 4036 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 4037 | .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1 |
| 4038 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 4039 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 4040 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 4041 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 4042 | .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 |
| 4043 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 4044 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4045 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4046 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4047 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4048 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4049 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4050 | .byte 255,224 // jmpq *%rax |
| 4051 | |
| 4052 | HIDDEN _sk_lighten_hsw |
| 4053 | .globl _sk_lighten_hsw |
| 4054 | _sk_lighten_hsw: |
| 4055 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 4056 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 4057 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 4058 | .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| 4059 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 4060 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 4061 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 4062 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 4063 | .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| 4064 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 4065 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 4066 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 4067 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 4068 | .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| 4069 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 4070 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4071 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4072 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4073 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4074 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4075 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4076 | .byte 255,224 // jmpq *%rax |
| 4077 | |
| 4078 | HIDDEN _sk_difference_hsw |
| 4079 | .globl _sk_difference_hsw |
| 4080 | _sk_difference_hsw: |
| 4081 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 4082 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 4083 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 4084 | .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| 4085 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 4086 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 4087 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 4088 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 4089 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 4090 | .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| 4091 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 4092 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 4093 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 4094 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 4095 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 4096 | .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| 4097 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 4098 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 4099 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4100 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4101 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4102 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4103 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4104 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4105 | .byte 255,224 // jmpq *%rax |
| 4106 | |
| 4107 | HIDDEN _sk_exclusion_hsw |
| 4108 | .globl _sk_exclusion_hsw |
| 4109 | _sk_exclusion_hsw: |
| 4110 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 4111 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 4112 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 4113 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 4114 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 4115 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 4116 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 4117 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 4118 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 4119 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 4120 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 4121 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 4122 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4123 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4124 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4125 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4126 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4127 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4128 | .byte 255,224 // jmpq *%rax |
| 4129 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 4130 | HIDDEN _sk_colorburn_hsw |
| 4131 | .globl _sk_colorburn_hsw |
| 4132 | _sk_colorburn_hsw: |
| 4133 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4134 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4135 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4136 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 4137 | .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 |
| 4138 | .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| 4139 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4140 | .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12 |
| 4141 | .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13 |
| 4142 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 4143 | .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13 |
| 4144 | .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| 4145 | .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| 4146 | .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13 |
| 4147 | .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| 4148 | .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12 |
| 4149 | .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0 |
| 4150 | .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| 4151 | .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12 |
| 4152 | .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| 4153 | .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| 4154 | .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11 |
| 4155 | .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| 4156 | .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13 |
| 4157 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 4158 | .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13 |
| 4159 | .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| 4160 | .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| 4161 | .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13 |
| 4162 | .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| 4163 | .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12 |
| 4164 | .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1 |
| 4165 | .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1 |
| 4166 | .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12 |
| 4167 | .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| 4168 | .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| 4169 | .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9 |
| 4170 | .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10 |
| 4171 | .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| 4172 | .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12 |
| 4173 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 4174 | .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12 |
| 4175 | .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| 4176 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 4177 | .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12 |
| 4178 | .byte 196,66,101,168,225 // vfmadd213ps %ymm9,%ymm3,%ymm12 |
| 4179 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 4180 | .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2 |
| 4181 | .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10 |
| 4182 | .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9 |
| 4183 | .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2 |
| 4184 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4185 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4186 | .byte 255,224 // jmpq *%rax |
| 4187 | |
| 4188 | HIDDEN _sk_colordodge_hsw |
| 4189 | .globl _sk_colordodge_hsw |
| 4190 | _sk_colordodge_hsw: |
| 4191 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 4192 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4193 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 4194 | .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| 4195 | .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 |
| 4196 | .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 |
| 4197 | .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 |
| 4198 | .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12 |
| 4199 | .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13 |
| 4200 | .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| 4201 | .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13 |
| 4202 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 4203 | .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12 |
| 4204 | .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| 4205 | .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13 |
| 4206 | .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0 |
| 4207 | .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0 |
| 4208 | .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12 |
| 4209 | .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| 4210 | .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| 4211 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 4212 | .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12 |
| 4213 | .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13 |
| 4214 | .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| 4215 | .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13 |
| 4216 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 4217 | .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12 |
| 4218 | .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| 4219 | .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13 |
| 4220 | .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1 |
| 4221 | .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| 4222 | .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12 |
| 4223 | .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| 4224 | .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| 4225 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 4226 | .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11 |
| 4227 | .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12 |
| 4228 | .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 |
| 4229 | .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12 |
| 4230 | .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11 |
| 4231 | .byte 196,66,101,168,218 // vfmadd213ps %ymm10,%ymm3,%ymm11 |
| 4232 | .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| 4233 | .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12 |
| 4234 | .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2 |
| 4235 | .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2 |
| 4236 | .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8 |
| 4237 | .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10 |
| 4238 | .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2 |
| 4239 | .byte 196,194,69,184,217 // vfmadd231ps %ymm9,%ymm7,%ymm3 |
| 4240 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4241 | .byte 255,224 // jmpq *%rax |
| 4242 | |
| 4243 | HIDDEN _sk_hardlight_hsw |
| 4244 | .globl _sk_hardlight_hsw |
| 4245 | _sk_hardlight_hsw: |
| 4246 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4247 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4248 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4249 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 4250 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4251 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 4252 | .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| 4253 | .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9 |
| 4254 | .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12 |
| 4255 | .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| 4256 | .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| 4257 | .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| 4258 | .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| 4259 | .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| 4260 | .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| 4261 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 4262 | .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| 4263 | .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| 4264 | .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| 4265 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 4266 | .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11 |
| 4267 | .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12 |
| 4268 | .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12 |
| 4269 | .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| 4270 | .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| 4271 | .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| 4272 | .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| 4273 | .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| 4274 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 4275 | .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| 4276 | .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| 4277 | .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| 4278 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 4279 | .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| 4280 | .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11 |
| 4281 | .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| 4282 | .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| 4283 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 4284 | .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| 4285 | .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| 4286 | .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| 4287 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 4288 | .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| 4289 | .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| 4290 | .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| 4291 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4292 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4293 | .byte 255,224 // jmpq *%rax |
| 4294 | |
| 4295 | HIDDEN _sk_overlay_hsw |
| 4296 | .globl _sk_overlay_hsw |
| 4297 | _sk_overlay_hsw: |
| 4298 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4299 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4300 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4301 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 4302 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4303 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 4304 | .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| 4305 | .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9 |
| 4306 | .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12 |
| 4307 | .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| 4308 | .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| 4309 | .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| 4310 | .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| 4311 | .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| 4312 | .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| 4313 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 4314 | .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| 4315 | .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| 4316 | .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| 4317 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 4318 | .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11 |
| 4319 | .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| 4320 | .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| 4321 | .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| 4322 | .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| 4323 | .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| 4324 | .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| 4325 | .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| 4326 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 4327 | .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| 4328 | .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| 4329 | .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| 4330 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 4331 | .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10 |
| 4332 | .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11 |
| 4333 | .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| 4334 | .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| 4335 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 4336 | .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| 4337 | .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| 4338 | .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| 4339 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 4340 | .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| 4341 | .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| 4342 | .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| 4343 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4344 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4345 | .byte 255,224 // jmpq *%rax |
| 4346 | |
| 4347 | HIDDEN _sk_softlight_hsw |
| 4348 | .globl _sk_softlight_hsw |
| 4349 | _sk_softlight_hsw: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4350 | .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 4351 | .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| 4352 | .byte 197,44,194,223,1 // vcmpltps %ymm7,%ymm10,%ymm11 |
| 4353 | .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8 |
| 4354 | .byte 196,67,45,74,224,176 // vblendvps %ymm11,%ymm8,%ymm10,%ymm12 |
| 4355 | .byte 196,65,28,88,196 // vaddps %ymm12,%ymm12,%ymm8 |
| 4356 | .byte 196,65,60,88,232 // vaddps %ymm8,%ymm8,%ymm13 |
| 4357 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4358 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4359 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4360 | .byte 196,66,21,168,237 // vfmadd213ps %ymm13,%ymm13,%ymm13 |
| 4361 | .byte 196,65,28,92,240 // vsubps %ymm8,%ymm12,%ymm14 |
| 4362 | .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| 4363 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 4364 | .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| 4365 | .byte 196,65,28,89,249 // vmulps %ymm9,%ymm12,%ymm15 |
| 4366 | .byte 196,66,21,184,254 // vfmadd231ps %ymm14,%ymm13,%ymm15 |
| 4367 | .byte 196,65,124,82,236 // vrsqrtps %ymm12,%ymm13 |
| 4368 | .byte 196,65,124,83,237 // vrcpps %ymm13,%ymm13 |
| 4369 | .byte 196,65,20,92,236 // vsubps %ymm12,%ymm13,%ymm13 |
| 4370 | .byte 197,92,88,244 // vaddps %ymm4,%ymm4,%ymm14 |
| 4371 | .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| 4372 | .byte 197,12,194,247,2 // vcmpleps %ymm7,%ymm14,%ymm14 |
| 4373 | .byte 196,67,21,74,239,224 // vblendvps %ymm14,%ymm15,%ymm13,%ymm13 |
| 4374 | .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14 |
| 4375 | .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| 4376 | .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12 |
| 4377 | .byte 196,98,5,168,227 // vfmadd213ps %ymm3,%ymm15,%ymm12 |
| 4378 | .byte 197,28,89,228 // vmulps %ymm4,%ymm12,%ymm12 |
| 4379 | .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| 4380 | .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| 4381 | .byte 196,98,101,184,236 // vfmadd231ps %ymm4,%ymm3,%ymm13 |
| 4382 | .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| 4383 | .byte 196,195,21,74,212,224 // vblendvps %ymm14,%ymm12,%ymm13,%ymm2 |
| 4384 | .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13 |
| 4385 | .byte 196,67,45,74,237,176 // vblendvps %ymm11,%ymm13,%ymm10,%ymm13 |
| 4386 | .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14 |
| 4387 | .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| 4388 | .byte 196,66,13,168,246 // vfmadd213ps %ymm14,%ymm14,%ymm14 |
| 4389 | .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15 |
| 4390 | .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14 |
| 4391 | .byte 196,66,53,184,245 // vfmadd231ps %ymm13,%ymm9,%ymm14 |
| 4392 | .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15 |
| 4393 | .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15 |
| 4394 | .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15 |
| 4395 | .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| 4396 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 4397 | .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| 4398 | .byte 196,67,5,74,230,192 // vblendvps %ymm12,%ymm14,%ymm15,%ymm12 |
| 4399 | .byte 197,116,88,241 // vaddps %ymm1,%ymm1,%ymm14 |
| 4400 | .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 |
| 4401 | .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| 4402 | .byte 196,98,5,168,235 // vfmadd213ps %ymm3,%ymm15,%ymm13 |
| 4403 | .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| 4404 | .byte 196,65,4,89,228 // vmulps %ymm12,%ymm15,%ymm12 |
| 4405 | .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13 |
| 4406 | .byte 196,98,101,184,229 // vfmadd231ps %ymm5,%ymm3,%ymm12 |
| 4407 | .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| 4408 | .byte 196,67,29,74,237,224 // vblendvps %ymm14,%ymm13,%ymm12,%ymm13 |
| 4409 | .byte 197,76,94,231 // vdivps %ymm7,%ymm6,%ymm12 |
| 4410 | .byte 196,67,45,74,212,176 // vblendvps %ymm11,%ymm12,%ymm10,%ymm10 |
| 4411 | .byte 196,65,44,88,218 // vaddps %ymm10,%ymm10,%ymm11 |
| 4412 | .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11 |
| 4413 | .byte 196,66,37,168,219 // vfmadd213ps %ymm11,%ymm11,%ymm11 |
| 4414 | .byte 196,65,44,92,224 // vsubps %ymm8,%ymm10,%ymm12 |
| 4415 | .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11 |
| 4416 | .byte 196,66,45,168,203 // vfmadd213ps %ymm11,%ymm10,%ymm9 |
| 4417 | .byte 196,65,124,82,218 // vrsqrtps %ymm10,%ymm11 |
| 4418 | .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11 |
| 4419 | .byte 196,65,36,92,218 // vsubps %ymm10,%ymm11,%ymm11 |
| 4420 | .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12 |
| 4421 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 4422 | .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| 4423 | .byte 196,67,37,74,201,192 // vblendvps %ymm12,%ymm9,%ymm11,%ymm9 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4424 | .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 4425 | .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11 |
| 4426 | .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12 |
| 4427 | .byte 196,65,60,92,210 // vsubps %ymm10,%ymm8,%ymm10 |
| 4428 | .byte 196,98,29,168,211 // vfmadd213ps %ymm3,%ymm12,%ymm10 |
| 4429 | .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12 |
| 4430 | .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9 |
| 4431 | .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10 |
| 4432 | .byte 196,98,101,184,206 // vfmadd231ps %ymm6,%ymm3,%ymm9 |
| 4433 | .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| 4434 | .byte 196,67,53,74,202,176 // vblendvps %ymm11,%ymm10,%ymm9,%ymm9 |
| 4435 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 4436 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 4437 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 4438 | .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11 |
| 4439 | .byte 196,193,108,88,195 // vaddps %ymm11,%ymm2,%ymm0 |
| 4440 | .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| 4441 | .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1 |
| 4442 | .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1 |
| 4443 | .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2 |
| 4444 | .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2 |
| 4445 | .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 |
| 4446 | .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3 |
| 4447 | .byte 72,173 // lods %ds:(%rsi),%rax |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 4448 | .byte 255,224 // jmpq *%rax |
| 4449 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4450 | HIDDEN _sk_clamp_0_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4451 | .globl _sk_clamp_0_hsw |
| 4452 | _sk_clamp_0_hsw: |
| 4453 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 4454 | .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| 4455 | .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| 4456 | .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| 4457 | .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 |
| 4458 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4459 | .byte 255,224 // jmpq *%rax |
| 4460 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4461 | HIDDEN _sk_clamp_1_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4462 | .globl _sk_clamp_1_hsw |
| 4463 | _sk_clamp_1_hsw: |
| 4464 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4465 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4466 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4467 | .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| 4468 | .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| 4469 | .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 |
| 4470 | .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| 4471 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4472 | .byte 255,224 // jmpq *%rax |
| 4473 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4474 | HIDDEN _sk_clamp_a_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4475 | .globl _sk_clamp_a_hsw |
| 4476 | _sk_clamp_a_hsw: |
| 4477 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4478 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4479 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4480 | .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| 4481 | .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 |
| 4482 | .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 |
| 4483 | .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2 |
| 4484 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4485 | .byte 255,224 // jmpq *%rax |
| 4486 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4487 | HIDDEN _sk_set_rgb_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4488 | .globl _sk_set_rgb_hsw |
| 4489 | _sk_set_rgb_hsw: |
| 4490 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4491 | .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| 4492 | .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| 4493 | .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| 4494 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4495 | .byte 255,224 // jmpq *%rax |
| 4496 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4497 | HIDDEN _sk_swap_rb_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4498 | .globl _sk_swap_rb_hsw |
| 4499 | _sk_swap_rb_hsw: |
| 4500 | .byte 197,124,40,192 // vmovaps %ymm0,%ymm8 |
| 4501 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4502 | .byte 197,252,40,194 // vmovaps %ymm2,%ymm0 |
| 4503 | .byte 197,124,41,194 // vmovaps %ymm8,%ymm2 |
| 4504 | .byte 255,224 // jmpq *%rax |
| 4505 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4506 | HIDDEN _sk_swap_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4507 | .globl _sk_swap_hsw |
| 4508 | _sk_swap_hsw: |
| 4509 | .byte 197,124,40,195 // vmovaps %ymm3,%ymm8 |
| 4510 | .byte 197,124,40,202 // vmovaps %ymm2,%ymm9 |
| 4511 | .byte 197,124,40,209 // vmovaps %ymm1,%ymm10 |
| 4512 | .byte 197,124,40,216 // vmovaps %ymm0,%ymm11 |
| 4513 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4514 | .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| 4515 | .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| 4516 | .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| 4517 | .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| 4518 | .byte 197,124,41,220 // vmovaps %ymm11,%ymm4 |
| 4519 | .byte 197,124,41,213 // vmovaps %ymm10,%ymm5 |
| 4520 | .byte 197,124,41,206 // vmovaps %ymm9,%ymm6 |
| 4521 | .byte 197,124,41,199 // vmovaps %ymm8,%ymm7 |
| 4522 | .byte 255,224 // jmpq *%rax |
| 4523 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4524 | HIDDEN _sk_move_src_dst_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4525 | .globl _sk_move_src_dst_hsw |
| 4526 | _sk_move_src_dst_hsw: |
| 4527 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4528 | .byte 197,252,40,224 // vmovaps %ymm0,%ymm4 |
| 4529 | .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| 4530 | .byte 197,252,40,242 // vmovaps %ymm2,%ymm6 |
| 4531 | .byte 197,252,40,251 // vmovaps %ymm3,%ymm7 |
| 4532 | .byte 255,224 // jmpq *%rax |
| 4533 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4534 | HIDDEN _sk_move_dst_src_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4535 | .globl _sk_move_dst_src_hsw |
| 4536 | _sk_move_dst_src_hsw: |
| 4537 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4538 | .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| 4539 | .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| 4540 | .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| 4541 | .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| 4542 | .byte 255,224 // jmpq *%rax |
| 4543 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4544 | HIDDEN _sk_premul_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4545 | .globl _sk_premul_hsw |
| 4546 | _sk_premul_hsw: |
| 4547 | .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 |
| 4548 | .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| 4549 | .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| 4550 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4551 | .byte 255,224 // jmpq *%rax |
| 4552 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4553 | HIDDEN _sk_unpremul_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4554 | .globl _sk_unpremul_hsw |
| 4555 | _sk_unpremul_hsw: |
| 4556 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 4557 | .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 |
| 4558 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4559 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 4560 | .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| 4561 | .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 |
| 4562 | .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 |
| 4563 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 4564 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 4565 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 4566 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4567 | .byte 255,224 // jmpq *%rax |
| 4568 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4569 | HIDDEN _sk_from_srgb_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4570 | .globl _sk_from_srgb_hsw |
| 4571 | _sk_from_srgb_hsw: |
| 4572 | .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| 4573 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4574 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4575 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 4576 | .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 |
| 4577 | .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| 4578 | .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| 4579 | .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11 |
| 4580 | .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| 4581 | .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| 4582 | .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12 |
| 4583 | .byte 196,65,125,111,235 // vmovdqa %ymm11,%ymm13 |
| 4584 | .byte 196,66,125,168,236 // vfmadd213ps %ymm12,%ymm0,%ymm13 |
| 4585 | .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| 4586 | .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| 4587 | .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14 |
| 4588 | .byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13 |
| 4589 | .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| 4590 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 4591 | .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| 4592 | .byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0 |
| 4593 | .byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0 |
| 4594 | .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9 |
| 4595 | .byte 197,116,89,233 // vmulps %ymm1,%ymm1,%ymm13 |
| 4596 | .byte 196,65,125,111,251 // vmovdqa %ymm11,%ymm15 |
| 4597 | .byte 196,66,117,168,252 // vfmadd213ps %ymm12,%ymm1,%ymm15 |
| 4598 | .byte 196,66,21,168,254 // vfmadd213ps %ymm14,%ymm13,%ymm15 |
| 4599 | .byte 196,193,116,194,202,1 // vcmpltps %ymm10,%ymm1,%ymm1 |
| 4600 | .byte 196,195,5,74,201,16 // vblendvps %ymm1,%ymm9,%ymm15,%ymm1 |
| 4601 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 4602 | .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9 |
| 4603 | .byte 196,66,109,168,220 // vfmadd213ps %ymm12,%ymm2,%ymm11 |
| 4604 | .byte 196,66,53,168,222 // vfmadd213ps %ymm14,%ymm9,%ymm11 |
| 4605 | .byte 196,193,108,194,210,1 // vcmpltps %ymm10,%ymm2,%ymm2 |
| 4606 | .byte 196,195,37,74,208,32 // vblendvps %ymm2,%ymm8,%ymm11,%ymm2 |
| 4607 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4608 | .byte 255,224 // jmpq *%rax |
| 4609 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4610 | HIDDEN _sk_to_srgb_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4611 | .globl _sk_to_srgb_hsw |
| 4612 | _sk_to_srgb_hsw: |
| 4613 | .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| 4614 | .byte 196,65,124,83,216 // vrcpps %ymm8,%ymm11 |
| 4615 | .byte 196,65,124,82,224 // vrsqrtps %ymm8,%ymm12 |
| 4616 | .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| 4617 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4618 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4619 | .byte 197,60,89,232 // vmulps %ymm0,%ymm8,%ymm13 |
| 4620 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4621 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 4622 | .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| 4623 | .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| 4624 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 4625 | .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| 4626 | .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| 4627 | .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| 4628 | .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14 |
| 4629 | .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| 4630 | .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| 4631 | .byte 197,121,110,248 // vmovd %eax,%xmm15 |
| 4632 | .byte 196,66,125,88,255 // vpbroadcastd %xmm15,%ymm15 |
| 4633 | .byte 196,66,13,168,223 // vfmadd213ps %ymm15,%ymm14,%ymm11 |
| 4634 | .byte 196,66,45,184,220 // vfmadd231ps %ymm12,%ymm10,%ymm11 |
| 4635 | .byte 196,65,52,93,219 // vminps %ymm11,%ymm9,%ymm11 |
| 4636 | .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| 4637 | .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| 4638 | .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12 |
| 4639 | .byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0 |
| 4640 | .byte 196,195,37,74,197,0 // vblendvps %ymm0,%ymm13,%ymm11,%ymm0 |
| 4641 | .byte 197,124,82,217 // vrsqrtps %ymm1,%ymm11 |
| 4642 | .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13 |
| 4643 | .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11 |
| 4644 | .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13 |
| 4645 | .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13 |
| 4646 | .byte 197,60,89,217 // vmulps %ymm1,%ymm8,%ymm11 |
| 4647 | .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13 |
| 4648 | .byte 196,193,116,194,204,1 // vcmpltps %ymm12,%ymm1,%ymm1 |
| 4649 | .byte 196,195,21,74,203,16 // vblendvps %ymm1,%ymm11,%ymm13,%ymm1 |
| 4650 | .byte 197,124,82,218 // vrsqrtps %ymm2,%ymm11 |
| 4651 | .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13 |
| 4652 | .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13 |
| 4653 | .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11 |
| 4654 | .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13 |
| 4655 | .byte 196,65,52,93,205 // vminps %ymm13,%ymm9,%ymm9 |
| 4656 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 4657 | .byte 196,193,108,194,212,1 // vcmpltps %ymm12,%ymm2,%ymm2 |
| 4658 | .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| 4659 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4660 | .byte 255,224 // jmpq *%rax |
| 4661 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4662 | HIDDEN _sk_scale_1_float_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4663 | .globl _sk_scale_1_float_hsw |
| 4664 | _sk_scale_1_float_hsw: |
| 4665 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4666 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 4667 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 4668 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 4669 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 4670 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 4671 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4672 | .byte 255,224 // jmpq *%rax |
| 4673 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4674 | HIDDEN _sk_scale_u8_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4675 | .globl _sk_scale_u8_hsw |
| 4676 | _sk_scale_u8_hsw: |
| 4677 | .byte 73,137,200 // mov %rcx,%r8 |
| 4678 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4679 | .byte 72,139,0 // mov (%rax),%rax |
| 4680 | .byte 72,1,248 // add %rdi,%rax |
| 4681 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4682 | .byte 117,56 // jne d47 <_sk_scale_u8_hsw+0x48> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4683 | .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| 4684 | .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 |
| 4685 | .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| 4686 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 4687 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 4688 | .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| 4689 | .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| 4690 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 4691 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 4692 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 4693 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 4694 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4695 | .byte 76,137,193 // mov %r8,%rcx |
| 4696 | .byte 255,224 // jmpq *%rax |
| 4697 | .byte 49,201 // xor %ecx,%ecx |
| 4698 | .byte 77,137,194 // mov %r8,%r10 |
| 4699 | .byte 69,49,201 // xor %r9d,%r9d |
| 4700 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 4701 | .byte 72,255,192 // inc %rax |
| 4702 | .byte 73,211,227 // shl %cl,%r11 |
| 4703 | .byte 77,9,217 // or %r11,%r9 |
| 4704 | .byte 72,131,193,8 // add $0x8,%rcx |
| 4705 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4706 | .byte 117,234 // jne d4f <_sk_scale_u8_hsw+0x50> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4707 | .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4708 | .byte 235,167 // jmp d13 <_sk_scale_u8_hsw+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4709 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4710 | HIDDEN _sk_lerp_1_float_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4711 | .globl _sk_lerp_1_float_hsw |
| 4712 | _sk_lerp_1_float_hsw: |
| 4713 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4714 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 4715 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 4716 | .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| 4717 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 4718 | .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| 4719 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 4720 | .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| 4721 | .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| 4722 | .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| 4723 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4724 | .byte 255,224 // jmpq *%rax |
| 4725 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4726 | HIDDEN _sk_lerp_u8_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4727 | .globl _sk_lerp_u8_hsw |
| 4728 | _sk_lerp_u8_hsw: |
| 4729 | .byte 73,137,200 // mov %rcx,%r8 |
| 4730 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4731 | .byte 72,139,0 // mov (%rax),%rax |
| 4732 | .byte 72,1,248 // add %rdi,%rax |
| 4733 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4734 | .byte 117,76 // jne df7 <_sk_lerp_u8_hsw+0x5c> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4735 | .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| 4736 | .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8 |
| 4737 | .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| 4738 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 4739 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 4740 | .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9 |
| 4741 | .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| 4742 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 4743 | .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0 |
| 4744 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 4745 | .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1 |
| 4746 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 4747 | .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2 |
| 4748 | .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| 4749 | .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3 |
| 4750 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4751 | .byte 76,137,193 // mov %r8,%rcx |
| 4752 | .byte 255,224 // jmpq *%rax |
| 4753 | .byte 49,201 // xor %ecx,%ecx |
| 4754 | .byte 77,137,194 // mov %r8,%r10 |
| 4755 | .byte 69,49,201 // xor %r9d,%r9d |
| 4756 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 4757 | .byte 72,255,192 // inc %rax |
| 4758 | .byte 73,211,227 // shl %cl,%r11 |
| 4759 | .byte 77,9,217 // or %r11,%r9 |
| 4760 | .byte 72,131,193,8 // add $0x8,%rcx |
| 4761 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4762 | .byte 117,234 // jne dff <_sk_lerp_u8_hsw+0x64> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4763 | .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4764 | .byte 235,147 // jmp daf <_sk_lerp_u8_hsw+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4765 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4766 | HIDDEN _sk_lerp_565_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4767 | .globl _sk_lerp_565_hsw |
| 4768 | _sk_lerp_565_hsw: |
| 4769 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4770 | .byte 76,139,16 // mov (%rax),%r10 |
| 4771 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4772 | .byte 15,133,179,0,0,0 // jne edd <_sk_lerp_565_hsw+0xc1> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4773 | .byte 196,193,122,111,28,122 // vmovdqu (%r10,%rdi,2),%xmm3 |
| 4774 | .byte 196,98,125,51,195 // vpmovzxwd %xmm3,%ymm8 |
| 4775 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 4776 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4777 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4778 | .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| 4779 | .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9 |
| 4780 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 4781 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4782 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4783 | .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| 4784 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 4785 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4786 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4787 | .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| 4788 | .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10 |
| 4789 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 4790 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4791 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4792 | .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10 |
| 4793 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 4794 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4795 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4796 | .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3 |
| 4797 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 4798 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 4799 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4800 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4801 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 4802 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 4803 | .byte 196,226,53,168,196 // vfmadd213ps %ymm4,%ymm9,%ymm0 |
| 4804 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 4805 | .byte 196,226,45,168,205 // vfmadd213ps %ymm5,%ymm10,%ymm1 |
| 4806 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 4807 | .byte 196,226,101,168,214 // vfmadd213ps %ymm6,%ymm3,%ymm2 |
| 4808 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 4809 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4810 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4811 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4812 | .byte 255,224 // jmpq *%rax |
| 4813 | .byte 65,137,200 // mov %ecx,%r8d |
| 4814 | .byte 65,128,224,7 // and $0x7,%r8b |
| 4815 | .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| 4816 | .byte 65,254,200 // dec %r8b |
| 4817 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4818 | .byte 15,135,59,255,255,255 // ja e30 <_sk_lerp_565_hsw+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4819 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4820 | .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # f4c <_sk_lerp_565_hsw+0x130> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4821 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 4822 | .byte 76,1,200 // add %r9,%rax |
| 4823 | .byte 255,224 // jmpq *%rax |
| 4824 | .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| 4825 | .byte 196,193,97,196,92,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm3 |
| 4826 | .byte 196,193,97,196,92,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm3,%xmm3 |
| 4827 | .byte 196,193,97,196,92,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm3,%xmm3 |
| 4828 | .byte 196,193,97,196,92,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm3,%xmm3 |
| 4829 | .byte 196,193,97,196,92,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3 |
| 4830 | .byte 196,193,97,196,92,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3 |
| 4831 | .byte 196,193,97,196,28,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4832 | .byte 233,231,254,255,255 // jmpq e30 <_sk_lerp_565_hsw+0x14> |
| 4833 | .byte 15,31,0 // nopl (%rax) |
| 4834 | .byte 241 // icebp |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4835 | .byte 255 // (bad) |
| 4836 | .byte 255 // (bad) |
| 4837 | .byte 255 // (bad) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 4838 | .byte 233,255,255,255,225 // jmpq ffffffffe2000f54 <_sk_linear_gradient_2stops_hsw+0xffffffffe1ffefbb> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4839 | .byte 255 // (bad) |
| 4840 | .byte 255 // (bad) |
| 4841 | .byte 255 // (bad) |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4842 | .byte 217,255 // fcos |
| 4843 | .byte 255 // (bad) |
| 4844 | .byte 255,209 // callq *%rcx |
| 4845 | .byte 255 // (bad) |
| 4846 | .byte 255 // (bad) |
| 4847 | .byte 255,201 // dec %ecx |
| 4848 | .byte 255 // (bad) |
| 4849 | .byte 255 // (bad) |
| 4850 | .byte 255 // (bad) |
| 4851 | .byte 189 // .byte 0xbd |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4852 | .byte 255 // (bad) |
| 4853 | .byte 255 // (bad) |
| 4854 | .byte 255 // .byte 0xff |
| 4855 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4856 | HIDDEN _sk_load_tables_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4857 | .globl _sk_load_tables_hsw |
| 4858 | _sk_load_tables_hsw: |
| 4859 | .byte 73,137,200 // mov %rcx,%r8 |
| 4860 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4861 | .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| 4862 | .byte 76,3,8 // add (%rax),%r9 |
| 4863 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4864 | .byte 117,121 // jne ff6 <_sk_load_tables_hsw+0x8e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4865 | .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 |
| 4866 | .byte 185,255,0,0,0 // mov $0xff,%ecx |
| 4867 | .byte 197,249,110,193 // vmovd %ecx,%xmm0 |
| 4868 | .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| 4869 | .byte 197,237,219,203 // vpand %ymm3,%ymm2,%ymm1 |
| 4870 | .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| 4871 | .byte 72,139,72,8 // mov 0x8(%rax),%rcx |
| 4872 | .byte 76,139,72,16 // mov 0x10(%rax),%r9 |
| 4873 | .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| 4874 | .byte 196,226,53,146,4,137 // vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0 |
| 4875 | .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| 4876 | .byte 197,109,219,201 // vpand %ymm1,%ymm2,%ymm9 |
| 4877 | .byte 196,65,45,118,210 // vpcmpeqd %ymm10,%ymm10,%ymm10 |
| 4878 | .byte 196,130,45,146,12,137 // vgatherdps %ymm10,(%r9,%ymm9,4),%ymm1 |
| 4879 | .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| 4880 | .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| 4881 | .byte 196,65,109,219,201 // vpand %ymm9,%ymm2,%ymm9 |
| 4882 | .byte 196,162,61,146,20,136 // vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2 |
| 4883 | .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| 4884 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 4885 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 4886 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 4887 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 4888 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 4889 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4890 | .byte 76,137,193 // mov %r8,%rcx |
| 4891 | .byte 255,224 // jmpq *%rax |
| 4892 | .byte 185,8,0,0,0 // mov $0x8,%ecx |
| 4893 | .byte 68,41,193 // sub %r8d,%ecx |
| 4894 | .byte 192,225,3 // shl $0x3,%cl |
| 4895 | .byte 73,199,194,255,255,255,255 // mov $0xffffffffffffffff,%r10 |
| 4896 | .byte 73,211,234 // shr %cl,%r10 |
| 4897 | .byte 196,193,249,110,194 // vmovq %r10,%xmm0 |
| 4898 | .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 |
| 4899 | .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4900 | .byte 233,99,255,255,255 // jmpq f82 <_sk_load_tables_hsw+0x1a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4901 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4902 | HIDDEN _sk_load_a8_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4903 | .globl _sk_load_a8_hsw |
| 4904 | _sk_load_a8_hsw: |
| 4905 | .byte 73,137,200 // mov %rcx,%r8 |
| 4906 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4907 | .byte 72,139,0 // mov (%rax),%rax |
| 4908 | .byte 72,1,248 // add %rdi,%rax |
| 4909 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4910 | .byte 117,50 // jne 1061 <_sk_load_a8_hsw+0x42> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4911 | .byte 197,250,126,0 // vmovq (%rax),%xmm0 |
| 4912 | .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 |
| 4913 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 4914 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 4915 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 4916 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 4917 | .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| 4918 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4919 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 4920 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 4921 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 4922 | .byte 76,137,193 // mov %r8,%rcx |
| 4923 | .byte 255,224 // jmpq *%rax |
| 4924 | .byte 49,201 // xor %ecx,%ecx |
| 4925 | .byte 77,137,194 // mov %r8,%r10 |
| 4926 | .byte 69,49,201 // xor %r9d,%r9d |
| 4927 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 4928 | .byte 72,255,192 // inc %rax |
| 4929 | .byte 73,211,227 // shl %cl,%r11 |
| 4930 | .byte 77,9,217 // or %r11,%r9 |
| 4931 | .byte 72,131,193,8 // add $0x8,%rcx |
| 4932 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4933 | .byte 117,234 // jne 1069 <_sk_load_a8_hsw+0x4a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4934 | .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4935 | .byte 235,173 // jmp 1033 <_sk_load_a8_hsw+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4936 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 4937 | HIDDEN _sk_store_a8_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4938 | .globl _sk_store_a8_hsw |
| 4939 | _sk_store_a8_hsw: |
| 4940 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4941 | .byte 76,139,8 // mov (%rax),%r9 |
| 4942 | .byte 184,0,0,127,67 // mov $0x437f0000,%eax |
| 4943 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 4944 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 4945 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 4946 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 4947 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 4948 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 4949 | .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 |
| 4950 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4951 | .byte 117,10 // jne 10c1 <_sk_store_a8_hsw+0x3b> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4952 | .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) |
| 4953 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 4954 | .byte 255,224 // jmpq *%rax |
| 4955 | .byte 65,137,200 // mov %ecx,%r8d |
| 4956 | .byte 65,128,224,7 // and $0x7,%r8b |
| 4957 | .byte 65,254,200 // dec %r8b |
| 4958 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4959 | .byte 119,236 // ja 10bd <_sk_store_a8_hsw+0x37> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4960 | .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 |
| 4961 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4962 | .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 1124 <_sk_store_a8_hsw+0x9e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4963 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 4964 | .byte 76,1,192 // add %r8,%rax |
| 4965 | .byte 255,224 // jmpq *%rax |
| 4966 | .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1) |
| 4967 | .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1) |
| 4968 | .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1) |
| 4969 | .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1) |
| 4970 | .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) |
| 4971 | .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) |
| 4972 | .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 4973 | .byte 235,154 // jmp 10bd <_sk_store_a8_hsw+0x37> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 4974 | .byte 144 // nop |
| 4975 | .byte 246,255 // idiv %bh |
| 4976 | .byte 255 // (bad) |
| 4977 | .byte 255 // (bad) |
| 4978 | .byte 238 // out %al,(%dx) |
| 4979 | .byte 255 // (bad) |
| 4980 | .byte 255 // (bad) |
| 4981 | .byte 255,230 // jmpq *%rsi |
| 4982 | .byte 255 // (bad) |
| 4983 | .byte 255 // (bad) |
| 4984 | .byte 255 // (bad) |
| 4985 | .byte 222,255 // fdivrp %st,%st(7) |
| 4986 | .byte 255 // (bad) |
| 4987 | .byte 255,214 // callq *%rsi |
| 4988 | .byte 255 // (bad) |
| 4989 | .byte 255 // (bad) |
| 4990 | .byte 255,206 // dec %esi |
| 4991 | .byte 255 // (bad) |
| 4992 | .byte 255 // (bad) |
| 4993 | .byte 255,198 // inc %esi |
| 4994 | .byte 255 // (bad) |
| 4995 | .byte 255 // (bad) |
| 4996 | .byte 255 // .byte 0xff |
| 4997 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 4998 | HIDDEN _sk_load_g8_hsw |
| 4999 | .globl _sk_load_g8_hsw |
| 5000 | _sk_load_g8_hsw: |
| 5001 | .byte 73,137,200 // mov %rcx,%r8 |
| 5002 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5003 | .byte 72,139,0 // mov (%rax),%rax |
| 5004 | .byte 72,1,248 // add %rdi,%rax |
| 5005 | .byte 77,133,192 // test %r8,%r8 |
| 5006 | .byte 117,60 // jne 118c <_sk_load_g8_hsw+0x4c> |
| 5007 | .byte 197,250,126,0 // vmovq (%rax),%xmm0 |
| 5008 | .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0 |
| 5009 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 5010 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 5011 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5012 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 5013 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 5014 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 5015 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5016 | .byte 196,226,125,88,217 // vpbroadcastd %xmm1,%ymm3 |
| 5017 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5018 | .byte 76,137,193 // mov %r8,%rcx |
| 5019 | .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 |
| 5020 | .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 |
| 5021 | .byte 255,224 // jmpq *%rax |
| 5022 | .byte 49,201 // xor %ecx,%ecx |
| 5023 | .byte 77,137,194 // mov %r8,%r10 |
| 5024 | .byte 69,49,201 // xor %r9d,%r9d |
| 5025 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 5026 | .byte 72,255,192 // inc %rax |
| 5027 | .byte 73,211,227 // shl %cl,%r11 |
| 5028 | .byte 77,9,217 // or %r11,%r9 |
| 5029 | .byte 72,131,193,8 // add $0x8,%rcx |
| 5030 | .byte 73,255,202 // dec %r10 |
| 5031 | .byte 117,234 // jne 1194 <_sk_load_g8_hsw+0x54> |
| 5032 | .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| 5033 | .byte 235,163 // jmp 1154 <_sk_load_g8_hsw+0x14> |
| 5034 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5035 | HIDDEN _sk_load_565_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5036 | .globl _sk_load_565_hsw |
| 5037 | _sk_load_565_hsw: |
| 5038 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5039 | .byte 76,139,16 // mov (%rax),%r10 |
| 5040 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5041 | .byte 15,133,149,0,0,0 // jne 1254 <_sk_load_565_hsw+0xa3> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5042 | .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 |
| 5043 | .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2 |
| 5044 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 5045 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 5046 | .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| 5047 | .byte 197,253,219,194 // vpand %ymm2,%ymm0,%ymm0 |
| 5048 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 5049 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 5050 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5051 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 5052 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 5053 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 5054 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5055 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 5056 | .byte 197,245,219,202 // vpand %ymm2,%ymm1,%ymm1 |
| 5057 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 5058 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 5059 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5060 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5061 | .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| 5062 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 5063 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5064 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5065 | .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2 |
| 5066 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 5067 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 5068 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5069 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5070 | .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| 5071 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 5072 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5073 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5074 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5075 | .byte 255,224 // jmpq *%rax |
| 5076 | .byte 65,137,200 // mov %ecx,%r8d |
| 5077 | .byte 65,128,224,7 // and $0x7,%r8b |
| 5078 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 5079 | .byte 65,254,200 // dec %r8b |
| 5080 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5081 | .byte 15,135,89,255,255,255 // ja 11c5 <_sk_load_565_hsw+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5082 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5083 | .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 12c0 <_sk_load_565_hsw+0x10f> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5084 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 5085 | .byte 76,1,200 // add %r9,%rax |
| 5086 | .byte 255,224 // jmpq *%rax |
| 5087 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 5088 | .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 |
| 5089 | .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 |
| 5090 | .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 |
| 5091 | .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 |
| 5092 | .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 |
| 5093 | .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 |
| 5094 | .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5095 | .byte 233,5,255,255,255 // jmpq 11c5 <_sk_load_565_hsw+0x14> |
| 5096 | .byte 244 // hlt |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5097 | .byte 255 // (bad) |
| 5098 | .byte 255 // (bad) |
| 5099 | .byte 255 // (bad) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5100 | .byte 236 // in (%dx),%al |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5101 | .byte 255 // (bad) |
| 5102 | .byte 255 // (bad) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5103 | .byte 255,228 // jmpq *%rsp |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5104 | .byte 255 // (bad) |
| 5105 | .byte 255 // (bad) |
| 5106 | .byte 255 // (bad) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5107 | .byte 220,255 // fdivr %st,%st(7) |
| 5108 | .byte 255 // (bad) |
| 5109 | .byte 255,212 // callq *%rsp |
| 5110 | .byte 255 // (bad) |
| 5111 | .byte 255 // (bad) |
| 5112 | .byte 255,204 // dec %esp |
| 5113 | .byte 255 // (bad) |
| 5114 | .byte 255 // (bad) |
| 5115 | .byte 255,192 // inc %eax |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5116 | .byte 255 // (bad) |
| 5117 | .byte 255 // (bad) |
| 5118 | .byte 255 // .byte 0xff |
| 5119 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5120 | HIDDEN _sk_store_565_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5121 | .globl _sk_store_565_hsw |
| 5122 | _sk_store_565_hsw: |
| 5123 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5124 | .byte 76,139,8 // mov (%rax),%r9 |
| 5125 | .byte 184,0,0,248,65 // mov $0x41f80000,%eax |
| 5126 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 5127 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 5128 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 5129 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 5130 | .byte 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9 |
| 5131 | .byte 184,0,0,124,66 // mov $0x427c0000,%eax |
| 5132 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 5133 | .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10 |
| 5134 | .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| 5135 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5136 | .byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10 |
| 5137 | .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9 |
| 5138 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 5139 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 5140 | .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 |
| 5141 | .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 |
| 5142 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 5143 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5144 | .byte 117,10 // jne 1348 <_sk_store_565_hsw+0x6c> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5145 | .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) |
| 5146 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5147 | .byte 255,224 // jmpq *%rax |
| 5148 | .byte 65,137,200 // mov %ecx,%r8d |
| 5149 | .byte 65,128,224,7 // and $0x7,%r8b |
| 5150 | .byte 65,254,200 // dec %r8b |
| 5151 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5152 | .byte 119,236 // ja 1344 <_sk_store_565_hsw+0x68> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5153 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5154 | .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 13a8 <_sk_store_565_hsw+0xcc> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5155 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 5156 | .byte 76,1,192 // add %r8,%rax |
| 5157 | .byte 255,224 // jmpq *%rax |
| 5158 | .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2) |
| 5159 | .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2) |
| 5160 | .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2) |
| 5161 | .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2) |
| 5162 | .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) |
| 5163 | .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) |
| 5164 | .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5165 | .byte 235,159 // jmp 1344 <_sk_store_565_hsw+0x68> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5166 | .byte 15,31,0 // nopl (%rax) |
| 5167 | .byte 244 // hlt |
| 5168 | .byte 255 // (bad) |
| 5169 | .byte 255 // (bad) |
| 5170 | .byte 255 // (bad) |
| 5171 | .byte 236 // in (%dx),%al |
| 5172 | .byte 255 // (bad) |
| 5173 | .byte 255 // (bad) |
| 5174 | .byte 255,228 // jmpq *%rsp |
| 5175 | .byte 255 // (bad) |
| 5176 | .byte 255 // (bad) |
| 5177 | .byte 255 // (bad) |
| 5178 | .byte 220,255 // fdivr %st,%st(7) |
| 5179 | .byte 255 // (bad) |
| 5180 | .byte 255,212 // callq *%rsp |
| 5181 | .byte 255 // (bad) |
| 5182 | .byte 255 // (bad) |
| 5183 | .byte 255,204 // dec %esp |
| 5184 | .byte 255 // (bad) |
| 5185 | .byte 255 // (bad) |
| 5186 | .byte 255,196 // inc %esp |
| 5187 | .byte 255 // (bad) |
| 5188 | .byte 255 // (bad) |
| 5189 | .byte 255 // .byte 0xff |
| 5190 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5191 | HIDDEN _sk_load_4444_hsw |
| 5192 | .globl _sk_load_4444_hsw |
| 5193 | _sk_load_4444_hsw: |
| 5194 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5195 | .byte 76,139,16 // mov (%rax),%r10 |
| 5196 | .byte 72,133,201 // test %rcx,%rcx |
| 5197 | .byte 15,133,179,0,0,0 // jne 1485 <_sk_load_4444_hsw+0xc1> |
| 5198 | .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 |
| 5199 | .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9 |
| 5200 | .byte 184,0,240,0,0 // mov $0xf000,%eax |
| 5201 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 5202 | .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0 |
| 5203 | .byte 196,193,125,219,193 // vpand %ymm9,%ymm0,%ymm0 |
| 5204 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 5205 | .byte 184,137,136,136,55 // mov $0x37888889,%eax |
| 5206 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5207 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 5208 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 5209 | .byte 184,0,15,0,0 // mov $0xf00,%eax |
| 5210 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5211 | .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1 |
| 5212 | .byte 196,193,117,219,201 // vpand %ymm9,%ymm1,%ymm1 |
| 5213 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 5214 | .byte 184,137,136,136,57 // mov $0x39888889,%eax |
| 5215 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 5216 | .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| 5217 | .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 |
| 5218 | .byte 184,240,0,0,0 // mov $0xf0,%eax |
| 5219 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 5220 | .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| 5221 | .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2 |
| 5222 | .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8 |
| 5223 | .byte 184,137,136,136,59 // mov $0x3b888889,%eax |
| 5224 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 5225 | .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2 |
| 5226 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 5227 | .byte 184,15,0,0,0 // mov $0xf,%eax |
| 5228 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5229 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5230 | .byte 196,193,101,219,217 // vpand %ymm9,%ymm3,%ymm3 |
| 5231 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 5232 | .byte 184,137,136,136,61 // mov $0x3d888889,%eax |
| 5233 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5234 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5235 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 5236 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5237 | .byte 255,224 // jmpq *%rax |
| 5238 | .byte 65,137,200 // mov %ecx,%r8d |
| 5239 | .byte 65,128,224,7 // and $0x7,%r8b |
| 5240 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 5241 | .byte 65,254,200 // dec %r8b |
| 5242 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| 5243 | .byte 15,135,59,255,255,255 // ja 13d8 <_sk_load_4444_hsw+0x14> |
| 5244 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| 5245 | .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 14f4 <_sk_load_4444_hsw+0x130> |
| 5246 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 5247 | .byte 76,1,200 // add %r9,%rax |
| 5248 | .byte 255,224 // jmpq *%rax |
| 5249 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 5250 | .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 |
| 5251 | .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 |
| 5252 | .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 |
| 5253 | .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 |
| 5254 | .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 |
| 5255 | .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 |
| 5256 | .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 |
| 5257 | .byte 233,231,254,255,255 // jmpq 13d8 <_sk_load_4444_hsw+0x14> |
| 5258 | .byte 15,31,0 // nopl (%rax) |
| 5259 | .byte 241 // icebp |
| 5260 | .byte 255 // (bad) |
| 5261 | .byte 255 // (bad) |
| 5262 | .byte 255 // (bad) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5263 | .byte 233,255,255,255,225 // jmpq ffffffffe20014fc <_sk_linear_gradient_2stops_hsw+0xffffffffe1fff563> |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5264 | .byte 255 // (bad) |
| 5265 | .byte 255 // (bad) |
| 5266 | .byte 255 // (bad) |
| 5267 | .byte 217,255 // fcos |
| 5268 | .byte 255 // (bad) |
| 5269 | .byte 255,209 // callq *%rcx |
| 5270 | .byte 255 // (bad) |
| 5271 | .byte 255 // (bad) |
| 5272 | .byte 255,201 // dec %ecx |
| 5273 | .byte 255 // (bad) |
| 5274 | .byte 255 // (bad) |
| 5275 | .byte 255 // (bad) |
| 5276 | .byte 189 // .byte 0xbd |
| 5277 | .byte 255 // (bad) |
| 5278 | .byte 255 // (bad) |
| 5279 | .byte 255 // .byte 0xff |
| 5280 | |
| 5281 | HIDDEN _sk_store_4444_hsw |
| 5282 | .globl _sk_store_4444_hsw |
| 5283 | _sk_store_4444_hsw: |
| 5284 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5285 | .byte 76,139,8 // mov (%rax),%r9 |
| 5286 | .byte 184,0,0,112,65 // mov $0x41700000,%eax |
| 5287 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 5288 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 5289 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 5290 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 5291 | .byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9 |
| 5292 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 5293 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5294 | .byte 196,193,45,114,242,8 // vpslld $0x8,%ymm10,%ymm10 |
| 5295 | .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9 |
| 5296 | .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10 |
| 5297 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5298 | .byte 196,193,45,114,242,4 // vpslld $0x4,%ymm10,%ymm10 |
| 5299 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 5300 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 5301 | .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8 |
| 5302 | .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 |
| 5303 | .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9 |
| 5304 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 5305 | .byte 72,133,201 // test %rcx,%rcx |
| 5306 | .byte 117,10 // jne 1582 <_sk_store_4444_hsw+0x72> |
| 5307 | .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) |
| 5308 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5309 | .byte 255,224 // jmpq *%rax |
| 5310 | .byte 65,137,200 // mov %ecx,%r8d |
| 5311 | .byte 65,128,224,7 // and $0x7,%r8b |
| 5312 | .byte 65,254,200 // dec %r8b |
| 5313 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| 5314 | .byte 119,236 // ja 157e <_sk_store_4444_hsw+0x6e> |
| 5315 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| 5316 | .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 15e0 <_sk_store_4444_hsw+0xd0> |
| 5317 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 5318 | .byte 76,1,192 // add %r8,%rax |
| 5319 | .byte 255,224 // jmpq *%rax |
| 5320 | .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2) |
| 5321 | .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2) |
| 5322 | .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2) |
| 5323 | .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2) |
| 5324 | .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) |
| 5325 | .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) |
| 5326 | .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) |
| 5327 | .byte 235,159 // jmp 157e <_sk_store_4444_hsw+0x6e> |
| 5328 | .byte 144 // nop |
| 5329 | .byte 246,255 // idiv %bh |
| 5330 | .byte 255 // (bad) |
| 5331 | .byte 255 // (bad) |
| 5332 | .byte 238 // out %al,(%dx) |
| 5333 | .byte 255 // (bad) |
| 5334 | .byte 255 // (bad) |
| 5335 | .byte 255,230 // jmpq *%rsi |
| 5336 | .byte 255 // (bad) |
| 5337 | .byte 255 // (bad) |
| 5338 | .byte 255 // (bad) |
| 5339 | .byte 222,255 // fdivrp %st,%st(7) |
| 5340 | .byte 255 // (bad) |
| 5341 | .byte 255,214 // callq *%rsi |
| 5342 | .byte 255 // (bad) |
| 5343 | .byte 255 // (bad) |
| 5344 | .byte 255,206 // dec %esi |
| 5345 | .byte 255 // (bad) |
| 5346 | .byte 255 // (bad) |
| 5347 | .byte 255,198 // inc %esi |
| 5348 | .byte 255 // (bad) |
| 5349 | .byte 255 // (bad) |
| 5350 | .byte 255 // .byte 0xff |
| 5351 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5352 | HIDDEN _sk_load_8888_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5353 | .globl _sk_load_8888_hsw |
| 5354 | _sk_load_8888_hsw: |
| 5355 | .byte 73,137,200 // mov %rcx,%r8 |
| 5356 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5357 | .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| 5358 | .byte 76,3,8 // add (%rax),%r9 |
| 5359 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5360 | .byte 117,104 // jne 1679 <_sk_load_8888_hsw+0x7d> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5361 | .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3 |
| 5362 | .byte 184,255,0,0,0 // mov $0xff,%eax |
| 5363 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 5364 | .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2 |
| 5365 | .byte 197,237,219,195 // vpand %ymm3,%ymm2,%ymm0 |
| 5366 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 5367 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 5368 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 5369 | .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8 |
| 5370 | .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| 5371 | .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1 |
| 5372 | .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1 |
| 5373 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 5374 | .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| 5375 | .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9 |
| 5376 | .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2 |
| 5377 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 5378 | .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| 5379 | .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3 |
| 5380 | .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| 5381 | .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| 5382 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5383 | .byte 76,137,193 // mov %r8,%rcx |
| 5384 | .byte 255,224 // jmpq *%rax |
| 5385 | .byte 185,8,0,0,0 // mov $0x8,%ecx |
| 5386 | .byte 68,41,193 // sub %r8d,%ecx |
| 5387 | .byte 192,225,3 // shl $0x3,%cl |
| 5388 | .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax |
| 5389 | .byte 72,211,232 // shr %cl,%rax |
| 5390 | .byte 196,225,249,110,192 // vmovq %rax,%xmm0 |
| 5391 | .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0 |
| 5392 | .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5393 | .byte 233,116,255,255,255 // jmpq 1616 <_sk_load_8888_hsw+0x1a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5394 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5395 | HIDDEN _sk_store_8888_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5396 | .globl _sk_store_8888_hsw |
| 5397 | _sk_store_8888_hsw: |
| 5398 | .byte 73,137,200 // mov %rcx,%r8 |
| 5399 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5400 | .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9 |
| 5401 | .byte 76,3,8 // add (%rax),%r9 |
| 5402 | .byte 184,0,0,127,67 // mov $0x437f0000,%eax |
| 5403 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 5404 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 5405 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 5406 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 5407 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 5408 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5409 | .byte 196,193,45,114,242,8 // vpslld $0x8,%ymm10,%ymm10 |
| 5410 | .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9 |
| 5411 | .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10 |
| 5412 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5413 | .byte 196,193,45,114,242,16 // vpslld $0x10,%ymm10,%ymm10 |
| 5414 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 5415 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 5416 | .byte 196,193,61,114,240,24 // vpslld $0x18,%ymm8,%ymm8 |
| 5417 | .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8 |
| 5418 | .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8 |
| 5419 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5420 | .byte 117,12 // jne 1716 <_sk_store_8888_hsw+0x74> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5421 | .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9) |
| 5422 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5423 | .byte 76,137,193 // mov %r8,%rcx |
| 5424 | .byte 255,224 // jmpq *%rax |
| 5425 | .byte 185,8,0,0,0 // mov $0x8,%ecx |
| 5426 | .byte 68,41,193 // sub %r8d,%ecx |
| 5427 | .byte 192,225,3 // shl $0x3,%cl |
| 5428 | .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax |
| 5429 | .byte 72,211,232 // shr %cl,%rax |
| 5430 | .byte 196,97,249,110,200 // vmovq %rax,%xmm9 |
| 5431 | .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9 |
| 5432 | .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5433 | .byte 235,211 // jmp 170f <_sk_store_8888_hsw+0x6d> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5434 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5435 | HIDDEN _sk_load_f16_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5436 | .globl _sk_load_f16_hsw |
| 5437 | _sk_load_f16_hsw: |
| 5438 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5439 | .byte 72,139,0 // mov (%rax),%rax |
| 5440 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5441 | .byte 117,97 // jne 17a7 <_sk_load_f16_hsw+0x6b> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5442 | .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 |
| 5443 | .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 |
| 5444 | .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 |
| 5445 | .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| 5446 | .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| 5447 | .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| 5448 | .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| 5449 | .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| 5450 | .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| 5451 | .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| 5452 | .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| 5453 | .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3 |
| 5454 | .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| 5455 | .byte 196,226,125,19,192 // vcvtph2ps %xmm0,%ymm0 |
| 5456 | .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| 5457 | .byte 196,226,125,19,201 // vcvtph2ps %xmm1,%ymm1 |
| 5458 | .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2 |
| 5459 | .byte 196,226,125,19,210 // vcvtph2ps %xmm2,%ymm2 |
| 5460 | .byte 197,177,109,219 // vpunpckhqdq %xmm3,%xmm9,%xmm3 |
| 5461 | .byte 196,226,125,19,219 // vcvtph2ps %xmm3,%ymm3 |
| 5462 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5463 | .byte 255,224 // jmpq *%rax |
| 5464 | .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 |
| 5465 | .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| 5466 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5467 | .byte 116,79 // je 1806 <_sk_load_f16_hsw+0xca> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5468 | .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 |
| 5469 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5470 | .byte 114,67 // jb 1806 <_sk_load_f16_hsw+0xca> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5471 | .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 |
| 5472 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5473 | .byte 116,68 // je 1813 <_sk_load_f16_hsw+0xd7> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5474 | .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 |
| 5475 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5476 | .byte 114,56 // jb 1813 <_sk_load_f16_hsw+0xd7> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5477 | .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 |
| 5478 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5479 | .byte 15,132,114,255,255,255 // je 175d <_sk_load_f16_hsw+0x21> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5480 | .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 |
| 5481 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5482 | .byte 15,130,98,255,255,255 // jb 175d <_sk_load_f16_hsw+0x21> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5483 | .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5484 | .byte 233,87,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5485 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 5486 | .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5487 | .byte 233,74,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5488 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5489 | .byte 233,65,255,255,255 // jmpq 175d <_sk_load_f16_hsw+0x21> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5490 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5491 | HIDDEN _sk_store_f16_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5492 | .globl _sk_store_f16_hsw |
| 5493 | _sk_store_f16_hsw: |
| 5494 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5495 | .byte 72,139,0 // mov (%rax),%rax |
| 5496 | .byte 196,195,125,29,192,4 // vcvtps2ph $0x4,%ymm0,%xmm8 |
| 5497 | .byte 196,195,125,29,201,4 // vcvtps2ph $0x4,%ymm1,%xmm9 |
| 5498 | .byte 196,195,125,29,210,4 // vcvtps2ph $0x4,%ymm2,%xmm10 |
| 5499 | .byte 196,195,125,29,219,4 // vcvtps2ph $0x4,%ymm3,%xmm11 |
| 5500 | .byte 196,65,57,97,225 // vpunpcklwd %xmm9,%xmm8,%xmm12 |
| 5501 | .byte 196,65,57,105,193 // vpunpckhwd %xmm9,%xmm8,%xmm8 |
| 5502 | .byte 196,65,41,97,203 // vpunpcklwd %xmm11,%xmm10,%xmm9 |
| 5503 | .byte 196,65,41,105,235 // vpunpckhwd %xmm11,%xmm10,%xmm13 |
| 5504 | .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| 5505 | .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| 5506 | .byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9 |
| 5507 | .byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8 |
| 5508 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5509 | .byte 117,27 // jne 1881 <_sk_store_f16_hsw+0x65> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5510 | .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8) |
| 5511 | .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8) |
| 5512 | .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8) |
| 5513 | .byte 197,122,127,68,248,48 // vmovdqu %xmm8,0x30(%rax,%rdi,8) |
| 5514 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5515 | .byte 255,224 // jmpq *%rax |
| 5516 | .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8) |
| 5517 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5518 | .byte 116,241 // je 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5519 | .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8) |
| 5520 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5521 | .byte 114,229 // jb 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5522 | .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5523 | .byte 116,221 // je 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5524 | .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8) |
| 5525 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5526 | .byte 114,209 // jb 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5527 | .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5528 | .byte 116,201 // je 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5529 | .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8) |
| 5530 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5531 | .byte 114,189 // jb 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5532 | .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 5533 | .byte 235,181 // jmp 187d <_sk_store_f16_hsw+0x61> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5534 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5535 | HIDDEN _sk_load_u16_be_hsw |
| 5536 | .globl _sk_load_u16_be_hsw |
| 5537 | _sk_load_u16_be_hsw: |
| 5538 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5539 | .byte 72,139,0 // mov (%rax),%rax |
| 5540 | .byte 72,133,201 // test %rcx,%rcx |
| 5541 | .byte 15,133,201,0,0,0 // jne 199f <_sk_load_u16_be_hsw+0xd7> |
| 5542 | .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 |
| 5543 | .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 |
| 5544 | .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 |
| 5545 | .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| 5546 | .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| 5547 | .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| 5548 | .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| 5549 | .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| 5550 | .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8 |
| 5551 | .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9 |
| 5552 | .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| 5553 | .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11 |
| 5554 | .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| 5555 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 5556 | .byte 196,98,125,88,208 // vpbroadcastd %xmm0,%ymm10 |
| 5557 | .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0 |
| 5558 | .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| 5559 | .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| 5560 | .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| 5561 | .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0 |
| 5562 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 5563 | .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| 5564 | .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1 |
| 5565 | .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| 5566 | .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| 5567 | .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| 5568 | .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1 |
| 5569 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 5570 | .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| 5571 | .byte 196,193,49,108,211 // vpunpcklqdq %xmm11,%xmm9,%xmm2 |
| 5572 | .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3 |
| 5573 | .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| 5574 | .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2 |
| 5575 | .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2 |
| 5576 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 5577 | .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2 |
| 5578 | .byte 196,193,49,109,219 // vpunpckhqdq %xmm11,%xmm9,%xmm3 |
| 5579 | .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8 |
| 5580 | .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| 5581 | .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 |
| 5582 | .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3 |
| 5583 | .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| 5584 | .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3 |
| 5585 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5586 | .byte 255,224 // jmpq *%rax |
| 5587 | .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 |
| 5588 | .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| 5589 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| 5590 | .byte 116,79 // je 19fe <_sk_load_u16_be_hsw+0x136> |
| 5591 | .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 |
| 5592 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 5593 | .byte 114,67 // jb 19fe <_sk_load_u16_be_hsw+0x136> |
| 5594 | .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 |
| 5595 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 5596 | .byte 116,68 // je 1a0b <_sk_load_u16_be_hsw+0x143> |
| 5597 | .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 |
| 5598 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 5599 | .byte 114,56 // jb 1a0b <_sk_load_u16_be_hsw+0x143> |
| 5600 | .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 |
| 5601 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 5602 | .byte 15,132,10,255,255,255 // je 18ed <_sk_load_u16_be_hsw+0x25> |
| 5603 | .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 |
| 5604 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| 5605 | .byte 15,130,250,254,255,255 // jb 18ed <_sk_load_u16_be_hsw+0x25> |
| 5606 | .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 |
| 5607 | .byte 233,239,254,255,255 // jmpq 18ed <_sk_load_u16_be_hsw+0x25> |
| 5608 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 5609 | .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| 5610 | .byte 233,226,254,255,255 // jmpq 18ed <_sk_load_u16_be_hsw+0x25> |
| 5611 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 5612 | .byte 233,217,254,255,255 // jmpq 18ed <_sk_load_u16_be_hsw+0x25> |
| 5613 | |
| 5614 | HIDDEN _sk_store_u16_be_hsw |
| 5615 | .globl _sk_store_u16_be_hsw |
| 5616 | _sk_store_u16_be_hsw: |
| 5617 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5618 | .byte 76,139,0 // mov (%rax),%r8 |
| 5619 | .byte 184,0,255,127,71 // mov $0x477fff00,%eax |
| 5620 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 5621 | .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8 |
| 5622 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 5623 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 5624 | .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 |
| 5625 | .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9 |
| 5626 | .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10 |
| 5627 | .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9 |
| 5628 | .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9 |
| 5629 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 5630 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 5631 | .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11 |
| 5632 | .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10 |
| 5633 | .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11 |
| 5634 | .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10 |
| 5635 | .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10 |
| 5636 | .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11 |
| 5637 | .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11 |
| 5638 | .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 |
| 5639 | .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11 |
| 5640 | .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12 |
| 5641 | .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11 |
| 5642 | .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11 |
| 5643 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 5644 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 5645 | .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12 |
| 5646 | .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8 |
| 5647 | .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12 |
| 5648 | .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8 |
| 5649 | .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8 |
| 5650 | .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12 |
| 5651 | .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13 |
| 5652 | .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9 |
| 5653 | .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8 |
| 5654 | .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| 5655 | .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| 5656 | .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 |
| 5657 | .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 |
| 5658 | .byte 72,133,201 // test %rcx,%rcx |
| 5659 | .byte 117,31 // jne 1b07 <_sk_store_u16_be_hsw+0xf3> |
| 5660 | .byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8) |
| 5661 | .byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8) |
| 5662 | .byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8) |
| 5663 | .byte 196,65,122,127,68,248,48 // vmovdqu %xmm8,0x30(%r8,%rdi,8) |
| 5664 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5665 | .byte 255,224 // jmpq *%rax |
| 5666 | .byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8) |
| 5667 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| 5668 | .byte 116,240 // je 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5669 | .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) |
| 5670 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 5671 | .byte 114,227 // jb 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5672 | .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) |
| 5673 | .byte 116,218 // je 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5674 | .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) |
| 5675 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 5676 | .byte 114,205 // jb 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5677 | .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) |
| 5678 | .byte 116,196 // je 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5679 | .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) |
| 5680 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| 5681 | .byte 114,183 // jb 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5682 | .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) |
| 5683 | .byte 235,174 // jmp 1b03 <_sk_store_u16_be_hsw+0xef> |
| 5684 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5685 | HIDDEN _sk_store_f32_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5686 | .globl _sk_store_f32_hsw |
| 5687 | _sk_store_f32_hsw: |
| 5688 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5689 | .byte 76,139,0 // mov (%rax),%r8 |
| 5690 | .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| 5691 | .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8 |
| 5692 | .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11 |
| 5693 | .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9 |
| 5694 | .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12 |
| 5695 | .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10 |
| 5696 | .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9 |
| 5697 | .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 |
| 5698 | .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 |
| 5699 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5700 | .byte 117,55 // jne 1bc2 <_sk_store_f32_hsw+0x6d> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5701 | .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 |
| 5702 | .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 |
| 5703 | .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 |
| 5704 | .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 |
| 5705 | .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4) |
| 5706 | .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4) |
| 5707 | .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4) |
| 5708 | .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4) |
| 5709 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5710 | .byte 255,224 // jmpq *%rax |
| 5711 | .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) |
| 5712 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5713 | .byte 116,240 // je 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5714 | .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) |
| 5715 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5716 | .byte 114,227 // jb 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5717 | .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5718 | .byte 116,218 // je 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5719 | .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) |
| 5720 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5721 | .byte 114,205 // jb 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5722 | .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5723 | .byte 116,195 // je 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5724 | .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) |
| 5725 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5726 | .byte 114,181 // jb 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5727 | .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 5728 | .byte 235,171 // jmp 1bbe <_sk_store_f32_hsw+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5729 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5730 | HIDDEN _sk_clamp_x_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5731 | .globl _sk_clamp_x_hsw |
| 5732 | _sk_clamp_x_hsw: |
| 5733 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5734 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 5735 | .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0 |
| 5736 | .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8 |
| 5737 | .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| 5738 | .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8 |
| 5739 | .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| 5740 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5741 | .byte 255,224 // jmpq *%rax |
| 5742 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5743 | HIDDEN _sk_clamp_y_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5744 | .globl _sk_clamp_y_hsw |
| 5745 | _sk_clamp_y_hsw: |
| 5746 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5747 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 5748 | .byte 197,188,95,201 // vmaxps %ymm1,%ymm8,%ymm1 |
| 5749 | .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8 |
| 5750 | .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9 |
| 5751 | .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8 |
| 5752 | .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| 5753 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5754 | .byte 255,224 // jmpq *%rax |
| 5755 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5756 | HIDDEN _sk_repeat_x_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5757 | .globl _sk_repeat_x_hsw |
| 5758 | _sk_repeat_x_hsw: |
| 5759 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5760 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 5761 | .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 |
| 5762 | .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| 5763 | .byte 196,98,61,172,200 // vfnmadd213ps %ymm0,%ymm8,%ymm9 |
| 5764 | .byte 197,253,118,192 // vpcmpeqd %ymm0,%ymm0,%ymm0 |
| 5765 | .byte 197,189,254,192 // vpaddd %ymm0,%ymm8,%ymm0 |
| 5766 | .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| 5767 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5768 | .byte 255,224 // jmpq *%rax |
| 5769 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5770 | HIDDEN _sk_repeat_y_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5771 | .globl _sk_repeat_y_hsw |
| 5772 | _sk_repeat_y_hsw: |
| 5773 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5774 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 5775 | .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 |
| 5776 | .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| 5777 | .byte 196,98,61,172,201 // vfnmadd213ps %ymm1,%ymm8,%ymm9 |
| 5778 | .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1 |
| 5779 | .byte 197,189,254,201 // vpaddd %ymm1,%ymm8,%ymm1 |
| 5780 | .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| 5781 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5782 | .byte 255,224 // jmpq *%rax |
| 5783 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5784 | HIDDEN _sk_mirror_x_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5785 | .globl _sk_mirror_x_hsw |
| 5786 | _sk_mirror_x_hsw: |
| 5787 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5788 | .byte 197,122,16,0 // vmovss (%rax),%xmm8 |
| 5789 | .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 |
| 5790 | .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 |
| 5791 | .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0 |
| 5792 | .byte 196,226,125,24,192 // vbroadcastss %xmm0,%ymm0 |
| 5793 | .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 |
| 5794 | .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| 5795 | .byte 196,66,125,172,194 // vfnmadd213ps %ymm10,%ymm0,%ymm8 |
| 5796 | .byte 196,193,60,92,193 // vsubps %ymm9,%ymm8,%ymm0 |
| 5797 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 5798 | .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 |
| 5799 | .byte 197,188,84,192 // vandps %ymm0,%ymm8,%ymm0 |
| 5800 | .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| 5801 | .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 |
| 5802 | .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| 5803 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5804 | .byte 255,224 // jmpq *%rax |
| 5805 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5806 | HIDDEN _sk_mirror_y_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5807 | .globl _sk_mirror_y_hsw |
| 5808 | _sk_mirror_y_hsw: |
| 5809 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5810 | .byte 197,122,16,0 // vmovss (%rax),%xmm8 |
| 5811 | .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9 |
| 5812 | .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 |
| 5813 | .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1 |
| 5814 | .byte 196,226,125,24,201 // vbroadcastss %xmm1,%ymm1 |
| 5815 | .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 |
| 5816 | .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| 5817 | .byte 196,66,117,172,194 // vfnmadd213ps %ymm10,%ymm1,%ymm8 |
| 5818 | .byte 196,193,60,92,201 // vsubps %ymm9,%ymm8,%ymm1 |
| 5819 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 5820 | .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 |
| 5821 | .byte 197,188,84,201 // vandps %ymm1,%ymm8,%ymm1 |
| 5822 | .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8 |
| 5823 | .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8 |
| 5824 | .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| 5825 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5826 | .byte 255,224 // jmpq *%rax |
| 5827 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5828 | HIDDEN _sk_luminance_to_alpha_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5829 | .globl _sk_luminance_to_alpha_hsw |
| 5830 | _sk_luminance_to_alpha_hsw: |
| 5831 | .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| 5832 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5833 | .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8 |
| 5834 | .byte 184,89,23,55,63 // mov $0x3f371759,%eax |
| 5835 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 5836 | .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3 |
| 5837 | .byte 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1 |
| 5838 | .byte 196,98,125,168,193 // vfmadd213ps %ymm1,%ymm0,%ymm8 |
| 5839 | .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| 5840 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 5841 | .byte 196,226,125,88,216 // vpbroadcastd %xmm0,%ymm3 |
| 5842 | .byte 196,194,109,168,216 // vfmadd213ps %ymm8,%ymm2,%ymm3 |
| 5843 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5844 | .byte 197,253,239,192 // vpxor %ymm0,%ymm0,%ymm0 |
| 5845 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 5846 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 5847 | .byte 255,224 // jmpq *%rax |
| 5848 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5849 | HIDDEN _sk_matrix_2x3_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5850 | .globl _sk_matrix_2x3_hsw |
| 5851 | _sk_matrix_2x3_hsw: |
| 5852 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5853 | .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| 5854 | .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| 5855 | .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8 |
| 5856 | .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| 5857 | .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| 5858 | .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| 5859 | .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11 |
| 5860 | .byte 196,98,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm9 |
| 5861 | .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| 5862 | .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| 5863 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5864 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 5865 | .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| 5866 | .byte 255,224 // jmpq *%rax |
| 5867 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5868 | HIDDEN _sk_matrix_3x4_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5869 | .globl _sk_matrix_3x4_hsw |
| 5870 | _sk_matrix_3x4_hsw: |
| 5871 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5872 | .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| 5873 | .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10 |
| 5874 | .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11 |
| 5875 | .byte 196,98,125,24,64,36 // vbroadcastss 0x24(%rax),%ymm8 |
| 5876 | .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8 |
| 5877 | .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| 5878 | .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| 5879 | .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| 5880 | .byte 196,98,125,24,88,16 // vbroadcastss 0x10(%rax),%ymm11 |
| 5881 | .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12 |
| 5882 | .byte 196,98,125,24,72,40 // vbroadcastss 0x28(%rax),%ymm9 |
| 5883 | .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9 |
| 5884 | .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| 5885 | .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| 5886 | .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| 5887 | .byte 196,98,125,24,96,20 // vbroadcastss 0x14(%rax),%ymm12 |
| 5888 | .byte 196,98,125,24,104,32 // vbroadcastss 0x20(%rax),%ymm13 |
| 5889 | .byte 196,98,125,24,80,44 // vbroadcastss 0x2c(%rax),%ymm10 |
| 5890 | .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10 |
| 5891 | .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10 |
| 5892 | .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10 |
| 5893 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5894 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 5895 | .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| 5896 | .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| 5897 | .byte 255,224 // jmpq *%rax |
| 5898 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5899 | HIDDEN _sk_matrix_4x5_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5900 | .globl _sk_matrix_4x5_hsw |
| 5901 | _sk_matrix_4x5_hsw: |
| 5902 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5903 | .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9 |
| 5904 | .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| 5905 | .byte 196,98,125,24,88,32 // vbroadcastss 0x20(%rax),%ymm11 |
| 5906 | .byte 196,98,125,24,96,48 // vbroadcastss 0x30(%rax),%ymm12 |
| 5907 | .byte 196,98,125,24,64,64 // vbroadcastss 0x40(%rax),%ymm8 |
| 5908 | .byte 196,66,101,184,196 // vfmadd231ps %ymm12,%ymm3,%ymm8 |
| 5909 | .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8 |
| 5910 | .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8 |
| 5911 | .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8 |
| 5912 | .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10 |
| 5913 | .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| 5914 | .byte 196,98,125,24,96,36 // vbroadcastss 0x24(%rax),%ymm12 |
| 5915 | .byte 196,98,125,24,104,52 // vbroadcastss 0x34(%rax),%ymm13 |
| 5916 | .byte 196,98,125,24,72,68 // vbroadcastss 0x44(%rax),%ymm9 |
| 5917 | .byte 196,66,101,184,205 // vfmadd231ps %ymm13,%ymm3,%ymm9 |
| 5918 | .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9 |
| 5919 | .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9 |
| 5920 | .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9 |
| 5921 | .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11 |
| 5922 | .byte 196,98,125,24,96,24 // vbroadcastss 0x18(%rax),%ymm12 |
| 5923 | .byte 196,98,125,24,104,40 // vbroadcastss 0x28(%rax),%ymm13 |
| 5924 | .byte 196,98,125,24,112,56 // vbroadcastss 0x38(%rax),%ymm14 |
| 5925 | .byte 196,98,125,24,80,72 // vbroadcastss 0x48(%rax),%ymm10 |
| 5926 | .byte 196,66,101,184,214 // vfmadd231ps %ymm14,%ymm3,%ymm10 |
| 5927 | .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10 |
| 5928 | .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10 |
| 5929 | .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10 |
| 5930 | .byte 196,98,125,24,96,12 // vbroadcastss 0xc(%rax),%ymm12 |
| 5931 | .byte 196,98,125,24,104,28 // vbroadcastss 0x1c(%rax),%ymm13 |
| 5932 | .byte 196,98,125,24,112,44 // vbroadcastss 0x2c(%rax),%ymm14 |
| 5933 | .byte 196,98,125,24,120,60 // vbroadcastss 0x3c(%rax),%ymm15 |
| 5934 | .byte 196,98,125,24,88,76 // vbroadcastss 0x4c(%rax),%ymm11 |
| 5935 | .byte 196,66,101,184,223 // vfmadd231ps %ymm15,%ymm3,%ymm11 |
| 5936 | .byte 196,66,109,184,222 // vfmadd231ps %ymm14,%ymm2,%ymm11 |
| 5937 | .byte 196,66,117,184,221 // vfmadd231ps %ymm13,%ymm1,%ymm11 |
| 5938 | .byte 196,66,125,184,220 // vfmadd231ps %ymm12,%ymm0,%ymm11 |
| 5939 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5940 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 5941 | .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| 5942 | .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| 5943 | .byte 197,124,41,219 // vmovaps %ymm11,%ymm3 |
| 5944 | .byte 255,224 // jmpq *%rax |
| 5945 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5946 | HIDDEN _sk_matrix_perspective_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5947 | .globl _sk_matrix_perspective_hsw |
| 5948 | _sk_matrix_perspective_hsw: |
| 5949 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5950 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 5951 | .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| 5952 | .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| 5953 | .byte 196,66,117,184,209 // vfmadd231ps %ymm9,%ymm1,%ymm10 |
| 5954 | .byte 196,66,125,184,208 // vfmadd231ps %ymm8,%ymm0,%ymm10 |
| 5955 | .byte 196,98,125,24,64,12 // vbroadcastss 0xc(%rax),%ymm8 |
| 5956 | .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9 |
| 5957 | .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| 5958 | .byte 196,66,117,184,217 // vfmadd231ps %ymm9,%ymm1,%ymm11 |
| 5959 | .byte 196,66,125,184,216 // vfmadd231ps %ymm8,%ymm0,%ymm11 |
| 5960 | .byte 196,98,125,24,64,24 // vbroadcastss 0x18(%rax),%ymm8 |
| 5961 | .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 |
| 5962 | .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| 5963 | .byte 196,66,117,184,225 // vfmadd231ps %ymm9,%ymm1,%ymm12 |
| 5964 | .byte 196,66,125,184,224 // vfmadd231ps %ymm8,%ymm0,%ymm12 |
| 5965 | .byte 196,193,124,83,204 // vrcpps %ymm12,%ymm1 |
| 5966 | .byte 197,172,89,193 // vmulps %ymm1,%ymm10,%ymm0 |
| 5967 | .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| 5968 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5969 | .byte 255,224 // jmpq *%rax |
| 5970 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5971 | HIDDEN _sk_linear_gradient_2stops_hsw |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5972 | .globl _sk_linear_gradient_2stops_hsw |
| 5973 | _sk_linear_gradient_2stops_hsw: |
| 5974 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5975 | .byte 196,226,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm1 |
| 5976 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 5977 | .byte 196,98,125,184,193 // vfmadd231ps %ymm1,%ymm0,%ymm8 |
| 5978 | .byte 196,226,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm2 |
| 5979 | .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| 5980 | .byte 196,226,125,184,202 // vfmadd231ps %ymm2,%ymm0,%ymm1 |
| 5981 | .byte 196,226,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm3 |
| 5982 | .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| 5983 | .byte 196,226,125,184,211 // vfmadd231ps %ymm3,%ymm0,%ymm2 |
| 5984 | .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9 |
| 5985 | .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| 5986 | .byte 196,194,125,184,217 // vfmadd231ps %ymm9,%ymm0,%ymm3 |
| 5987 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 5988 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 5989 | .byte 255,224 // jmpq *%rax |
| 5990 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 5991 | HIDDEN _sk_start_pipeline_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 5992 | .globl _sk_start_pipeline_avx |
| 5993 | _sk_start_pipeline_avx: |
| 5994 | .byte 65,87 // push %r15 |
| 5995 | .byte 65,86 // push %r14 |
| 5996 | .byte 65,85 // push %r13 |
| 5997 | .byte 65,84 // push %r12 |
| 5998 | .byte 83 // push %rbx |
| 5999 | .byte 73,137,205 // mov %rcx,%r13 |
| 6000 | .byte 73,137,214 // mov %rdx,%r14 |
| 6001 | .byte 72,137,251 // mov %rdi,%rbx |
| 6002 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6003 | .byte 73,137,199 // mov %rax,%r15 |
| 6004 | .byte 73,137,244 // mov %rsi,%r12 |
| 6005 | .byte 72,141,67,8 // lea 0x8(%rbx),%rax |
| 6006 | .byte 76,57,232 // cmp %r13,%rax |
| 6007 | .byte 118,5 // jbe 28 <_sk_start_pipeline_avx+0x28> |
| 6008 | .byte 72,137,223 // mov %rbx,%rdi |
| 6009 | .byte 235,65 // jmp 69 <_sk_start_pipeline_avx+0x69> |
| 6010 | .byte 185,0,0,0,0 // mov $0x0,%ecx |
| 6011 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 6012 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 6013 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 6014 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 6015 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 6016 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 6017 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 6018 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 6019 | .byte 72,137,223 // mov %rbx,%rdi |
| 6020 | .byte 76,137,230 // mov %r12,%rsi |
| 6021 | .byte 76,137,242 // mov %r14,%rdx |
| 6022 | .byte 65,255,215 // callq *%r15 |
| 6023 | .byte 72,141,123,8 // lea 0x8(%rbx),%rdi |
| 6024 | .byte 72,131,195,16 // add $0x10,%rbx |
| 6025 | .byte 76,57,235 // cmp %r13,%rbx |
| 6026 | .byte 72,137,251 // mov %rdi,%rbx |
| 6027 | .byte 118,191 // jbe 28 <_sk_start_pipeline_avx+0x28> |
| 6028 | .byte 76,137,233 // mov %r13,%rcx |
| 6029 | .byte 72,41,249 // sub %rdi,%rcx |
| 6030 | .byte 116,41 // je 9a <_sk_start_pipeline_avx+0x9a> |
| 6031 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 6032 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 6033 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 6034 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 6035 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 6036 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 6037 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 6038 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 6039 | .byte 76,137,230 // mov %r12,%rsi |
| 6040 | .byte 76,137,242 // mov %r14,%rdx |
| 6041 | .byte 65,255,215 // callq *%r15 |
| 6042 | .byte 76,137,232 // mov %r13,%rax |
| 6043 | .byte 91 // pop %rbx |
| 6044 | .byte 65,92 // pop %r12 |
| 6045 | .byte 65,93 // pop %r13 |
| 6046 | .byte 65,94 // pop %r14 |
| 6047 | .byte 65,95 // pop %r15 |
| 6048 | .byte 197,248,119 // vzeroupper |
| 6049 | .byte 195 // retq |
| 6050 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6051 | HIDDEN _sk_just_return_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6052 | .globl _sk_just_return_avx |
| 6053 | _sk_just_return_avx: |
| 6054 | .byte 195 // retq |
| 6055 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6056 | HIDDEN _sk_seed_shader_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6057 | .globl _sk_seed_shader_avx |
| 6058 | _sk_seed_shader_avx: |
| 6059 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6060 | .byte 197,249,110,199 // vmovd %edi,%xmm0 |
| 6061 | .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| 6062 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 6063 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 6064 | .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d |
| 6065 | .byte 196,193,121,110,200 // vmovd %r8d,%xmm1 |
| 6066 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 6067 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 6068 | .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| 6069 | .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0 |
| 6070 | .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 |
| 6071 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 6072 | .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| 6073 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6074 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 6075 | .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| 6076 | .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| 6077 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6078 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 6079 | .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4 |
| 6080 | .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5 |
| 6081 | .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6 |
| 6082 | .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7 |
| 6083 | .byte 255,224 // jmpq *%rax |
| 6084 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6085 | HIDDEN _sk_constant_color_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6086 | .globl _sk_constant_color_avx |
| 6087 | _sk_constant_color_avx: |
| 6088 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6089 | .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| 6090 | .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| 6091 | .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| 6092 | .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3 |
| 6093 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6094 | .byte 255,224 // jmpq *%rax |
| 6095 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6096 | HIDDEN _sk_clear_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6097 | .globl _sk_clear_avx |
| 6098 | _sk_clear_avx: |
| 6099 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6100 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 6101 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 6102 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 6103 | .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3 |
| 6104 | .byte 255,224 // jmpq *%rax |
| 6105 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 6106 | HIDDEN _sk_srcatop_avx |
| 6107 | .globl _sk_srcatop_avx |
| 6108 | _sk_srcatop_avx: |
| 6109 | .byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8 |
| 6110 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6111 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 6112 | .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| 6113 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 6114 | .byte 197,124,92,203 // vsubps %ymm3,%ymm0,%ymm9 |
| 6115 | .byte 197,180,89,196 // vmulps %ymm4,%ymm9,%ymm0 |
| 6116 | .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 |
| 6117 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 6118 | .byte 197,52,89,197 // vmulps %ymm5,%ymm9,%ymm8 |
| 6119 | .byte 196,193,116,88,200 // vaddps %ymm8,%ymm1,%ymm1 |
| 6120 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 6121 | .byte 197,52,89,198 // vmulps %ymm6,%ymm9,%ymm8 |
| 6122 | .byte 196,193,108,88,208 // vaddps %ymm8,%ymm2,%ymm2 |
| 6123 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6124 | .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8 |
| 6125 | .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3 |
| 6126 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6127 | .byte 255,224 // jmpq *%rax |
| 6128 | |
| 6129 | HIDDEN _sk_dstatop_avx |
| 6130 | .globl _sk_dstatop_avx |
| 6131 | _sk_dstatop_avx: |
| 6132 | .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8 |
| 6133 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6134 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 6135 | .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| 6136 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 6137 | .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9 |
| 6138 | .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| 6139 | .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0 |
| 6140 | .byte 197,100,89,197 // vmulps %ymm5,%ymm3,%ymm8 |
| 6141 | .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| 6142 | .byte 197,188,88,201 // vaddps %ymm1,%ymm8,%ymm1 |
| 6143 | .byte 197,100,89,198 // vmulps %ymm6,%ymm3,%ymm8 |
| 6144 | .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| 6145 | .byte 197,188,88,210 // vaddps %ymm2,%ymm8,%ymm2 |
| 6146 | .byte 197,100,89,199 // vmulps %ymm7,%ymm3,%ymm8 |
| 6147 | .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| 6148 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6149 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6150 | .byte 255,224 // jmpq *%rax |
| 6151 | |
| 6152 | HIDDEN _sk_srcin_avx |
| 6153 | .globl _sk_srcin_avx |
| 6154 | _sk_srcin_avx: |
| 6155 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 6156 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 6157 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 6158 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6159 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6160 | .byte 255,224 // jmpq *%rax |
| 6161 | |
| 6162 | HIDDEN _sk_dstin_avx |
| 6163 | .globl _sk_dstin_avx |
| 6164 | _sk_dstin_avx: |
| 6165 | .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| 6166 | .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| 6167 | .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| 6168 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6169 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6170 | .byte 255,224 // jmpq *%rax |
| 6171 | |
| 6172 | HIDDEN _sk_srcout_avx |
| 6173 | .globl _sk_srcout_avx |
| 6174 | _sk_srcout_avx: |
| 6175 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6176 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6177 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6178 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6179 | .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| 6180 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 6181 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 6182 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 6183 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 6184 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6185 | .byte 255,224 // jmpq *%rax |
| 6186 | |
| 6187 | HIDDEN _sk_dstout_avx |
| 6188 | .globl _sk_dstout_avx |
| 6189 | _sk_dstout_avx: |
| 6190 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6191 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 6192 | .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| 6193 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 6194 | .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3 |
| 6195 | .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0 |
| 6196 | .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1 |
| 6197 | .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2 |
| 6198 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6199 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6200 | .byte 255,224 // jmpq *%rax |
| 6201 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6202 | HIDDEN _sk_srcover_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6203 | .globl _sk_srcover_avx |
| 6204 | _sk_srcover_avx: |
| 6205 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6206 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6207 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6208 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6209 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6210 | .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9 |
| 6211 | .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0 |
| 6212 | .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9 |
| 6213 | .byte 197,180,88,201 // vaddps %ymm1,%ymm9,%ymm1 |
| 6214 | .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9 |
| 6215 | .byte 197,180,88,210 // vaddps %ymm2,%ymm9,%ymm2 |
| 6216 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6217 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6218 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6219 | .byte 255,224 // jmpq *%rax |
| 6220 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6221 | HIDDEN _sk_dstover_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6222 | .globl _sk_dstover_avx |
| 6223 | _sk_dstover_avx: |
| 6224 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6225 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6226 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6227 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6228 | .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8 |
| 6229 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 6230 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 6231 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 6232 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 6233 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 6234 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 6235 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 6236 | .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| 6237 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6238 | .byte 255,224 // jmpq *%rax |
| 6239 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 6240 | HIDDEN _sk_modulate_avx |
| 6241 | .globl _sk_modulate_avx |
| 6242 | _sk_modulate_avx: |
| 6243 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 6244 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 6245 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 6246 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6247 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6248 | .byte 255,224 // jmpq *%rax |
| 6249 | |
| 6250 | HIDDEN _sk_multiply_avx |
| 6251 | .globl _sk_multiply_avx |
| 6252 | _sk_multiply_avx: |
| 6253 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6254 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6255 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6256 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6257 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 6258 | .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10 |
| 6259 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6260 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 6261 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 6262 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 6263 | .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| 6264 | .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10 |
| 6265 | .byte 197,60,89,221 // vmulps %ymm5,%ymm8,%ymm11 |
| 6266 | .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| 6267 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 6268 | .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1 |
| 6269 | .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10 |
| 6270 | .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| 6271 | .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| 6272 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 6273 | .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| 6274 | .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| 6275 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6276 | .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| 6277 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6278 | .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3 |
| 6279 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6280 | .byte 255,224 // jmpq *%rax |
| 6281 | |
| 6282 | HIDDEN _sk_plus__avx |
| 6283 | .globl _sk_plus__avx |
| 6284 | _sk_plus__avx: |
| 6285 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 6286 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 6287 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 6288 | .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| 6289 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6290 | .byte 255,224 // jmpq *%rax |
| 6291 | |
| 6292 | HIDDEN _sk_screen_avx |
| 6293 | .globl _sk_screen_avx |
| 6294 | _sk_screen_avx: |
| 6295 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 6296 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 6297 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 6298 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 6299 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 6300 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 6301 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 6302 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 6303 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 6304 | .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8 |
| 6305 | .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3 |
| 6306 | .byte 197,188,92,219 // vsubps %ymm3,%ymm8,%ymm3 |
| 6307 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6308 | .byte 255,224 // jmpq *%rax |
| 6309 | |
| 6310 | HIDDEN _sk_xor__avx |
| 6311 | .globl _sk_xor__avx |
| 6312 | _sk_xor__avx: |
| 6313 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6314 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6315 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6316 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6317 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 6318 | .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| 6319 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6320 | .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10 |
| 6321 | .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0 |
| 6322 | .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| 6323 | .byte 197,60,89,213 // vmulps %ymm5,%ymm8,%ymm10 |
| 6324 | .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1 |
| 6325 | .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2 |
| 6326 | .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10 |
| 6327 | .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| 6328 | .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3 |
| 6329 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6330 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6331 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6332 | .byte 255,224 // jmpq *%rax |
| 6333 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 6334 | HIDDEN _sk_darken_avx |
| 6335 | .globl _sk_darken_avx |
| 6336 | _sk_darken_avx: |
| 6337 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 6338 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 6339 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 6340 | .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0 |
| 6341 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 6342 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 6343 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 6344 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 6345 | .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1 |
| 6346 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 6347 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 6348 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 6349 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 6350 | .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2 |
| 6351 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 6352 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6353 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6354 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6355 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6356 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6357 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6358 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6359 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6360 | .byte 255,224 // jmpq *%rax |
| 6361 | |
| 6362 | HIDDEN _sk_lighten_avx |
| 6363 | .globl _sk_lighten_avx |
| 6364 | _sk_lighten_avx: |
| 6365 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 6366 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 6367 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 6368 | .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| 6369 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 6370 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 6371 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 6372 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 6373 | .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| 6374 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 6375 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 6376 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 6377 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 6378 | .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| 6379 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 6380 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6381 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6382 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6383 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6384 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6385 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6386 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6387 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6388 | .byte 255,224 // jmpq *%rax |
| 6389 | |
| 6390 | HIDDEN _sk_difference_avx |
| 6391 | .globl _sk_difference_avx |
| 6392 | _sk_difference_avx: |
| 6393 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 6394 | .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0 |
| 6395 | .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9 |
| 6396 | .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0 |
| 6397 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 6398 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 6399 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 6400 | .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1 |
| 6401 | .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9 |
| 6402 | .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1 |
| 6403 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 6404 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 6405 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 6406 | .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2 |
| 6407 | .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9 |
| 6408 | .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2 |
| 6409 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 6410 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 6411 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6412 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6413 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6414 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6415 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6416 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6417 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6418 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6419 | .byte 255,224 // jmpq *%rax |
| 6420 | |
| 6421 | HIDDEN _sk_exclusion_avx |
| 6422 | .globl _sk_exclusion_avx |
| 6423 | _sk_exclusion_avx: |
| 6424 | .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8 |
| 6425 | .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0 |
| 6426 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 6427 | .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0 |
| 6428 | .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8 |
| 6429 | .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1 |
| 6430 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 6431 | .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1 |
| 6432 | .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8 |
| 6433 | .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2 |
| 6434 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 6435 | .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2 |
| 6436 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6437 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6438 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6439 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6440 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6441 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6442 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6443 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6444 | .byte 255,224 // jmpq *%rax |
| 6445 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 6446 | HIDDEN _sk_colorburn_avx |
| 6447 | .globl _sk_colorburn_avx |
| 6448 | _sk_colorburn_avx: |
| 6449 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6450 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6451 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6452 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6453 | .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9 |
| 6454 | .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11 |
| 6455 | .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10 |
| 6456 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6457 | .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12 |
| 6458 | .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13 |
| 6459 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 6460 | .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13 |
| 6461 | .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| 6462 | .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| 6463 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 6464 | .byte 196,65,20,88,235 // vaddps %ymm11,%ymm13,%ymm13 |
| 6465 | .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| 6466 | .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12 |
| 6467 | .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0 |
| 6468 | .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| 6469 | .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12 |
| 6470 | .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| 6471 | .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| 6472 | .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11 |
| 6473 | .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| 6474 | .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13 |
| 6475 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 6476 | .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13 |
| 6477 | .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13 |
| 6478 | .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13 |
| 6479 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 6480 | .byte 196,65,36,88,237 // vaddps %ymm13,%ymm11,%ymm13 |
| 6481 | .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13 |
| 6482 | .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12 |
| 6483 | .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1 |
| 6484 | .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1 |
| 6485 | .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12 |
| 6486 | .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| 6487 | .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| 6488 | .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9 |
| 6489 | .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10 |
| 6490 | .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| 6491 | .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12 |
| 6492 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 6493 | .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12 |
| 6494 | .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2 |
| 6495 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 6496 | .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12 |
| 6497 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 6498 | .byte 196,65,52,88,228 // vaddps %ymm12,%ymm9,%ymm12 |
| 6499 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 6500 | .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2 |
| 6501 | .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10 |
| 6502 | .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9 |
| 6503 | .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2 |
| 6504 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6505 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6506 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6507 | .byte 255,224 // jmpq *%rax |
| 6508 | |
| 6509 | HIDDEN _sk_colordodge_avx |
| 6510 | .globl _sk_colordodge_avx |
| 6511 | _sk_colordodge_avx: |
| 6512 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 6513 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6514 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 6515 | .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| 6516 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 6517 | .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10 |
| 6518 | .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11 |
| 6519 | .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9 |
| 6520 | .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12 |
| 6521 | .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13 |
| 6522 | .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| 6523 | .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13 |
| 6524 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 6525 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 6526 | .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12 |
| 6527 | .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| 6528 | .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13 |
| 6529 | .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0 |
| 6530 | .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0 |
| 6531 | .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12 |
| 6532 | .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11 |
| 6533 | .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0 |
| 6534 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 6535 | .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12 |
| 6536 | .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13 |
| 6537 | .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12 |
| 6538 | .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13 |
| 6539 | .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12 |
| 6540 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 6541 | .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12 |
| 6542 | .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12 |
| 6543 | .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13 |
| 6544 | .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1 |
| 6545 | .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| 6546 | .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12 |
| 6547 | .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11 |
| 6548 | .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1 |
| 6549 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 6550 | .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11 |
| 6551 | .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12 |
| 6552 | .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11 |
| 6553 | .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12 |
| 6554 | .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11 |
| 6555 | .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 |
| 6556 | .byte 196,65,44,88,219 // vaddps %ymm11,%ymm10,%ymm11 |
| 6557 | .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| 6558 | .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12 |
| 6559 | .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2 |
| 6560 | .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2 |
| 6561 | .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8 |
| 6562 | .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10 |
| 6563 | .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2 |
| 6564 | .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8 |
| 6565 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6566 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6567 | .byte 255,224 // jmpq *%rax |
| 6568 | |
| 6569 | HIDDEN _sk_hardlight_avx |
| 6570 | .globl _sk_hardlight_avx |
| 6571 | _sk_hardlight_avx: |
| 6572 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6573 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6574 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6575 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6576 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 6577 | .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 |
| 6578 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6579 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 6580 | .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11 |
| 6581 | .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9 |
| 6582 | .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12 |
| 6583 | .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| 6584 | .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| 6585 | .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| 6586 | .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| 6587 | .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| 6588 | .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| 6589 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 6590 | .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| 6591 | .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| 6592 | .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| 6593 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 6594 | .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| 6595 | .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| 6596 | .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12 |
| 6597 | .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12 |
| 6598 | .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| 6599 | .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| 6600 | .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| 6601 | .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| 6602 | .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| 6603 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 6604 | .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| 6605 | .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| 6606 | .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| 6607 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 6608 | .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| 6609 | .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| 6610 | .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11 |
| 6611 | .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| 6612 | .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| 6613 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 6614 | .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| 6615 | .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| 6616 | .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| 6617 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 6618 | .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| 6619 | .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| 6620 | .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| 6621 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6622 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6623 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6624 | .byte 255,224 // jmpq *%rax |
| 6625 | |
| 6626 | HIDDEN _sk_overlay_avx |
| 6627 | .globl _sk_overlay_avx |
| 6628 | _sk_overlay_avx: |
| 6629 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6630 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6631 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6632 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6633 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 6634 | .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9 |
| 6635 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6636 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 6637 | .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11 |
| 6638 | .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9 |
| 6639 | .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12 |
| 6640 | .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9 |
| 6641 | .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13 |
| 6642 | .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9 |
| 6643 | .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14 |
| 6644 | .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0 |
| 6645 | .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0 |
| 6646 | .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0 |
| 6647 | .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0 |
| 6648 | .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0 |
| 6649 | .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| 6650 | .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11 |
| 6651 | .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12 |
| 6652 | .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11 |
| 6653 | .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12 |
| 6654 | .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| 6655 | .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13 |
| 6656 | .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13 |
| 6657 | .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14 |
| 6658 | .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1 |
| 6659 | .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1 |
| 6660 | .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1 |
| 6661 | .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1 |
| 6662 | .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1 |
| 6663 | .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| 6664 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 6665 | .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11 |
| 6666 | .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| 6667 | .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11 |
| 6668 | .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| 6669 | .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12 |
| 6670 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 6671 | .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13 |
| 6672 | .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2 |
| 6673 | .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2 |
| 6674 | .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2 |
| 6675 | .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2 |
| 6676 | .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2 |
| 6677 | .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2 |
| 6678 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6679 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6680 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6681 | .byte 255,224 // jmpq *%rax |
| 6682 | |
| 6683 | HIDDEN _sk_softlight_avx |
| 6684 | .globl _sk_softlight_avx |
| 6685 | _sk_softlight_avx: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 6686 | .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 6687 | .byte 197,252,40,209 // vmovaps %ymm1,%ymm2 |
| 6688 | .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 |
| 6689 | .byte 197,52,194,215,1 // vcmpltps %ymm7,%ymm9,%ymm10 |
| 6690 | .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8 |
| 6691 | .byte 196,67,53,74,216,160 // vblendvps %ymm10,%ymm8,%ymm9,%ymm11 |
| 6692 | .byte 196,65,36,88,195 // vaddps %ymm11,%ymm11,%ymm8 |
| 6693 | .byte 196,65,60,88,224 // vaddps %ymm8,%ymm8,%ymm12 |
| 6694 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6695 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6696 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6697 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6698 | .byte 196,65,28,89,236 // vmulps %ymm12,%ymm12,%ymm13 |
| 6699 | .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| 6700 | .byte 196,65,36,92,232 // vsubps %ymm8,%ymm11,%ymm13 |
| 6701 | .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13 |
| 6702 | .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| 6703 | .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| 6704 | .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12 |
| 6705 | .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12 |
| 6706 | .byte 196,65,36,89,244 // vmulps %ymm12,%ymm11,%ymm14 |
| 6707 | .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 |
| 6708 | .byte 196,65,124,82,243 // vrsqrtps %ymm11,%ymm14 |
| 6709 | .byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14 |
| 6710 | .byte 196,65,12,92,243 // vsubps %ymm11,%ymm14,%ymm14 |
| 6711 | .byte 197,92,88,252 // vaddps %ymm4,%ymm4,%ymm15 |
| 6712 | .byte 196,65,4,88,255 // vaddps %ymm15,%ymm15,%ymm15 |
| 6713 | .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15 |
| 6714 | .byte 196,67,13,74,237,240 // vblendvps %ymm15,%ymm13,%ymm14,%ymm13 |
| 6715 | .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14 |
| 6716 | .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| 6717 | .byte 196,65,60,92,219 // vsubps %ymm11,%ymm8,%ymm11 |
| 6718 | .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11 |
| 6719 | .byte 197,36,88,219 // vaddps %ymm3,%ymm11,%ymm11 |
| 6720 | .byte 197,36,89,220 // vmulps %ymm4,%ymm11,%ymm11 |
| 6721 | .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| 6722 | .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| 6723 | .byte 197,100,89,252 // vmulps %ymm4,%ymm3,%ymm15 |
| 6724 | .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13 |
| 6725 | .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| 6726 | .byte 196,195,21,74,203,224 // vblendvps %ymm14,%ymm11,%ymm13,%ymm1 |
| 6727 | .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13 |
| 6728 | .byte 196,67,53,74,237,160 // vblendvps %ymm10,%ymm13,%ymm9,%ymm13 |
| 6729 | .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14 |
| 6730 | .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14 |
| 6731 | .byte 196,65,12,89,254 // vmulps %ymm14,%ymm14,%ymm15 |
| 6732 | .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14 |
| 6733 | .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15 |
| 6734 | .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14 |
| 6735 | .byte 196,65,28,89,253 // vmulps %ymm13,%ymm12,%ymm15 |
| 6736 | .byte 196,65,4,88,246 // vaddps %ymm14,%ymm15,%ymm14 |
| 6737 | .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15 |
| 6738 | .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15 |
| 6739 | .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15 |
| 6740 | .byte 197,84,88,221 // vaddps %ymm5,%ymm5,%ymm11 |
| 6741 | .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11 |
| 6742 | .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11 |
| 6743 | .byte 196,67,5,74,222,176 // vblendvps %ymm11,%ymm14,%ymm15,%ymm11 |
| 6744 | .byte 197,108,88,242 // vaddps %ymm2,%ymm2,%ymm14 |
| 6745 | .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13 |
| 6746 | .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15 |
| 6747 | .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13 |
| 6748 | .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15 |
| 6749 | .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11 |
| 6750 | .byte 197,100,89,253 // vmulps %ymm5,%ymm3,%ymm15 |
| 6751 | .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11 |
| 6752 | .byte 197,20,88,235 // vaddps %ymm3,%ymm13,%ymm13 |
| 6753 | .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13 |
| 6754 | .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14 |
| 6755 | .byte 196,67,37,74,237,224 // vblendvps %ymm14,%ymm13,%ymm11,%ymm13 |
| 6756 | .byte 197,76,94,223 // vdivps %ymm7,%ymm6,%ymm11 |
| 6757 | .byte 196,67,53,74,203,160 // vblendvps %ymm10,%ymm11,%ymm9,%ymm9 |
| 6758 | .byte 196,65,52,88,209 // vaddps %ymm9,%ymm9,%ymm10 |
| 6759 | .byte 196,65,44,88,210 // vaddps %ymm10,%ymm10,%ymm10 |
| 6760 | .byte 196,65,44,89,218 // vmulps %ymm10,%ymm10,%ymm11 |
| 6761 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 6762 | .byte 196,65,52,92,216 // vsubps %ymm8,%ymm9,%ymm11 |
| 6763 | .byte 196,65,36,89,210 // vmulps %ymm10,%ymm11,%ymm10 |
| 6764 | .byte 196,65,28,89,217 // vmulps %ymm9,%ymm12,%ymm11 |
| 6765 | .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10 |
| 6766 | .byte 196,65,124,82,217 // vrsqrtps %ymm9,%ymm11 |
| 6767 | .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11 |
| 6768 | .byte 196,65,36,92,217 // vsubps %ymm9,%ymm11,%ymm11 |
| 6769 | .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12 |
| 6770 | .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12 |
| 6771 | .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12 |
| 6772 | .byte 196,67,37,74,210,192 // vblendvps %ymm12,%ymm10,%ymm11,%ymm10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 6773 | .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 6774 | .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11 |
| 6775 | .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12 |
| 6776 | .byte 196,65,60,92,201 // vsubps %ymm9,%ymm8,%ymm9 |
| 6777 | .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9 |
| 6778 | .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12 |
| 6779 | .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10 |
| 6780 | .byte 197,100,89,230 // vmulps %ymm6,%ymm3,%ymm12 |
| 6781 | .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10 |
| 6782 | .byte 197,52,88,203 // vaddps %ymm3,%ymm9,%ymm9 |
| 6783 | .byte 197,52,89,206 // vmulps %ymm6,%ymm9,%ymm9 |
| 6784 | .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11 |
| 6785 | .byte 196,67,45,74,201,176 // vblendvps %ymm11,%ymm9,%ymm10,%ymm9 |
| 6786 | .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10 |
| 6787 | .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| 6788 | .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8 |
| 6789 | .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11 |
| 6790 | .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0 |
| 6791 | .byte 197,244,88,192 // vaddps %ymm0,%ymm1,%ymm0 |
| 6792 | .byte 197,172,89,202 // vmulps %ymm2,%ymm10,%ymm1 |
| 6793 | .byte 197,188,89,213 // vmulps %ymm5,%ymm8,%ymm2 |
| 6794 | .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| 6795 | .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1 |
| 6796 | .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2 |
| 6797 | .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10 |
| 6798 | .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2 |
| 6799 | .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2 |
| 6800 | .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8 |
| 6801 | .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3 |
| 6802 | .byte 72,173 // lods %ds:(%rsi),%rax |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 6803 | .byte 255,224 // jmpq *%rax |
| 6804 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6805 | HIDDEN _sk_clamp_0_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6806 | .globl _sk_clamp_0_avx |
| 6807 | _sk_clamp_0_avx: |
| 6808 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 6809 | .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0 |
| 6810 | .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1 |
| 6811 | .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2 |
| 6812 | .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3 |
| 6813 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6814 | .byte 255,224 // jmpq *%rax |
| 6815 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6816 | HIDDEN _sk_clamp_1_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6817 | .globl _sk_clamp_1_avx |
| 6818 | _sk_clamp_1_avx: |
| 6819 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6820 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6821 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6822 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6823 | .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0 |
| 6824 | .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1 |
| 6825 | .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2 |
| 6826 | .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| 6827 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6828 | .byte 255,224 // jmpq *%rax |
| 6829 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6830 | HIDDEN _sk_clamp_a_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6831 | .globl _sk_clamp_a_avx |
| 6832 | _sk_clamp_a_avx: |
| 6833 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6834 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6835 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6836 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6837 | .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3 |
| 6838 | .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0 |
| 6839 | .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1 |
| 6840 | .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2 |
| 6841 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6842 | .byte 255,224 // jmpq *%rax |
| 6843 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6844 | HIDDEN _sk_set_rgb_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6845 | .globl _sk_set_rgb_avx |
| 6846 | _sk_set_rgb_avx: |
| 6847 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6848 | .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0 |
| 6849 | .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1 |
| 6850 | .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2 |
| 6851 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6852 | .byte 255,224 // jmpq *%rax |
| 6853 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6854 | HIDDEN _sk_swap_rb_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6855 | .globl _sk_swap_rb_avx |
| 6856 | _sk_swap_rb_avx: |
| 6857 | .byte 197,124,40,192 // vmovaps %ymm0,%ymm8 |
| 6858 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6859 | .byte 197,252,40,194 // vmovaps %ymm2,%ymm0 |
| 6860 | .byte 197,124,41,194 // vmovaps %ymm8,%ymm2 |
| 6861 | .byte 255,224 // jmpq *%rax |
| 6862 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6863 | HIDDEN _sk_swap_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6864 | .globl _sk_swap_avx |
| 6865 | _sk_swap_avx: |
| 6866 | .byte 197,124,40,195 // vmovaps %ymm3,%ymm8 |
| 6867 | .byte 197,124,40,202 // vmovaps %ymm2,%ymm9 |
| 6868 | .byte 197,124,40,209 // vmovaps %ymm1,%ymm10 |
| 6869 | .byte 197,124,40,216 // vmovaps %ymm0,%ymm11 |
| 6870 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6871 | .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| 6872 | .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| 6873 | .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| 6874 | .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| 6875 | .byte 197,124,41,220 // vmovaps %ymm11,%ymm4 |
| 6876 | .byte 197,124,41,213 // vmovaps %ymm10,%ymm5 |
| 6877 | .byte 197,124,41,206 // vmovaps %ymm9,%ymm6 |
| 6878 | .byte 197,124,41,199 // vmovaps %ymm8,%ymm7 |
| 6879 | .byte 255,224 // jmpq *%rax |
| 6880 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6881 | HIDDEN _sk_move_src_dst_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6882 | .globl _sk_move_src_dst_avx |
| 6883 | _sk_move_src_dst_avx: |
| 6884 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6885 | .byte 197,252,40,224 // vmovaps %ymm0,%ymm4 |
| 6886 | .byte 197,252,40,233 // vmovaps %ymm1,%ymm5 |
| 6887 | .byte 197,252,40,242 // vmovaps %ymm2,%ymm6 |
| 6888 | .byte 197,252,40,251 // vmovaps %ymm3,%ymm7 |
| 6889 | .byte 255,224 // jmpq *%rax |
| 6890 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6891 | HIDDEN _sk_move_dst_src_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6892 | .globl _sk_move_dst_src_avx |
| 6893 | _sk_move_dst_src_avx: |
| 6894 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6895 | .byte 197,252,40,196 // vmovaps %ymm4,%ymm0 |
| 6896 | .byte 197,252,40,205 // vmovaps %ymm5,%ymm1 |
| 6897 | .byte 197,252,40,214 // vmovaps %ymm6,%ymm2 |
| 6898 | .byte 197,252,40,223 // vmovaps %ymm7,%ymm3 |
| 6899 | .byte 255,224 // jmpq *%rax |
| 6900 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6901 | HIDDEN _sk_premul_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6902 | .globl _sk_premul_avx |
| 6903 | _sk_premul_avx: |
| 6904 | .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0 |
| 6905 | .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| 6906 | .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| 6907 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6908 | .byte 255,224 // jmpq *%rax |
| 6909 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6910 | HIDDEN _sk_unpremul_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6911 | .globl _sk_unpremul_avx |
| 6912 | _sk_unpremul_avx: |
| 6913 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 6914 | .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9 |
| 6915 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6916 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 6917 | .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| 6918 | .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| 6919 | .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10 |
| 6920 | .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8 |
| 6921 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 6922 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 6923 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 6924 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6925 | .byte 255,224 // jmpq *%rax |
| 6926 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6927 | HIDDEN _sk_from_srgb_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6928 | .globl _sk_from_srgb_avx |
| 6929 | _sk_from_srgb_avx: |
| 6930 | .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| 6931 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6932 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6933 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6934 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 6935 | .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10 |
| 6936 | .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| 6937 | .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| 6938 | .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| 6939 | .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| 6940 | .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| 6941 | .byte 197,121,110,224 // vmovd %eax,%xmm12 |
| 6942 | .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12 |
| 6943 | .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12 |
| 6944 | .byte 197,36,89,232 // vmulps %ymm0,%ymm11,%ymm13 |
| 6945 | .byte 196,65,20,88,236 // vaddps %ymm12,%ymm13,%ymm13 |
| 6946 | .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| 6947 | .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| 6948 | .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| 6949 | .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| 6950 | .byte 196,65,44,89,213 // vmulps %ymm13,%ymm10,%ymm10 |
| 6951 | .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10 |
| 6952 | .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| 6953 | .byte 197,121,110,232 // vmovd %eax,%xmm13 |
| 6954 | .byte 196,67,121,4,237,0 // vpermilps $0x0,%xmm13,%xmm13 |
| 6955 | .byte 196,67,21,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm13,%ymm13 |
| 6956 | .byte 196,193,124,194,197,1 // vcmpltps %ymm13,%ymm0,%ymm0 |
| 6957 | .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0 |
| 6958 | .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9 |
| 6959 | .byte 197,116,89,209 // vmulps %ymm1,%ymm1,%ymm10 |
| 6960 | .byte 197,36,89,249 // vmulps %ymm1,%ymm11,%ymm15 |
| 6961 | .byte 196,65,28,88,255 // vaddps %ymm15,%ymm12,%ymm15 |
| 6962 | .byte 196,65,44,89,215 // vmulps %ymm15,%ymm10,%ymm10 |
| 6963 | .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10 |
| 6964 | .byte 196,193,116,194,205,1 // vcmpltps %ymm13,%ymm1,%ymm1 |
| 6965 | .byte 196,195,45,74,201,16 // vblendvps %ymm1,%ymm9,%ymm10,%ymm1 |
| 6966 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 6967 | .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9 |
| 6968 | .byte 197,36,89,210 // vmulps %ymm2,%ymm11,%ymm10 |
| 6969 | .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10 |
| 6970 | .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9 |
| 6971 | .byte 196,65,12,88,201 // vaddps %ymm9,%ymm14,%ymm9 |
| 6972 | .byte 196,193,108,194,213,1 // vcmpltps %ymm13,%ymm2,%ymm2 |
| 6973 | .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| 6974 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 6975 | .byte 255,224 // jmpq *%rax |
| 6976 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 6977 | HIDDEN _sk_to_srgb_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 6978 | .globl _sk_to_srgb_avx |
| 6979 | _sk_to_srgb_avx: |
| 6980 | .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8 |
| 6981 | .byte 196,65,124,83,232 // vrcpps %ymm8,%ymm13 |
| 6982 | .byte 196,65,124,82,240 // vrsqrtps %ymm8,%ymm14 |
| 6983 | .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| 6984 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 6985 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 6986 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 6987 | .byte 197,60,89,224 // vmulps %ymm0,%ymm8,%ymm12 |
| 6988 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 6989 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 6990 | .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| 6991 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 6992 | .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| 6993 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 6994 | .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| 6995 | .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| 6996 | .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| 6997 | .byte 197,121,110,216 // vmovd %eax,%xmm11 |
| 6998 | .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11 |
| 6999 | .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11 |
| 7000 | .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| 7001 | .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| 7002 | .byte 197,121,110,248 // vmovd %eax,%xmm15 |
| 7003 | .byte 196,67,121,4,255,0 // vpermilps $0x0,%xmm15,%xmm15 |
| 7004 | .byte 196,67,5,24,255,1 // vinsertf128 $0x1,%xmm15,%ymm15,%ymm15 |
| 7005 | .byte 196,65,20,89,235 // vmulps %ymm11,%ymm13,%ymm13 |
| 7006 | .byte 196,65,20,88,239 // vaddps %ymm15,%ymm13,%ymm13 |
| 7007 | .byte 196,65,12,89,242 // vmulps %ymm10,%ymm14,%ymm14 |
| 7008 | .byte 196,65,12,88,237 // vaddps %ymm13,%ymm14,%ymm13 |
| 7009 | .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13 |
| 7010 | .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| 7011 | .byte 197,121,110,240 // vmovd %eax,%xmm14 |
| 7012 | .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14 |
| 7013 | .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14 |
| 7014 | .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0 |
| 7015 | .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0 |
| 7016 | .byte 197,124,82,225 // vrsqrtps %ymm1,%ymm12 |
| 7017 | .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13 |
| 7018 | .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12 |
| 7019 | .byte 196,65,36,89,237 // vmulps %ymm13,%ymm11,%ymm13 |
| 7020 | .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13 |
| 7021 | .byte 196,65,44,89,228 // vmulps %ymm12,%ymm10,%ymm12 |
| 7022 | .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| 7023 | .byte 197,60,89,233 // vmulps %ymm1,%ymm8,%ymm13 |
| 7024 | .byte 196,65,52,93,228 // vminps %ymm12,%ymm9,%ymm12 |
| 7025 | .byte 196,193,116,194,206,1 // vcmpltps %ymm14,%ymm1,%ymm1 |
| 7026 | .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1 |
| 7027 | .byte 197,124,82,226 // vrsqrtps %ymm2,%ymm12 |
| 7028 | .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13 |
| 7029 | .byte 196,65,36,89,221 // vmulps %ymm13,%ymm11,%ymm11 |
| 7030 | .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11 |
| 7031 | .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12 |
| 7032 | .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10 |
| 7033 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 7034 | .byte 196,65,52,93,202 // vminps %ymm10,%ymm9,%ymm9 |
| 7035 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 7036 | .byte 196,193,108,194,214,1 // vcmpltps %ymm14,%ymm2,%ymm2 |
| 7037 | .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2 |
| 7038 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7039 | .byte 255,224 // jmpq *%rax |
| 7040 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7041 | HIDDEN _sk_scale_1_float_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7042 | .globl _sk_scale_1_float_avx |
| 7043 | _sk_scale_1_float_avx: |
| 7044 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7045 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 7046 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 7047 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 7048 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 7049 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 7050 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7051 | .byte 255,224 // jmpq *%rax |
| 7052 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7053 | HIDDEN _sk_scale_u8_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7054 | .globl _sk_scale_u8_avx |
| 7055 | _sk_scale_u8_avx: |
| 7056 | .byte 73,137,200 // mov %rcx,%r8 |
| 7057 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7058 | .byte 72,139,0 // mov (%rax),%rax |
| 7059 | .byte 72,1,248 // add %rdi,%rax |
| 7060 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7061 | .byte 117,80 // jne f78 <_sk_scale_u8_avx+0x60> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7062 | .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| 7063 | .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 |
| 7064 | .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 |
| 7065 | .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 |
| 7066 | .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 |
| 7067 | .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| 7068 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7069 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 7070 | .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| 7071 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 7072 | .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| 7073 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 7074 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 7075 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 7076 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 7077 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7078 | .byte 76,137,193 // mov %r8,%rcx |
| 7079 | .byte 255,224 // jmpq *%rax |
| 7080 | .byte 49,201 // xor %ecx,%ecx |
| 7081 | .byte 77,137,194 // mov %r8,%r10 |
| 7082 | .byte 69,49,201 // xor %r9d,%r9d |
| 7083 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 7084 | .byte 72,255,192 // inc %rax |
| 7085 | .byte 73,211,227 // shl %cl,%r11 |
| 7086 | .byte 77,9,217 // or %r11,%r9 |
| 7087 | .byte 72,131,193,8 // add $0x8,%rcx |
| 7088 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7089 | .byte 117,234 // jne f80 <_sk_scale_u8_avx+0x68> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7090 | .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7091 | .byte 235,143 // jmp f2c <_sk_scale_u8_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7092 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7093 | HIDDEN _sk_lerp_1_float_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7094 | .globl _sk_lerp_1_float_avx |
| 7095 | _sk_lerp_1_float_avx: |
| 7096 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7097 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 7098 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 7099 | .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| 7100 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 7101 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 7102 | .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| 7103 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 7104 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 7105 | .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| 7106 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 7107 | .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| 7108 | .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| 7109 | .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| 7110 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7111 | .byte 255,224 // jmpq *%rax |
| 7112 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7113 | HIDDEN _sk_lerp_u8_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7114 | .globl _sk_lerp_u8_avx |
| 7115 | _sk_lerp_u8_avx: |
| 7116 | .byte 73,137,200 // mov %rcx,%r8 |
| 7117 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7118 | .byte 72,139,0 // mov (%rax),%rax |
| 7119 | .byte 72,1,248 // add %rdi,%rax |
| 7120 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7121 | .byte 117,116 // jne 1060 <_sk_lerp_u8_avx+0x84> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7122 | .byte 197,122,126,0 // vmovq (%rax),%xmm8 |
| 7123 | .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9 |
| 7124 | .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8 |
| 7125 | .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8 |
| 7126 | .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8 |
| 7127 | .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8 |
| 7128 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7129 | .byte 197,121,110,200 // vmovd %eax,%xmm9 |
| 7130 | .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9 |
| 7131 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 7132 | .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8 |
| 7133 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 7134 | .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| 7135 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 7136 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 7137 | .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| 7138 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 7139 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 7140 | .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| 7141 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 7142 | .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3 |
| 7143 | .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| 7144 | .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3 |
| 7145 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7146 | .byte 76,137,193 // mov %r8,%rcx |
| 7147 | .byte 255,224 // jmpq *%rax |
| 7148 | .byte 49,201 // xor %ecx,%ecx |
| 7149 | .byte 77,137,194 // mov %r8,%r10 |
| 7150 | .byte 69,49,201 // xor %r9d,%r9d |
| 7151 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 7152 | .byte 72,255,192 // inc %rax |
| 7153 | .byte 73,211,227 // shl %cl,%r11 |
| 7154 | .byte 77,9,217 // or %r11,%r9 |
| 7155 | .byte 72,131,193,8 // add $0x8,%rcx |
| 7156 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7157 | .byte 117,234 // jne 1068 <_sk_lerp_u8_avx+0x8c> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7158 | .byte 196,65,249,110,193 // vmovq %r9,%xmm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7159 | .byte 233,104,255,255,255 // jmpq ff0 <_sk_lerp_u8_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7160 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7161 | HIDDEN _sk_lerp_565_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7162 | .globl _sk_lerp_565_avx |
| 7163 | _sk_lerp_565_avx: |
| 7164 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7165 | .byte 76,139,16 // mov (%rax),%r10 |
| 7166 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7167 | .byte 15,133,250,0,0,0 // jne 1190 <_sk_lerp_565_avx+0x108> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7168 | .byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8 |
| 7169 | .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| 7170 | .byte 197,185,105,219 // vpunpckhwd %xmm3,%xmm8,%xmm3 |
| 7171 | .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8 |
| 7172 | .byte 196,99,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm8 |
| 7173 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 7174 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7175 | .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| 7176 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7177 | .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| 7178 | .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9 |
| 7179 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 7180 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7181 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7182 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7183 | .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9 |
| 7184 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 7185 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7186 | .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| 7187 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7188 | .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| 7189 | .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10 |
| 7190 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 7191 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7192 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7193 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7194 | .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10 |
| 7195 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 7196 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7197 | .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| 7198 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7199 | .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3 |
| 7200 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 7201 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 7202 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7203 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7204 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7205 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 7206 | .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0 |
| 7207 | .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0 |
| 7208 | .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0 |
| 7209 | .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1 |
| 7210 | .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1 |
| 7211 | .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1 |
| 7212 | .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2 |
| 7213 | .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| 7214 | .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2 |
| 7215 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 7216 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7217 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7218 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7219 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7220 | .byte 255,224 // jmpq *%rax |
| 7221 | .byte 65,137,200 // mov %ecx,%r8d |
| 7222 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7223 | .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8 |
| 7224 | .byte 65,254,200 // dec %r8b |
| 7225 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7226 | .byte 15,135,243,254,255,255 // ja 109c <_sk_lerp_565_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7227 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7228 | .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 1200 <_sk_lerp_565_avx+0x178> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7229 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 7230 | .byte 76,1,200 // add %r9,%rax |
| 7231 | .byte 255,224 // jmpq *%rax |
| 7232 | .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3 |
| 7233 | .byte 196,65,97,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm8 |
| 7234 | .byte 196,65,57,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8 |
| 7235 | .byte 196,65,57,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8 |
| 7236 | .byte 196,65,57,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8 |
| 7237 | .byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8 |
| 7238 | .byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8 |
| 7239 | .byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7240 | .byte 233,159,254,255,255 // jmpq 109c <_sk_lerp_565_avx+0x14> |
| 7241 | .byte 15,31,0 // nopl (%rax) |
| 7242 | .byte 241 // icebp |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7243 | .byte 255 // (bad) |
| 7244 | .byte 255 // (bad) |
| 7245 | .byte 255 // (bad) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 7246 | .byte 233,255,255,255,225 // jmpq ffffffffe2001208 <_sk_linear_gradient_2stops_avx+0xffffffffe1ffe847> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7247 | .byte 255 // (bad) |
| 7248 | .byte 255 // (bad) |
| 7249 | .byte 255 // (bad) |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7250 | .byte 217,255 // fcos |
| 7251 | .byte 255 // (bad) |
| 7252 | .byte 255,209 // callq *%rcx |
| 7253 | .byte 255 // (bad) |
| 7254 | .byte 255 // (bad) |
| 7255 | .byte 255,201 // dec %ecx |
| 7256 | .byte 255 // (bad) |
| 7257 | .byte 255 // (bad) |
| 7258 | .byte 255 // (bad) |
| 7259 | .byte 189 // .byte 0xbd |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7260 | .byte 255 // (bad) |
| 7261 | .byte 255 // (bad) |
| 7262 | .byte 255 // .byte 0xff |
| 7263 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7264 | HIDDEN _sk_load_tables_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7265 | .globl _sk_load_tables_avx |
| 7266 | _sk_load_tables_avx: |
| 7267 | .byte 85 // push %rbp |
| 7268 | .byte 65,87 // push %r15 |
| 7269 | .byte 65,86 // push %r14 |
| 7270 | .byte 65,85 // push %r13 |
| 7271 | .byte 65,84 // push %r12 |
| 7272 | .byte 83 // push %rbx |
| 7273 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7274 | .byte 76,139,0 // mov (%rax),%r8 |
| 7275 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7276 | .byte 15,133,56,2,0,0 // jne 146c <_sk_load_tables_avx+0x250> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7277 | .byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8 |
| 7278 | .byte 187,255,0,0,0 // mov $0xff,%ebx |
| 7279 | .byte 197,249,110,195 // vmovd %ebx,%xmm0 |
| 7280 | .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| 7281 | .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9 |
| 7282 | .byte 196,193,52,84,192 // vandps %ymm8,%ymm9,%ymm0 |
| 7283 | .byte 196,193,249,126,193 // vmovq %xmm0,%r9 |
| 7284 | .byte 69,137,203 // mov %r9d,%r11d |
| 7285 | .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10 |
| 7286 | .byte 69,137,214 // mov %r10d,%r14d |
| 7287 | .byte 73,193,234,32 // shr $0x20,%r10 |
| 7288 | .byte 73,193,233,32 // shr $0x20,%r9 |
| 7289 | .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0 |
| 7290 | .byte 196,193,249,126,196 // vmovq %xmm0,%r12 |
| 7291 | .byte 69,137,231 // mov %r12d,%r15d |
| 7292 | .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx |
| 7293 | .byte 65,137,221 // mov %ebx,%r13d |
| 7294 | .byte 72,193,235,32 // shr $0x20,%rbx |
| 7295 | .byte 73,193,236,32 // shr $0x20,%r12 |
| 7296 | .byte 72,139,104,8 // mov 0x8(%rax),%rbp |
| 7297 | .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| 7298 | .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0 |
| 7299 | .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0 |
| 7300 | .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1 |
| 7301 | .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0 |
| 7302 | .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1 |
| 7303 | .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0 |
| 7304 | .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1 |
| 7305 | .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1 |
| 7306 | .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3 |
| 7307 | .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1 |
| 7308 | .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3 |
| 7309 | .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1 |
| 7310 | .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| 7311 | .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1 |
| 7312 | .byte 196,67,125,25,194,1 // vextractf128 $0x1,%ymm8,%xmm10 |
| 7313 | .byte 196,193,105,114,210,8 // vpsrld $0x8,%xmm10,%xmm2 |
| 7314 | .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| 7315 | .byte 197,180,84,201 // vandps %ymm1,%ymm9,%ymm1 |
| 7316 | .byte 196,193,249,126,201 // vmovq %xmm1,%r9 |
| 7317 | .byte 69,137,203 // mov %r9d,%r11d |
| 7318 | .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10 |
| 7319 | .byte 69,137,214 // mov %r10d,%r14d |
| 7320 | .byte 73,193,234,32 // shr $0x20,%r10 |
| 7321 | .byte 73,193,233,32 // shr $0x20,%r9 |
| 7322 | .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1 |
| 7323 | .byte 196,225,249,126,205 // vmovq %xmm1,%rbp |
| 7324 | .byte 65,137,239 // mov %ebp,%r15d |
| 7325 | .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx |
| 7326 | .byte 65,137,220 // mov %ebx,%r12d |
| 7327 | .byte 72,193,235,32 // shr $0x20,%rbx |
| 7328 | .byte 72,193,237,32 // shr $0x20,%rbp |
| 7329 | .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1 |
| 7330 | .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1 |
| 7331 | .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2 |
| 7332 | .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1 |
| 7333 | .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2 |
| 7334 | .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1 |
| 7335 | .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2 |
| 7336 | .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2 |
| 7337 | .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3 |
| 7338 | .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| 7339 | .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3 |
| 7340 | .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| 7341 | .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 |
| 7342 | .byte 72,139,64,24 // mov 0x18(%rax),%rax |
| 7343 | .byte 196,193,105,114,208,16 // vpsrld $0x10,%xmm8,%xmm2 |
| 7344 | .byte 196,193,97,114,210,16 // vpsrld $0x10,%xmm10,%xmm3 |
| 7345 | .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2 |
| 7346 | .byte 197,180,84,210 // vandps %ymm2,%ymm9,%ymm2 |
| 7347 | .byte 196,193,249,126,208 // vmovq %xmm2,%r8 |
| 7348 | .byte 69,137,194 // mov %r8d,%r10d |
| 7349 | .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9 |
| 7350 | .byte 69,137,203 // mov %r9d,%r11d |
| 7351 | .byte 73,193,233,32 // shr $0x20,%r9 |
| 7352 | .byte 73,193,232,32 // shr $0x20,%r8 |
| 7353 | .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2 |
| 7354 | .byte 196,225,249,126,213 // vmovq %xmm2,%rbp |
| 7355 | .byte 65,137,238 // mov %ebp,%r14d |
| 7356 | .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx |
| 7357 | .byte 65,137,223 // mov %ebx,%r15d |
| 7358 | .byte 72,193,235,32 // shr $0x20,%rbx |
| 7359 | .byte 72,193,237,32 // shr $0x20,%rbp |
| 7360 | .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2 |
| 7361 | .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2 |
| 7362 | .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3 |
| 7363 | .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2 |
| 7364 | .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3 |
| 7365 | .byte 196,99,105,33,203,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm9 |
| 7366 | .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3 |
| 7367 | .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3 |
| 7368 | .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2 |
| 7369 | .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2 |
| 7370 | .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3 |
| 7371 | .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2 |
| 7372 | .byte 196,195,109,24,209,1 // vinsertf128 $0x1,%xmm9,%ymm2,%ymm2 |
| 7373 | .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8 |
| 7374 | .byte 196,193,97,114,210,24 // vpsrld $0x18,%xmm10,%xmm3 |
| 7375 | .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| 7376 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 7377 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7378 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7379 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7380 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7381 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 7382 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7383 | .byte 91 // pop %rbx |
| 7384 | .byte 65,92 // pop %r12 |
| 7385 | .byte 65,93 // pop %r13 |
| 7386 | .byte 65,94 // pop %r14 |
| 7387 | .byte 65,95 // pop %r15 |
| 7388 | .byte 93 // pop %rbp |
| 7389 | .byte 255,224 // jmpq *%rax |
| 7390 | .byte 137,203 // mov %ecx,%ebx |
| 7391 | .byte 128,227,7 // and $0x7,%bl |
| 7392 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 7393 | .byte 254,203 // dec %bl |
| 7394 | .byte 128,251,6 // cmp $0x6,%bl |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7395 | .byte 15,135,185,253,255,255 // ja 123a <_sk_load_tables_avx+0x1e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7396 | .byte 15,182,219 // movzbl %bl,%ebx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7397 | .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 1514 <_sk_load_tables_avx+0x2f8> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7398 | .byte 73,99,28,153 // movslq (%r9,%rbx,4),%rbx |
| 7399 | .byte 76,1,203 // add %r9,%rbx |
| 7400 | .byte 255,227 // jmpq *%rbx |
| 7401 | .byte 196,193,121,110,68,184,24 // vmovd 0x18(%r8,%rdi,4),%xmm0 |
| 7402 | .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 |
| 7403 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 7404 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 7405 | .byte 196,99,117,12,192,64 // vblendps $0x40,%ymm0,%ymm1,%ymm8 |
| 7406 | .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| 7407 | .byte 196,195,121,34,68,184,20,1 // vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0 |
| 7408 | .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 |
| 7409 | .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| 7410 | .byte 196,195,121,34,68,184,16,0 // vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0 |
| 7411 | .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8 |
| 7412 | .byte 196,195,57,34,68,184,12,3 // vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0 |
| 7413 | .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| 7414 | .byte 196,195,57,34,68,184,8,2 // vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0 |
| 7415 | .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| 7416 | .byte 196,195,57,34,68,184,4,1 // vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0 |
| 7417 | .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| 7418 | .byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0 |
| 7419 | .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7420 | .byte 233,38,253,255,255 // jmpq 123a <_sk_load_tables_avx+0x1e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7421 | .byte 238 // out %al,(%dx) |
| 7422 | .byte 255 // (bad) |
| 7423 | .byte 255 // (bad) |
| 7424 | .byte 255,224 // jmpq *%rax |
| 7425 | .byte 255 // (bad) |
| 7426 | .byte 255 // (bad) |
| 7427 | .byte 255,210 // callq *%rdx |
| 7428 | .byte 255 // (bad) |
| 7429 | .byte 255 // (bad) |
| 7430 | .byte 255,196 // inc %esp |
| 7431 | .byte 255 // (bad) |
| 7432 | .byte 255 // (bad) |
| 7433 | .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax) |
| 7434 | .byte 255 // (bad) |
| 7435 | .byte 255 // (bad) |
| 7436 | .byte 255 // .byte 0xff |
| 7437 | .byte 128,255,255 // cmp $0xff,%bh |
| 7438 | .byte 255 // .byte 0xff |
| 7439 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7440 | HIDDEN _sk_load_a8_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7441 | .globl _sk_load_a8_avx |
| 7442 | _sk_load_a8_avx: |
| 7443 | .byte 73,137,200 // mov %rcx,%r8 |
| 7444 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7445 | .byte 72,139,0 // mov (%rax),%rax |
| 7446 | .byte 72,1,248 // add %rdi,%rax |
| 7447 | .byte 77,133,192 // test %r8,%r8 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7448 | .byte 117,74 // jne 158a <_sk_load_a8_avx+0x5a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7449 | .byte 197,250,126,0 // vmovq (%rax),%xmm0 |
| 7450 | .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 |
| 7451 | .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 |
| 7452 | .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| 7453 | .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| 7454 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 7455 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7456 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7457 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7458 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7459 | .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3 |
| 7460 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7461 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 7462 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 7463 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 7464 | .byte 76,137,193 // mov %r8,%rcx |
| 7465 | .byte 255,224 // jmpq *%rax |
| 7466 | .byte 49,201 // xor %ecx,%ecx |
| 7467 | .byte 77,137,194 // mov %r8,%r10 |
| 7468 | .byte 69,49,201 // xor %r9d,%r9d |
| 7469 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 7470 | .byte 72,255,192 // inc %rax |
| 7471 | .byte 73,211,227 // shl %cl,%r11 |
| 7472 | .byte 77,9,217 // or %r11,%r9 |
| 7473 | .byte 72,131,193,8 // add $0x8,%rcx |
| 7474 | .byte 73,255,202 // dec %r10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7475 | .byte 117,234 // jne 1592 <_sk_load_a8_avx+0x62> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7476 | .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7477 | .byte 235,149 // jmp 1544 <_sk_load_a8_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7478 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7479 | HIDDEN _sk_store_a8_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7480 | .globl _sk_store_a8_avx |
| 7481 | _sk_store_a8_avx: |
| 7482 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7483 | .byte 76,139,8 // mov (%rax),%r9 |
| 7484 | .byte 184,0,0,127,67 // mov $0x437f0000,%eax |
| 7485 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 7486 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 7487 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 7488 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 7489 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 7490 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 7491 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 7492 | .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8 |
| 7493 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7494 | .byte 117,10 // jne 15f1 <_sk_store_a8_avx+0x42> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7495 | .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1) |
| 7496 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7497 | .byte 255,224 // jmpq *%rax |
| 7498 | .byte 65,137,200 // mov %ecx,%r8d |
| 7499 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7500 | .byte 65,254,200 // dec %r8b |
| 7501 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7502 | .byte 119,236 // ja 15ed <_sk_store_a8_avx+0x3e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7503 | .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8 |
| 7504 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7505 | .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 1654 <_sk_store_a8_avx+0xa5> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7506 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 7507 | .byte 76,1,192 // add %r8,%rax |
| 7508 | .byte 255,224 // jmpq *%rax |
| 7509 | .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1) |
| 7510 | .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1) |
| 7511 | .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1) |
| 7512 | .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1) |
| 7513 | .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1) |
| 7514 | .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1) |
| 7515 | .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1) |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 7516 | .byte 235,154 // jmp 15ed <_sk_store_a8_avx+0x3e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7517 | .byte 144 // nop |
| 7518 | .byte 246,255 // idiv %bh |
| 7519 | .byte 255 // (bad) |
| 7520 | .byte 255 // (bad) |
| 7521 | .byte 238 // out %al,(%dx) |
| 7522 | .byte 255 // (bad) |
| 7523 | .byte 255 // (bad) |
| 7524 | .byte 255,230 // jmpq *%rsi |
| 7525 | .byte 255 // (bad) |
| 7526 | .byte 255 // (bad) |
| 7527 | .byte 255 // (bad) |
| 7528 | .byte 222,255 // fdivrp %st,%st(7) |
| 7529 | .byte 255 // (bad) |
| 7530 | .byte 255,214 // callq *%rsi |
| 7531 | .byte 255 // (bad) |
| 7532 | .byte 255 // (bad) |
| 7533 | .byte 255,206 // dec %esi |
| 7534 | .byte 255 // (bad) |
| 7535 | .byte 255 // (bad) |
| 7536 | .byte 255,198 // inc %esi |
| 7537 | .byte 255 // (bad) |
| 7538 | .byte 255 // (bad) |
| 7539 | .byte 255 // .byte 0xff |
| 7540 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7541 | HIDDEN _sk_load_g8_avx |
| 7542 | .globl _sk_load_g8_avx |
| 7543 | _sk_load_g8_avx: |
| 7544 | .byte 73,137,200 // mov %rcx,%r8 |
| 7545 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7546 | .byte 72,139,0 // mov (%rax),%rax |
| 7547 | .byte 72,1,248 // add %rdi,%rax |
| 7548 | .byte 77,133,192 // test %r8,%r8 |
| 7549 | .byte 117,91 // jne 16db <_sk_load_g8_avx+0x6b> |
| 7550 | .byte 197,250,126,0 // vmovq (%rax),%xmm0 |
| 7551 | .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1 |
| 7552 | .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0 |
| 7553 | .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0 |
| 7554 | .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| 7555 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 7556 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7557 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7558 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7559 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7560 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 7561 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 7562 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7563 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7564 | .byte 196,227,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm3 |
| 7565 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7566 | .byte 76,137,193 // mov %r8,%rcx |
| 7567 | .byte 197,252,40,200 // vmovaps %ymm0,%ymm1 |
| 7568 | .byte 197,252,40,208 // vmovaps %ymm0,%ymm2 |
| 7569 | .byte 255,224 // jmpq *%rax |
| 7570 | .byte 49,201 // xor %ecx,%ecx |
| 7571 | .byte 77,137,194 // mov %r8,%r10 |
| 7572 | .byte 69,49,201 // xor %r9d,%r9d |
| 7573 | .byte 68,15,182,24 // movzbl (%rax),%r11d |
| 7574 | .byte 72,255,192 // inc %rax |
| 7575 | .byte 73,211,227 // shl %cl,%r11 |
| 7576 | .byte 77,9,217 // or %r11,%r9 |
| 7577 | .byte 72,131,193,8 // add $0x8,%rcx |
| 7578 | .byte 73,255,202 // dec %r10 |
| 7579 | .byte 117,234 // jne 16e3 <_sk_load_g8_avx+0x73> |
| 7580 | .byte 196,193,249,110,193 // vmovq %r9,%xmm0 |
| 7581 | .byte 235,132 // jmp 1684 <_sk_load_g8_avx+0x14> |
| 7582 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7583 | HIDDEN _sk_load_565_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7584 | .globl _sk_load_565_avx |
| 7585 | _sk_load_565_avx: |
| 7586 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7587 | .byte 76,139,16 // mov (%rax),%r10 |
| 7588 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7589 | .byte 15,133,209,0,0,0 // jne 17df <_sk_load_565_avx+0xdf> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7590 | .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 |
| 7591 | .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 |
| 7592 | .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 |
| 7593 | .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| 7594 | .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2 |
| 7595 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 7596 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 7597 | .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| 7598 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 7599 | .byte 197,252,84,194 // vandps %ymm2,%ymm0,%ymm0 |
| 7600 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 7601 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 7602 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7603 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7604 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7605 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 7606 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 7607 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7608 | .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| 7609 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7610 | .byte 197,244,84,202 // vandps %ymm2,%ymm1,%ymm1 |
| 7611 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 7612 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 7613 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7614 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7615 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7616 | .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1 |
| 7617 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 7618 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7619 | .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| 7620 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7621 | .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2 |
| 7622 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 7623 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 7624 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7625 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7626 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7627 | .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2 |
| 7628 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 7629 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7630 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7631 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7632 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7633 | .byte 255,224 // jmpq *%rax |
| 7634 | .byte 65,137,200 // mov %ecx,%r8d |
| 7635 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7636 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 7637 | .byte 65,254,200 // dec %r8b |
| 7638 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7639 | .byte 15,135,29,255,255,255 // ja 1714 <_sk_load_565_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7640 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7641 | .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 184c <_sk_load_565_avx+0x14c> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7642 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 7643 | .byte 76,1,200 // add %r9,%rax |
| 7644 | .byte 255,224 // jmpq *%rax |
| 7645 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 7646 | .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 |
| 7647 | .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 |
| 7648 | .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 |
| 7649 | .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 |
| 7650 | .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 |
| 7651 | .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 |
| 7652 | .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7653 | .byte 233,201,254,255,255 // jmpq 1714 <_sk_load_565_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7654 | .byte 144 // nop |
| 7655 | .byte 243,255 // repz (bad) |
| 7656 | .byte 255 // (bad) |
| 7657 | .byte 255 // (bad) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7658 | .byte 235,255 // jmp 1851 <_sk_load_565_avx+0x151> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7659 | .byte 255 // (bad) |
| 7660 | .byte 255,227 // jmpq *%rbx |
| 7661 | .byte 255 // (bad) |
| 7662 | .byte 255 // (bad) |
| 7663 | .byte 255 // (bad) |
| 7664 | .byte 219,255 // (bad) |
| 7665 | .byte 255 // (bad) |
| 7666 | .byte 255,211 // callq *%rbx |
| 7667 | .byte 255 // (bad) |
| 7668 | .byte 255 // (bad) |
| 7669 | .byte 255,203 // dec %ebx |
| 7670 | .byte 255 // (bad) |
| 7671 | .byte 255 // (bad) |
| 7672 | .byte 255 // (bad) |
| 7673 | .byte 191 // .byte 0xbf |
| 7674 | .byte 255 // (bad) |
| 7675 | .byte 255 // (bad) |
| 7676 | .byte 255 // .byte 0xff |
| 7677 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7678 | HIDDEN _sk_store_565_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7679 | .globl _sk_store_565_avx |
| 7680 | _sk_store_565_avx: |
| 7681 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7682 | .byte 76,139,8 // mov (%rax),%r9 |
| 7683 | .byte 184,0,0,248,65 // mov $0x41f80000,%eax |
| 7684 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 7685 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 7686 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 7687 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 7688 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 7689 | .byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10 |
| 7690 | .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 |
| 7691 | .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9 |
| 7692 | .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 |
| 7693 | .byte 184,0,0,124,66 // mov $0x427c0000,%eax |
| 7694 | .byte 197,121,110,208 // vmovd %eax,%xmm10 |
| 7695 | .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10 |
| 7696 | .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10 |
| 7697 | .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| 7698 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 7699 | .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11 |
| 7700 | .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| 7701 | .byte 196,193,41,114,242,5 // vpslld $0x5,%xmm10,%xmm10 |
| 7702 | .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 |
| 7703 | .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9 |
| 7704 | .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8 |
| 7705 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 7706 | .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 |
| 7707 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 7708 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 7709 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7710 | .byte 117,10 // jne 1906 <_sk_store_565_avx+0x9e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7711 | .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) |
| 7712 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7713 | .byte 255,224 // jmpq *%rax |
| 7714 | .byte 65,137,200 // mov %ecx,%r8d |
| 7715 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7716 | .byte 65,254,200 // dec %r8b |
| 7717 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7718 | .byte 119,236 // ja 1902 <_sk_store_565_avx+0x9a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7719 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7720 | .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 1964 <_sk_store_565_avx+0xfc> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7721 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 7722 | .byte 76,1,192 // add %r8,%rax |
| 7723 | .byte 255,224 // jmpq *%rax |
| 7724 | .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2) |
| 7725 | .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2) |
| 7726 | .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2) |
| 7727 | .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2) |
| 7728 | .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) |
| 7729 | .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) |
| 7730 | .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7731 | .byte 235,159 // jmp 1902 <_sk_store_565_avx+0x9a> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7732 | .byte 144 // nop |
| 7733 | .byte 246,255 // idiv %bh |
| 7734 | .byte 255 // (bad) |
| 7735 | .byte 255 // (bad) |
| 7736 | .byte 238 // out %al,(%dx) |
| 7737 | .byte 255 // (bad) |
| 7738 | .byte 255 // (bad) |
| 7739 | .byte 255,230 // jmpq *%rsi |
| 7740 | .byte 255 // (bad) |
| 7741 | .byte 255 // (bad) |
| 7742 | .byte 255 // (bad) |
| 7743 | .byte 222,255 // fdivrp %st,%st(7) |
| 7744 | .byte 255 // (bad) |
| 7745 | .byte 255,214 // callq *%rsi |
| 7746 | .byte 255 // (bad) |
| 7747 | .byte 255 // (bad) |
| 7748 | .byte 255,206 // dec %esi |
| 7749 | .byte 255 // (bad) |
| 7750 | .byte 255 // (bad) |
| 7751 | .byte 255,198 // inc %esi |
| 7752 | .byte 255 // (bad) |
| 7753 | .byte 255 // (bad) |
| 7754 | .byte 255 // .byte 0xff |
| 7755 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7756 | HIDDEN _sk_load_4444_avx |
| 7757 | .globl _sk_load_4444_avx |
| 7758 | _sk_load_4444_avx: |
| 7759 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7760 | .byte 76,139,16 // mov (%rax),%r10 |
| 7761 | .byte 72,133,201 // test %rcx,%rcx |
| 7762 | .byte 15,133,245,0,0,0 // jne 1a83 <_sk_load_4444_avx+0x103> |
| 7763 | .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0 |
| 7764 | .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1 |
| 7765 | .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1 |
| 7766 | .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| 7767 | .byte 196,99,125,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm9 |
| 7768 | .byte 184,0,240,0,0 // mov $0xf000,%eax |
| 7769 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 7770 | .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| 7771 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 7772 | .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0 |
| 7773 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 7774 | .byte 184,137,136,136,55 // mov $0x37888889,%eax |
| 7775 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7776 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7777 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7778 | .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0 |
| 7779 | .byte 184,0,15,0,0 // mov $0xf00,%eax |
| 7780 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7781 | .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| 7782 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 7783 | .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1 |
| 7784 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 7785 | .byte 184,137,136,136,57 // mov $0x39888889,%eax |
| 7786 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 7787 | .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| 7788 | .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| 7789 | .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 |
| 7790 | .byte 184,240,0,0,0 // mov $0xf0,%eax |
| 7791 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 7792 | .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2 |
| 7793 | .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| 7794 | .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2 |
| 7795 | .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8 |
| 7796 | .byte 184,137,136,136,59 // mov $0x3b888889,%eax |
| 7797 | .byte 197,249,110,208 // vmovd %eax,%xmm2 |
| 7798 | .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2 |
| 7799 | .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2 |
| 7800 | .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2 |
| 7801 | .byte 184,15,0,0,0 // mov $0xf,%eax |
| 7802 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7803 | .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3 |
| 7804 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7805 | .byte 196,193,100,84,217 // vandps %ymm9,%ymm3,%ymm3 |
| 7806 | .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8 |
| 7807 | .byte 184,137,136,136,61 // mov $0x3d888889,%eax |
| 7808 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 7809 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 7810 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 7811 | .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3 |
| 7812 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7813 | .byte 255,224 // jmpq *%rax |
| 7814 | .byte 65,137,200 // mov %ecx,%r8d |
| 7815 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7816 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 7817 | .byte 65,254,200 // dec %r8b |
| 7818 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| 7819 | .byte 15,135,249,254,255,255 // ja 1994 <_sk_load_4444_avx+0x14> |
| 7820 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| 7821 | .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 1af0 <_sk_load_4444_avx+0x170> |
| 7822 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 7823 | .byte 76,1,200 // add %r9,%rax |
| 7824 | .byte 255,224 // jmpq *%rax |
| 7825 | .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0 |
| 7826 | .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0 |
| 7827 | .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0 |
| 7828 | .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0 |
| 7829 | .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0 |
| 7830 | .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0 |
| 7831 | .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0 |
| 7832 | .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0 |
| 7833 | .byte 233,165,254,255,255 // jmpq 1994 <_sk_load_4444_avx+0x14> |
| 7834 | .byte 144 // nop |
| 7835 | .byte 243,255 // repz (bad) |
| 7836 | .byte 255 // (bad) |
| 7837 | .byte 255 // (bad) |
| 7838 | .byte 235,255 // jmp 1af5 <_sk_load_4444_avx+0x175> |
| 7839 | .byte 255 // (bad) |
| 7840 | .byte 255,227 // jmpq *%rbx |
| 7841 | .byte 255 // (bad) |
| 7842 | .byte 255 // (bad) |
| 7843 | .byte 255 // (bad) |
| 7844 | .byte 219,255 // (bad) |
| 7845 | .byte 255 // (bad) |
| 7846 | .byte 255,211 // callq *%rbx |
| 7847 | .byte 255 // (bad) |
| 7848 | .byte 255 // (bad) |
| 7849 | .byte 255,203 // dec %ebx |
| 7850 | .byte 255 // (bad) |
| 7851 | .byte 255 // (bad) |
| 7852 | .byte 255 // (bad) |
| 7853 | .byte 191 // .byte 0xbf |
| 7854 | .byte 255 // (bad) |
| 7855 | .byte 255 // (bad) |
| 7856 | .byte 255 // .byte 0xff |
| 7857 | |
| 7858 | HIDDEN _sk_store_4444_avx |
| 7859 | .globl _sk_store_4444_avx |
| 7860 | _sk_store_4444_avx: |
| 7861 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7862 | .byte 76,139,8 // mov (%rax),%r9 |
| 7863 | .byte 184,0,0,112,65 // mov $0x41700000,%eax |
| 7864 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 7865 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 7866 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 7867 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 7868 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 7869 | .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10 |
| 7870 | .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9 |
| 7871 | .byte 196,193,49,114,241,12 // vpslld $0xc,%xmm9,%xmm9 |
| 7872 | .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9 |
| 7873 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 7874 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 7875 | .byte 196,193,33,114,242,8 // vpslld $0x8,%xmm10,%xmm11 |
| 7876 | .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| 7877 | .byte 196,193,41,114,242,8 // vpslld $0x8,%xmm10,%xmm10 |
| 7878 | .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 |
| 7879 | .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9 |
| 7880 | .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10 |
| 7881 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 7882 | .byte 196,193,33,114,242,4 // vpslld $0x4,%xmm10,%xmm11 |
| 7883 | .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| 7884 | .byte 196,193,41,114,242,4 // vpslld $0x4,%xmm10,%xmm10 |
| 7885 | .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 |
| 7886 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 7887 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 7888 | .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8 |
| 7889 | .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 |
| 7890 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 7891 | .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8 |
| 7892 | .byte 72,133,201 // test %rcx,%rcx |
| 7893 | .byte 117,10 // jne 1bbb <_sk_store_4444_avx+0xaf> |
| 7894 | .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2) |
| 7895 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7896 | .byte 255,224 // jmpq *%rax |
| 7897 | .byte 65,137,200 // mov %ecx,%r8d |
| 7898 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7899 | .byte 65,254,200 // dec %r8b |
| 7900 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| 7901 | .byte 119,236 // ja 1bb7 <_sk_store_4444_avx+0xab> |
| 7902 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| 7903 | .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 1c18 <_sk_store_4444_avx+0x10c> |
| 7904 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 7905 | .byte 76,1,192 // add %r8,%rax |
| 7906 | .byte 255,224 // jmpq *%rax |
| 7907 | .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2) |
| 7908 | .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2) |
| 7909 | .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2) |
| 7910 | .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2) |
| 7911 | .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2) |
| 7912 | .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2) |
| 7913 | .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2) |
| 7914 | .byte 235,159 // jmp 1bb7 <_sk_store_4444_avx+0xab> |
| 7915 | .byte 247,255 // idiv %edi |
| 7916 | .byte 255 // (bad) |
| 7917 | .byte 255 // (bad) |
| 7918 | .byte 239 // out %eax,(%dx) |
| 7919 | .byte 255 // (bad) |
| 7920 | .byte 255 // (bad) |
| 7921 | .byte 255,231 // jmpq *%rdi |
| 7922 | .byte 255 // (bad) |
| 7923 | .byte 255 // (bad) |
| 7924 | .byte 255 // (bad) |
| 7925 | .byte 223,255 // (bad) |
| 7926 | .byte 255 // (bad) |
| 7927 | .byte 255,215 // callq *%rdi |
| 7928 | .byte 255 // (bad) |
| 7929 | .byte 255 // (bad) |
| 7930 | .byte 255,207 // dec %edi |
| 7931 | .byte 255 // (bad) |
| 7932 | .byte 255 // (bad) |
| 7933 | .byte 255,199 // inc %edi |
| 7934 | .byte 255 // (bad) |
| 7935 | .byte 255 // (bad) |
| 7936 | .byte 255 // .byte 0xff |
| 7937 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 7938 | HIDDEN _sk_load_8888_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7939 | .globl _sk_load_8888_avx |
| 7940 | _sk_load_8888_avx: |
| 7941 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7942 | .byte 76,139,16 // mov (%rax),%r10 |
| 7943 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7944 | .byte 15,133,157,0,0,0 // jne 1cdf <_sk_load_8888_avx+0xab> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7945 | .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9 |
| 7946 | .byte 184,255,0,0,0 // mov $0xff,%eax |
| 7947 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 7948 | .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0 |
| 7949 | .byte 196,99,125,24,216,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm11 |
| 7950 | .byte 196,193,36,84,193 // vandps %ymm9,%ymm11,%ymm0 |
| 7951 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 7952 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 7953 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 7954 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 7955 | .byte 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8 |
| 7956 | .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0 |
| 7957 | .byte 196,193,41,114,209,8 // vpsrld $0x8,%xmm9,%xmm10 |
| 7958 | .byte 196,99,125,25,203,1 // vextractf128 $0x1,%ymm9,%xmm3 |
| 7959 | .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1 |
| 7960 | .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1 |
| 7961 | .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1 |
| 7962 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 7963 | .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1 |
| 7964 | .byte 196,193,41,114,209,16 // vpsrld $0x10,%xmm9,%xmm10 |
| 7965 | .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2 |
| 7966 | .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2 |
| 7967 | .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2 |
| 7968 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 7969 | .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2 |
| 7970 | .byte 196,193,49,114,209,24 // vpsrld $0x18,%xmm9,%xmm9 |
| 7971 | .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3 |
| 7972 | .byte 196,227,53,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm9,%ymm3 |
| 7973 | .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| 7974 | .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3 |
| 7975 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 7976 | .byte 255,224 // jmpq *%rax |
| 7977 | .byte 65,137,200 // mov %ecx,%r8d |
| 7978 | .byte 65,128,224,7 // and $0x7,%r8b |
| 7979 | .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9 |
| 7980 | .byte 65,254,200 // dec %r8b |
| 7981 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7982 | .byte 15,135,80,255,255,255 // ja 1c48 <_sk_load_8888_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7983 | .byte 69,15,182,192 // movzbl %r8b,%r8d |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 7984 | .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 1d8c <_sk_load_8888_avx+0x158> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 7985 | .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax |
| 7986 | .byte 76,1,200 // add %r9,%rax |
| 7987 | .byte 255,224 // jmpq *%rax |
| 7988 | .byte 196,193,121,110,68,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm0 |
| 7989 | .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0 |
| 7990 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 7991 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 7992 | .byte 196,99,117,12,200,64 // vblendps $0x40,%ymm0,%ymm1,%ymm9 |
| 7993 | .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| 7994 | .byte 196,195,121,34,68,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0 |
| 7995 | .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 |
| 7996 | .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| 7997 | .byte 196,195,121,34,68,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0 |
| 7998 | .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9 |
| 7999 | .byte 196,195,49,34,68,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0 |
| 8000 | .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| 8001 | .byte 196,195,49,34,68,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0 |
| 8002 | .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| 8003 | .byte 196,195,49,34,68,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0 |
| 8004 | .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| 8005 | .byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0 |
| 8006 | .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9 |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 8007 | .byte 233,188,254,255,255 // jmpq 1c48 <_sk_load_8888_avx+0x14> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8008 | .byte 238 // out %al,(%dx) |
| 8009 | .byte 255 // (bad) |
| 8010 | .byte 255 // (bad) |
| 8011 | .byte 255,224 // jmpq *%rax |
| 8012 | .byte 255 // (bad) |
| 8013 | .byte 255 // (bad) |
| 8014 | .byte 255,210 // callq *%rdx |
| 8015 | .byte 255 // (bad) |
| 8016 | .byte 255 // (bad) |
| 8017 | .byte 255,196 // inc %esp |
| 8018 | .byte 255 // (bad) |
| 8019 | .byte 255 // (bad) |
| 8020 | .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax) |
| 8021 | .byte 255 // (bad) |
| 8022 | .byte 255 // (bad) |
| 8023 | .byte 255 // .byte 0xff |
| 8024 | .byte 128,255,255 // cmp $0xff,%bh |
| 8025 | .byte 255 // .byte 0xff |
| 8026 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8027 | HIDDEN _sk_store_8888_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8028 | .globl _sk_store_8888_avx |
| 8029 | _sk_store_8888_avx: |
| 8030 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8031 | .byte 76,139,8 // mov (%rax),%r9 |
| 8032 | .byte 184,0,0,127,67 // mov $0x437f0000,%eax |
| 8033 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 8034 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 8035 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 8036 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 8037 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 8038 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 8039 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 8040 | .byte 196,193,33,114,242,8 // vpslld $0x8,%xmm10,%xmm11 |
| 8041 | .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| 8042 | .byte 196,193,41,114,242,8 // vpslld $0x8,%xmm10,%xmm10 |
| 8043 | .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 |
| 8044 | .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9 |
| 8045 | .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10 |
| 8046 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 8047 | .byte 196,193,33,114,242,16 // vpslld $0x10,%xmm10,%xmm11 |
| 8048 | .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10 |
| 8049 | .byte 196,193,41,114,242,16 // vpslld $0x10,%xmm10,%xmm10 |
| 8050 | .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10 |
| 8051 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 8052 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 8053 | .byte 196,193,33,114,240,24 // vpslld $0x18,%xmm8,%xmm11 |
| 8054 | .byte 196,67,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm8 |
| 8055 | .byte 196,193,57,114,240,24 // vpslld $0x18,%xmm8,%xmm8 |
| 8056 | .byte 196,67,37,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm11,%ymm8 |
| 8057 | .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8 |
| 8058 | .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8 |
| 8059 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 8060 | .byte 117,10 // jne 1e4c <_sk_store_8888_avx+0xa4> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8061 | .byte 196,65,124,17,4,185 // vmovups %ymm8,(%r9,%rdi,4) |
| 8062 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8063 | .byte 255,224 // jmpq *%rax |
| 8064 | .byte 65,137,200 // mov %ecx,%r8d |
| 8065 | .byte 65,128,224,7 // and $0x7,%r8b |
| 8066 | .byte 65,254,200 // dec %r8b |
| 8067 | .byte 65,128,248,6 // cmp $0x6,%r8b |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 8068 | .byte 119,236 // ja 1e48 <_sk_store_8888_avx+0xa0> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8069 | .byte 65,15,182,192 // movzbl %r8b,%eax |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 8070 | .byte 76,141,5,85,0,0,0 // lea 0x55(%rip),%r8 # 1ebc <_sk_store_8888_avx+0x114> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8071 | .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax |
| 8072 | .byte 76,1,192 // add %r8,%rax |
| 8073 | .byte 255,224 // jmpq *%rax |
| 8074 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 8075 | .byte 196,67,121,22,76,185,24,2 // vpextrd $0x2,%xmm9,0x18(%r9,%rdi,4) |
| 8076 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 8077 | .byte 196,67,121,22,76,185,20,1 // vpextrd $0x1,%xmm9,0x14(%r9,%rdi,4) |
| 8078 | .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9 |
| 8079 | .byte 196,65,122,17,76,185,16 // vmovss %xmm9,0x10(%r9,%rdi,4) |
| 8080 | .byte 196,67,121,22,68,185,12,3 // vpextrd $0x3,%xmm8,0xc(%r9,%rdi,4) |
| 8081 | .byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4) |
| 8082 | .byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4) |
| 8083 | .byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4) |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 8084 | .byte 235,143 // jmp 1e48 <_sk_store_8888_avx+0xa0> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8085 | .byte 15,31,0 // nopl (%rax) |
| 8086 | .byte 245 // cmc |
| 8087 | .byte 255 // (bad) |
| 8088 | .byte 255 // (bad) |
| 8089 | .byte 255 // (bad) |
| 8090 | .byte 237 // in (%dx),%eax |
| 8091 | .byte 255 // (bad) |
| 8092 | .byte 255 // (bad) |
| 8093 | .byte 255,229 // jmpq *%rbp |
| 8094 | .byte 255 // (bad) |
| 8095 | .byte 255 // (bad) |
| 8096 | .byte 255 // (bad) |
| 8097 | .byte 221,255 // (bad) |
| 8098 | .byte 255 // (bad) |
| 8099 | .byte 255,208 // callq *%rax |
| 8100 | .byte 255 // (bad) |
| 8101 | .byte 255 // (bad) |
| 8102 | .byte 255,194 // inc %edx |
| 8103 | .byte 255 // (bad) |
| 8104 | .byte 255 // (bad) |
| 8105 | .byte 255 // .byte 0xff |
| 8106 | .byte 180,255 // mov $0xff,%ah |
| 8107 | .byte 255 // (bad) |
| 8108 | .byte 255 // .byte 0xff |
| 8109 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8110 | HIDDEN _sk_load_f16_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8111 | .globl _sk_load_f16_avx |
| 8112 | _sk_load_f16_avx: |
| 8113 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8114 | .byte 72,139,0 // mov (%rax),%rax |
| 8115 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8116 | .byte 15,133,17,1,0,0 // jne 1ff7 <_sk_load_f16_avx+0x11f> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8117 | .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 |
| 8118 | .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 |
| 8119 | .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 |
| 8120 | .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| 8121 | .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| 8122 | .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| 8123 | .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| 8124 | .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8125 | .byte 197,121,97,218 // vpunpcklwd %xmm2,%xmm0,%xmm11 |
| 8126 | .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8127 | .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| 8128 | .byte 197,113,105,203 // vpunpckhwd %xmm3,%xmm1,%xmm9 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8129 | .byte 197,161,108,194 // vpunpcklqdq %xmm2,%xmm11,%xmm0 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8130 | .byte 184,0,4,0,4 // mov $0x4000400,%eax |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8131 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 8132 | .byte 197,121,112,233,0 // vpshufd $0x0,%xmm1,%xmm13 |
| 8133 | .byte 197,145,101,200 // vpcmpgtw %xmm0,%xmm13,%xmm1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8134 | .byte 197,241,223,192 // vpandn %xmm0,%xmm1,%xmm0 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8135 | .byte 196,226,121,51,200 // vpmovzxwd %xmm0,%xmm1 |
| 8136 | .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10 |
| 8137 | .byte 196,193,121,105,194 // vpunpckhwd %xmm10,%xmm0,%xmm0 |
| 8138 | .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 |
| 8139 | .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0 |
| 8140 | .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8141 | .byte 184,0,0,128,119 // mov $0x77800000,%eax |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8142 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 8143 | .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1 |
| 8144 | .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12 |
| 8145 | .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 |
| 8146 | .byte 197,161,109,202 // vpunpckhqdq %xmm2,%xmm11,%xmm1 |
| 8147 | .byte 197,145,101,209 // vpcmpgtw %xmm1,%xmm13,%xmm2 |
| 8148 | .byte 197,233,223,201 // vpandn %xmm1,%xmm2,%xmm1 |
| 8149 | .byte 196,226,121,51,209 // vpmovzxwd %xmm1,%xmm2 |
| 8150 | .byte 196,193,113,105,202 // vpunpckhwd %xmm10,%xmm1,%xmm1 |
| 8151 | .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8152 | .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1 |
| 8153 | .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8154 | .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| 8155 | .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2 |
| 8156 | .byte 197,145,101,218 // vpcmpgtw %xmm2,%xmm13,%xmm3 |
| 8157 | .byte 197,225,223,210 // vpandn %xmm2,%xmm3,%xmm2 |
| 8158 | .byte 196,226,121,51,218 // vpmovzxwd %xmm2,%xmm3 |
| 8159 | .byte 196,193,105,105,210 // vpunpckhwd %xmm10,%xmm2,%xmm2 |
| 8160 | .byte 197,225,114,243,13 // vpslld $0xd,%xmm3,%xmm3 |
| 8161 | .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2 |
| 8162 | .byte 196,227,101,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm3,%ymm2 |
| 8163 | .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| 8164 | .byte 196,65,57,109,193 // vpunpckhqdq %xmm9,%xmm8,%xmm8 |
| 8165 | .byte 196,193,17,101,216 // vpcmpgtw %xmm8,%xmm13,%xmm3 |
| 8166 | .byte 196,193,97,223,216 // vpandn %xmm8,%xmm3,%xmm3 |
| 8167 | .byte 196,98,121,51,195 // vpmovzxwd %xmm3,%xmm8 |
| 8168 | .byte 196,193,97,105,218 // vpunpckhwd %xmm10,%xmm3,%xmm3 |
| 8169 | .byte 196,193,57,114,240,13 // vpslld $0xd,%xmm8,%xmm8 |
| 8170 | .byte 197,225,114,243,13 // vpslld $0xd,%xmm3,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8171 | .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8172 | .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8173 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8174 | .byte 255,224 // jmpq *%rax |
| 8175 | .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 |
| 8176 | .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| 8177 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8178 | .byte 116,79 // je 2056 <_sk_load_f16_avx+0x17e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8179 | .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 |
| 8180 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8181 | .byte 114,67 // jb 2056 <_sk_load_f16_avx+0x17e> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8182 | .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 |
| 8183 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8184 | .byte 116,68 // je 2063 <_sk_load_f16_avx+0x18b> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8185 | .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 |
| 8186 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8187 | .byte 114,56 // jb 2063 <_sk_load_f16_avx+0x18b> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8188 | .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 |
| 8189 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8190 | .byte 15,132,194,254,255,255 // je 1efd <_sk_load_f16_avx+0x25> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8191 | .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 |
| 8192 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8193 | .byte 15,130,178,254,255,255 // jb 1efd <_sk_load_f16_avx+0x25> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8194 | .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8195 | .byte 233,167,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8196 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 8197 | .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8198 | .byte 233,154,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8199 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 8200 | .byte 233,145,254,255,255 // jmpq 1efd <_sk_load_f16_avx+0x25> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8201 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8202 | HIDDEN _sk_store_f16_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8203 | .globl _sk_store_f16_avx |
| 8204 | _sk_store_f16_avx: |
| 8205 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8206 | .byte 76,139,0 // mov (%rax),%r8 |
| 8207 | .byte 184,0,0,128,7 // mov $0x7800000,%eax |
| 8208 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 8209 | .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8 |
| 8210 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 8211 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 8212 | .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 |
| 8213 | .byte 196,193,41,114,210,13 // vpsrld $0xd,%xmm10,%xmm10 |
| 8214 | .byte 196,193,49,114,209,13 // vpsrld $0xd,%xmm9,%xmm9 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8215 | .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9 |
| 8216 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 8217 | .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11 |
| 8218 | .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 |
| 8219 | .byte 196,193,41,114,210,13 // vpsrld $0xd,%xmm10,%xmm10 |
| 8220 | .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10 |
| 8221 | .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8222 | .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 |
| 8223 | .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12 |
| 8224 | .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8225 | .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8226 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8227 | .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12 |
| 8228 | .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8229 | .byte 196,193,57,114,208,13 // vpsrld $0xd,%xmm8,%xmm8 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8230 | .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8 |
| 8231 | .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12 |
| 8232 | .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13 |
| 8233 | .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9 |
| 8234 | .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8 |
| 8235 | .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| 8236 | .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| 8237 | .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 |
| 8238 | .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8239 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8240 | .byte 117,31 // jne 213e <_sk_store_f16_avx+0xd2> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8241 | .byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8) |
| 8242 | .byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8) |
| 8243 | .byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8) |
| 8244 | .byte 196,65,122,127,68,248,48 // vmovdqu %xmm8,0x30(%r8,%rdi,8) |
| 8245 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8246 | .byte 255,224 // jmpq *%rax |
| 8247 | .byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8) |
| 8248 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8249 | .byte 116,240 // je 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8250 | .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) |
| 8251 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8252 | .byte 114,227 // jb 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8253 | .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8254 | .byte 116,218 // je 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8255 | .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) |
| 8256 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8257 | .byte 114,205 // jb 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8258 | .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8259 | .byte 116,196 // je 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8260 | .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) |
| 8261 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8262 | .byte 114,183 // jb 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8263 | .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 8264 | .byte 235,174 // jmp 213a <_sk_store_f16_avx+0xce> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8265 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8266 | HIDDEN _sk_load_u16_be_avx |
| 8267 | .globl _sk_load_u16_be_avx |
| 8268 | _sk_load_u16_be_avx: |
| 8269 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8270 | .byte 72,139,0 // mov (%rax),%rax |
| 8271 | .byte 72,133,201 // test %rcx,%rcx |
| 8272 | .byte 15,133,1,1,0,0 // jne 229b <_sk_load_u16_be_avx+0x10f> |
| 8273 | .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8 |
| 8274 | .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2 |
| 8275 | .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3 |
| 8276 | .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9 |
| 8277 | .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0 |
| 8278 | .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2 |
| 8279 | .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1 |
| 8280 | .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3 |
| 8281 | .byte 197,121,97,210 // vpunpcklwd %xmm2,%xmm0,%xmm10 |
| 8282 | .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8 |
| 8283 | .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2 |
| 8284 | .byte 197,113,105,203 // vpunpckhwd %xmm3,%xmm1,%xmm9 |
| 8285 | .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| 8286 | .byte 197,249,110,192 // vmovd %eax,%xmm0 |
| 8287 | .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| 8288 | .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12 |
| 8289 | .byte 197,169,108,194 // vpunpcklqdq %xmm2,%xmm10,%xmm0 |
| 8290 | .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1 |
| 8291 | .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0 |
| 8292 | .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0 |
| 8293 | .byte 196,65,33,239,219 // vpxor %xmm11,%xmm11,%xmm11 |
| 8294 | .byte 196,193,121,105,203 // vpunpckhwd %xmm11,%xmm0,%xmm1 |
| 8295 | .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0 |
| 8296 | .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0 |
| 8297 | .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0 |
| 8298 | .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0 |
| 8299 | .byte 197,169,109,202 // vpunpckhqdq %xmm2,%xmm10,%xmm1 |
| 8300 | .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2 |
| 8301 | .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1 |
| 8302 | .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1 |
| 8303 | .byte 196,193,113,105,211 // vpunpckhwd %xmm11,%xmm1,%xmm2 |
| 8304 | .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1 |
| 8305 | .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1 |
| 8306 | .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1 |
| 8307 | .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| 8308 | .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2 |
| 8309 | .byte 197,169,113,242,8 // vpsllw $0x8,%xmm2,%xmm10 |
| 8310 | .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2 |
| 8311 | .byte 197,169,235,210 // vpor %xmm2,%xmm10,%xmm2 |
| 8312 | .byte 196,65,105,105,211 // vpunpckhwd %xmm11,%xmm2,%xmm10 |
| 8313 | .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2 |
| 8314 | .byte 196,195,109,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm2,%ymm2 |
| 8315 | .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2 |
| 8316 | .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| 8317 | .byte 196,193,57,109,217 // vpunpckhqdq %xmm9,%xmm8,%xmm3 |
| 8318 | .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8 |
| 8319 | .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3 |
| 8320 | .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3 |
| 8321 | .byte 196,65,97,105,195 // vpunpckhwd %xmm11,%xmm3,%xmm8 |
| 8322 | .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3 |
| 8323 | .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3 |
| 8324 | .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3 |
| 8325 | .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3 |
| 8326 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8327 | .byte 255,224 // jmpq *%rax |
| 8328 | .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8 |
| 8329 | .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9 |
| 8330 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| 8331 | .byte 116,79 // je 22fa <_sk_load_u16_be_avx+0x16e> |
| 8332 | .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8 |
| 8333 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 8334 | .byte 114,67 // jb 22fa <_sk_load_u16_be_avx+0x16e> |
| 8335 | .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2 |
| 8336 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 8337 | .byte 116,68 // je 2307 <_sk_load_u16_be_avx+0x17b> |
| 8338 | .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2 |
| 8339 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 8340 | .byte 114,56 // jb 2307 <_sk_load_u16_be_avx+0x17b> |
| 8341 | .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3 |
| 8342 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 8343 | .byte 15,132,210,254,255,255 // je 21b1 <_sk_load_u16_be_avx+0x25> |
| 8344 | .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3 |
| 8345 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| 8346 | .byte 15,130,194,254,255,255 // jb 21b1 <_sk_load_u16_be_avx+0x25> |
| 8347 | .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9 |
| 8348 | .byte 233,183,254,255,255 // jmpq 21b1 <_sk_load_u16_be_avx+0x25> |
| 8349 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 8350 | .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2 |
| 8351 | .byte 233,170,254,255,255 // jmpq 21b1 <_sk_load_u16_be_avx+0x25> |
| 8352 | .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3 |
| 8353 | .byte 233,161,254,255,255 // jmpq 21b1 <_sk_load_u16_be_avx+0x25> |
| 8354 | |
| 8355 | HIDDEN _sk_store_u16_be_avx |
| 8356 | .globl _sk_store_u16_be_avx |
| 8357 | _sk_store_u16_be_avx: |
| 8358 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8359 | .byte 76,139,0 // mov (%rax),%r8 |
| 8360 | .byte 184,0,255,127,71 // mov $0x477fff00,%eax |
| 8361 | .byte 197,121,110,192 // vmovd %eax,%xmm8 |
| 8362 | .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8 |
| 8363 | .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8 |
| 8364 | .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9 |
| 8365 | .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9 |
| 8366 | .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10 |
| 8367 | .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9 |
| 8368 | .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10 |
| 8369 | .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9 |
| 8370 | .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9 |
| 8371 | .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10 |
| 8372 | .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10 |
| 8373 | .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11 |
| 8374 | .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10 |
| 8375 | .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11 |
| 8376 | .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10 |
| 8377 | .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10 |
| 8378 | .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11 |
| 8379 | .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11 |
| 8380 | .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12 |
| 8381 | .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11 |
| 8382 | .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12 |
| 8383 | .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11 |
| 8384 | .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11 |
| 8385 | .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8 |
| 8386 | .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8 |
| 8387 | .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12 |
| 8388 | .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8 |
| 8389 | .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12 |
| 8390 | .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8 |
| 8391 | .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8 |
| 8392 | .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12 |
| 8393 | .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13 |
| 8394 | .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9 |
| 8395 | .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8 |
| 8396 | .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11 |
| 8397 | .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10 |
| 8398 | .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9 |
| 8399 | .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8 |
| 8400 | .byte 72,133,201 // test %rcx,%rcx |
| 8401 | .byte 117,31 // jne 240a <_sk_store_u16_be_avx+0xfa> |
| 8402 | .byte 196,65,120,17,28,248 // vmovups %xmm11,(%r8,%rdi,8) |
| 8403 | .byte 196,65,120,17,84,248,16 // vmovups %xmm10,0x10(%r8,%rdi,8) |
| 8404 | .byte 196,65,120,17,76,248,32 // vmovups %xmm9,0x20(%r8,%rdi,8) |
| 8405 | .byte 196,65,122,127,68,248,48 // vmovdqu %xmm8,0x30(%r8,%rdi,8) |
| 8406 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8407 | .byte 255,224 // jmpq *%rax |
| 8408 | .byte 196,65,121,214,28,248 // vmovq %xmm11,(%r8,%rdi,8) |
| 8409 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| 8410 | .byte 116,240 // je 2406 <_sk_store_u16_be_avx+0xf6> |
| 8411 | .byte 196,65,121,23,92,248,8 // vmovhpd %xmm11,0x8(%r8,%rdi,8) |
| 8412 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| 8413 | .byte 114,227 // jb 2406 <_sk_store_u16_be_avx+0xf6> |
| 8414 | .byte 196,65,121,214,84,248,16 // vmovq %xmm10,0x10(%r8,%rdi,8) |
| 8415 | .byte 116,218 // je 2406 <_sk_store_u16_be_avx+0xf6> |
| 8416 | .byte 196,65,121,23,84,248,24 // vmovhpd %xmm10,0x18(%r8,%rdi,8) |
| 8417 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| 8418 | .byte 114,205 // jb 2406 <_sk_store_u16_be_avx+0xf6> |
| 8419 | .byte 196,65,121,214,76,248,32 // vmovq %xmm9,0x20(%r8,%rdi,8) |
| 8420 | .byte 116,196 // je 2406 <_sk_store_u16_be_avx+0xf6> |
| 8421 | .byte 196,65,121,23,76,248,40 // vmovhpd %xmm9,0x28(%r8,%rdi,8) |
| 8422 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| 8423 | .byte 114,183 // jb 2406 <_sk_store_u16_be_avx+0xf6> |
| 8424 | .byte 196,65,121,214,68,248,48 // vmovq %xmm8,0x30(%r8,%rdi,8) |
| 8425 | .byte 235,174 // jmp 2406 <_sk_store_u16_be_avx+0xf6> |
| 8426 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8427 | HIDDEN _sk_store_f32_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8428 | .globl _sk_store_f32_avx |
| 8429 | _sk_store_f32_avx: |
| 8430 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8431 | .byte 76,139,0 // mov (%rax),%r8 |
| 8432 | .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax |
| 8433 | .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8 |
| 8434 | .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11 |
| 8435 | .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9 |
| 8436 | .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12 |
| 8437 | .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10 |
| 8438 | .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9 |
| 8439 | .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8 |
| 8440 | .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11 |
| 8441 | .byte 72,133,201 // test %rcx,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8442 | .byte 117,55 // jne 24c5 <_sk_store_f32_avx+0x6d> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8443 | .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12 |
| 8444 | .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13 |
| 8445 | .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9 |
| 8446 | .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8 |
| 8447 | .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4) |
| 8448 | .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4) |
| 8449 | .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4) |
| 8450 | .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4) |
| 8451 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8452 | .byte 255,224 // jmpq *%rax |
| 8453 | .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4) |
| 8454 | .byte 72,131,249,1 // cmp $0x1,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8455 | .byte 116,240 // je 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8456 | .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4) |
| 8457 | .byte 72,131,249,3 // cmp $0x3,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8458 | .byte 114,227 // jb 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8459 | .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8460 | .byte 116,218 // je 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8461 | .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4) |
| 8462 | .byte 72,131,249,5 // cmp $0x5,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8463 | .byte 114,205 // jb 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8464 | .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8465 | .byte 116,195 // je 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8466 | .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4) |
| 8467 | .byte 72,131,249,7 // cmp $0x7,%rcx |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8468 | .byte 114,181 // jb 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8469 | .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4) |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 8470 | .byte 235,171 // jmp 24c1 <_sk_store_f32_avx+0x69> |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8471 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8472 | HIDDEN _sk_clamp_x_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8473 | .globl _sk_clamp_x_avx |
| 8474 | _sk_clamp_x_avx: |
| 8475 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8476 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 8477 | .byte 197,60,95,200 // vmaxps %ymm0,%ymm8,%ymm9 |
| 8478 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8479 | .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| 8480 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8481 | .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| 8482 | .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| 8483 | .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 |
| 8484 | .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| 8485 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8486 | .byte 255,224 // jmpq *%rax |
| 8487 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8488 | HIDDEN _sk_clamp_y_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8489 | .globl _sk_clamp_y_avx |
| 8490 | _sk_clamp_y_avx: |
| 8491 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8492 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 8493 | .byte 197,60,95,201 // vmaxps %ymm1,%ymm8,%ymm9 |
| 8494 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8495 | .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1 |
| 8496 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8497 | .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| 8498 | .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| 8499 | .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 |
| 8500 | .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| 8501 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8502 | .byte 255,224 // jmpq *%rax |
| 8503 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8504 | HIDDEN _sk_repeat_x_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8505 | .globl _sk_repeat_x_avx |
| 8506 | _sk_repeat_x_avx: |
| 8507 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8508 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8509 | .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9 |
| 8510 | .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| 8511 | .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| 8512 | .byte 196,65,124,92,201 // vsubps %ymm9,%ymm0,%ymm9 |
| 8513 | .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0 |
| 8514 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8515 | .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| 8516 | .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| 8517 | .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0 |
| 8518 | .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0 |
| 8519 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8520 | .byte 255,224 // jmpq *%rax |
| 8521 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8522 | HIDDEN _sk_repeat_y_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8523 | .globl _sk_repeat_y_avx |
| 8524 | _sk_repeat_y_avx: |
| 8525 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8526 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8527 | .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9 |
| 8528 | .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9 |
| 8529 | .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9 |
| 8530 | .byte 196,65,116,92,201 // vsubps %ymm9,%ymm1,%ymm9 |
| 8531 | .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1 |
| 8532 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8533 | .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| 8534 | .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8 |
| 8535 | .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1 |
| 8536 | .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1 |
| 8537 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8538 | .byte 255,224 // jmpq *%rax |
| 8539 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8540 | HIDDEN _sk_mirror_x_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8541 | .globl _sk_mirror_x_avx |
| 8542 | _sk_mirror_x_avx: |
| 8543 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8544 | .byte 197,121,110,0 // vmovd (%rax),%xmm8 |
| 8545 | .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 |
| 8546 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 8547 | .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10 |
| 8548 | .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0 |
| 8549 | .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0 |
| 8550 | .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0 |
| 8551 | .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8 |
| 8552 | .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| 8553 | .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0 |
| 8554 | .byte 197,172,92,192 // vsubps %ymm0,%ymm10,%ymm0 |
| 8555 | .byte 196,193,124,92,193 // vsubps %ymm9,%ymm0,%ymm0 |
| 8556 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 8557 | .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8 |
| 8558 | .byte 197,60,84,192 // vandps %ymm0,%ymm8,%ymm8 |
| 8559 | .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0 |
| 8560 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8561 | .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0 |
| 8562 | .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 |
| 8563 | .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0 |
| 8564 | .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0 |
| 8565 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8566 | .byte 255,224 // jmpq *%rax |
| 8567 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8568 | HIDDEN _sk_mirror_y_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8569 | .globl _sk_mirror_y_avx |
| 8570 | _sk_mirror_y_avx: |
| 8571 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8572 | .byte 197,121,110,0 // vmovd (%rax),%xmm8 |
| 8573 | .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9 |
| 8574 | .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9 |
| 8575 | .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10 |
| 8576 | .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1 |
| 8577 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 8578 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 8579 | .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8 |
| 8580 | .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8 |
| 8581 | .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1 |
| 8582 | .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1 |
| 8583 | .byte 196,193,116,92,201 // vsubps %ymm9,%ymm1,%ymm1 |
| 8584 | .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8 |
| 8585 | .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8 |
| 8586 | .byte 197,60,84,193 // vandps %ymm1,%ymm8,%ymm8 |
| 8587 | .byte 196,99,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm1 |
| 8588 | .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10 |
| 8589 | .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1 |
| 8590 | .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9 |
| 8591 | .byte 196,227,53,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm9,%ymm1 |
| 8592 | .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1 |
| 8593 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8594 | .byte 255,224 // jmpq *%rax |
| 8595 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8596 | HIDDEN _sk_luminance_to_alpha_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8597 | .globl _sk_luminance_to_alpha_avx |
| 8598 | _sk_luminance_to_alpha_avx: |
| 8599 | .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| 8600 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 8601 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 8602 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 8603 | .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 |
| 8604 | .byte 184,89,23,55,63 // mov $0x3f371759,%eax |
| 8605 | .byte 197,249,110,216 // vmovd %eax,%xmm3 |
| 8606 | .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3 |
| 8607 | .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3 |
| 8608 | .byte 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1 |
| 8609 | .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| 8610 | .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| 8611 | .byte 197,249,110,200 // vmovd %eax,%xmm1 |
| 8612 | .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1 |
| 8613 | .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 |
| 8614 | .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1 |
| 8615 | .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 |
| 8616 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8617 | .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0 |
| 8618 | .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1 |
| 8619 | .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2 |
| 8620 | .byte 255,224 // jmpq *%rax |
| 8621 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8622 | HIDDEN _sk_matrix_2x3_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8623 | .globl _sk_matrix_2x3_avx |
| 8624 | _sk_matrix_2x3_avx: |
| 8625 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8626 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8627 | .byte 196,98,125,24,72,8 // vbroadcastss 0x8(%rax),%ymm9 |
| 8628 | .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| 8629 | .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| 8630 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8631 | .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| 8632 | .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| 8633 | .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| 8634 | .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10 |
| 8635 | .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| 8636 | .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1 |
| 8637 | .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1 |
| 8638 | .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0 |
| 8639 | .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1 |
| 8640 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8641 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 8642 | .byte 255,224 // jmpq *%rax |
| 8643 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8644 | HIDDEN _sk_matrix_3x4_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8645 | .globl _sk_matrix_3x4_avx |
| 8646 | _sk_matrix_3x4_avx: |
| 8647 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8648 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8649 | .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| 8650 | .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| 8651 | .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11 |
| 8652 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 8653 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8654 | .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| 8655 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8656 | .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| 8657 | .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| 8658 | .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| 8659 | .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| 8660 | .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11 |
| 8661 | .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12 |
| 8662 | .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 |
| 8663 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 8664 | .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| 8665 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8666 | .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| 8667 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8668 | .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| 8669 | .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| 8670 | .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| 8671 | .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13 |
| 8672 | .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2 |
| 8673 | .byte 196,193,108,88,213 // vaddps %ymm13,%ymm2,%ymm2 |
| 8674 | .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| 8675 | .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1 |
| 8676 | .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| 8677 | .byte 197,252,88,209 // vaddps %ymm1,%ymm0,%ymm2 |
| 8678 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8679 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 8680 | .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| 8681 | .byte 255,224 // jmpq *%rax |
| 8682 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8683 | HIDDEN _sk_matrix_4x5_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8684 | .globl _sk_matrix_4x5_avx |
| 8685 | _sk_matrix_4x5_avx: |
| 8686 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8687 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8688 | .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9 |
| 8689 | .byte 196,98,125,24,80,32 // vbroadcastss 0x20(%rax),%ymm10 |
| 8690 | .byte 196,98,125,24,88,48 // vbroadcastss 0x30(%rax),%ymm11 |
| 8691 | .byte 196,98,125,24,96,64 // vbroadcastss 0x40(%rax),%ymm12 |
| 8692 | .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11 |
| 8693 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 8694 | .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10 |
| 8695 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8696 | .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| 8697 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8698 | .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| 8699 | .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| 8700 | .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| 8701 | .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10 |
| 8702 | .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11 |
| 8703 | .byte 196,98,125,24,96,52 // vbroadcastss 0x34(%rax),%ymm12 |
| 8704 | .byte 196,98,125,24,104,68 // vbroadcastss 0x44(%rax),%ymm13 |
| 8705 | .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12 |
| 8706 | .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| 8707 | .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11 |
| 8708 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 8709 | .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| 8710 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8711 | .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| 8712 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8713 | .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| 8714 | .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11 |
| 8715 | .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12 |
| 8716 | .byte 196,98,125,24,104,56 // vbroadcastss 0x38(%rax),%ymm13 |
| 8717 | .byte 196,98,125,24,112,72 // vbroadcastss 0x48(%rax),%ymm14 |
| 8718 | .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13 |
| 8719 | .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13 |
| 8720 | .byte 197,28,89,226 // vmulps %ymm2,%ymm12,%ymm12 |
| 8721 | .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12 |
| 8722 | .byte 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11 |
| 8723 | .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11 |
| 8724 | .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10 |
| 8725 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8726 | .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11 |
| 8727 | .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12 |
| 8728 | .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13 |
| 8729 | .byte 196,98,125,24,112,60 // vbroadcastss 0x3c(%rax),%ymm14 |
| 8730 | .byte 196,98,125,24,120,76 // vbroadcastss 0x4c(%rax),%ymm15 |
| 8731 | .byte 197,140,89,219 // vmulps %ymm3,%ymm14,%ymm3 |
| 8732 | .byte 196,193,100,88,223 // vaddps %ymm15,%ymm3,%ymm3 |
| 8733 | .byte 197,148,89,210 // vmulps %ymm2,%ymm13,%ymm2 |
| 8734 | .byte 197,236,88,211 // vaddps %ymm3,%ymm2,%ymm2 |
| 8735 | .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1 |
| 8736 | .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1 |
| 8737 | .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0 |
| 8738 | .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3 |
| 8739 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8740 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 8741 | .byte 197,124,41,201 // vmovaps %ymm9,%ymm1 |
| 8742 | .byte 197,124,41,210 // vmovaps %ymm10,%ymm2 |
| 8743 | .byte 255,224 // jmpq *%rax |
| 8744 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8745 | HIDDEN _sk_matrix_perspective_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8746 | .globl _sk_matrix_perspective_avx |
| 8747 | _sk_matrix_perspective_avx: |
| 8748 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8749 | .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8 |
| 8750 | .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9 |
| 8751 | .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10 |
| 8752 | .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9 |
| 8753 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8754 | .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8 |
| 8755 | .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8 |
| 8756 | .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| 8757 | .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10 |
| 8758 | .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11 |
| 8759 | .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10 |
| 8760 | .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10 |
| 8761 | .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9 |
| 8762 | .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9 |
| 8763 | .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10 |
| 8764 | .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11 |
| 8765 | .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12 |
| 8766 | .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1 |
| 8767 | .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1 |
| 8768 | .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0 |
| 8769 | .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0 |
| 8770 | .byte 197,252,83,200 // vrcpps %ymm0,%ymm1 |
| 8771 | .byte 197,188,89,193 // vmulps %ymm1,%ymm8,%ymm0 |
| 8772 | .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1 |
| 8773 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8774 | .byte 255,224 // jmpq *%rax |
| 8775 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8776 | HIDDEN _sk_linear_gradient_2stops_avx |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8777 | .globl _sk_linear_gradient_2stops_avx |
| 8778 | _sk_linear_gradient_2stops_avx: |
| 8779 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8780 | .byte 196,226,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm1 |
| 8781 | .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2 |
| 8782 | .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1 |
| 8783 | .byte 197,108,88,193 // vaddps %ymm1,%ymm2,%ymm8 |
| 8784 | .byte 196,226,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm1 |
| 8785 | .byte 196,226,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm2 |
| 8786 | .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1 |
| 8787 | .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1 |
| 8788 | .byte 196,226,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm2 |
| 8789 | .byte 196,226,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm3 |
| 8790 | .byte 197,236,89,208 // vmulps %ymm0,%ymm2,%ymm2 |
| 8791 | .byte 197,228,88,210 // vaddps %ymm2,%ymm3,%ymm2 |
| 8792 | .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3 |
| 8793 | .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9 |
| 8794 | .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0 |
| 8795 | .byte 197,180,88,216 // vaddps %ymm0,%ymm9,%ymm3 |
| 8796 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8797 | .byte 197,124,41,192 // vmovaps %ymm8,%ymm0 |
| 8798 | .byte 255,224 // jmpq *%rax |
| 8799 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8800 | HIDDEN _sk_start_pipeline_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8801 | .globl _sk_start_pipeline_sse41 |
| 8802 | _sk_start_pipeline_sse41: |
| 8803 | .byte 65,87 // push %r15 |
| 8804 | .byte 65,86 // push %r14 |
| 8805 | .byte 65,85 // push %r13 |
| 8806 | .byte 65,84 // push %r12 |
| 8807 | .byte 83 // push %rbx |
| 8808 | .byte 73,137,207 // mov %rcx,%r15 |
| 8809 | .byte 73,137,214 // mov %rdx,%r14 |
| 8810 | .byte 72,137,251 // mov %rdi,%rbx |
| 8811 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8812 | .byte 73,137,196 // mov %rax,%r12 |
| 8813 | .byte 73,137,245 // mov %rsi,%r13 |
| 8814 | .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| 8815 | .byte 76,57,248 // cmp %r15,%rax |
| 8816 | .byte 118,5 // jbe 28 <_sk_start_pipeline_sse41+0x28> |
| 8817 | .byte 72,137,216 // mov %rbx,%rax |
| 8818 | .byte 235,52 // jmp 5c <_sk_start_pipeline_sse41+0x5c> |
| 8819 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 8820 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 8821 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 8822 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 8823 | .byte 15,87,228 // xorps %xmm4,%xmm4 |
| 8824 | .byte 15,87,237 // xorps %xmm5,%xmm5 |
| 8825 | .byte 15,87,246 // xorps %xmm6,%xmm6 |
| 8826 | .byte 15,87,255 // xorps %xmm7,%xmm7 |
| 8827 | .byte 72,137,223 // mov %rbx,%rdi |
| 8828 | .byte 76,137,238 // mov %r13,%rsi |
| 8829 | .byte 76,137,242 // mov %r14,%rdx |
| 8830 | .byte 65,255,212 // callq *%r12 |
| 8831 | .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| 8832 | .byte 72,131,195,8 // add $0x8,%rbx |
| 8833 | .byte 76,57,251 // cmp %r15,%rbx |
| 8834 | .byte 72,137,195 // mov %rax,%rbx |
| 8835 | .byte 118,204 // jbe 28 <_sk_start_pipeline_sse41+0x28> |
| 8836 | .byte 91 // pop %rbx |
| 8837 | .byte 65,92 // pop %r12 |
| 8838 | .byte 65,93 // pop %r13 |
| 8839 | .byte 65,94 // pop %r14 |
| 8840 | .byte 65,95 // pop %r15 |
| 8841 | .byte 195 // retq |
| 8842 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8843 | HIDDEN _sk_just_return_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8844 | .globl _sk_just_return_sse41 |
| 8845 | _sk_just_return_sse41: |
| 8846 | .byte 195 // retq |
| 8847 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8848 | HIDDEN _sk_seed_shader_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8849 | .globl _sk_seed_shader_sse41 |
| 8850 | _sk_seed_shader_sse41: |
| 8851 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8852 | .byte 102,15,110,199 // movd %edi,%xmm0 |
| 8853 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 8854 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 8855 | .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| 8856 | .byte 102,15,110,209 // movd %ecx,%xmm2 |
| 8857 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 8858 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 8859 | .byte 15,16,2 // movups (%rdx),%xmm0 |
| 8860 | .byte 15,88,193 // addps %xmm1,%xmm0 |
| 8861 | .byte 102,15,110,8 // movd (%rax),%xmm1 |
| 8862 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 8863 | .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| 8864 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 8865 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 8866 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 8867 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 8868 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8869 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 8870 | .byte 15,87,228 // xorps %xmm4,%xmm4 |
| 8871 | .byte 15,87,237 // xorps %xmm5,%xmm5 |
| 8872 | .byte 15,87,246 // xorps %xmm6,%xmm6 |
| 8873 | .byte 15,87,255 // xorps %xmm7,%xmm7 |
| 8874 | .byte 255,224 // jmpq *%rax |
| 8875 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8876 | HIDDEN _sk_constant_color_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8877 | .globl _sk_constant_color_sse41 |
| 8878 | _sk_constant_color_sse41: |
| 8879 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8880 | .byte 15,16,24 // movups (%rax),%xmm3 |
| 8881 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 8882 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 8883 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 8884 | .byte 15,198,201,85 // shufps $0x55,%xmm1,%xmm1 |
| 8885 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 8886 | .byte 15,198,210,170 // shufps $0xaa,%xmm2,%xmm2 |
| 8887 | .byte 15,198,219,255 // shufps $0xff,%xmm3,%xmm3 |
| 8888 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8889 | .byte 255,224 // jmpq *%rax |
| 8890 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 8891 | HIDDEN _sk_clear_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8892 | .globl _sk_clear_sse41 |
| 8893 | _sk_clear_sse41: |
| 8894 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8895 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 8896 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 8897 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 8898 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 8899 | .byte 255,224 // jmpq *%rax |
| 8900 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 8901 | HIDDEN _sk_srcatop_sse41 |
| 8902 | .globl _sk_srcatop_sse41 |
| 8903 | _sk_srcatop_sse41: |
| 8904 | .byte 15,89,199 // mulps %xmm7,%xmm0 |
| 8905 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 8906 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 8907 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 8908 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 8909 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 8910 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 8911 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 8912 | .byte 15,89,207 // mulps %xmm7,%xmm1 |
| 8913 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 8914 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 8915 | .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| 8916 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 8917 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 8918 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 8919 | .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| 8920 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 8921 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 8922 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 8923 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8924 | .byte 255,224 // jmpq *%rax |
| 8925 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 8926 | HIDDEN _sk_dstatop_sse41 |
| 8927 | .globl _sk_dstatop_sse41 |
| 8928 | _sk_dstatop_sse41: |
| 8929 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 8930 | .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| 8931 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 8932 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 8933 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 8934 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 8935 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 8936 | .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| 8937 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 8938 | .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| 8939 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 8940 | .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| 8941 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 8942 | .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| 8943 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 8944 | .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| 8945 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 8946 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 8947 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 8948 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8949 | .byte 255,224 // jmpq *%rax |
| 8950 | |
| 8951 | HIDDEN _sk_srcin_sse41 |
| 8952 | .globl _sk_srcin_sse41 |
| 8953 | _sk_srcin_sse41: |
| 8954 | .byte 15,89,199 // mulps %xmm7,%xmm0 |
| 8955 | .byte 15,89,207 // mulps %xmm7,%xmm1 |
| 8956 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 8957 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 8958 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8959 | .byte 255,224 // jmpq *%rax |
| 8960 | |
| 8961 | HIDDEN _sk_dstin_sse41 |
| 8962 | .globl _sk_dstin_sse41 |
| 8963 | _sk_dstin_sse41: |
| 8964 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 8965 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 8966 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 8967 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 8968 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 8969 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 8970 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 8971 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8972 | .byte 255,224 // jmpq *%rax |
| 8973 | |
| 8974 | HIDDEN _sk_srcout_sse41 |
| 8975 | .globl _sk_srcout_sse41 |
| 8976 | _sk_srcout_sse41: |
| 8977 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 8978 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 8979 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 8980 | .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| 8981 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 8982 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 8983 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 8984 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 8985 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 8986 | .byte 255,224 // jmpq *%rax |
| 8987 | |
| 8988 | HIDDEN _sk_dstout_sse41 |
| 8989 | .globl _sk_dstout_sse41 |
| 8990 | _sk_dstout_sse41: |
| 8991 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 8992 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 8993 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 8994 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 8995 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 8996 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 8997 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 8998 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 8999 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9000 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9001 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9002 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9003 | .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| 9004 | .byte 255,224 // jmpq *%rax |
| 9005 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9006 | HIDDEN _sk_srcover_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9007 | .globl _sk_srcover_sse41 |
| 9008 | _sk_srcover_sse41: |
| 9009 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9010 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9011 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9012 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 9013 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9014 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 9015 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 9016 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9017 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9018 | .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| 9019 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9020 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9021 | .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| 9022 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9023 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| 9024 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9025 | .byte 255,224 // jmpq *%rax |
| 9026 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9027 | HIDDEN _sk_dstover_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9028 | .globl _sk_dstover_sse41 |
| 9029 | _sk_dstover_sse41: |
| 9030 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9031 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9032 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9033 | .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| 9034 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 9035 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9036 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 9037 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9038 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 9039 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 9040 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 9041 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 9042 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9043 | .byte 255,224 // jmpq *%rax |
| 9044 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 9045 | HIDDEN _sk_modulate_sse41 |
| 9046 | .globl _sk_modulate_sse41 |
| 9047 | _sk_modulate_sse41: |
| 9048 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 9049 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9050 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9051 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 9052 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9053 | .byte 255,224 // jmpq *%rax |
| 9054 | |
| 9055 | HIDDEN _sk_multiply_sse41 |
| 9056 | .globl _sk_multiply_sse41 |
| 9057 | _sk_multiply_sse41: |
| 9058 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9059 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9060 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9061 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9062 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 9063 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 9064 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 9065 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 9066 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 9067 | .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| 9068 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 9069 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 9070 | .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| 9071 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 9072 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 9073 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 9074 | .byte 68,15,89,221 // mulps %xmm5,%xmm11 |
| 9075 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 9076 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9077 | .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| 9078 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 9079 | .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| 9080 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 9081 | .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| 9082 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 9083 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9084 | .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| 9085 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 9086 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9087 | .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| 9088 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 9089 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| 9090 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9091 | .byte 255,224 // jmpq *%rax |
| 9092 | |
| 9093 | HIDDEN _sk_plus__sse41 |
| 9094 | .globl _sk_plus__sse41 |
| 9095 | _sk_plus__sse41: |
| 9096 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9097 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9098 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 9099 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 9100 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9101 | .byte 255,224 // jmpq *%rax |
| 9102 | |
| 9103 | HIDDEN _sk_screen_sse41 |
| 9104 | .globl _sk_screen_sse41 |
| 9105 | _sk_screen_sse41: |
| 9106 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 9107 | .byte 68,15,88,196 // addps %xmm4,%xmm8 |
| 9108 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 9109 | .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| 9110 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 9111 | .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| 9112 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9113 | .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| 9114 | .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| 9115 | .byte 68,15,88,214 // addps %xmm6,%xmm10 |
| 9116 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9117 | .byte 68,15,92,210 // subps %xmm2,%xmm10 |
| 9118 | .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| 9119 | .byte 68,15,88,223 // addps %xmm7,%xmm11 |
| 9120 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 9121 | .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| 9122 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9123 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9124 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 9125 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9126 | .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| 9127 | .byte 255,224 // jmpq *%rax |
| 9128 | |
| 9129 | HIDDEN _sk_xor__sse41 |
| 9130 | .globl _sk_xor__sse41 |
| 9131 | _sk_xor__sse41: |
| 9132 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 9133 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9134 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 9135 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 9136 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9137 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 9138 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 9139 | .byte 65,15,92,216 // subps %xmm8,%xmm3 |
| 9140 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 9141 | .byte 68,15,89,212 // mulps %xmm4,%xmm10 |
| 9142 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 9143 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 9144 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 9145 | .byte 68,15,89,213 // mulps %xmm5,%xmm10 |
| 9146 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 9147 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 9148 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 9149 | .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| 9150 | .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| 9151 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 9152 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 9153 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 9154 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9155 | .byte 255,224 // jmpq *%rax |
| 9156 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 9157 | HIDDEN _sk_darken_sse41 |
| 9158 | .globl _sk_darken_sse41 |
| 9159 | _sk_darken_sse41: |
| 9160 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 9161 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 9162 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9163 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 9164 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 9165 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 9166 | .byte 68,15,95,201 // maxps %xmm1,%xmm9 |
| 9167 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 9168 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 9169 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9170 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9171 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9172 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9173 | .byte 69,15,95,193 // maxps %xmm9,%xmm8 |
| 9174 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 9175 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 9176 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 9177 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9178 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9179 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9180 | .byte 65,15,95,209 // maxps %xmm9,%xmm2 |
| 9181 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 9182 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9183 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 9184 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 9185 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9186 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9187 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 9188 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9189 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9190 | .byte 255,224 // jmpq *%rax |
| 9191 | |
| 9192 | HIDDEN _sk_lighten_sse41 |
| 9193 | .globl _sk_lighten_sse41 |
| 9194 | _sk_lighten_sse41: |
| 9195 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 9196 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 9197 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9198 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 9199 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 9200 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 9201 | .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| 9202 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 9203 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 9204 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9205 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9206 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9207 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9208 | .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| 9209 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 9210 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 9211 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 9212 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9213 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9214 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9215 | .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| 9216 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 9217 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9218 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 9219 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 9220 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9221 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9222 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 9223 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9224 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9225 | .byte 255,224 // jmpq *%rax |
| 9226 | |
| 9227 | HIDDEN _sk_difference_sse41 |
| 9228 | .globl _sk_difference_sse41 |
| 9229 | _sk_difference_sse41: |
| 9230 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 9231 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 9232 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9233 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 9234 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 9235 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 9236 | .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| 9237 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 9238 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 9239 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 9240 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9241 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 9242 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9243 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9244 | .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| 9245 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 9246 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 9247 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 9248 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 9249 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9250 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9251 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9252 | .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| 9253 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 9254 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 9255 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9256 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 9257 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 9258 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9259 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9260 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 9261 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9262 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9263 | .byte 255,224 // jmpq *%rax |
| 9264 | |
| 9265 | HIDDEN _sk_exclusion_sse41 |
| 9266 | .globl _sk_exclusion_sse41 |
| 9267 | _sk_exclusion_sse41: |
| 9268 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 9269 | .byte 15,40,200 // movaps %xmm0,%xmm1 |
| 9270 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 9271 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 9272 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 9273 | .byte 15,92,193 // subps %xmm1,%xmm0 |
| 9274 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 9275 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 9276 | .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| 9277 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 9278 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 9279 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 9280 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 9281 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9282 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 9283 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 9284 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9285 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 9286 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 9287 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9288 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9289 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 9290 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9291 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9292 | .byte 255,224 // jmpq *%rax |
| 9293 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9294 | HIDDEN _sk_colorburn_sse41 |
| 9295 | .globl _sk_colorburn_sse41 |
| 9296 | _sk_colorburn_sse41: |
| 9297 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 9298 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9299 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 9300 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 9301 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 9302 | .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| 9303 | .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| 9304 | .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| 9305 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 9306 | .byte 15,40,199 // movaps %xmm7,%xmm0 |
| 9307 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 9308 | .byte 15,89,195 // mulps %xmm3,%xmm0 |
| 9309 | .byte 65,15,94,193 // divps %xmm9,%xmm0 |
| 9310 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9311 | .byte 68,15,93,232 // minps %xmm0,%xmm13 |
| 9312 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 9313 | .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| 9314 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9315 | .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 |
| 9316 | .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| 9317 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 9318 | .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| 9319 | .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14 |
| 9320 | .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| 9321 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 9322 | .byte 68,15,88,228 // addps %xmm4,%xmm12 |
| 9323 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9324 | .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| 9325 | .byte 69,15,88,206 // addps %xmm14,%xmm9 |
| 9326 | .byte 102,69,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm9 |
| 9327 | .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| 9328 | .byte 68,15,89,225 // mulps %xmm1,%xmm12 |
| 9329 | .byte 15,40,199 // movaps %xmm7,%xmm0 |
| 9330 | .byte 15,92,197 // subps %xmm5,%xmm0 |
| 9331 | .byte 15,89,195 // mulps %xmm3,%xmm0 |
| 9332 | .byte 15,94,193 // divps %xmm1,%xmm0 |
| 9333 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9334 | .byte 68,15,93,232 // minps %xmm0,%xmm13 |
| 9335 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 9336 | .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| 9337 | .byte 15,40,193 // movaps %xmm1,%xmm0 |
| 9338 | .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0 |
| 9339 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 9340 | .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| 9341 | .byte 102,68,15,56,20,241 // blendvps %xmm0,%xmm1,%xmm14 |
| 9342 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 9343 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9344 | .byte 68,15,88,229 // addps %xmm5,%xmm12 |
| 9345 | .byte 65,15,88,206 // addps %xmm14,%xmm1 |
| 9346 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9347 | .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| 9348 | .byte 102,65,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm1 |
| 9349 | .byte 15,40,199 // movaps %xmm7,%xmm0 |
| 9350 | .byte 15,92,198 // subps %xmm6,%xmm0 |
| 9351 | .byte 15,89,195 // mulps %xmm3,%xmm0 |
| 9352 | .byte 15,94,194 // divps %xmm2,%xmm0 |
| 9353 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 9354 | .byte 68,15,93,224 // minps %xmm0,%xmm12 |
| 9355 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9356 | .byte 69,15,92,236 // subps %xmm12,%xmm13 |
| 9357 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 9358 | .byte 68,15,194,194,0 // cmpeqps %xmm2,%xmm8 |
| 9359 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 9360 | .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| 9361 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9362 | .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13 |
| 9363 | .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| 9364 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9365 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9366 | .byte 65,15,88,213 // addps %xmm13,%xmm2 |
| 9367 | .byte 15,40,198 // movaps %xmm6,%xmm0 |
| 9368 | .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0 |
| 9369 | .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2 |
| 9370 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 9371 | .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| 9372 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9373 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9374 | .byte 255,224 // jmpq *%rax |
| 9375 | |
| 9376 | HIDDEN _sk_colordodge_sse41 |
| 9377 | .globl _sk_colordodge_sse41 |
| 9378 | _sk_colordodge_sse41: |
| 9379 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 9380 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9381 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 9382 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 9383 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 9384 | .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| 9385 | .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| 9386 | .byte 69,15,89,224 // mulps %xmm8,%xmm12 |
| 9387 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 9388 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 9389 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 9390 | .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| 9391 | .byte 68,15,94,200 // divps %xmm0,%xmm9 |
| 9392 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9393 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 9394 | .byte 69,15,93,241 // minps %xmm9,%xmm14 |
| 9395 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9396 | .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| 9397 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 9398 | .byte 69,15,88,244 // addps %xmm12,%xmm14 |
| 9399 | .byte 102,69,15,56,20,240 // blendvps %xmm0,%xmm8,%xmm14 |
| 9400 | .byte 69,15,87,201 // xorps %xmm9,%xmm9 |
| 9401 | .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| 9402 | .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| 9403 | .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| 9404 | .byte 68,15,88,228 // addps %xmm4,%xmm12 |
| 9405 | .byte 69,15,88,198 // addps %xmm14,%xmm8 |
| 9406 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9407 | .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0 |
| 9408 | .byte 102,69,15,56,20,196 // blendvps %xmm0,%xmm12,%xmm8 |
| 9409 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 9410 | .byte 68,15,89,229 // mulps %xmm5,%xmm12 |
| 9411 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 9412 | .byte 15,92,193 // subps %xmm1,%xmm0 |
| 9413 | .byte 68,15,94,224 // divps %xmm0,%xmm12 |
| 9414 | .byte 69,15,40,243 // movaps %xmm11,%xmm14 |
| 9415 | .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| 9416 | .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| 9417 | .byte 15,40,193 // movaps %xmm1,%xmm0 |
| 9418 | .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| 9419 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 9420 | .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| 9421 | .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13 |
| 9422 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 9423 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9424 | .byte 68,15,88,245 // addps %xmm5,%xmm14 |
| 9425 | .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| 9426 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9427 | .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0 |
| 9428 | .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1 |
| 9429 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 9430 | .byte 68,15,89,230 // mulps %xmm6,%xmm12 |
| 9431 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 9432 | .byte 15,92,194 // subps %xmm2,%xmm0 |
| 9433 | .byte 68,15,94,224 // divps %xmm0,%xmm12 |
| 9434 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9435 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 9436 | .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| 9437 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9438 | .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0 |
| 9439 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 9440 | .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| 9441 | .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13 |
| 9442 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9443 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9444 | .byte 65,15,88,213 // addps %xmm13,%xmm2 |
| 9445 | .byte 68,15,194,206,0 // cmpeqps %xmm6,%xmm9 |
| 9446 | .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| 9447 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9448 | .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2 |
| 9449 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 9450 | .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| 9451 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9452 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9453 | .byte 255,224 // jmpq *%rax |
| 9454 | |
| 9455 | HIDDEN _sk_hardlight_sse41 |
| 9456 | .globl _sk_hardlight_sse41 |
| 9457 | _sk_hardlight_sse41: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9458 | .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9459 | .byte 68,15,40,229 // movaps %xmm5,%xmm12 |
| 9460 | .byte 15,40,244 // movaps %xmm4,%xmm6 |
| 9461 | .byte 15,40,227 // movaps %xmm3,%xmm4 |
| 9462 | .byte 15,40,234 // movaps %xmm2,%xmm5 |
| 9463 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 9464 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9465 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 9466 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 9467 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9468 | .byte 15,92,215 // subps %xmm7,%xmm2 |
| 9469 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9470 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 9471 | .byte 68,15,92,212 // subps %xmm4,%xmm10 |
| 9472 | .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| 9473 | .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| 9474 | .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| 9475 | .byte 68,15,40,252 // movaps %xmm4,%xmm15 |
| 9476 | .byte 69,15,92,249 // subps %xmm9,%xmm15 |
| 9477 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 9478 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9479 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 9480 | .byte 15,40,199 // movaps %xmm7,%xmm0 |
| 9481 | .byte 15,92,198 // subps %xmm6,%xmm0 |
| 9482 | .byte 65,15,89,199 // mulps %xmm15,%xmm0 |
| 9483 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9484 | .byte 68,15,40,251 // movaps %xmm3,%xmm15 |
| 9485 | .byte 68,15,92,248 // subps %xmm0,%xmm15 |
| 9486 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9487 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9488 | .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| 9489 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9490 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 9491 | .byte 102,69,15,56,20,249 // blendvps %xmm0,%xmm9,%xmm15 |
| 9492 | .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| 9493 | .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| 9494 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9495 | .byte 15,92,193 // subps %xmm1,%xmm0 |
| 9496 | .byte 69,15,40,204 // movaps %xmm12,%xmm9 |
| 9497 | .byte 69,15,92,233 // subps %xmm9,%xmm13 |
| 9498 | .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| 9499 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 9500 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 9501 | .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| 9502 | .byte 15,40,193 // movaps %xmm1,%xmm0 |
| 9503 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9504 | .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| 9505 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 9506 | .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| 9507 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 9508 | .byte 102,68,15,56,20,225 // blendvps %xmm0,%xmm1,%xmm12 |
| 9509 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 9510 | .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| 9511 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 9512 | .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| 9513 | .byte 65,15,89,205 // mulps %xmm13,%xmm1 |
| 9514 | .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| 9515 | .byte 65,15,88,204 // addps %xmm12,%xmm1 |
| 9516 | .byte 15,89,213 // mulps %xmm5,%xmm2 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9517 | .byte 68,15,40,92,36,232 // movaps -0x18(%rsp),%xmm11 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9518 | .byte 69,15,89,203 // mulps %xmm11,%xmm9 |
| 9519 | .byte 68,15,88,202 // addps %xmm2,%xmm9 |
| 9520 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9521 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9522 | .byte 15,194,196,2 // cmpleps %xmm4,%xmm0 |
| 9523 | .byte 15,40,212 // movaps %xmm4,%xmm2 |
| 9524 | .byte 15,92,213 // subps %xmm5,%xmm2 |
| 9525 | .byte 65,15,89,235 // mulps %xmm11,%xmm5 |
| 9526 | .byte 15,88,237 // addps %xmm5,%xmm5 |
| 9527 | .byte 69,15,92,243 // subps %xmm11,%xmm14 |
| 9528 | .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| 9529 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 9530 | .byte 65,15,92,222 // subps %xmm14,%xmm3 |
| 9531 | .byte 102,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm3 |
| 9532 | .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| 9533 | .byte 65,15,88,226 // addps %xmm10,%xmm4 |
| 9534 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9535 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9536 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| 9537 | .byte 15,40,220 // movaps %xmm4,%xmm3 |
| 9538 | .byte 15,40,230 // movaps %xmm6,%xmm4 |
| 9539 | .byte 65,15,40,237 // movaps %xmm13,%xmm5 |
| 9540 | .byte 65,15,40,243 // movaps %xmm11,%xmm6 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9541 | .byte 255,224 // jmpq *%rax |
| 9542 | |
| 9543 | HIDDEN _sk_overlay_sse41 |
| 9544 | .globl _sk_overlay_sse41 |
| 9545 | _sk_overlay_sse41: |
| 9546 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 9547 | .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| 9548 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9549 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 9550 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 9551 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 9552 | .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| 9553 | .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| 9554 | .byte 65,15,89,198 // mulps %xmm14,%xmm0 |
| 9555 | .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| 9556 | .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| 9557 | .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| 9558 | .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| 9559 | .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| 9560 | .byte 69,15,92,238 // subps %xmm14,%xmm13 |
| 9561 | .byte 68,15,89,244 // mulps %xmm4,%xmm14 |
| 9562 | .byte 15,40,207 // movaps %xmm7,%xmm1 |
| 9563 | .byte 15,92,204 // subps %xmm4,%xmm1 |
| 9564 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9565 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9566 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9567 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 9568 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 9569 | .byte 68,15,89,231 // mulps %xmm7,%xmm12 |
| 9570 | .byte 65,15,89,205 // mulps %xmm13,%xmm1 |
| 9571 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 9572 | .byte 69,15,40,236 // movaps %xmm12,%xmm13 |
| 9573 | .byte 68,15,92,233 // subps %xmm1,%xmm13 |
| 9574 | .byte 102,69,15,56,20,238 // blendvps %xmm0,%xmm14,%xmm13 |
| 9575 | .byte 69,15,88,197 // addps %xmm13,%xmm8 |
| 9576 | .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| 9577 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 9578 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 9579 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9580 | .byte 15,88,200 // addps %xmm0,%xmm1 |
| 9581 | .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| 9582 | .byte 69,15,92,233 // subps %xmm9,%xmm13 |
| 9583 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9584 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 9585 | .byte 68,15,92,245 // subps %xmm5,%xmm14 |
| 9586 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9587 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9588 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9589 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 9590 | .byte 69,15,89,245 // mulps %xmm13,%xmm14 |
| 9591 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 9592 | .byte 69,15,40,236 // movaps %xmm12,%xmm13 |
| 9593 | .byte 69,15,92,238 // subps %xmm14,%xmm13 |
| 9594 | .byte 102,69,15,56,20,233 // blendvps %xmm0,%xmm9,%xmm13 |
| 9595 | .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| 9596 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 9597 | .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| 9598 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 9599 | .byte 69,15,88,203 // addps %xmm11,%xmm9 |
| 9600 | .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| 9601 | .byte 68,15,92,218 // subps %xmm2,%xmm11 |
| 9602 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9603 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 9604 | .byte 68,15,92,238 // subps %xmm6,%xmm13 |
| 9605 | .byte 15,40,198 // movaps %xmm6,%xmm0 |
| 9606 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9607 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9608 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 9609 | .byte 69,15,89,235 // mulps %xmm11,%xmm13 |
| 9610 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 9611 | .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| 9612 | .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12 |
| 9613 | .byte 69,15,88,204 // addps %xmm12,%xmm9 |
| 9614 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 9615 | .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| 9616 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9617 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9618 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| 9619 | .byte 255,224 // jmpq *%rax |
| 9620 | |
| 9621 | HIDDEN _sk_softlight_sse41 |
| 9622 | .globl _sk_softlight_sse41 |
| 9623 | _sk_softlight_sse41: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9624 | .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9625 | .byte 15,40,244 // movaps %xmm4,%xmm6 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9626 | .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9627 | .byte 68,15,40,225 // movaps %xmm1,%xmm12 |
| 9628 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 9629 | .byte 15,87,228 // xorps %xmm4,%xmm4 |
| 9630 | .byte 15,194,231,1 // cmpltps %xmm7,%xmm4 |
| 9631 | .byte 15,40,198 // movaps %xmm6,%xmm0 |
| 9632 | .byte 15,94,199 // divps %xmm7,%xmm0 |
| 9633 | .byte 15,84,196 // andps %xmm4,%xmm0 |
| 9634 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9635 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 9636 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 9637 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| 9638 | .byte 68,15,92,208 // subps %xmm0,%xmm10 |
| 9639 | .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| 9640 | .byte 68,15,40,248 // movaps %xmm0,%xmm15 |
| 9641 | .byte 15,82,208 // rsqrtps %xmm0,%xmm2 |
| 9642 | .byte 68,15,83,218 // rcpps %xmm2,%xmm11 |
| 9643 | .byte 68,15,92,216 // subps %xmm0,%xmm11 |
| 9644 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9645 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9646 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 9647 | .byte 15,89,210 // mulps %xmm2,%xmm2 |
| 9648 | .byte 15,88,208 // addps %xmm0,%xmm2 |
| 9649 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9650 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 9651 | .byte 68,15,92,241 // subps %xmm1,%xmm14 |
| 9652 | .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| 9653 | .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| 9654 | .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| 9655 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 9656 | .byte 69,15,89,253 // mulps %xmm13,%xmm15 |
| 9657 | .byte 69,15,88,254 // addps %xmm14,%xmm15 |
| 9658 | .byte 15,40,198 // movaps %xmm6,%xmm0 |
| 9659 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9660 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9661 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9662 | .byte 102,69,15,56,20,223 // blendvps %xmm0,%xmm15,%xmm11 |
| 9663 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9664 | .byte 15,92,195 // subps %xmm3,%xmm0 |
| 9665 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 9666 | .byte 68,15,88,211 // addps %xmm3,%xmm10 |
| 9667 | .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| 9668 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 9669 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 9670 | .byte 15,89,199 // mulps %xmm7,%xmm0 |
| 9671 | .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| 9672 | .byte 68,15,88,218 // addps %xmm2,%xmm11 |
| 9673 | .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9 |
| 9674 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9675 | .byte 102,69,15,56,20,218 // blendvps %xmm0,%xmm10,%xmm11 |
| 9676 | .byte 68,15,40,213 // movaps %xmm5,%xmm10 |
| 9677 | .byte 68,15,94,215 // divps %xmm7,%xmm10 |
| 9678 | .byte 68,15,84,212 // andps %xmm4,%xmm10 |
| 9679 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 9680 | .byte 69,15,92,202 // subps %xmm10,%xmm9 |
| 9681 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9682 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 9683 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 9684 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9685 | .byte 15,89,192 // mulps %xmm0,%xmm0 |
| 9686 | .byte 15,88,194 // addps %xmm2,%xmm0 |
| 9687 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 9688 | .byte 15,92,209 // subps %xmm1,%xmm2 |
| 9689 | .byte 15,89,208 // mulps %xmm0,%xmm2 |
| 9690 | .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0 |
| 9691 | .byte 68,15,83,240 // rcpps %xmm0,%xmm14 |
| 9692 | .byte 69,15,92,242 // subps %xmm10,%xmm14 |
| 9693 | .byte 69,15,89,213 // mulps %xmm13,%xmm10 |
| 9694 | .byte 68,15,88,210 // addps %xmm2,%xmm10 |
| 9695 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9696 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9697 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9698 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9699 | .byte 102,69,15,56,20,242 // blendvps %xmm0,%xmm10,%xmm14 |
| 9700 | .byte 65,15,40,196 // movaps %xmm12,%xmm0 |
| 9701 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9702 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 9703 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9704 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 9705 | .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9706 | .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9707 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 9708 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9709 | .byte 68,15,89,242 // mulps %xmm2,%xmm14 |
| 9710 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 9711 | .byte 15,89,213 // mulps %xmm5,%xmm2 |
| 9712 | .byte 68,15,88,242 // addps %xmm2,%xmm14 |
| 9713 | .byte 68,15,40,249 // movaps %xmm1,%xmm15 |
| 9714 | .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| 9715 | .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14 |
| 9716 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9717 | .byte 15,40,108,36,216 // movaps -0x28(%rsp),%xmm5 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9718 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9719 | .byte 15,94,199 // divps %xmm7,%xmm0 |
| 9720 | .byte 15,84,196 // andps %xmm4,%xmm0 |
| 9721 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 9722 | .byte 15,92,209 // subps %xmm1,%xmm2 |
| 9723 | .byte 15,92,200 // subps %xmm0,%xmm1 |
| 9724 | .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| 9725 | .byte 15,82,224 // rsqrtps %xmm0,%xmm4 |
| 9726 | .byte 68,15,83,204 // rcpps %xmm4,%xmm9 |
| 9727 | .byte 68,15,92,200 // subps %xmm0,%xmm9 |
| 9728 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9729 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9730 | .byte 15,40,224 // movaps %xmm0,%xmm4 |
| 9731 | .byte 15,89,228 // mulps %xmm4,%xmm4 |
| 9732 | .byte 15,88,224 // addps %xmm0,%xmm4 |
| 9733 | .byte 15,89,226 // mulps %xmm2,%xmm4 |
| 9734 | .byte 68,15,88,236 // addps %xmm4,%xmm13 |
| 9735 | .byte 15,40,197 // movaps %xmm5,%xmm0 |
| 9736 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9737 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9738 | .byte 15,194,199,2 // cmpleps %xmm7,%xmm0 |
| 9739 | .byte 102,69,15,56,20,205 // blendvps %xmm0,%xmm13,%xmm9 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9740 | .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9741 | .byte 65,15,40,197 // movaps %xmm13,%xmm0 |
| 9742 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 9743 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 9744 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 9745 | .byte 15,89,202 // mulps %xmm2,%xmm1 |
| 9746 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 9747 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 9748 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 9749 | .byte 15,89,213 // mulps %xmm5,%xmm2 |
| 9750 | .byte 68,15,88,202 // addps %xmm2,%xmm9 |
| 9751 | .byte 15,88,203 // addps %xmm3,%xmm1 |
| 9752 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 9753 | .byte 15,40,213 // movaps %xmm5,%xmm2 |
| 9754 | .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| 9755 | .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9 |
| 9756 | .byte 68,15,92,255 // subps %xmm7,%xmm15 |
| 9757 | .byte 69,15,89,199 // mulps %xmm15,%xmm8 |
| 9758 | .byte 69,15,89,231 // mulps %xmm15,%xmm12 |
| 9759 | .byte 69,15,89,253 // mulps %xmm13,%xmm15 |
| 9760 | .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| 9761 | .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| 9762 | .byte 15,89,198 // mulps %xmm6,%xmm0 |
| 9763 | .byte 68,15,88,192 // addps %xmm0,%xmm8 |
| 9764 | .byte 69,15,88,195 // addps %xmm11,%xmm8 |
| 9765 | .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9766 | .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9767 | .byte 15,89,197 // mulps %xmm5,%xmm0 |
| 9768 | .byte 68,15,88,224 // addps %xmm0,%xmm12 |
| 9769 | .byte 69,15,88,230 // addps %xmm14,%xmm12 |
| 9770 | .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| 9771 | .byte 15,89,194 // mulps %xmm2,%xmm0 |
| 9772 | .byte 65,15,88,199 // addps %xmm15,%xmm0 |
| 9773 | .byte 68,15,88,200 // addps %xmm0,%xmm9 |
| 9774 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 9775 | .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| 9776 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9777 | .byte 15,40,230 // movaps %xmm6,%xmm4 |
| 9778 | .byte 15,40,242 // movaps %xmm2,%xmm6 |
| 9779 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 9780 | .byte 65,15,40,204 // movaps %xmm12,%xmm1 |
| 9781 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 9782 | .byte 255,224 // jmpq *%rax |
| 9783 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9784 | HIDDEN _sk_clamp_0_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9785 | .globl _sk_clamp_0_sse41 |
| 9786 | _sk_clamp_0_sse41: |
| 9787 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 9788 | .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| 9789 | .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| 9790 | .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| 9791 | .byte 65,15,95,216 // maxps %xmm8,%xmm3 |
| 9792 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9793 | .byte 255,224 // jmpq *%rax |
| 9794 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9795 | HIDDEN _sk_clamp_1_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9796 | .globl _sk_clamp_1_sse41 |
| 9797 | _sk_clamp_1_sse41: |
| 9798 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9799 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9800 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9801 | .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| 9802 | .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| 9803 | .byte 65,15,93,208 // minps %xmm8,%xmm2 |
| 9804 | .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| 9805 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9806 | .byte 255,224 // jmpq *%rax |
| 9807 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9808 | HIDDEN _sk_clamp_a_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9809 | .globl _sk_clamp_a_sse41 |
| 9810 | _sk_clamp_a_sse41: |
| 9811 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9812 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9813 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9814 | .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| 9815 | .byte 15,93,195 // minps %xmm3,%xmm0 |
| 9816 | .byte 15,93,203 // minps %xmm3,%xmm1 |
| 9817 | .byte 15,93,211 // minps %xmm3,%xmm2 |
| 9818 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9819 | .byte 255,224 // jmpq *%rax |
| 9820 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9821 | HIDDEN _sk_set_rgb_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9822 | .globl _sk_set_rgb_sse41 |
| 9823 | _sk_set_rgb_sse41: |
| 9824 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9825 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 9826 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 9827 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 9828 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 9829 | .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| 9830 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 9831 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9832 | .byte 255,224 // jmpq *%rax |
| 9833 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9834 | HIDDEN _sk_swap_rb_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9835 | .globl _sk_swap_rb_sse41 |
| 9836 | _sk_swap_rb_sse41: |
| 9837 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 9838 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9839 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9840 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9841 | .byte 255,224 // jmpq *%rax |
| 9842 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9843 | HIDDEN _sk_swap_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9844 | .globl _sk_swap_sse41 |
| 9845 | _sk_swap_sse41: |
| 9846 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 9847 | .byte 68,15,40,202 // movaps %xmm2,%xmm9 |
| 9848 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| 9849 | .byte 68,15,40,216 // movaps %xmm0,%xmm11 |
| 9850 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9851 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9852 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 9853 | .byte 15,40,214 // movaps %xmm6,%xmm2 |
| 9854 | .byte 15,40,223 // movaps %xmm7,%xmm3 |
| 9855 | .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| 9856 | .byte 65,15,40,234 // movaps %xmm10,%xmm5 |
| 9857 | .byte 65,15,40,241 // movaps %xmm9,%xmm6 |
| 9858 | .byte 65,15,40,248 // movaps %xmm8,%xmm7 |
| 9859 | .byte 255,224 // jmpq *%rax |
| 9860 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9861 | HIDDEN _sk_move_src_dst_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9862 | .globl _sk_move_src_dst_sse41 |
| 9863 | _sk_move_src_dst_sse41: |
| 9864 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9865 | .byte 15,40,224 // movaps %xmm0,%xmm4 |
| 9866 | .byte 15,40,233 // movaps %xmm1,%xmm5 |
| 9867 | .byte 15,40,242 // movaps %xmm2,%xmm6 |
| 9868 | .byte 15,40,251 // movaps %xmm3,%xmm7 |
| 9869 | .byte 255,224 // jmpq *%rax |
| 9870 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9871 | HIDDEN _sk_move_dst_src_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9872 | .globl _sk_move_dst_src_sse41 |
| 9873 | _sk_move_dst_src_sse41: |
| 9874 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9875 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 9876 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 9877 | .byte 15,40,214 // movaps %xmm6,%xmm2 |
| 9878 | .byte 15,40,223 // movaps %xmm7,%xmm3 |
| 9879 | .byte 255,224 // jmpq *%rax |
| 9880 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9881 | HIDDEN _sk_premul_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9882 | .globl _sk_premul_sse41 |
| 9883 | _sk_premul_sse41: |
| 9884 | .byte 15,89,195 // mulps %xmm3,%xmm0 |
| 9885 | .byte 15,89,203 // mulps %xmm3,%xmm1 |
| 9886 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 9887 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9888 | .byte 255,224 // jmpq *%rax |
| 9889 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9890 | HIDDEN _sk_unpremul_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9891 | .globl _sk_unpremul_sse41 |
| 9892 | _sk_unpremul_sse41: |
| 9893 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 9894 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9895 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 9896 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 9897 | .byte 68,15,94,203 // divps %xmm3,%xmm9 |
| 9898 | .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 |
| 9899 | .byte 69,15,84,193 // andps %xmm9,%xmm8 |
| 9900 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 9901 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 9902 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 9903 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9904 | .byte 255,224 // jmpq *%rax |
| 9905 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9906 | HIDDEN _sk_from_srgb_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9907 | .globl _sk_from_srgb_sse41 |
| 9908 | _sk_from_srgb_sse41: |
| 9909 | .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| 9910 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 9911 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 9912 | .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| 9913 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 9914 | .byte 68,15,40,240 // movaps %xmm0,%xmm14 |
| 9915 | .byte 69,15,89,246 // mulps %xmm14,%xmm14 |
| 9916 | .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| 9917 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9918 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9919 | .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| 9920 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 9921 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 9922 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9923 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 9924 | .byte 69,15,88,204 // addps %xmm12,%xmm9 |
| 9925 | .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| 9926 | .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| 9927 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 9928 | .byte 69,15,89,206 // mulps %xmm14,%xmm9 |
| 9929 | .byte 69,15,88,205 // addps %xmm13,%xmm9 |
| 9930 | .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| 9931 | .byte 102,68,15,110,240 // movd %eax,%xmm14 |
| 9932 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 9933 | .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0 |
| 9934 | .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 |
| 9935 | .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| 9936 | .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| 9937 | .byte 15,40,193 // movaps %xmm1,%xmm0 |
| 9938 | .byte 15,89,192 // mulps %xmm0,%xmm0 |
| 9939 | .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| 9940 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 9941 | .byte 69,15,88,212 // addps %xmm12,%xmm10 |
| 9942 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 9943 | .byte 69,15,88,213 // addps %xmm13,%xmm10 |
| 9944 | .byte 65,15,194,206,1 // cmpltps %xmm14,%xmm1 |
| 9945 | .byte 15,40,193 // movaps %xmm1,%xmm0 |
| 9946 | .byte 102,69,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm10 |
| 9947 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 9948 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9949 | .byte 15,89,192 // mulps %xmm0,%xmm0 |
| 9950 | .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| 9951 | .byte 69,15,88,196 // addps %xmm12,%xmm8 |
| 9952 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 9953 | .byte 69,15,88,197 // addps %xmm13,%xmm8 |
| 9954 | .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2 |
| 9955 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 9956 | .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8 |
| 9957 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 9958 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 9959 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 9960 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 9961 | .byte 255,224 // jmpq *%rax |
| 9962 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 9963 | HIDDEN _sk_to_srgb_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9964 | .globl _sk_to_srgb_sse41 |
| 9965 | _sk_to_srgb_sse41: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 9966 | .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp) |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 9967 | .byte 15,40,254 // movaps %xmm6,%xmm7 |
| 9968 | .byte 15,40,245 // movaps %xmm5,%xmm6 |
| 9969 | .byte 15,40,236 // movaps %xmm4,%xmm5 |
| 9970 | .byte 15,40,227 // movaps %xmm3,%xmm4 |
| 9971 | .byte 15,40,218 // movaps %xmm2,%xmm3 |
| 9972 | .byte 15,40,209 // movaps %xmm1,%xmm2 |
| 9973 | .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| 9974 | .byte 69,15,83,200 // rcpps %xmm8,%xmm9 |
| 9975 | .byte 69,15,82,248 // rsqrtps %xmm8,%xmm15 |
| 9976 | .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| 9977 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 9978 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 9979 | .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| 9980 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 9981 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 9982 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 9983 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 9984 | .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| 9985 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 9986 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 9987 | .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| 9988 | .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| 9989 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 9990 | .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| 9991 | .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| 9992 | .byte 102,68,15,110,240 // movd %eax,%xmm14 |
| 9993 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 9994 | .byte 69,15,89,205 // mulps %xmm13,%xmm9 |
| 9995 | .byte 69,15,88,206 // addps %xmm14,%xmm9 |
| 9996 | .byte 69,15,89,252 // mulps %xmm12,%xmm15 |
| 9997 | .byte 69,15,88,249 // addps %xmm9,%xmm15 |
| 9998 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 9999 | .byte 69,15,93,207 // minps %xmm15,%xmm9 |
| 10000 | .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| 10001 | .byte 102,68,15,110,248 // movd %eax,%xmm15 |
| 10002 | .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| 10003 | .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0 |
| 10004 | .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9 |
| 10005 | .byte 68,15,82,210 // rsqrtps %xmm2,%xmm10 |
| 10006 | .byte 65,15,83,194 // rcpps %xmm10,%xmm0 |
| 10007 | .byte 69,15,82,210 // rsqrtps %xmm10,%xmm10 |
| 10008 | .byte 65,15,89,197 // mulps %xmm13,%xmm0 |
| 10009 | .byte 65,15,88,198 // addps %xmm14,%xmm0 |
| 10010 | .byte 69,15,89,212 // mulps %xmm12,%xmm10 |
| 10011 | .byte 68,15,88,208 // addps %xmm0,%xmm10 |
| 10012 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 10013 | .byte 65,15,93,202 // minps %xmm10,%xmm1 |
| 10014 | .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| 10015 | .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| 10016 | .byte 65,15,194,215,1 // cmpltps %xmm15,%xmm2 |
| 10017 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 10018 | .byte 102,65,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm1 |
| 10019 | .byte 15,82,195 // rsqrtps %xmm3,%xmm0 |
| 10020 | .byte 15,83,208 // rcpps %xmm0,%xmm2 |
| 10021 | .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| 10022 | .byte 65,15,88,214 // addps %xmm14,%xmm2 |
| 10023 | .byte 15,82,192 // rsqrtps %xmm0,%xmm0 |
| 10024 | .byte 65,15,89,196 // mulps %xmm12,%xmm0 |
| 10025 | .byte 15,88,194 // addps %xmm2,%xmm0 |
| 10026 | .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| 10027 | .byte 68,15,89,219 // mulps %xmm3,%xmm11 |
| 10028 | .byte 65,15,194,223,1 // cmpltps %xmm15,%xmm3 |
| 10029 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 10030 | .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8 |
| 10031 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10032 | .byte 65,15,40,193 // movaps %xmm9,%xmm0 |
| 10033 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 10034 | .byte 15,40,220 // movaps %xmm4,%xmm3 |
| 10035 | .byte 15,40,229 // movaps %xmm5,%xmm4 |
| 10036 | .byte 15,40,238 // movaps %xmm6,%xmm5 |
| 10037 | .byte 15,40,247 // movaps %xmm7,%xmm6 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 10038 | .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10039 | .byte 255,224 // jmpq *%rax |
| 10040 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10041 | HIDDEN _sk_scale_1_float_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10042 | .globl _sk_scale_1_float_sse41 |
| 10043 | _sk_scale_1_float_sse41: |
| 10044 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10045 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10046 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10047 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10048 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10049 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 10050 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10051 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10052 | .byte 255,224 // jmpq *%rax |
| 10053 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10054 | HIDDEN _sk_scale_u8_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10055 | .globl _sk_scale_u8_sse41 |
| 10056 | _sk_scale_u8_sse41: |
| 10057 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10058 | .byte 72,139,0 // mov (%rax),%rax |
| 10059 | .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 |
| 10060 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 10061 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10062 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 10063 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10064 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 10065 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 10066 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 10067 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 10068 | .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| 10069 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10070 | .byte 255,224 // jmpq *%rax |
| 10071 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10072 | HIDDEN _sk_lerp_1_float_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10073 | .globl _sk_lerp_1_float_sse41 |
| 10074 | _sk_lerp_1_float_sse41: |
| 10075 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10076 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10077 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10078 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 10079 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10080 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 10081 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 10082 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10083 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 10084 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 10085 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 10086 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 10087 | .byte 15,92,223 // subps %xmm7,%xmm3 |
| 10088 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10089 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 10090 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10091 | .byte 255,224 // jmpq *%rax |
| 10092 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10093 | HIDDEN _sk_lerp_u8_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10094 | .globl _sk_lerp_u8_sse41 |
| 10095 | _sk_lerp_u8_sse41: |
| 10096 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10097 | .byte 72,139,0 // mov (%rax),%rax |
| 10098 | .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8 |
| 10099 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 10100 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10101 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 10102 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10103 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 10104 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 10105 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 10106 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 10107 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 10108 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 10109 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 10110 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 10111 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 10112 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 10113 | .byte 15,92,223 // subps %xmm7,%xmm3 |
| 10114 | .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| 10115 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 10116 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10117 | .byte 255,224 // jmpq *%rax |
| 10118 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10119 | HIDDEN _sk_lerp_565_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10120 | .globl _sk_lerp_565_sse41 |
| 10121 | _sk_lerp_565_sse41: |
| 10122 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10123 | .byte 72,139,0 // mov (%rax),%rax |
| 10124 | .byte 102,68,15,56,51,4,120 // pmovzxwd (%rax,%rdi,2),%xmm8 |
| 10125 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 10126 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10127 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 10128 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 10129 | .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| 10130 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 10131 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 10132 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10133 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10134 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 10135 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10136 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 10137 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 10138 | .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| 10139 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 10140 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 10141 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10142 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 10143 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 10144 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10145 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 10146 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 10147 | .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| 10148 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 10149 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10150 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10151 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10152 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 10153 | .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| 10154 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 10155 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 10156 | .byte 65,15,89,203 // mulps %xmm11,%xmm1 |
| 10157 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 10158 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 10159 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 10160 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 10161 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 10162 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10163 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10164 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10165 | .byte 255,224 // jmpq *%rax |
| 10166 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10167 | HIDDEN _sk_load_tables_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10168 | .globl _sk_load_tables_sse41 |
| 10169 | _sk_load_tables_sse41: |
| 10170 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10171 | .byte 72,139,8 // mov (%rax),%rcx |
| 10172 | .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| 10173 | .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8 |
| 10174 | .byte 185,255,0,0,0 // mov $0xff,%ecx |
| 10175 | .byte 102,15,110,193 // movd %ecx,%xmm0 |
| 10176 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 10177 | .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 |
| 10178 | .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| 10179 | .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| 10180 | .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2 |
| 10181 | .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| 10182 | .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| 10183 | .byte 102,65,15,219,192 // pand %xmm8,%xmm0 |
| 10184 | .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx |
| 10185 | .byte 65,137,201 // mov %ecx,%r9d |
| 10186 | .byte 72,193,233,32 // shr $0x20,%rcx |
| 10187 | .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| 10188 | .byte 69,137,211 // mov %r10d,%r11d |
| 10189 | .byte 73,193,234,32 // shr $0x20,%r10 |
| 10190 | .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| 10191 | .byte 102,67,15,58,33,4,144,16 // insertps $0x10,(%r8,%r10,4),%xmm0 |
| 10192 | .byte 102,67,15,58,33,4,136,32 // insertps $0x20,(%r8,%r9,4),%xmm0 |
| 10193 | .byte 102,65,15,58,33,4,136,48 // insertps $0x30,(%r8,%rcx,4),%xmm0 |
| 10194 | .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| 10195 | .byte 102,73,15,58,22,202,1 // pextrq $0x1,%xmm1,%r10 |
| 10196 | .byte 77,137,209 // mov %r10,%r9 |
| 10197 | .byte 73,193,233,32 // shr $0x20,%r9 |
| 10198 | .byte 102,72,15,126,201 // movq %xmm1,%rcx |
| 10199 | .byte 65,137,203 // mov %ecx,%r11d |
| 10200 | .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d |
| 10201 | .byte 72,193,233,30 // shr $0x1e,%rcx |
| 10202 | .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d |
| 10203 | .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1 |
| 10204 | .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1 |
| 10205 | .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3 |
| 10206 | .byte 102,15,58,33,203,32 // insertps $0x20,%xmm3,%xmm1 |
| 10207 | .byte 243,67,15,16,28,136 // movss (%r8,%r9,4),%xmm3 |
| 10208 | .byte 102,15,58,33,203,48 // insertps $0x30,%xmm3,%xmm1 |
| 10209 | .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| 10210 | .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx |
| 10211 | .byte 68,15,183,193 // movzwl %cx,%r8d |
| 10212 | .byte 72,193,233,32 // shr $0x20,%rcx |
| 10213 | .byte 102,72,15,126,208 // movq %xmm2,%rax |
| 10214 | .byte 68,15,183,208 // movzwl %ax,%r10d |
| 10215 | .byte 72,193,232,30 // shr $0x1e,%rax |
| 10216 | .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| 10217 | .byte 102,65,15,58,33,20,1,16 // insertps $0x10,(%r9,%rax,1),%xmm2 |
| 10218 | .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| 10219 | .byte 102,15,58,33,211,32 // insertps $0x20,%xmm3,%xmm2 |
| 10220 | .byte 243,65,15,16,28,137 // movss (%r9,%rcx,4),%xmm3 |
| 10221 | .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2 |
| 10222 | .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 |
| 10223 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 10224 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10225 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10226 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10227 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10228 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10229 | .byte 255,224 // jmpq *%rax |
| 10230 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10231 | HIDDEN _sk_load_a8_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10232 | .globl _sk_load_a8_sse41 |
| 10233 | _sk_load_a8_sse41: |
| 10234 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10235 | .byte 72,139,0 // mov (%rax),%rax |
| 10236 | .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0 |
| 10237 | .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| 10238 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10239 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10240 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10241 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 10242 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10243 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 10244 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 10245 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 10246 | .byte 255,224 // jmpq *%rax |
| 10247 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10248 | HIDDEN _sk_store_a8_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10249 | .globl _sk_store_a8_sse41 |
| 10250 | _sk_store_a8_sse41: |
| 10251 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10252 | .byte 72,139,0 // mov (%rax),%rax |
| 10253 | .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| 10254 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 10255 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10256 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 10257 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 10258 | .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| 10259 | .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8 |
| 10260 | .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1) |
| 10261 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10262 | .byte 255,224 // jmpq *%rax |
| 10263 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 10264 | HIDDEN _sk_load_g8_sse41 |
| 10265 | .globl _sk_load_g8_sse41 |
| 10266 | _sk_load_g8_sse41: |
| 10267 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10268 | .byte 72,139,0 // mov (%rax),%rax |
| 10269 | .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0 |
| 10270 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 10271 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10272 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10273 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10274 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 10275 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 10276 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10277 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10278 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10279 | .byte 15,40,200 // movaps %xmm0,%xmm1 |
| 10280 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 10281 | .byte 255,224 // jmpq *%rax |
| 10282 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10283 | HIDDEN _sk_load_565_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10284 | .globl _sk_load_565_sse41 |
| 10285 | _sk_load_565_sse41: |
| 10286 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10287 | .byte 72,139,0 // mov (%rax),%rax |
| 10288 | .byte 102,15,56,51,20,120 // pmovzxwd (%rax,%rdi,2),%xmm2 |
| 10289 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 10290 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10291 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 10292 | .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| 10293 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 10294 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 10295 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10296 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10297 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 10298 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 10299 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 10300 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 10301 | .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| 10302 | .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| 10303 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 10304 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 10305 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 10306 | .byte 15,89,203 // mulps %xmm3,%xmm1 |
| 10307 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 10308 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10309 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 10310 | .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| 10311 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 10312 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 10313 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 10314 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 10315 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 10316 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 10317 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10318 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10319 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10320 | .byte 255,224 // jmpq *%rax |
| 10321 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10322 | HIDDEN _sk_store_565_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10323 | .globl _sk_store_565_sse41 |
| 10324 | _sk_store_565_sse41: |
| 10325 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10326 | .byte 72,139,0 // mov (%rax),%rax |
| 10327 | .byte 185,0,0,248,65 // mov $0x41f80000,%ecx |
| 10328 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 10329 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10330 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10331 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 10332 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10333 | .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9 |
| 10334 | .byte 185,0,0,124,66 // mov $0x427c0000,%ecx |
| 10335 | .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| 10336 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10337 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 10338 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 10339 | .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 |
| 10340 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 10341 | .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| 10342 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 10343 | .byte 102,69,15,86,194 // orpd %xmm10,%xmm8 |
| 10344 | .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| 10345 | .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2) |
| 10346 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10347 | .byte 255,224 // jmpq *%rax |
| 10348 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 10349 | HIDDEN _sk_load_4444_sse41 |
| 10350 | .globl _sk_load_4444_sse41 |
| 10351 | _sk_load_4444_sse41: |
| 10352 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10353 | .byte 72,139,0 // mov (%rax),%rax |
| 10354 | .byte 102,68,15,56,51,12,120 // pmovzxwd (%rax,%rdi,2),%xmm9 |
| 10355 | .byte 184,0,240,0,0 // mov $0xf000,%eax |
| 10356 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10357 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 10358 | .byte 102,65,15,219,193 // pand %xmm9,%xmm0 |
| 10359 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 10360 | .byte 184,137,136,136,55 // mov $0x37888889,%eax |
| 10361 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10362 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10363 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 10364 | .byte 184,0,15,0,0 // mov $0xf00,%eax |
| 10365 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 10366 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 10367 | .byte 102,65,15,219,201 // pand %xmm9,%xmm1 |
| 10368 | .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2 |
| 10369 | .byte 184,137,136,136,57 // mov $0x39888889,%eax |
| 10370 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 10371 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 10372 | .byte 15,89,202 // mulps %xmm2,%xmm1 |
| 10373 | .byte 184,240,0,0,0 // mov $0xf0,%eax |
| 10374 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 10375 | .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| 10376 | .byte 102,65,15,219,209 // pand %xmm9,%xmm2 |
| 10377 | .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8 |
| 10378 | .byte 184,137,136,136,59 // mov $0x3b888889,%eax |
| 10379 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 10380 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 10381 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 10382 | .byte 184,15,0,0,0 // mov $0xf,%eax |
| 10383 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10384 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 10385 | .byte 102,65,15,219,217 // pand %xmm9,%xmm3 |
| 10386 | .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| 10387 | .byte 184,137,136,136,61 // mov $0x3d888889,%eax |
| 10388 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10389 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10390 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10391 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10392 | .byte 255,224 // jmpq *%rax |
| 10393 | |
| 10394 | HIDDEN _sk_store_4444_sse41 |
| 10395 | .globl _sk_store_4444_sse41 |
| 10396 | _sk_store_4444_sse41: |
| 10397 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10398 | .byte 72,139,0 // mov (%rax),%rax |
| 10399 | .byte 185,0,0,112,65 // mov $0x41700000,%ecx |
| 10400 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 10401 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10402 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10403 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 10404 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10405 | .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9 |
| 10406 | .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| 10407 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 10408 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 10409 | .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10 |
| 10410 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 10411 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10412 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 10413 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10414 | .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9 |
| 10415 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 10416 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 10417 | .byte 102,69,15,86,193 // orpd %xmm9,%xmm8 |
| 10418 | .byte 102,69,15,86,194 // orpd %xmm10,%xmm8 |
| 10419 | .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| 10420 | .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2) |
| 10421 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10422 | .byte 255,224 // jmpq *%rax |
| 10423 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10424 | HIDDEN _sk_load_8888_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10425 | .globl _sk_load_8888_sse41 |
| 10426 | _sk_load_8888_sse41: |
| 10427 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10428 | .byte 72,139,0 // mov (%rax),%rax |
| 10429 | .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3 |
| 10430 | .byte 184,255,0,0,0 // mov $0xff,%eax |
| 10431 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10432 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 10433 | .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| 10434 | .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| 10435 | .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| 10436 | .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| 10437 | .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| 10438 | .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| 10439 | .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| 10440 | .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| 10441 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 10442 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 10443 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10444 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10445 | .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| 10446 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10447 | .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| 10448 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 10449 | .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| 10450 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 10451 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10452 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10453 | .byte 255,224 // jmpq *%rax |
| 10454 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10455 | HIDDEN _sk_store_8888_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10456 | .globl _sk_store_8888_sse41 |
| 10457 | _sk_store_8888_sse41: |
| 10458 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10459 | .byte 72,139,0 // mov (%rax),%rax |
| 10460 | .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| 10461 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 10462 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10463 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10464 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 10465 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10466 | .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| 10467 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 10468 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 10469 | .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10 |
| 10470 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 10471 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10472 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 10473 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10474 | .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| 10475 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 10476 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 10477 | .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8 |
| 10478 | .byte 102,69,15,235,193 // por %xmm9,%xmm8 |
| 10479 | .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| 10480 | .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4) |
| 10481 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10482 | .byte 255,224 // jmpq *%rax |
| 10483 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10484 | HIDDEN _sk_load_f16_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10485 | .globl _sk_load_f16_sse41 |
| 10486 | _sk_load_f16_sse41: |
| 10487 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10488 | .byte 72,139,0 // mov (%rax),%rax |
| 10489 | .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| 10490 | .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10491 | .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 |
| 10492 | .byte 102,68,15,97,193 // punpcklwd %xmm1,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10493 | .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10494 | .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 |
| 10495 | .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| 10496 | .byte 102,68,15,105,192 // punpckhwd %xmm0,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10497 | .byte 184,0,4,0,4 // mov $0x4000400,%eax |
| 10498 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10499 | .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10500 | .byte 102,15,111,195 // movdqa %xmm3,%xmm0 |
| 10501 | .byte 102,15,101,193 // pcmpgtw %xmm1,%xmm0 |
| 10502 | .byte 102,15,223,193 // pandn %xmm1,%xmm0 |
| 10503 | .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10504 | .byte 102,15,114,240,13 // pslld $0xd,%xmm0 |
| 10505 | .byte 184,0,0,128,119 // mov $0x77800000,%eax |
| 10506 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10507 | .byte 102,68,15,112,202,0 // pshufd $0x0,%xmm2,%xmm9 |
| 10508 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 10509 | .byte 102,15,112,201,78 // pshufd $0x4e,%xmm1,%xmm1 |
| 10510 | .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| 10511 | .byte 102,15,101,209 // pcmpgtw %xmm1,%xmm2 |
| 10512 | .byte 102,15,223,209 // pandn %xmm1,%xmm2 |
| 10513 | .byte 102,15,56,51,202 // pmovzxwd %xmm2,%xmm1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10514 | .byte 102,15,114,241,13 // pslld $0xd,%xmm1 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10515 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 10516 | .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| 10517 | .byte 102,65,15,101,208 // pcmpgtw %xmm8,%xmm2 |
| 10518 | .byte 102,65,15,223,208 // pandn %xmm8,%xmm2 |
| 10519 | .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10520 | .byte 102,15,114,242,13 // pslld $0xd,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10521 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 10522 | .byte 102,69,15,112,192,78 // pshufd $0x4e,%xmm8,%xmm8 |
| 10523 | .byte 102,65,15,101,216 // pcmpgtw %xmm8,%xmm3 |
| 10524 | .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 |
| 10525 | .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10526 | .byte 102,15,114,243,13 // pslld $0xd,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 10527 | .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10528 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10529 | .byte 255,224 // jmpq *%rax |
| 10530 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10531 | HIDDEN _sk_store_f16_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10532 | .globl _sk_store_f16_sse41 |
| 10533 | _sk_store_f16_sse41: |
| 10534 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10535 | .byte 72,139,0 // mov (%rax),%rax |
| 10536 | .byte 185,0,0,128,7 // mov $0x7800000,%ecx |
| 10537 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 10538 | .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 |
| 10539 | .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 |
| 10540 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 10541 | .byte 102,65,15,114,208,13 // psrld $0xd,%xmm8 |
| 10542 | .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| 10543 | .byte 102,69,15,111,209 // movdqa %xmm9,%xmm10 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10544 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 10545 | .byte 102,65,15,114,210,13 // psrld $0xd,%xmm10 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 10546 | .byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10 |
| 10547 | .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10548 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10549 | .byte 102,65,15,114,211,13 // psrld $0xd,%xmm11 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 10550 | .byte 102,69,15,56,43,219 // packusdw %xmm11,%xmm11 |
| 10551 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 10552 | .byte 102,65,15,114,209,13 // psrld $0xd,%xmm9 |
| 10553 | .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9 |
| 10554 | .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| 10555 | .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| 10556 | .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| 10557 | .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10558 | .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 10559 | .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| 10560 | .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10561 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10562 | .byte 255,224 // jmpq *%rax |
| 10563 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 10564 | HIDDEN _sk_load_u16_be_sse41 |
| 10565 | .globl _sk_load_u16_be_sse41 |
| 10566 | _sk_load_u16_be_sse41: |
| 10567 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10568 | .byte 72,139,0 // mov (%rax),%rax |
| 10569 | .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| 10570 | .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| 10571 | .byte 102,15,111,208 // movdqa %xmm0,%xmm2 |
| 10572 | .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2 |
| 10573 | .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| 10574 | .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| 10575 | .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| 10576 | .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2 |
| 10577 | .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| 10578 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 10579 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10580 | .byte 102,15,111,193 // movdqa %xmm1,%xmm0 |
| 10581 | .byte 102,15,113,240,8 // psllw $0x8,%xmm0 |
| 10582 | .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3 |
| 10583 | .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 |
| 10584 | .byte 102,15,235,200 // por %xmm0,%xmm1 |
| 10585 | .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0 |
| 10586 | .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| 10587 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10588 | .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| 10589 | .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| 10590 | .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| 10591 | .byte 102,15,235,217 // por %xmm1,%xmm3 |
| 10592 | .byte 102,15,56,51,203 // pmovzxwd %xmm3,%xmm1 |
| 10593 | .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| 10594 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10595 | .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9 |
| 10596 | .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9 |
| 10597 | .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3 |
| 10598 | .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| 10599 | .byte 102,65,15,235,209 // por %xmm9,%xmm2 |
| 10600 | .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2 |
| 10601 | .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| 10602 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 10603 | .byte 102,68,15,111,203 // movdqa %xmm3,%xmm9 |
| 10604 | .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9 |
| 10605 | .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| 10606 | .byte 102,65,15,235,217 // por %xmm9,%xmm3 |
| 10607 | .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3 |
| 10608 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 10609 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 10610 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10611 | .byte 255,224 // jmpq *%rax |
| 10612 | |
| 10613 | HIDDEN _sk_store_u16_be_sse41 |
| 10614 | .globl _sk_store_u16_be_sse41 |
| 10615 | _sk_store_u16_be_sse41: |
| 10616 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10617 | .byte 72,139,0 // mov (%rax),%rax |
| 10618 | .byte 185,0,255,127,71 // mov $0x477fff00,%ecx |
| 10619 | .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| 10620 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10621 | .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| 10622 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 10623 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 10624 | .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8 |
| 10625 | .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| 10626 | .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| 10627 | .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8 |
| 10628 | .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| 10629 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 10630 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 10631 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 10632 | .byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10 |
| 10633 | .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11 |
| 10634 | .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11 |
| 10635 | .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 |
| 10636 | .byte 102,69,15,235,211 // por %xmm11,%xmm10 |
| 10637 | .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| 10638 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10639 | .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11 |
| 10640 | .byte 102,69,15,56,43,219 // packusdw %xmm11,%xmm11 |
| 10641 | .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 |
| 10642 | .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| 10643 | .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 |
| 10644 | .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| 10645 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 10646 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 10647 | .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9 |
| 10648 | .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| 10649 | .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| 10650 | .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9 |
| 10651 | .byte 102,69,15,235,204 // por %xmm12,%xmm9 |
| 10652 | .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| 10653 | .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| 10654 | .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| 10655 | .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| 10656 | .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| 10657 | .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| 10658 | .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| 10659 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10660 | .byte 255,224 // jmpq *%rax |
| 10661 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10662 | HIDDEN _sk_store_f32_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10663 | .globl _sk_store_f32_sse41 |
| 10664 | _sk_store_f32_sse41: |
| 10665 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10666 | .byte 72,139,0 // mov (%rax),%rax |
| 10667 | .byte 72,137,249 // mov %rdi,%rcx |
| 10668 | .byte 72,193,225,4 // shl $0x4,%rcx |
| 10669 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 10670 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 10671 | .byte 68,15,20,201 // unpcklps %xmm1,%xmm9 |
| 10672 | .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| 10673 | .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| 10674 | .byte 68,15,20,219 // unpcklps %xmm3,%xmm11 |
| 10675 | .byte 68,15,21,193 // unpckhps %xmm1,%xmm8 |
| 10676 | .byte 68,15,21,211 // unpckhps %xmm3,%xmm10 |
| 10677 | .byte 69,15,40,225 // movaps %xmm9,%xmm12 |
| 10678 | .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12 |
| 10679 | .byte 69,15,18,217 // movhlps %xmm9,%xmm11 |
| 10680 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10681 | .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9 |
| 10682 | .byte 69,15,18,208 // movhlps %xmm8,%xmm10 |
| 10683 | .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1) |
| 10684 | .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1) |
| 10685 | .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1) |
| 10686 | .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1) |
| 10687 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10688 | .byte 255,224 // jmpq *%rax |
| 10689 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10690 | HIDDEN _sk_clamp_x_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10691 | .globl _sk_clamp_x_sse41 |
| 10692 | _sk_clamp_x_sse41: |
| 10693 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10694 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 10695 | .byte 68,15,95,192 // maxps %xmm0,%xmm8 |
| 10696 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 10697 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10698 | .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0 |
| 10699 | .byte 102,65,15,254,193 // paddd %xmm9,%xmm0 |
| 10700 | .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| 10701 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10702 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 10703 | .byte 255,224 // jmpq *%rax |
| 10704 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10705 | HIDDEN _sk_clamp_y_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10706 | .globl _sk_clamp_y_sse41 |
| 10707 | _sk_clamp_y_sse41: |
| 10708 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10709 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 10710 | .byte 68,15,95,193 // maxps %xmm1,%xmm8 |
| 10711 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 10712 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10713 | .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1 |
| 10714 | .byte 102,65,15,254,201 // paddd %xmm9,%xmm1 |
| 10715 | .byte 68,15,93,193 // minps %xmm1,%xmm8 |
| 10716 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10717 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 10718 | .byte 255,224 // jmpq *%rax |
| 10719 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10720 | HIDDEN _sk_repeat_x_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10721 | .globl _sk_repeat_x_sse41 |
| 10722 | _sk_repeat_x_sse41: |
| 10723 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10724 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10725 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10726 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 10727 | .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| 10728 | .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 |
| 10729 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 10730 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 10731 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 10732 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 10733 | .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| 10734 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10735 | .byte 255,224 // jmpq *%rax |
| 10736 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10737 | HIDDEN _sk_repeat_y_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10738 | .globl _sk_repeat_y_sse41 |
| 10739 | _sk_repeat_y_sse41: |
| 10740 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10741 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10742 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10743 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 10744 | .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| 10745 | .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9 |
| 10746 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 10747 | .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| 10748 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 10749 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 10750 | .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| 10751 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10752 | .byte 255,224 // jmpq *%rax |
| 10753 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10754 | HIDDEN _sk_mirror_x_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10755 | .globl _sk_mirror_x_sse41 |
| 10756 | _sk_mirror_x_sse41: |
| 10757 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10758 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10759 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10760 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10761 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 10762 | .byte 243,69,15,88,192 // addss %xmm8,%xmm8 |
| 10763 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10764 | .byte 68,15,40,208 // movaps %xmm0,%xmm10 |
| 10765 | .byte 69,15,94,208 // divps %xmm8,%xmm10 |
| 10766 | .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 |
| 10767 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 10768 | .byte 65,15,92,194 // subps %xmm10,%xmm0 |
| 10769 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 10770 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 10771 | .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| 10772 | .byte 65,15,84,192 // andps %xmm8,%xmm0 |
| 10773 | .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 |
| 10774 | .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| 10775 | .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| 10776 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10777 | .byte 255,224 // jmpq *%rax |
| 10778 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10779 | HIDDEN _sk_mirror_y_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10780 | .globl _sk_mirror_y_sse41 |
| 10781 | _sk_mirror_y_sse41: |
| 10782 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10783 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 10784 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 10785 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 10786 | .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| 10787 | .byte 243,69,15,88,192 // addss %xmm8,%xmm8 |
| 10788 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 10789 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| 10790 | .byte 69,15,94,208 // divps %xmm8,%xmm10 |
| 10791 | .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10 |
| 10792 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 10793 | .byte 65,15,92,202 // subps %xmm10,%xmm1 |
| 10794 | .byte 65,15,92,201 // subps %xmm9,%xmm1 |
| 10795 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 10796 | .byte 68,15,92,193 // subps %xmm1,%xmm8 |
| 10797 | .byte 65,15,84,200 // andps %xmm8,%xmm1 |
| 10798 | .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8 |
| 10799 | .byte 102,69,15,254,193 // paddd %xmm9,%xmm8 |
| 10800 | .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| 10801 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10802 | .byte 255,224 // jmpq *%rax |
| 10803 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10804 | HIDDEN _sk_luminance_to_alpha_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10805 | .globl _sk_luminance_to_alpha_sse41 |
| 10806 | _sk_luminance_to_alpha_sse41: |
| 10807 | .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| 10808 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10809 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10810 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 10811 | .byte 184,89,23,55,63 // mov $0x3f371759,%eax |
| 10812 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 10813 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10814 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 10815 | .byte 15,88,195 // addps %xmm3,%xmm0 |
| 10816 | .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| 10817 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 10818 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 10819 | .byte 15,89,218 // mulps %xmm2,%xmm3 |
| 10820 | .byte 15,88,216 // addps %xmm0,%xmm3 |
| 10821 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10822 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 10823 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 10824 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 10825 | .byte 255,224 // jmpq *%rax |
| 10826 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10827 | HIDDEN _sk_matrix_2x3_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10828 | .globl _sk_matrix_2x3_sse41 |
| 10829 | _sk_matrix_2x3_sse41: |
| 10830 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 10831 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 10832 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10833 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 10834 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 10835 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10836 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 10837 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10838 | .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11 |
| 10839 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10840 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10841 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10842 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10843 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 10844 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 10845 | .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| 10846 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10847 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 10848 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10849 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10850 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10851 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10852 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 10853 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10854 | .byte 255,224 // jmpq *%rax |
| 10855 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10856 | HIDDEN _sk_matrix_3x4_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10857 | .globl _sk_matrix_3x4_sse41 |
| 10858 | _sk_matrix_3x4_sse41: |
| 10859 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 10860 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 10861 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10862 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 10863 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 10864 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10865 | .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| 10866 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10867 | .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| 10868 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10869 | .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12 |
| 10870 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10871 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10872 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10873 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10874 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10875 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10876 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 10877 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 10878 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 10879 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10880 | .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| 10881 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10882 | .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| 10883 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10884 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10885 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10886 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10887 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10888 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10889 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 10890 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 10891 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10892 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 10893 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10894 | .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| 10895 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10896 | .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| 10897 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 10898 | .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| 10899 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 10900 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 10901 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10902 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 10903 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10904 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10905 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 10906 | .byte 255,224 // jmpq *%rax |
| 10907 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10908 | HIDDEN _sk_matrix_4x5_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10909 | .globl _sk_matrix_4x5_sse41 |
| 10910 | _sk_matrix_4x5_sse41: |
| 10911 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 10912 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 10913 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10914 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 10915 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 10916 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10917 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 10918 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10919 | .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11 |
| 10920 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10921 | .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12 |
| 10922 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10923 | .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13 |
| 10924 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 10925 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 10926 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 10927 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10928 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10929 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10930 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10931 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 10932 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 10933 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 10934 | .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10 |
| 10935 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10936 | .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11 |
| 10937 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10938 | .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12 |
| 10939 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10940 | .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13 |
| 10941 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 10942 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 10943 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 10944 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 10945 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10946 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 10947 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10948 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 10949 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 10950 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 10951 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 10952 | .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| 10953 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10954 | .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| 10955 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10956 | .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13 |
| 10957 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 10958 | .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14 |
| 10959 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 10960 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 10961 | .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| 10962 | .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| 10963 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 10964 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 10965 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10966 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 10967 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 10968 | .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11 |
| 10969 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 10970 | .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12 |
| 10971 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 10972 | .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| 10973 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 10974 | .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14 |
| 10975 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 10976 | .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15 |
| 10977 | .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| 10978 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 10979 | .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| 10980 | .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| 10981 | .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| 10982 | .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| 10983 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 10984 | .byte 69,15,89,216 // mulps %xmm8,%xmm11 |
| 10985 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 10986 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10987 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 10988 | .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| 10989 | .byte 255,224 // jmpq *%rax |
| 10990 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 10991 | HIDDEN _sk_matrix_perspective_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 10992 | .globl _sk_matrix_perspective_sse41 |
| 10993 | _sk_matrix_perspective_sse41: |
| 10994 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 10995 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 10996 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 10997 | .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 |
| 10998 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 10999 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 11000 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 11001 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 11002 | .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| 11003 | .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| 11004 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 11005 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 11006 | .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| 11007 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 11008 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 11009 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 11010 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 11011 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 11012 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 11013 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 11014 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 11015 | .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| 11016 | .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10 |
| 11017 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 11018 | .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| 11019 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 11020 | .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| 11021 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 11022 | .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| 11023 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 11024 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 11025 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 11026 | .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| 11027 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 11028 | .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| 11029 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11030 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 11031 | .byte 255,224 // jmpq *%rax |
| 11032 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11033 | HIDDEN _sk_linear_gradient_2stops_sse41 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11034 | .globl _sk_linear_gradient_2stops_sse41 |
| 11035 | _sk_linear_gradient_2stops_sse41: |
| 11036 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11037 | .byte 68,15,16,8 // movups (%rax),%xmm9 |
| 11038 | .byte 15,16,88,16 // movups 0x10(%rax),%xmm3 |
| 11039 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11040 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11041 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 11042 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 11043 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 11044 | .byte 68,15,88,193 // addps %xmm1,%xmm8 |
| 11045 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11046 | .byte 15,198,201,85 // shufps $0x55,%xmm1,%xmm1 |
| 11047 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| 11048 | .byte 15,198,210,85 // shufps $0x55,%xmm2,%xmm2 |
| 11049 | .byte 15,89,200 // mulps %xmm0,%xmm1 |
| 11050 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 11051 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 11052 | .byte 15,198,210,170 // shufps $0xaa,%xmm2,%xmm2 |
| 11053 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 11054 | .byte 69,15,198,210,170 // shufps $0xaa,%xmm10,%xmm10 |
| 11055 | .byte 15,89,208 // mulps %xmm0,%xmm2 |
| 11056 | .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| 11057 | .byte 15,198,219,255 // shufps $0xff,%xmm3,%xmm3 |
| 11058 | .byte 69,15,198,201,255 // shufps $0xff,%xmm9,%xmm9 |
| 11059 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 11060 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 11061 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11062 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 11063 | .byte 255,224 // jmpq *%rax |
| 11064 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11065 | HIDDEN _sk_start_pipeline_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11066 | .globl _sk_start_pipeline_sse2 |
| 11067 | _sk_start_pipeline_sse2: |
| 11068 | .byte 65,87 // push %r15 |
| 11069 | .byte 65,86 // push %r14 |
| 11070 | .byte 65,85 // push %r13 |
| 11071 | .byte 65,84 // push %r12 |
| 11072 | .byte 83 // push %rbx |
| 11073 | .byte 73,137,207 // mov %rcx,%r15 |
| 11074 | .byte 73,137,214 // mov %rdx,%r14 |
| 11075 | .byte 72,137,251 // mov %rdi,%rbx |
| 11076 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11077 | .byte 73,137,196 // mov %rax,%r12 |
| 11078 | .byte 73,137,245 // mov %rsi,%r13 |
| 11079 | .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| 11080 | .byte 76,57,248 // cmp %r15,%rax |
| 11081 | .byte 118,5 // jbe 28 <_sk_start_pipeline_sse2+0x28> |
| 11082 | .byte 72,137,216 // mov %rbx,%rax |
| 11083 | .byte 235,52 // jmp 5c <_sk_start_pipeline_sse2+0x5c> |
| 11084 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 11085 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 11086 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 11087 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 11088 | .byte 15,87,228 // xorps %xmm4,%xmm4 |
| 11089 | .byte 15,87,237 // xorps %xmm5,%xmm5 |
| 11090 | .byte 15,87,246 // xorps %xmm6,%xmm6 |
| 11091 | .byte 15,87,255 // xorps %xmm7,%xmm7 |
| 11092 | .byte 72,137,223 // mov %rbx,%rdi |
| 11093 | .byte 76,137,238 // mov %r13,%rsi |
| 11094 | .byte 76,137,242 // mov %r14,%rdx |
| 11095 | .byte 65,255,212 // callq *%r12 |
| 11096 | .byte 72,141,67,4 // lea 0x4(%rbx),%rax |
| 11097 | .byte 72,131,195,8 // add $0x8,%rbx |
| 11098 | .byte 76,57,251 // cmp %r15,%rbx |
| 11099 | .byte 72,137,195 // mov %rax,%rbx |
| 11100 | .byte 118,204 // jbe 28 <_sk_start_pipeline_sse2+0x28> |
| 11101 | .byte 91 // pop %rbx |
| 11102 | .byte 65,92 // pop %r12 |
| 11103 | .byte 65,93 // pop %r13 |
| 11104 | .byte 65,94 // pop %r14 |
| 11105 | .byte 65,95 // pop %r15 |
| 11106 | .byte 195 // retq |
| 11107 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11108 | HIDDEN _sk_just_return_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11109 | .globl _sk_just_return_sse2 |
| 11110 | _sk_just_return_sse2: |
| 11111 | .byte 195 // retq |
| 11112 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11113 | HIDDEN _sk_seed_shader_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11114 | .globl _sk_seed_shader_sse2 |
| 11115 | _sk_seed_shader_sse2: |
| 11116 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11117 | .byte 102,15,110,199 // movd %edi,%xmm0 |
| 11118 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 11119 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 11120 | .byte 185,0,0,0,63 // mov $0x3f000000,%ecx |
| 11121 | .byte 102,15,110,209 // movd %ecx,%xmm2 |
| 11122 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11123 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 11124 | .byte 15,16,2 // movups (%rdx),%xmm0 |
| 11125 | .byte 15,88,193 // addps %xmm1,%xmm0 |
| 11126 | .byte 102,15,110,8 // movd (%rax),%xmm1 |
| 11127 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 11128 | .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| 11129 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 11130 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11131 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 11132 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11133 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11134 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 11135 | .byte 15,87,228 // xorps %xmm4,%xmm4 |
| 11136 | .byte 15,87,237 // xorps %xmm5,%xmm5 |
| 11137 | .byte 15,87,246 // xorps %xmm6,%xmm6 |
| 11138 | .byte 15,87,255 // xorps %xmm7,%xmm7 |
| 11139 | .byte 255,224 // jmpq *%rax |
| 11140 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11141 | HIDDEN _sk_constant_color_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11142 | .globl _sk_constant_color_sse2 |
| 11143 | _sk_constant_color_sse2: |
| 11144 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11145 | .byte 15,16,24 // movups (%rax),%xmm3 |
| 11146 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 11147 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 11148 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11149 | .byte 15,198,201,85 // shufps $0x55,%xmm1,%xmm1 |
| 11150 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 11151 | .byte 15,198,210,170 // shufps $0xaa,%xmm2,%xmm2 |
| 11152 | .byte 15,198,219,255 // shufps $0xff,%xmm3,%xmm3 |
| 11153 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11154 | .byte 255,224 // jmpq *%rax |
| 11155 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11156 | HIDDEN _sk_clear_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11157 | .globl _sk_clear_sse2 |
| 11158 | _sk_clear_sse2: |
| 11159 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11160 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 11161 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 11162 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 11163 | .byte 15,87,219 // xorps %xmm3,%xmm3 |
| 11164 | .byte 255,224 // jmpq *%rax |
| 11165 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 11166 | HIDDEN _sk_srcatop_sse2 |
| 11167 | .globl _sk_srcatop_sse2 |
| 11168 | _sk_srcatop_sse2: |
| 11169 | .byte 15,89,199 // mulps %xmm7,%xmm0 |
| 11170 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11171 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11172 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11173 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 11174 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11175 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 11176 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 11177 | .byte 15,89,207 // mulps %xmm7,%xmm1 |
| 11178 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11179 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11180 | .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| 11181 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11182 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11183 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 11184 | .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| 11185 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11186 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11187 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11188 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11189 | .byte 255,224 // jmpq *%rax |
| 11190 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 11191 | HIDDEN _sk_dstatop_sse2 |
| 11192 | .globl _sk_dstatop_sse2 |
| 11193 | _sk_dstatop_sse2: |
| 11194 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11195 | .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| 11196 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11197 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 11198 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 11199 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 11200 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 11201 | .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| 11202 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11203 | .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| 11204 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 11205 | .byte 65,15,88,200 // addps %xmm8,%xmm1 |
| 11206 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11207 | .byte 68,15,89,198 // mulps %xmm6,%xmm8 |
| 11208 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 11209 | .byte 65,15,88,208 // addps %xmm8,%xmm2 |
| 11210 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 11211 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11212 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 11213 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11214 | .byte 255,224 // jmpq *%rax |
| 11215 | |
| 11216 | HIDDEN _sk_srcin_sse2 |
| 11217 | .globl _sk_srcin_sse2 |
| 11218 | _sk_srcin_sse2: |
| 11219 | .byte 15,89,199 // mulps %xmm7,%xmm0 |
| 11220 | .byte 15,89,207 // mulps %xmm7,%xmm1 |
| 11221 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11222 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11223 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11224 | .byte 255,224 // jmpq *%rax |
| 11225 | |
| 11226 | HIDDEN _sk_dstin_sse2 |
| 11227 | .globl _sk_dstin_sse2 |
| 11228 | _sk_dstin_sse2: |
| 11229 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 11230 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11231 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11232 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11233 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 11234 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11235 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11236 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11237 | .byte 255,224 // jmpq *%rax |
| 11238 | |
| 11239 | HIDDEN _sk_srcout_sse2 |
| 11240 | .globl _sk_srcout_sse2 |
| 11241 | _sk_srcout_sse2: |
| 11242 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11243 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11244 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11245 | .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| 11246 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 11247 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 11248 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 11249 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 11250 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11251 | .byte 255,224 // jmpq *%rax |
| 11252 | |
| 11253 | HIDDEN _sk_dstout_sse2 |
| 11254 | .globl _sk_dstout_sse2 |
| 11255 | _sk_dstout_sse2: |
| 11256 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11257 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11258 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11259 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 11260 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 11261 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11262 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 11263 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11264 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11265 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11266 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11267 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11268 | .byte 65,15,40,216 // movaps %xmm8,%xmm3 |
| 11269 | .byte 255,224 // jmpq *%rax |
| 11270 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11271 | HIDDEN _sk_srcover_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11272 | .globl _sk_srcover_sse2 |
| 11273 | _sk_srcover_sse2: |
| 11274 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11275 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11276 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11277 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 11278 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11279 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 11280 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 11281 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11282 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11283 | .byte 65,15,88,201 // addps %xmm9,%xmm1 |
| 11284 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11285 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 11286 | .byte 65,15,88,209 // addps %xmm9,%xmm2 |
| 11287 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11288 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| 11289 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11290 | .byte 255,224 // jmpq *%rax |
| 11291 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 11292 | HIDDEN _sk_dstover_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 11293 | .globl _sk_dstover_sse2 |
| 11294 | _sk_dstover_sse2: |
| 11295 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11296 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11297 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11298 | .byte 68,15,92,199 // subps %xmm7,%xmm8 |
| 11299 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 11300 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11301 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 11302 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11303 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 11304 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 11305 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 11306 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 11307 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11308 | .byte 255,224 // jmpq *%rax |
| 11309 | |
| Mike Klein | aaca1e4 | 2017-03-31 09:29:01 -0400 | [diff] [blame] | 11310 | HIDDEN _sk_modulate_sse2 |
| 11311 | .globl _sk_modulate_sse2 |
| 11312 | _sk_modulate_sse2: |
| 11313 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11314 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11315 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11316 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11317 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11318 | .byte 255,224 // jmpq *%rax |
| 11319 | |
| 11320 | HIDDEN _sk_multiply_sse2 |
| 11321 | .globl _sk_multiply_sse2 |
| 11322 | _sk_multiply_sse2: |
| 11323 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11324 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11325 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11326 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11327 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 11328 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 11329 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 11330 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 11331 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 11332 | .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| 11333 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 11334 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11335 | .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| 11336 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 11337 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 11338 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 11339 | .byte 68,15,89,221 // mulps %xmm5,%xmm11 |
| 11340 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 11341 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11342 | .byte 65,15,88,203 // addps %xmm11,%xmm1 |
| 11343 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 11344 | .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| 11345 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 11346 | .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| 11347 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 11348 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11349 | .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| 11350 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 11351 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11352 | .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| 11353 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11354 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| 11355 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11356 | .byte 255,224 // jmpq *%rax |
| 11357 | |
| 11358 | HIDDEN _sk_plus__sse2 |
| 11359 | .globl _sk_plus__sse2 |
| 11360 | _sk_plus__sse2: |
| 11361 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11362 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11363 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 11364 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 11365 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11366 | .byte 255,224 // jmpq *%rax |
| 11367 | |
| 11368 | HIDDEN _sk_screen_sse2 |
| 11369 | .globl _sk_screen_sse2 |
| 11370 | _sk_screen_sse2: |
| 11371 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 11372 | .byte 68,15,88,196 // addps %xmm4,%xmm8 |
| 11373 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11374 | .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| 11375 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 11376 | .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| 11377 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11378 | .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| 11379 | .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| 11380 | .byte 68,15,88,214 // addps %xmm6,%xmm10 |
| 11381 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11382 | .byte 68,15,92,210 // subps %xmm2,%xmm10 |
| 11383 | .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| 11384 | .byte 68,15,88,223 // addps %xmm7,%xmm11 |
| 11385 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11386 | .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| 11387 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11388 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 11389 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 11390 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 11391 | .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| 11392 | .byte 255,224 // jmpq *%rax |
| 11393 | |
| 11394 | HIDDEN _sk_xor__sse2 |
| 11395 | .globl _sk_xor__sse2 |
| 11396 | _sk_xor__sse2: |
| 11397 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11398 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11399 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 11400 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 11401 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11402 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 11403 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 11404 | .byte 65,15,92,216 // subps %xmm8,%xmm3 |
| 11405 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 11406 | .byte 68,15,89,212 // mulps %xmm4,%xmm10 |
| 11407 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 11408 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 11409 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 11410 | .byte 68,15,89,213 // mulps %xmm5,%xmm10 |
| 11411 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 11412 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 11413 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 11414 | .byte 68,15,89,214 // mulps %xmm6,%xmm10 |
| 11415 | .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| 11416 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 11417 | .byte 15,89,223 // mulps %xmm7,%xmm3 |
| 11418 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 11419 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11420 | .byte 255,224 // jmpq *%rax |
| 11421 | |
| Mike Klein | 66b09ab | 2017-03-31 10:29:40 -0400 | [diff] [blame] | 11422 | HIDDEN _sk_darken_sse2 |
| 11423 | .globl _sk_darken_sse2 |
| 11424 | _sk_darken_sse2: |
| 11425 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 11426 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 11427 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11428 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 11429 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11430 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 11431 | .byte 68,15,95,201 // maxps %xmm1,%xmm9 |
| 11432 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 11433 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 11434 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11435 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11436 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11437 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11438 | .byte 69,15,95,193 // maxps %xmm9,%xmm8 |
| 11439 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 11440 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 11441 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 11442 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11443 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11444 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 11445 | .byte 65,15,95,209 // maxps %xmm9,%xmm2 |
| 11446 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 11447 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11448 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 11449 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11450 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 11451 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11452 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 11453 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11454 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11455 | .byte 255,224 // jmpq *%rax |
| 11456 | |
| 11457 | HIDDEN _sk_lighten_sse2 |
| 11458 | .globl _sk_lighten_sse2 |
| 11459 | _sk_lighten_sse2: |
| 11460 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 11461 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 11462 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11463 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 11464 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11465 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 11466 | .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| 11467 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 11468 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 11469 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11470 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11471 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11472 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11473 | .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| 11474 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 11475 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 11476 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 11477 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11478 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11479 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 11480 | .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| 11481 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 11482 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11483 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 11484 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11485 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 11486 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11487 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 11488 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11489 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11490 | .byte 255,224 // jmpq *%rax |
| 11491 | |
| 11492 | HIDDEN _sk_difference_sse2 |
| 11493 | .globl _sk_difference_sse2 |
| 11494 | _sk_difference_sse2: |
| 11495 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 11496 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 11497 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11498 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 11499 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11500 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 11501 | .byte 68,15,93,201 // minps %xmm1,%xmm9 |
| 11502 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 11503 | .byte 65,15,92,193 // subps %xmm9,%xmm0 |
| 11504 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 11505 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11506 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11507 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11508 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11509 | .byte 69,15,93,193 // minps %xmm9,%xmm8 |
| 11510 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 11511 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 11512 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 11513 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 11514 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11515 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11516 | .byte 68,15,89,206 // mulps %xmm6,%xmm9 |
| 11517 | .byte 65,15,93,209 // minps %xmm9,%xmm2 |
| 11518 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 11519 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 11520 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11521 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 11522 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11523 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 11524 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11525 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 11526 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11527 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11528 | .byte 255,224 // jmpq *%rax |
| 11529 | |
| 11530 | HIDDEN _sk_exclusion_sse2 |
| 11531 | .globl _sk_exclusion_sse2 |
| 11532 | _sk_exclusion_sse2: |
| 11533 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 11534 | .byte 15,40,200 // movaps %xmm0,%xmm1 |
| 11535 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11536 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 11537 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 11538 | .byte 15,92,193 // subps %xmm1,%xmm0 |
| 11539 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 11540 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11541 | .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| 11542 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 11543 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 11544 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 11545 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 11546 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11547 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 11548 | .byte 68,15,92,194 // subps %xmm2,%xmm8 |
| 11549 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11550 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 11551 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 11552 | .byte 15,92,211 // subps %xmm3,%xmm2 |
| 11553 | .byte 15,89,215 // mulps %xmm7,%xmm2 |
| 11554 | .byte 15,88,218 // addps %xmm2,%xmm3 |
| 11555 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11556 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11557 | .byte 255,224 // jmpq *%rax |
| 11558 | |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11559 | HIDDEN _sk_colorburn_sse2 |
| 11560 | .globl _sk_colorburn_sse2 |
| 11561 | _sk_colorburn_sse2: |
| 11562 | .byte 68,15,40,193 // movaps %xmm1,%xmm8 |
| 11563 | .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| 11564 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11565 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 11566 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 11567 | .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| 11568 | .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| 11569 | .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| 11570 | .byte 65,15,89,196 // mulps %xmm12,%xmm0 |
| 11571 | .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| 11572 | .byte 15,40,207 // movaps %xmm7,%xmm1 |
| 11573 | .byte 15,92,204 // subps %xmm4,%xmm1 |
| 11574 | .byte 15,89,203 // mulps %xmm3,%xmm1 |
| 11575 | .byte 65,15,94,204 // divps %xmm12,%xmm1 |
| 11576 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11577 | .byte 68,15,93,233 // minps %xmm1,%xmm13 |
| 11578 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 11579 | .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| 11580 | .byte 65,15,40,204 // movaps %xmm12,%xmm1 |
| 11581 | .byte 65,15,194,202,0 // cmpeqps %xmm10,%xmm1 |
| 11582 | .byte 68,15,92,203 // subps %xmm3,%xmm9 |
| 11583 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 11584 | .byte 68,15,88,240 // addps %xmm0,%xmm14 |
| 11585 | .byte 68,15,84,225 // andps %xmm1,%xmm12 |
| 11586 | .byte 65,15,85,206 // andnps %xmm14,%xmm1 |
| 11587 | .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| 11588 | .byte 68,15,89,236 // mulps %xmm4,%xmm13 |
| 11589 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11590 | .byte 65,15,86,204 // orps %xmm12,%xmm1 |
| 11591 | .byte 68,15,40,228 // movaps %xmm4,%xmm12 |
| 11592 | .byte 68,15,194,231,0 // cmpeqps %xmm7,%xmm12 |
| 11593 | .byte 65,15,88,205 // addps %xmm13,%xmm1 |
| 11594 | .byte 65,15,84,196 // andps %xmm12,%xmm0 |
| 11595 | .byte 68,15,85,225 // andnps %xmm1,%xmm12 |
| 11596 | .byte 65,15,86,196 // orps %xmm12,%xmm0 |
| 11597 | .byte 65,15,40,203 // movaps %xmm11,%xmm1 |
| 11598 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 11599 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 11600 | .byte 68,15,92,229 // subps %xmm5,%xmm12 |
| 11601 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 11602 | .byte 69,15,94,224 // divps %xmm8,%xmm12 |
| 11603 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11604 | .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| 11605 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 11606 | .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| 11607 | .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| 11608 | .byte 69,15,194,234,0 // cmpeqps %xmm10,%xmm13 |
| 11609 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 11610 | .byte 68,15,88,225 // addps %xmm1,%xmm12 |
| 11611 | .byte 69,15,84,197 // andps %xmm13,%xmm8 |
| 11612 | .byte 69,15,85,236 // andnps %xmm12,%xmm13 |
| 11613 | .byte 69,15,86,232 // orps %xmm8,%xmm13 |
| 11614 | .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| 11615 | .byte 68,15,89,197 // mulps %xmm5,%xmm8 |
| 11616 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 11617 | .byte 69,15,88,232 // addps %xmm8,%xmm13 |
| 11618 | .byte 68,15,40,197 // movaps %xmm5,%xmm8 |
| 11619 | .byte 68,15,194,199,0 // cmpeqps %xmm7,%xmm8 |
| 11620 | .byte 65,15,84,200 // andps %xmm8,%xmm1 |
| 11621 | .byte 69,15,85,197 // andnps %xmm13,%xmm8 |
| 11622 | .byte 65,15,86,200 // orps %xmm8,%xmm1 |
| 11623 | .byte 68,15,40,199 // movaps %xmm7,%xmm8 |
| 11624 | .byte 68,15,92,198 // subps %xmm6,%xmm8 |
| 11625 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 11626 | .byte 68,15,94,194 // divps %xmm2,%xmm8 |
| 11627 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 11628 | .byte 69,15,93,224 // minps %xmm8,%xmm12 |
| 11629 | .byte 68,15,40,199 // movaps %xmm7,%xmm8 |
| 11630 | .byte 69,15,92,196 // subps %xmm12,%xmm8 |
| 11631 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 11632 | .byte 68,15,194,210,0 // cmpeqps %xmm2,%xmm10 |
| 11633 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 11634 | .byte 69,15,88,195 // addps %xmm11,%xmm8 |
| 11635 | .byte 65,15,84,210 // andps %xmm10,%xmm2 |
| 11636 | .byte 69,15,85,208 // andnps %xmm8,%xmm10 |
| 11637 | .byte 69,15,40,195 // movaps %xmm11,%xmm8 |
| 11638 | .byte 68,15,88,198 // addps %xmm6,%xmm8 |
| 11639 | .byte 68,15,86,210 // orps %xmm2,%xmm10 |
| 11640 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| 11641 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11642 | .byte 68,15,88,210 // addps %xmm2,%xmm10 |
| 11643 | .byte 15,40,214 // movaps %xmm6,%xmm2 |
| 11644 | .byte 15,194,215,0 // cmpeqps %xmm7,%xmm2 |
| 11645 | .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| 11646 | .byte 65,15,85,210 // andnps %xmm10,%xmm2 |
| 11647 | .byte 68,15,86,194 // orps %xmm2,%xmm8 |
| 11648 | .byte 68,15,89,207 // mulps %xmm7,%xmm9 |
| 11649 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 11650 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11651 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11652 | .byte 255,224 // jmpq *%rax |
| 11653 | |
| 11654 | HIDDEN _sk_colordodge_sse2 |
| 11655 | .globl _sk_colordodge_sse2 |
| 11656 | _sk_colordodge_sse2: |
| 11657 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 11658 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11659 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 11660 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 11661 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 11662 | .byte 68,15,92,223 // subps %xmm7,%xmm11 |
| 11663 | .byte 65,15,40,195 // movaps %xmm11,%xmm0 |
| 11664 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 11665 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11666 | .byte 68,15,89,196 // mulps %xmm4,%xmm8 |
| 11667 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 11668 | .byte 69,15,92,225 // subps %xmm9,%xmm12 |
| 11669 | .byte 69,15,94,196 // divps %xmm12,%xmm8 |
| 11670 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 11671 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11672 | .byte 69,15,93,232 // minps %xmm8,%xmm13 |
| 11673 | .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| 11674 | .byte 68,15,194,243,0 // cmpeqps %xmm3,%xmm14 |
| 11675 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 11676 | .byte 68,15,88,232 // addps %xmm0,%xmm13 |
| 11677 | .byte 69,15,84,206 // andps %xmm14,%xmm9 |
| 11678 | .byte 69,15,85,245 // andnps %xmm13,%xmm14 |
| 11679 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 11680 | .byte 68,15,92,211 // subps %xmm3,%xmm10 |
| 11681 | .byte 69,15,86,241 // orps %xmm9,%xmm14 |
| 11682 | .byte 69,15,40,202 // movaps %xmm10,%xmm9 |
| 11683 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 11684 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 11685 | .byte 69,15,88,241 // addps %xmm9,%xmm14 |
| 11686 | .byte 68,15,40,204 // movaps %xmm4,%xmm9 |
| 11687 | .byte 69,15,194,200,0 // cmpeqps %xmm8,%xmm9 |
| 11688 | .byte 65,15,84,193 // andps %xmm9,%xmm0 |
| 11689 | .byte 69,15,85,206 // andnps %xmm14,%xmm9 |
| 11690 | .byte 65,15,86,193 // orps %xmm9,%xmm0 |
| 11691 | .byte 68,15,40,235 // movaps %xmm3,%xmm13 |
| 11692 | .byte 68,15,89,237 // mulps %xmm5,%xmm13 |
| 11693 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11694 | .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| 11695 | .byte 69,15,94,233 // divps %xmm9,%xmm13 |
| 11696 | .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| 11697 | .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| 11698 | .byte 69,15,93,229 // minps %xmm13,%xmm12 |
| 11699 | .byte 68,15,40,233 // movaps %xmm1,%xmm13 |
| 11700 | .byte 68,15,194,235,0 // cmpeqps %xmm3,%xmm13 |
| 11701 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 11702 | .byte 69,15,88,225 // addps %xmm9,%xmm12 |
| 11703 | .byte 65,15,84,205 // andps %xmm13,%xmm1 |
| 11704 | .byte 69,15,85,236 // andnps %xmm12,%xmm13 |
| 11705 | .byte 68,15,86,233 // orps %xmm1,%xmm13 |
| 11706 | .byte 65,15,40,202 // movaps %xmm10,%xmm1 |
| 11707 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11708 | .byte 68,15,88,205 // addps %xmm5,%xmm9 |
| 11709 | .byte 68,15,88,233 // addps %xmm1,%xmm13 |
| 11710 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 11711 | .byte 65,15,194,200,0 // cmpeqps %xmm8,%xmm1 |
| 11712 | .byte 68,15,84,201 // andps %xmm1,%xmm9 |
| 11713 | .byte 65,15,85,205 // andnps %xmm13,%xmm1 |
| 11714 | .byte 68,15,86,201 // orps %xmm1,%xmm9 |
| 11715 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 11716 | .byte 68,15,89,230 // mulps %xmm6,%xmm12 |
| 11717 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 11718 | .byte 15,92,202 // subps %xmm2,%xmm1 |
| 11719 | .byte 68,15,94,225 // divps %xmm1,%xmm12 |
| 11720 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11721 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 11722 | .byte 69,15,93,236 // minps %xmm12,%xmm13 |
| 11723 | .byte 15,40,202 // movaps %xmm2,%xmm1 |
| 11724 | .byte 15,194,203,0 // cmpeqps %xmm3,%xmm1 |
| 11725 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 11726 | .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| 11727 | .byte 15,84,209 // andps %xmm1,%xmm2 |
| 11728 | .byte 65,15,85,205 // andnps %xmm13,%xmm1 |
| 11729 | .byte 15,86,202 // orps %xmm2,%xmm1 |
| 11730 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 11731 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11732 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 11733 | .byte 68,15,194,198,0 // cmpeqps %xmm6,%xmm8 |
| 11734 | .byte 68,15,88,222 // addps %xmm6,%xmm11 |
| 11735 | .byte 69,15,84,216 // andps %xmm8,%xmm11 |
| 11736 | .byte 68,15,85,193 // andnps %xmm1,%xmm8 |
| 11737 | .byte 69,15,86,195 // orps %xmm11,%xmm8 |
| 11738 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 11739 | .byte 65,15,88,218 // addps %xmm10,%xmm3 |
| 11740 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11741 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 11742 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 11743 | .byte 255,224 // jmpq *%rax |
| 11744 | |
| 11745 | HIDDEN _sk_hardlight_sse2 |
| 11746 | .globl _sk_hardlight_sse2 |
| 11747 | _sk_hardlight_sse2: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 11748 | .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11749 | .byte 15,40,245 // movaps %xmm5,%xmm6 |
| 11750 | .byte 15,40,236 // movaps %xmm4,%xmm5 |
| 11751 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11752 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 11753 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 11754 | .byte 69,15,40,211 // movaps %xmm11,%xmm10 |
| 11755 | .byte 68,15,92,215 // subps %xmm7,%xmm10 |
| 11756 | .byte 69,15,40,194 // movaps %xmm10,%xmm8 |
| 11757 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 11758 | .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| 11759 | .byte 69,15,40,203 // movaps %xmm11,%xmm9 |
| 11760 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 11761 | .byte 69,15,88,200 // addps %xmm8,%xmm9 |
| 11762 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 11763 | .byte 68,15,92,192 // subps %xmm0,%xmm8 |
| 11764 | .byte 15,40,227 // movaps %xmm3,%xmm4 |
| 11765 | .byte 15,89,231 // mulps %xmm7,%xmm4 |
| 11766 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11767 | .byte 68,15,40,247 // movaps %xmm7,%xmm14 |
| 11768 | .byte 68,15,40,255 // movaps %xmm7,%xmm15 |
| 11769 | .byte 68,15,92,253 // subps %xmm5,%xmm15 |
| 11770 | .byte 69,15,89,248 // mulps %xmm8,%xmm15 |
| 11771 | .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| 11772 | .byte 68,15,40,228 // movaps %xmm4,%xmm12 |
| 11773 | .byte 69,15,92,231 // subps %xmm15,%xmm12 |
| 11774 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 11775 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 11776 | .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8 |
| 11777 | .byte 15,89,197 // mulps %xmm5,%xmm0 |
| 11778 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 11779 | .byte 65,15,84,192 // andps %xmm8,%xmm0 |
| 11780 | .byte 69,15,85,196 // andnps %xmm12,%xmm8 |
| 11781 | .byte 68,15,86,192 // orps %xmm0,%xmm8 |
| 11782 | .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| 11783 | .byte 69,15,40,227 // movaps %xmm11,%xmm12 |
| 11784 | .byte 68,15,89,223 // mulps %xmm7,%xmm11 |
| 11785 | .byte 69,15,88,193 // addps %xmm9,%xmm8 |
| 11786 | .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| 11787 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 11788 | .byte 68,15,89,254 // mulps %xmm6,%xmm15 |
| 11789 | .byte 68,15,88,248 // addps %xmm0,%xmm15 |
| 11790 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 11791 | .byte 15,92,193 // subps %xmm1,%xmm0 |
| 11792 | .byte 68,15,92,238 // subps %xmm6,%xmm13 |
| 11793 | .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| 11794 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 11795 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 11796 | .byte 65,15,92,197 // subps %xmm13,%xmm0 |
| 11797 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 11798 | .byte 69,15,88,201 // addps %xmm9,%xmm9 |
| 11799 | .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9 |
| 11800 | .byte 15,89,206 // mulps %xmm6,%xmm1 |
| 11801 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 11802 | .byte 65,15,84,201 // andps %xmm9,%xmm1 |
| 11803 | .byte 68,15,85,200 // andnps %xmm0,%xmm9 |
| 11804 | .byte 68,15,86,201 // orps %xmm1,%xmm9 |
| 11805 | .byte 69,15,88,207 // addps %xmm15,%xmm9 |
| 11806 | .byte 68,15,89,210 // mulps %xmm2,%xmm10 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 11807 | .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11808 | .byte 69,15,89,229 // mulps %xmm13,%xmm12 |
| 11809 | .byte 69,15,88,226 // addps %xmm10,%xmm12 |
| 11810 | .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| 11811 | .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| 11812 | .byte 68,15,194,211,2 // cmpleps %xmm3,%xmm10 |
| 11813 | .byte 15,40,195 // movaps %xmm3,%xmm0 |
| 11814 | .byte 15,92,194 // subps %xmm2,%xmm0 |
| 11815 | .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| 11816 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 11817 | .byte 69,15,92,245 // subps %xmm13,%xmm14 |
| 11818 | .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| 11819 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 11820 | .byte 65,15,92,230 // subps %xmm14,%xmm4 |
| 11821 | .byte 65,15,84,210 // andps %xmm10,%xmm2 |
| 11822 | .byte 68,15,85,212 // andnps %xmm4,%xmm10 |
| 11823 | .byte 68,15,86,210 // orps %xmm2,%xmm10 |
| 11824 | .byte 69,15,88,212 // addps %xmm12,%xmm10 |
| 11825 | .byte 65,15,88,219 // addps %xmm11,%xmm3 |
| 11826 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11827 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 11828 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 11829 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 11830 | .byte 15,40,229 // movaps %xmm5,%xmm4 |
| 11831 | .byte 15,40,238 // movaps %xmm6,%xmm5 |
| 11832 | .byte 65,15,40,245 // movaps %xmm13,%xmm6 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11833 | .byte 255,224 // jmpq *%rax |
| 11834 | |
| 11835 | HIDDEN _sk_overlay_sse2 |
| 11836 | .globl _sk_overlay_sse2 |
| 11837 | _sk_overlay_sse2: |
| 11838 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11839 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 11840 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 11841 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 11842 | .byte 68,15,92,207 // subps %xmm7,%xmm9 |
| 11843 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 11844 | .byte 68,15,89,208 // mulps %xmm0,%xmm10 |
| 11845 | .byte 68,15,92,195 // subps %xmm3,%xmm8 |
| 11846 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 11847 | .byte 68,15,89,220 // mulps %xmm4,%xmm11 |
| 11848 | .byte 69,15,88,218 // addps %xmm10,%xmm11 |
| 11849 | .byte 68,15,40,227 // movaps %xmm3,%xmm12 |
| 11850 | .byte 68,15,92,224 // subps %xmm0,%xmm12 |
| 11851 | .byte 15,89,196 // mulps %xmm4,%xmm0 |
| 11852 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11853 | .byte 68,15,92,236 // subps %xmm4,%xmm13 |
| 11854 | .byte 68,15,40,244 // movaps %xmm4,%xmm14 |
| 11855 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 11856 | .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14 |
| 11857 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 11858 | .byte 68,15,40,211 // movaps %xmm3,%xmm10 |
| 11859 | .byte 68,15,89,215 // mulps %xmm7,%xmm10 |
| 11860 | .byte 69,15,89,236 // mulps %xmm12,%xmm13 |
| 11861 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 11862 | .byte 69,15,40,226 // movaps %xmm10,%xmm12 |
| 11863 | .byte 69,15,92,229 // subps %xmm13,%xmm12 |
| 11864 | .byte 65,15,84,198 // andps %xmm14,%xmm0 |
| 11865 | .byte 69,15,85,244 // andnps %xmm12,%xmm14 |
| 11866 | .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| 11867 | .byte 65,15,88,195 // addps %xmm11,%xmm0 |
| 11868 | .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| 11869 | .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| 11870 | .byte 69,15,40,224 // movaps %xmm8,%xmm12 |
| 11871 | .byte 68,15,89,229 // mulps %xmm5,%xmm12 |
| 11872 | .byte 69,15,88,227 // addps %xmm11,%xmm12 |
| 11873 | .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| 11874 | .byte 68,15,92,217 // subps %xmm1,%xmm11 |
| 11875 | .byte 15,89,205 // mulps %xmm5,%xmm1 |
| 11876 | .byte 68,15,40,239 // movaps %xmm7,%xmm13 |
| 11877 | .byte 68,15,92,237 // subps %xmm5,%xmm13 |
| 11878 | .byte 68,15,40,245 // movaps %xmm5,%xmm14 |
| 11879 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 11880 | .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14 |
| 11881 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 11882 | .byte 69,15,89,235 // mulps %xmm11,%xmm13 |
| 11883 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 11884 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 11885 | .byte 69,15,92,221 // subps %xmm13,%xmm11 |
| 11886 | .byte 65,15,84,206 // andps %xmm14,%xmm1 |
| 11887 | .byte 69,15,85,243 // andnps %xmm11,%xmm14 |
| 11888 | .byte 65,15,86,206 // orps %xmm14,%xmm1 |
| 11889 | .byte 65,15,88,204 // addps %xmm12,%xmm1 |
| 11890 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 11891 | .byte 69,15,40,216 // movaps %xmm8,%xmm11 |
| 11892 | .byte 68,15,89,222 // mulps %xmm6,%xmm11 |
| 11893 | .byte 69,15,88,217 // addps %xmm9,%xmm11 |
| 11894 | .byte 68,15,40,203 // movaps %xmm3,%xmm9 |
| 11895 | .byte 68,15,92,202 // subps %xmm2,%xmm9 |
| 11896 | .byte 15,89,214 // mulps %xmm6,%xmm2 |
| 11897 | .byte 68,15,40,231 // movaps %xmm7,%xmm12 |
| 11898 | .byte 68,15,92,230 // subps %xmm6,%xmm12 |
| 11899 | .byte 68,15,40,238 // movaps %xmm6,%xmm13 |
| 11900 | .byte 69,15,88,237 // addps %xmm13,%xmm13 |
| 11901 | .byte 68,15,194,239,2 // cmpleps %xmm7,%xmm13 |
| 11902 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 11903 | .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| 11904 | .byte 69,15,88,228 // addps %xmm12,%xmm12 |
| 11905 | .byte 69,15,92,212 // subps %xmm12,%xmm10 |
| 11906 | .byte 65,15,84,213 // andps %xmm13,%xmm2 |
| 11907 | .byte 69,15,85,234 // andnps %xmm10,%xmm13 |
| 11908 | .byte 65,15,86,213 // orps %xmm13,%xmm2 |
| 11909 | .byte 65,15,88,211 // addps %xmm11,%xmm2 |
| 11910 | .byte 68,15,89,199 // mulps %xmm7,%xmm8 |
| 11911 | .byte 65,15,88,216 // addps %xmm8,%xmm3 |
| 11912 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 11913 | .byte 255,224 // jmpq *%rax |
| 11914 | |
| 11915 | HIDDEN _sk_softlight_sse2 |
| 11916 | .globl _sk_softlight_sse2 |
| 11917 | _sk_softlight_sse2: |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 11918 | .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11919 | .byte 15,40,209 // movaps %xmm1,%xmm2 |
| 11920 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 11921 | .byte 69,15,87,228 // xorps %xmm12,%xmm12 |
| 11922 | .byte 68,15,194,231,1 // cmpltps %xmm7,%xmm12 |
| 11923 | .byte 68,15,40,212 // movaps %xmm4,%xmm10 |
| 11924 | .byte 68,15,94,215 // divps %xmm7,%xmm10 |
| 11925 | .byte 69,15,84,212 // andps %xmm12,%xmm10 |
| 11926 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 11927 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 11928 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 11929 | .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| 11930 | .byte 69,15,92,242 // subps %xmm10,%xmm14 |
| 11931 | .byte 69,15,40,218 // movaps %xmm10,%xmm11 |
| 11932 | .byte 69,15,40,234 // movaps %xmm10,%xmm13 |
| 11933 | .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0 |
| 11934 | .byte 68,15,83,248 // rcpps %xmm0,%xmm15 |
| 11935 | .byte 69,15,92,250 // subps %xmm10,%xmm15 |
| 11936 | .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| 11937 | .byte 69,15,88,210 // addps %xmm10,%xmm10 |
| 11938 | .byte 65,15,40,194 // movaps %xmm10,%xmm0 |
| 11939 | .byte 15,89,192 // mulps %xmm0,%xmm0 |
| 11940 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 11941 | .byte 69,15,92,217 // subps %xmm9,%xmm11 |
| 11942 | .byte 68,15,89,216 // mulps %xmm0,%xmm11 |
| 11943 | .byte 184,0,0,224,64 // mov $0x40e00000,%eax |
| 11944 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 11945 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 11946 | .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| 11947 | .byte 69,15,88,235 // addps %xmm11,%xmm13 |
| 11948 | .byte 68,15,40,219 // movaps %xmm3,%xmm11 |
| 11949 | .byte 15,40,204 // movaps %xmm4,%xmm1 |
| 11950 | .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| 11951 | .byte 15,88,228 // addps %xmm4,%xmm4 |
| 11952 | .byte 15,88,228 // addps %xmm4,%xmm4 |
| 11953 | .byte 15,194,231,2 // cmpleps %xmm7,%xmm4 |
| 11954 | .byte 68,15,84,236 // andps %xmm4,%xmm13 |
| 11955 | .byte 65,15,85,231 // andnps %xmm15,%xmm4 |
| 11956 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 11957 | .byte 15,88,192 // addps %xmm0,%xmm0 |
| 11958 | .byte 65,15,86,229 // orps %xmm13,%xmm4 |
| 11959 | .byte 68,15,40,232 // movaps %xmm0,%xmm13 |
| 11960 | .byte 68,15,92,235 // subps %xmm3,%xmm13 |
| 11961 | .byte 69,15,89,245 // mulps %xmm13,%xmm14 |
| 11962 | .byte 68,15,89,239 // mulps %xmm7,%xmm13 |
| 11963 | .byte 65,15,89,229 // mulps %xmm13,%xmm4 |
| 11964 | .byte 65,15,88,227 // addps %xmm11,%xmm4 |
| 11965 | .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| 11966 | .byte 68,15,92,219 // subps %xmm3,%xmm11 |
| 11967 | .byte 69,15,40,251 // movaps %xmm11,%xmm15 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 11968 | .byte 15,41,76,36,216 // movaps %xmm1,-0x28(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11969 | .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| 11970 | .byte 69,15,40,233 // movaps %xmm9,%xmm13 |
| 11971 | .byte 68,15,92,239 // subps %xmm7,%xmm13 |
| 11972 | .byte 69,15,89,197 // mulps %xmm13,%xmm8 |
| 11973 | .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| 11974 | .byte 68,15,88,243 // addps %xmm3,%xmm14 |
| 11975 | .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| 11976 | .byte 15,194,195,2 // cmpleps %xmm3,%xmm0 |
| 11977 | .byte 68,15,84,240 // andps %xmm0,%xmm14 |
| 11978 | .byte 15,85,196 // andnps %xmm4,%xmm0 |
| 11979 | .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| 11980 | .byte 65,15,88,192 // addps %xmm8,%xmm0 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 11981 | .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp) |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 11982 | .byte 68,15,40,197 // movaps %xmm5,%xmm8 |
| 11983 | .byte 68,15,94,199 // divps %xmm7,%xmm8 |
| 11984 | .byte 69,15,84,196 // andps %xmm12,%xmm8 |
| 11985 | .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| 11986 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 11987 | .byte 69,15,88,246 // addps %xmm14,%xmm14 |
| 11988 | .byte 65,15,40,230 // movaps %xmm14,%xmm4 |
| 11989 | .byte 15,89,228 // mulps %xmm4,%xmm4 |
| 11990 | .byte 65,15,88,230 // addps %xmm14,%xmm4 |
| 11991 | .byte 69,15,40,248 // movaps %xmm8,%xmm15 |
| 11992 | .byte 69,15,92,249 // subps %xmm9,%xmm15 |
| 11993 | .byte 68,15,89,252 // mulps %xmm4,%xmm15 |
| 11994 | .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| 11995 | .byte 69,15,92,240 // subps %xmm8,%xmm14 |
| 11996 | .byte 65,15,82,224 // rsqrtps %xmm8,%xmm4 |
| 11997 | .byte 15,83,228 // rcpps %xmm4,%xmm4 |
| 11998 | .byte 65,15,92,224 // subps %xmm8,%xmm4 |
| 11999 | .byte 69,15,89,194 // mulps %xmm10,%xmm8 |
| 12000 | .byte 69,15,88,199 // addps %xmm15,%xmm8 |
| 12001 | .byte 68,15,40,253 // movaps %xmm5,%xmm15 |
| 12002 | .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| 12003 | .byte 69,15,88,255 // addps %xmm15,%xmm15 |
| 12004 | .byte 68,15,194,255,2 // cmpleps %xmm7,%xmm15 |
| 12005 | .byte 69,15,84,199 // andps %xmm15,%xmm8 |
| 12006 | .byte 68,15,85,252 // andnps %xmm4,%xmm15 |
| 12007 | .byte 69,15,86,248 // orps %xmm8,%xmm15 |
| 12008 | .byte 68,15,40,194 // movaps %xmm2,%xmm8 |
| 12009 | .byte 69,15,88,192 // addps %xmm8,%xmm8 |
| 12010 | .byte 65,15,40,224 // movaps %xmm8,%xmm4 |
| 12011 | .byte 15,92,227 // subps %xmm3,%xmm4 |
| 12012 | .byte 68,15,89,244 // mulps %xmm4,%xmm14 |
| 12013 | .byte 15,89,231 // mulps %xmm7,%xmm4 |
| 12014 | .byte 68,15,89,252 // mulps %xmm4,%xmm15 |
| 12015 | .byte 15,40,227 // movaps %xmm3,%xmm4 |
| 12016 | .byte 15,89,229 // mulps %xmm5,%xmm4 |
| 12017 | .byte 68,15,88,252 // addps %xmm4,%xmm15 |
| 12018 | .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| 12019 | .byte 15,89,229 // mulps %xmm5,%xmm4 |
| 12020 | .byte 65,15,89,213 // mulps %xmm13,%xmm2 |
| 12021 | .byte 15,88,212 // addps %xmm4,%xmm2 |
| 12022 | .byte 68,15,88,243 // addps %xmm3,%xmm14 |
| 12023 | .byte 68,15,89,245 // mulps %xmm5,%xmm14 |
| 12024 | .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8 |
| 12025 | .byte 69,15,84,240 // andps %xmm8,%xmm14 |
| 12026 | .byte 69,15,85,199 // andnps %xmm15,%xmm8 |
| 12027 | .byte 69,15,86,198 // orps %xmm14,%xmm8 |
| 12028 | .byte 68,15,88,194 // addps %xmm2,%xmm8 |
| 12029 | .byte 68,15,40,246 // movaps %xmm6,%xmm14 |
| 12030 | .byte 65,15,40,206 // movaps %xmm14,%xmm1 |
| 12031 | .byte 15,94,207 // divps %xmm7,%xmm1 |
| 12032 | .byte 65,15,84,204 // andps %xmm12,%xmm1 |
| 12033 | .byte 15,40,225 // movaps %xmm1,%xmm4 |
| 12034 | .byte 65,15,92,225 // subps %xmm9,%xmm4 |
| 12035 | .byte 68,15,92,201 // subps %xmm1,%xmm9 |
| 12036 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12037 | .byte 15,82,241 // rsqrtps %xmm1,%xmm6 |
| 12038 | .byte 15,83,246 // rcpps %xmm6,%xmm6 |
| 12039 | .byte 15,92,241 // subps %xmm1,%xmm6 |
| 12040 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 12041 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 12042 | .byte 15,40,233 // movaps %xmm1,%xmm5 |
| 12043 | .byte 15,89,237 // mulps %xmm5,%xmm5 |
| 12044 | .byte 15,88,233 // addps %xmm1,%xmm5 |
| 12045 | .byte 15,89,236 // mulps %xmm4,%xmm5 |
| 12046 | .byte 68,15,88,213 // addps %xmm5,%xmm10 |
| 12047 | .byte 65,15,40,238 // movaps %xmm14,%xmm5 |
| 12048 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 12049 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 12050 | .byte 15,88,201 // addps %xmm1,%xmm1 |
| 12051 | .byte 15,194,207,2 // cmpleps %xmm7,%xmm1 |
| 12052 | .byte 68,15,84,209 // andps %xmm1,%xmm10 |
| 12053 | .byte 15,85,206 // andnps %xmm6,%xmm1 |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 12054 | .byte 15,40,84,36,232 // movaps -0x18(%rsp),%xmm2 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 12055 | .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| 12056 | .byte 15,88,210 // addps %xmm2,%xmm2 |
| 12057 | .byte 65,15,86,202 // orps %xmm10,%xmm1 |
| 12058 | .byte 15,40,226 // movaps %xmm2,%xmm4 |
| 12059 | .byte 15,92,227 // subps %xmm3,%xmm4 |
| 12060 | .byte 68,15,89,204 // mulps %xmm4,%xmm9 |
| 12061 | .byte 15,89,231 // mulps %xmm7,%xmm4 |
| 12062 | .byte 15,89,204 // mulps %xmm4,%xmm1 |
| 12063 | .byte 15,40,227 // movaps %xmm3,%xmm4 |
| 12064 | .byte 15,89,229 // mulps %xmm5,%xmm4 |
| 12065 | .byte 15,88,204 // addps %xmm4,%xmm1 |
| 12066 | .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| 12067 | .byte 15,89,229 // mulps %xmm5,%xmm4 |
| 12068 | .byte 65,15,88,229 // addps %xmm13,%xmm4 |
| 12069 | .byte 68,15,88,203 // addps %xmm3,%xmm9 |
| 12070 | .byte 68,15,89,205 // mulps %xmm5,%xmm9 |
| 12071 | .byte 15,40,245 // movaps %xmm5,%xmm6 |
| 12072 | .byte 15,194,211,2 // cmpleps %xmm3,%xmm2 |
| 12073 | .byte 68,15,84,202 // andps %xmm2,%xmm9 |
| 12074 | .byte 15,85,209 // andnps %xmm1,%xmm2 |
| 12075 | .byte 65,15,86,209 // orps %xmm9,%xmm2 |
| 12076 | .byte 15,88,212 // addps %xmm4,%xmm2 |
| 12077 | .byte 68,15,89,223 // mulps %xmm7,%xmm11 |
| 12078 | .byte 65,15,88,219 // addps %xmm11,%xmm3 |
| 12079 | .byte 72,173 // lods %ds:(%rsi),%rax |
| Mike Klein | 15bf678 | 2017-03-31 12:21:46 -0400 | [diff] [blame] | 12080 | .byte 15,40,100,36,216 // movaps -0x28(%rsp),%xmm4 |
| 12081 | .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 12082 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| Mike Klein | 61b8416 | 2017-03-31 11:48:14 -0400 | [diff] [blame] | 12083 | .byte 255,224 // jmpq *%rax |
| 12084 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12085 | HIDDEN _sk_clamp_0_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12086 | .globl _sk_clamp_0_sse2 |
| 12087 | _sk_clamp_0_sse2: |
| 12088 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 12089 | .byte 65,15,95,192 // maxps %xmm8,%xmm0 |
| 12090 | .byte 65,15,95,200 // maxps %xmm8,%xmm1 |
| 12091 | .byte 65,15,95,208 // maxps %xmm8,%xmm2 |
| 12092 | .byte 65,15,95,216 // maxps %xmm8,%xmm3 |
| 12093 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12094 | .byte 255,224 // jmpq *%rax |
| 12095 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12096 | HIDDEN _sk_clamp_1_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12097 | .globl _sk_clamp_1_sse2 |
| 12098 | _sk_clamp_1_sse2: |
| 12099 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12100 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12101 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12102 | .byte 65,15,93,192 // minps %xmm8,%xmm0 |
| 12103 | .byte 65,15,93,200 // minps %xmm8,%xmm1 |
| 12104 | .byte 65,15,93,208 // minps %xmm8,%xmm2 |
| 12105 | .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| 12106 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12107 | .byte 255,224 // jmpq *%rax |
| 12108 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12109 | HIDDEN _sk_clamp_a_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12110 | .globl _sk_clamp_a_sse2 |
| 12111 | _sk_clamp_a_sse2: |
| 12112 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12113 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12114 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12115 | .byte 65,15,93,216 // minps %xmm8,%xmm3 |
| 12116 | .byte 15,93,195 // minps %xmm3,%xmm0 |
| 12117 | .byte 15,93,203 // minps %xmm3,%xmm1 |
| 12118 | .byte 15,93,211 // minps %xmm3,%xmm2 |
| 12119 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12120 | .byte 255,224 // jmpq *%rax |
| 12121 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12122 | HIDDEN _sk_set_rgb_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12123 | .globl _sk_set_rgb_sse2 |
| 12124 | _sk_set_rgb_sse2: |
| 12125 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12126 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 12127 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 12128 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 12129 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 12130 | .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2 |
| 12131 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 12132 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12133 | .byte 255,224 // jmpq *%rax |
| 12134 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12135 | HIDDEN _sk_swap_rb_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12136 | .globl _sk_swap_rb_sse2 |
| 12137 | _sk_swap_rb_sse2: |
| 12138 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 12139 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12140 | .byte 15,40,194 // movaps %xmm2,%xmm0 |
| 12141 | .byte 65,15,40,208 // movaps %xmm8,%xmm2 |
| 12142 | .byte 255,224 // jmpq *%rax |
| 12143 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12144 | HIDDEN _sk_swap_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12145 | .globl _sk_swap_sse2 |
| 12146 | _sk_swap_sse2: |
| 12147 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 12148 | .byte 68,15,40,202 // movaps %xmm2,%xmm9 |
| 12149 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| 12150 | .byte 68,15,40,216 // movaps %xmm0,%xmm11 |
| 12151 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12152 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 12153 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 12154 | .byte 15,40,214 // movaps %xmm6,%xmm2 |
| 12155 | .byte 15,40,223 // movaps %xmm7,%xmm3 |
| 12156 | .byte 65,15,40,227 // movaps %xmm11,%xmm4 |
| 12157 | .byte 65,15,40,234 // movaps %xmm10,%xmm5 |
| 12158 | .byte 65,15,40,241 // movaps %xmm9,%xmm6 |
| 12159 | .byte 65,15,40,248 // movaps %xmm8,%xmm7 |
| 12160 | .byte 255,224 // jmpq *%rax |
| 12161 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12162 | HIDDEN _sk_move_src_dst_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12163 | .globl _sk_move_src_dst_sse2 |
| 12164 | _sk_move_src_dst_sse2: |
| 12165 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12166 | .byte 15,40,224 // movaps %xmm0,%xmm4 |
| 12167 | .byte 15,40,233 // movaps %xmm1,%xmm5 |
| 12168 | .byte 15,40,242 // movaps %xmm2,%xmm6 |
| 12169 | .byte 15,40,251 // movaps %xmm3,%xmm7 |
| 12170 | .byte 255,224 // jmpq *%rax |
| 12171 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12172 | HIDDEN _sk_move_dst_src_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12173 | .globl _sk_move_dst_src_sse2 |
| 12174 | _sk_move_dst_src_sse2: |
| 12175 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12176 | .byte 15,40,196 // movaps %xmm4,%xmm0 |
| 12177 | .byte 15,40,205 // movaps %xmm5,%xmm1 |
| 12178 | .byte 15,40,214 // movaps %xmm6,%xmm2 |
| 12179 | .byte 15,40,223 // movaps %xmm7,%xmm3 |
| 12180 | .byte 255,224 // jmpq *%rax |
| 12181 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12182 | HIDDEN _sk_premul_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12183 | .globl _sk_premul_sse2 |
| 12184 | _sk_premul_sse2: |
| 12185 | .byte 15,89,195 // mulps %xmm3,%xmm0 |
| 12186 | .byte 15,89,203 // mulps %xmm3,%xmm1 |
| 12187 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 12188 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12189 | .byte 255,224 // jmpq *%rax |
| 12190 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12191 | HIDDEN _sk_unpremul_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12192 | .globl _sk_unpremul_sse2 |
| 12193 | _sk_unpremul_sse2: |
| 12194 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 12195 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12196 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 12197 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12198 | .byte 68,15,94,203 // divps %xmm3,%xmm9 |
| 12199 | .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8 |
| 12200 | .byte 69,15,84,193 // andps %xmm9,%xmm8 |
| 12201 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 12202 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 12203 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12204 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12205 | .byte 255,224 // jmpq *%rax |
| 12206 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12207 | HIDDEN _sk_from_srgb_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12208 | .globl _sk_from_srgb_sse2 |
| 12209 | _sk_from_srgb_sse2: |
| 12210 | .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax |
| 12211 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12212 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12213 | .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| 12214 | .byte 68,15,89,232 // mulps %xmm0,%xmm13 |
| 12215 | .byte 68,15,40,224 // movaps %xmm0,%xmm12 |
| 12216 | .byte 69,15,89,228 // mulps %xmm12,%xmm12 |
| 12217 | .byte 184,154,153,153,62 // mov $0x3e99999a,%eax |
| 12218 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 12219 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12220 | .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax |
| 12221 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 12222 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 12223 | .byte 69,15,40,241 // movaps %xmm9,%xmm14 |
| 12224 | .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| 12225 | .byte 69,15,88,242 // addps %xmm10,%xmm14 |
| 12226 | .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax |
| 12227 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 12228 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 12229 | .byte 69,15,89,244 // mulps %xmm12,%xmm14 |
| 12230 | .byte 69,15,88,243 // addps %xmm11,%xmm14 |
| 12231 | .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax |
| 12232 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 12233 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 12234 | .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0 |
| 12235 | .byte 68,15,84,232 // andps %xmm0,%xmm13 |
| 12236 | .byte 65,15,85,198 // andnps %xmm14,%xmm0 |
| 12237 | .byte 65,15,86,197 // orps %xmm13,%xmm0 |
| 12238 | .byte 69,15,40,232 // movaps %xmm8,%xmm13 |
| 12239 | .byte 68,15,89,233 // mulps %xmm1,%xmm13 |
| 12240 | .byte 68,15,40,241 // movaps %xmm1,%xmm14 |
| 12241 | .byte 69,15,89,246 // mulps %xmm14,%xmm14 |
| 12242 | .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| 12243 | .byte 68,15,89,249 // mulps %xmm1,%xmm15 |
| 12244 | .byte 69,15,88,250 // addps %xmm10,%xmm15 |
| 12245 | .byte 69,15,89,254 // mulps %xmm14,%xmm15 |
| 12246 | .byte 69,15,88,251 // addps %xmm11,%xmm15 |
| 12247 | .byte 65,15,194,204,1 // cmpltps %xmm12,%xmm1 |
| 12248 | .byte 68,15,84,233 // andps %xmm1,%xmm13 |
| 12249 | .byte 65,15,85,207 // andnps %xmm15,%xmm1 |
| 12250 | .byte 65,15,86,205 // orps %xmm13,%xmm1 |
| 12251 | .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| 12252 | .byte 68,15,40,234 // movaps %xmm2,%xmm13 |
| 12253 | .byte 69,15,89,237 // mulps %xmm13,%xmm13 |
| 12254 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 12255 | .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| 12256 | .byte 69,15,89,205 // mulps %xmm13,%xmm9 |
| 12257 | .byte 69,15,88,203 // addps %xmm11,%xmm9 |
| 12258 | .byte 65,15,194,212,1 // cmpltps %xmm12,%xmm2 |
| 12259 | .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| 12260 | .byte 65,15,85,209 // andnps %xmm9,%xmm2 |
| 12261 | .byte 65,15,86,208 // orps %xmm8,%xmm2 |
| 12262 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12263 | .byte 255,224 // jmpq *%rax |
| 12264 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12265 | HIDDEN _sk_to_srgb_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12266 | .globl _sk_to_srgb_sse2 |
| 12267 | _sk_to_srgb_sse2: |
| 12268 | .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8 |
| 12269 | .byte 69,15,83,248 // rcpps %xmm8,%xmm15 |
| 12270 | .byte 69,15,82,232 // rsqrtps %xmm8,%xmm13 |
| 12271 | .byte 184,41,92,71,65 // mov $0x41475c29,%eax |
| 12272 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12273 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12274 | .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| 12275 | .byte 68,15,89,240 // mulps %xmm0,%xmm14 |
| 12276 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12277 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 12278 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12279 | .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax |
| 12280 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 12281 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 12282 | .byte 184,206,111,48,63 // mov $0x3f306fce,%eax |
| 12283 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 12284 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 12285 | .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax |
| 12286 | .byte 53,0,0,0,128 // xor $0x80000000,%eax |
| 12287 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 12288 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 12289 | .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| 12290 | .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| 12291 | .byte 69,15,89,234 // mulps %xmm10,%xmm13 |
| 12292 | .byte 69,15,88,239 // addps %xmm15,%xmm13 |
| 12293 | .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| 12294 | .byte 69,15,93,253 // minps %xmm13,%xmm15 |
| 12295 | .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax |
| 12296 | .byte 102,68,15,110,232 // movd %eax,%xmm13 |
| 12297 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 12298 | .byte 65,15,194,197,1 // cmpltps %xmm13,%xmm0 |
| 12299 | .byte 68,15,84,240 // andps %xmm0,%xmm14 |
| 12300 | .byte 65,15,85,199 // andnps %xmm15,%xmm0 |
| 12301 | .byte 65,15,86,198 // orps %xmm14,%xmm0 |
| 12302 | .byte 68,15,82,241 // rsqrtps %xmm1,%xmm14 |
| 12303 | .byte 69,15,83,254 // rcpps %xmm14,%xmm15 |
| 12304 | .byte 69,15,82,246 // rsqrtps %xmm14,%xmm14 |
| 12305 | .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| 12306 | .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| 12307 | .byte 69,15,89,242 // mulps %xmm10,%xmm14 |
| 12308 | .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| 12309 | .byte 69,15,40,249 // movaps %xmm9,%xmm15 |
| 12310 | .byte 69,15,93,254 // minps %xmm14,%xmm15 |
| 12311 | .byte 69,15,40,240 // movaps %xmm8,%xmm14 |
| 12312 | .byte 68,15,89,241 // mulps %xmm1,%xmm14 |
| 12313 | .byte 65,15,194,205,1 // cmpltps %xmm13,%xmm1 |
| 12314 | .byte 68,15,84,241 // andps %xmm1,%xmm14 |
| 12315 | .byte 65,15,85,207 // andnps %xmm15,%xmm1 |
| 12316 | .byte 65,15,86,206 // orps %xmm14,%xmm1 |
| 12317 | .byte 68,15,82,242 // rsqrtps %xmm2,%xmm14 |
| 12318 | .byte 69,15,83,254 // rcpps %xmm14,%xmm15 |
| 12319 | .byte 69,15,89,251 // mulps %xmm11,%xmm15 |
| 12320 | .byte 69,15,88,252 // addps %xmm12,%xmm15 |
| 12321 | .byte 69,15,82,222 // rsqrtps %xmm14,%xmm11 |
| 12322 | .byte 69,15,89,218 // mulps %xmm10,%xmm11 |
| 12323 | .byte 69,15,88,223 // addps %xmm15,%xmm11 |
| 12324 | .byte 69,15,93,203 // minps %xmm11,%xmm9 |
| 12325 | .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| 12326 | .byte 65,15,194,213,1 // cmpltps %xmm13,%xmm2 |
| 12327 | .byte 68,15,84,194 // andps %xmm2,%xmm8 |
| 12328 | .byte 65,15,85,209 // andnps %xmm9,%xmm2 |
| 12329 | .byte 65,15,86,208 // orps %xmm8,%xmm2 |
| 12330 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12331 | .byte 255,224 // jmpq *%rax |
| 12332 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12333 | HIDDEN _sk_scale_1_float_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12334 | .globl _sk_scale_1_float_sse2 |
| 12335 | _sk_scale_1_float_sse2: |
| 12336 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12337 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 12338 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12339 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 12340 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 12341 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12342 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12343 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12344 | .byte 255,224 // jmpq *%rax |
| 12345 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12346 | HIDDEN _sk_scale_u8_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12347 | .globl _sk_scale_u8_sse2 |
| 12348 | _sk_scale_u8_sse2: |
| 12349 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12350 | .byte 72,139,0 // mov (%rax),%rax |
| 12351 | .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8 |
| 12352 | .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| 12353 | .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 |
| 12354 | .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| 12355 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 12356 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12357 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 12358 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12359 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 12360 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 12361 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 12362 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 12363 | .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| 12364 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12365 | .byte 255,224 // jmpq *%rax |
| 12366 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12367 | HIDDEN _sk_lerp_1_float_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12368 | .globl _sk_lerp_1_float_sse2 |
| 12369 | _sk_lerp_1_float_sse2: |
| 12370 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12371 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 12372 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12373 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 12374 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 12375 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 12376 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 12377 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 12378 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 12379 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 12380 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12381 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 12382 | .byte 15,92,223 // subps %xmm7,%xmm3 |
| 12383 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12384 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 12385 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12386 | .byte 255,224 // jmpq *%rax |
| 12387 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12388 | HIDDEN _sk_lerp_u8_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12389 | .globl _sk_lerp_u8_sse2 |
| 12390 | _sk_lerp_u8_sse2: |
| 12391 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12392 | .byte 72,139,0 // mov (%rax),%rax |
| 12393 | .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8 |
| 12394 | .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| 12395 | .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8 |
| 12396 | .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8 |
| 12397 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 12398 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12399 | .byte 102,68,15,110,200 // movd %eax,%xmm9 |
| 12400 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12401 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 12402 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 12403 | .byte 65,15,89,193 // mulps %xmm9,%xmm0 |
| 12404 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 12405 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 12406 | .byte 65,15,89,201 // mulps %xmm9,%xmm1 |
| 12407 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 12408 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 12409 | .byte 65,15,89,209 // mulps %xmm9,%xmm2 |
| 12410 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 12411 | .byte 15,92,223 // subps %xmm7,%xmm3 |
| 12412 | .byte 65,15,89,217 // mulps %xmm9,%xmm3 |
| 12413 | .byte 15,88,223 // addps %xmm7,%xmm3 |
| 12414 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12415 | .byte 255,224 // jmpq *%rax |
| 12416 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12417 | HIDDEN _sk_lerp_565_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12418 | .globl _sk_lerp_565_sse2 |
| 12419 | _sk_lerp_565_sse2: |
| 12420 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12421 | .byte 72,139,0 // mov (%rax),%rax |
| 12422 | .byte 243,68,15,126,4,120 // movq (%rax,%rdi,2),%xmm8 |
| 12423 | .byte 102,15,239,219 // pxor %xmm3,%xmm3 |
| 12424 | .byte 102,68,15,97,195 // punpcklwd %xmm3,%xmm8 |
| 12425 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 12426 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12427 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 12428 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 12429 | .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| 12430 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 12431 | .byte 102,68,15,110,208 // movd %eax,%xmm10 |
| 12432 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 12433 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 12434 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 12435 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12436 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 12437 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 12438 | .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9 |
| 12439 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 12440 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 12441 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 12442 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 12443 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 12444 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12445 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 12446 | .byte 102,65,15,219,216 // pand %xmm8,%xmm3 |
| 12447 | .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| 12448 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 12449 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12450 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12451 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12452 | .byte 15,92,196 // subps %xmm4,%xmm0 |
| 12453 | .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| 12454 | .byte 15,88,196 // addps %xmm4,%xmm0 |
| 12455 | .byte 15,92,205 // subps %xmm5,%xmm1 |
| 12456 | .byte 65,15,89,203 // mulps %xmm11,%xmm1 |
| 12457 | .byte 15,88,205 // addps %xmm5,%xmm1 |
| 12458 | .byte 15,92,214 // subps %xmm6,%xmm2 |
| 12459 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 12460 | .byte 15,88,214 // addps %xmm6,%xmm2 |
| 12461 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12462 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12463 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12464 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12465 | .byte 255,224 // jmpq *%rax |
| 12466 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12467 | HIDDEN _sk_load_tables_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12468 | .globl _sk_load_tables_sse2 |
| 12469 | _sk_load_tables_sse2: |
| 12470 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12471 | .byte 72,139,8 // mov (%rax),%rcx |
| 12472 | .byte 76,139,64,8 // mov 0x8(%rax),%r8 |
| 12473 | .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8 |
| 12474 | .byte 185,255,0,0,0 // mov $0xff,%ecx |
| 12475 | .byte 102,15,110,193 // movd %ecx,%xmm0 |
| 12476 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 12477 | .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| 12478 | .byte 102,65,15,114,209,8 // psrld $0x8,%xmm9 |
| 12479 | .byte 102,68,15,219,200 // pand %xmm0,%xmm9 |
| 12480 | .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| 12481 | .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10 |
| 12482 | .byte 102,68,15,219,208 // pand %xmm0,%xmm10 |
| 12483 | .byte 102,65,15,219,192 // pand %xmm8,%xmm0 |
| 12484 | .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3 |
| 12485 | .byte 102,72,15,126,217 // movq %xmm3,%rcx |
| 12486 | .byte 65,137,201 // mov %ecx,%r9d |
| 12487 | .byte 72,193,233,32 // shr $0x20,%rcx |
| 12488 | .byte 102,73,15,126,194 // movq %xmm0,%r10 |
| 12489 | .byte 69,137,211 // mov %r10d,%r11d |
| 12490 | .byte 73,193,234,32 // shr $0x20,%r10 |
| 12491 | .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3 |
| 12492 | .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0 |
| 12493 | .byte 15,20,216 // unpcklps %xmm0,%xmm3 |
| 12494 | .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0 |
| 12495 | .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1 |
| 12496 | .byte 15,20,193 // unpcklps %xmm1,%xmm0 |
| 12497 | .byte 15,20,195 // unpcklps %xmm3,%xmm0 |
| 12498 | .byte 76,139,64,16 // mov 0x10(%rax),%r8 |
| 12499 | .byte 102,65,15,112,201,78 // pshufd $0x4e,%xmm9,%xmm1 |
| 12500 | .byte 102,73,15,126,202 // movq %xmm1,%r10 |
| 12501 | .byte 77,137,209 // mov %r10,%r9 |
| 12502 | .byte 73,193,233,32 // shr $0x20,%r9 |
| 12503 | .byte 102,76,15,126,201 // movq %xmm9,%rcx |
| 12504 | .byte 65,137,203 // mov %ecx,%r11d |
| 12505 | .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d |
| 12506 | .byte 72,193,233,30 // shr $0x1e,%rcx |
| 12507 | .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d |
| 12508 | .byte 243,65,15,16,28,8 // movss (%r8,%rcx,1),%xmm3 |
| 12509 | .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1 |
| 12510 | .byte 15,20,217 // unpcklps %xmm1,%xmm3 |
| 12511 | .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1 |
| 12512 | .byte 243,67,15,16,20,144 // movss (%r8,%r10,4),%xmm2 |
| 12513 | .byte 15,20,202 // unpcklps %xmm2,%xmm1 |
| 12514 | .byte 15,20,203 // unpcklps %xmm3,%xmm1 |
| 12515 | .byte 76,139,72,24 // mov 0x18(%rax),%r9 |
| 12516 | .byte 102,65,15,112,210,78 // pshufd $0x4e,%xmm10,%xmm2 |
| 12517 | .byte 102,72,15,126,209 // movq %xmm2,%rcx |
| 12518 | .byte 68,15,183,193 // movzwl %cx,%r8d |
| 12519 | .byte 72,193,233,32 // shr $0x20,%rcx |
| 12520 | .byte 102,76,15,126,208 // movq %xmm10,%rax |
| 12521 | .byte 68,15,183,208 // movzwl %ax,%r10d |
| 12522 | .byte 72,193,232,30 // shr $0x1e,%rax |
| 12523 | .byte 243,69,15,16,12,1 // movss (%r9,%rax,1),%xmm9 |
| 12524 | .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2 |
| 12525 | .byte 68,15,20,202 // unpcklps %xmm2,%xmm9 |
| 12526 | .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2 |
| 12527 | .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3 |
| 12528 | .byte 15,20,211 // unpcklps %xmm3,%xmm2 |
| 12529 | .byte 65,15,20,209 // unpcklps %xmm9,%xmm2 |
| 12530 | .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8 |
| 12531 | .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8 |
| 12532 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12533 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12534 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12535 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12536 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12537 | .byte 255,224 // jmpq *%rax |
| 12538 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12539 | HIDDEN _sk_load_a8_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12540 | .globl _sk_load_a8_sse2 |
| 12541 | _sk_load_a8_sse2: |
| 12542 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12543 | .byte 72,139,0 // mov (%rax),%rax |
| 12544 | .byte 102,15,110,4,56 // movd (%rax,%rdi,1),%xmm0 |
| 12545 | .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| 12546 | .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 |
| 12547 | .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 |
| 12548 | .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| 12549 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12550 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12551 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12552 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 12553 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12554 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 12555 | .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| 12556 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 12557 | .byte 255,224 // jmpq *%rax |
| 12558 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12559 | HIDDEN _sk_store_a8_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12560 | .globl _sk_store_a8_sse2 |
| 12561 | _sk_store_a8_sse2: |
| 12562 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12563 | .byte 72,139,0 // mov (%rax),%rax |
| 12564 | .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| 12565 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 12566 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12567 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 12568 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 12569 | .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| 12570 | .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8 |
| 12571 | .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| 12572 | .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8 |
| 12573 | .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1) |
| 12574 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12575 | .byte 255,224 // jmpq *%rax |
| 12576 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 12577 | HIDDEN _sk_load_g8_sse2 |
| 12578 | .globl _sk_load_g8_sse2 |
| 12579 | _sk_load_g8_sse2: |
| 12580 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12581 | .byte 72,139,0 // mov (%rax),%rax |
| 12582 | .byte 102,15,110,4,56 // movd (%rax,%rdi,1),%xmm0 |
| 12583 | .byte 102,15,239,201 // pxor %xmm1,%xmm1 |
| 12584 | .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0 |
| 12585 | .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0 |
| 12586 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 12587 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12588 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12589 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 12590 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 12591 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12592 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12593 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12594 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12595 | .byte 15,40,200 // movaps %xmm0,%xmm1 |
| 12596 | .byte 15,40,208 // movaps %xmm0,%xmm2 |
| 12597 | .byte 255,224 // jmpq *%rax |
| 12598 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12599 | HIDDEN _sk_load_565_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12600 | .globl _sk_load_565_sse2 |
| 12601 | _sk_load_565_sse2: |
| 12602 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12603 | .byte 72,139,0 // mov (%rax),%rax |
| 12604 | .byte 243,15,126,20,120 // movq (%rax,%rdi,2),%xmm2 |
| 12605 | .byte 102,15,239,192 // pxor %xmm0,%xmm0 |
| 12606 | .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2 |
| 12607 | .byte 184,0,248,0,0 // mov $0xf800,%eax |
| 12608 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12609 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 12610 | .byte 102,15,219,194 // pand %xmm2,%xmm0 |
| 12611 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 12612 | .byte 184,8,33,132,55 // mov $0x37842108,%eax |
| 12613 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12614 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 12615 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 12616 | .byte 184,224,7,0,0 // mov $0x7e0,%eax |
| 12617 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 12618 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 12619 | .byte 102,15,219,202 // pand %xmm2,%xmm1 |
| 12620 | .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3 |
| 12621 | .byte 184,33,8,2,58 // mov $0x3a020821,%eax |
| 12622 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 12623 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 12624 | .byte 15,89,203 // mulps %xmm3,%xmm1 |
| 12625 | .byte 184,31,0,0,0 // mov $0x1f,%eax |
| 12626 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12627 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 12628 | .byte 102,15,219,218 // pand %xmm2,%xmm3 |
| 12629 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 12630 | .byte 184,8,33,4,61 // mov $0x3d042108,%eax |
| 12631 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 12632 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 12633 | .byte 15,89,211 // mulps %xmm3,%xmm2 |
| 12634 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 12635 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12636 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12637 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12638 | .byte 255,224 // jmpq *%rax |
| 12639 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12640 | HIDDEN _sk_store_565_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12641 | .globl _sk_store_565_sse2 |
| 12642 | _sk_store_565_sse2: |
| 12643 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12644 | .byte 72,139,0 // mov (%rax),%rax |
| 12645 | .byte 185,0,0,248,65 // mov $0x41f80000,%ecx |
| 12646 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 12647 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12648 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 12649 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 12650 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12651 | .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9 |
| 12652 | .byte 185,0,0,124,66 // mov $0x427c0000,%ecx |
| 12653 | .byte 102,68,15,110,209 // movd %ecx,%xmm10 |
| 12654 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 12655 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12656 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 12657 | .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10 |
| 12658 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 12659 | .byte 68,15,89,194 // mulps %xmm2,%xmm8 |
| 12660 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 12661 | .byte 102,69,15,86,194 // orpd %xmm10,%xmm8 |
| 12662 | .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| 12663 | .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8 |
| 12664 | .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| 12665 | .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2) |
| 12666 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12667 | .byte 255,224 // jmpq *%rax |
| 12668 | |
| Mike Klein | f809fef | 2017-03-31 13:52:45 -0400 | [diff] [blame] | 12669 | HIDDEN _sk_load_4444_sse2 |
| 12670 | .globl _sk_load_4444_sse2 |
| 12671 | _sk_load_4444_sse2: |
| 12672 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12673 | .byte 72,139,0 // mov (%rax),%rax |
| 12674 | .byte 243,68,15,126,12,120 // movq (%rax,%rdi,2),%xmm9 |
| 12675 | .byte 102,15,239,192 // pxor %xmm0,%xmm0 |
| 12676 | .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9 |
| 12677 | .byte 184,0,240,0,0 // mov $0xf000,%eax |
| 12678 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12679 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 12680 | .byte 102,65,15,219,193 // pand %xmm9,%xmm0 |
| 12681 | .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1 |
| 12682 | .byte 184,137,136,136,55 // mov $0x37888889,%eax |
| 12683 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12684 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 12685 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 12686 | .byte 184,0,15,0,0 // mov $0xf00,%eax |
| 12687 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 12688 | .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1 |
| 12689 | .byte 102,65,15,219,201 // pand %xmm9,%xmm1 |
| 12690 | .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2 |
| 12691 | .byte 184,137,136,136,57 // mov $0x39888889,%eax |
| 12692 | .byte 102,15,110,200 // movd %eax,%xmm1 |
| 12693 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 12694 | .byte 15,89,202 // mulps %xmm2,%xmm1 |
| 12695 | .byte 184,240,0,0,0 // mov $0xf0,%eax |
| 12696 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 12697 | .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2 |
| 12698 | .byte 102,65,15,219,209 // pand %xmm9,%xmm2 |
| 12699 | .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8 |
| 12700 | .byte 184,137,136,136,59 // mov $0x3b888889,%eax |
| 12701 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| 12702 | .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2 |
| 12703 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12704 | .byte 184,15,0,0,0 // mov $0xf,%eax |
| 12705 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12706 | .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3 |
| 12707 | .byte 102,65,15,219,217 // pand %xmm9,%xmm3 |
| 12708 | .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8 |
| 12709 | .byte 184,137,136,136,61 // mov $0x3d888889,%eax |
| 12710 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 12711 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 12712 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12713 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12714 | .byte 255,224 // jmpq *%rax |
| 12715 | |
| 12716 | HIDDEN _sk_store_4444_sse2 |
| 12717 | .globl _sk_store_4444_sse2 |
| 12718 | _sk_store_4444_sse2: |
| 12719 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12720 | .byte 72,139,0 // mov (%rax),%rax |
| 12721 | .byte 185,0,0,112,65 // mov $0x41700000,%ecx |
| 12722 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 12723 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12724 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 12725 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 12726 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12727 | .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9 |
| 12728 | .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| 12729 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12730 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 12731 | .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10 |
| 12732 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 12733 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 12734 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 12735 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12736 | .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9 |
| 12737 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 12738 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 12739 | .byte 102,69,15,86,193 // orpd %xmm9,%xmm8 |
| 12740 | .byte 102,69,15,86,194 // orpd %xmm10,%xmm8 |
| 12741 | .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| 12742 | .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8 |
| 12743 | .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| 12744 | .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2) |
| 12745 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12746 | .byte 255,224 // jmpq *%rax |
| 12747 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12748 | HIDDEN _sk_load_8888_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12749 | .globl _sk_load_8888_sse2 |
| 12750 | _sk_load_8888_sse2: |
| 12751 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12752 | .byte 72,139,0 // mov (%rax),%rax |
| 12753 | .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3 |
| 12754 | .byte 184,255,0,0,0 // mov $0xff,%eax |
| 12755 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12756 | .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0 |
| 12757 | .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| 12758 | .byte 102,15,114,209,8 // psrld $0x8,%xmm1 |
| 12759 | .byte 102,15,219,200 // pand %xmm0,%xmm1 |
| 12760 | .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| 12761 | .byte 102,15,114,210,16 // psrld $0x10,%xmm2 |
| 12762 | .byte 102,15,219,208 // pand %xmm0,%xmm2 |
| 12763 | .byte 102,15,219,195 // pand %xmm3,%xmm0 |
| 12764 | .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0 |
| 12765 | .byte 184,129,128,128,59 // mov $0x3b808081,%eax |
| 12766 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12767 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12768 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 12769 | .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1 |
| 12770 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 12771 | .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| 12772 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12773 | .byte 102,15,114,211,24 // psrld $0x18,%xmm3 |
| 12774 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 12775 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12776 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12777 | .byte 255,224 // jmpq *%rax |
| 12778 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12779 | HIDDEN _sk_store_8888_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12780 | .globl _sk_store_8888_sse2 |
| 12781 | _sk_store_8888_sse2: |
| 12782 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12783 | .byte 72,139,0 // mov (%rax),%rax |
| 12784 | .byte 185,0,0,127,67 // mov $0x437f0000,%ecx |
| 12785 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| 12786 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12787 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 12788 | .byte 68,15,89,200 // mulps %xmm0,%xmm9 |
| 12789 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12790 | .byte 69,15,40,208 // movaps %xmm8,%xmm10 |
| 12791 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12792 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 12793 | .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10 |
| 12794 | .byte 102,69,15,235,209 // por %xmm9,%xmm10 |
| 12795 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 12796 | .byte 68,15,89,202 // mulps %xmm2,%xmm9 |
| 12797 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12798 | .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| 12799 | .byte 68,15,89,195 // mulps %xmm3,%xmm8 |
| 12800 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 12801 | .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8 |
| 12802 | .byte 102,69,15,235,193 // por %xmm9,%xmm8 |
| 12803 | .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| 12804 | .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4) |
| 12805 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12806 | .byte 255,224 // jmpq *%rax |
| 12807 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12808 | HIDDEN _sk_load_f16_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12809 | .globl _sk_load_f16_sse2 |
| 12810 | _sk_load_f16_sse2: |
| 12811 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12812 | .byte 72,139,0 // mov (%rax),%rax |
| 12813 | .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| 12814 | .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12815 | .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8 |
| 12816 | .byte 102,68,15,97,193 // punpcklwd %xmm1,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12817 | .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12818 | .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1 |
| 12819 | .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| 12820 | .byte 102,68,15,105,192 // punpckhwd %xmm0,%xmm8 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12821 | .byte 184,0,4,0,4 // mov $0x4000400,%eax |
| 12822 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 12823 | .byte 102,15,112,216,0 // pshufd $0x0,%xmm0,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12824 | .byte 102,15,111,195 // movdqa %xmm3,%xmm0 |
| 12825 | .byte 102,15,101,193 // pcmpgtw %xmm1,%xmm0 |
| 12826 | .byte 102,15,223,193 // pandn %xmm1,%xmm0 |
| 12827 | .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| 12828 | .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12829 | .byte 102,15,114,240,13 // pslld $0xd,%xmm0 |
| 12830 | .byte 184,0,0,128,119 // mov $0x77800000,%eax |
| 12831 | .byte 102,15,110,208 // movd %eax,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12832 | .byte 102,68,15,112,210,0 // pshufd $0x0,%xmm2,%xmm10 |
| 12833 | .byte 65,15,89,194 // mulps %xmm10,%xmm0 |
| 12834 | .byte 102,15,112,209,78 // pshufd $0x4e,%xmm1,%xmm2 |
| 12835 | .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| 12836 | .byte 102,15,101,202 // pcmpgtw %xmm2,%xmm1 |
| 12837 | .byte 102,15,223,202 // pandn %xmm2,%xmm1 |
| 12838 | .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12839 | .byte 102,15,114,241,13 // pslld $0xd,%xmm1 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12840 | .byte 65,15,89,202 // mulps %xmm10,%xmm1 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12841 | .byte 102,15,111,211 // movdqa %xmm3,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12842 | .byte 102,65,15,101,208 // pcmpgtw %xmm8,%xmm2 |
| 12843 | .byte 102,65,15,223,208 // pandn %xmm8,%xmm2 |
| 12844 | .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12845 | .byte 102,15,114,242,13 // pslld $0xd,%xmm2 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12846 | .byte 65,15,89,210 // mulps %xmm10,%xmm2 |
| 12847 | .byte 102,69,15,112,192,78 // pshufd $0x4e,%xmm8,%xmm8 |
| 12848 | .byte 102,65,15,101,216 // pcmpgtw %xmm8,%xmm3 |
| 12849 | .byte 102,65,15,223,216 // pandn %xmm8,%xmm3 |
| 12850 | .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12851 | .byte 102,15,114,243,13 // pslld $0xd,%xmm3 |
| Mike Klein | 114e6b3 | 2017-04-03 22:21:15 -0400 | [diff] [blame] | 12852 | .byte 65,15,89,218 // mulps %xmm10,%xmm3 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12853 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12854 | .byte 255,224 // jmpq *%rax |
| 12855 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 12856 | HIDDEN _sk_store_f16_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12857 | .globl _sk_store_f16_sse2 |
| 12858 | _sk_store_f16_sse2: |
| 12859 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12860 | .byte 72,139,0 // mov (%rax),%rax |
| 12861 | .byte 185,0,0,128,7 // mov $0x7800000,%ecx |
| 12862 | .byte 102,68,15,110,193 // movd %ecx,%xmm8 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 12863 | .byte 102,69,15,112,200,0 // pshufd $0x0,%xmm8,%xmm9 |
| 12864 | .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8 |
| 12865 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 12866 | .byte 102,65,15,114,208,13 // psrld $0xd,%xmm8 |
| 12867 | .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| 12868 | .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8 |
| 12869 | .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| 12870 | .byte 102,69,15,111,209 // movdqa %xmm9,%xmm10 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12871 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12872 | .byte 102,65,15,114,210,13 // psrld $0xd,%xmm10 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 12873 | .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 |
| 12874 | .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10 |
| 12875 | .byte 102,69,15,107,210 // packssdw %xmm10,%xmm10 |
| 12876 | .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12877 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 12878 | .byte 102,65,15,114,211,13 // psrld $0xd,%xmm11 |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 12879 | .byte 102,65,15,114,243,16 // pslld $0x10,%xmm11 |
| 12880 | .byte 102,65,15,114,227,16 // psrad $0x10,%xmm11 |
| 12881 | .byte 102,69,15,107,219 // packssdw %xmm11,%xmm11 |
| 12882 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 12883 | .byte 102,65,15,114,209,13 // psrld $0xd,%xmm9 |
| 12884 | .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| 12885 | .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9 |
| 12886 | .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9 |
| 12887 | .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| 12888 | .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| 12889 | .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| 12890 | .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12891 | .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| Mike Klein | 95f53be | 2017-04-04 10:24:56 -0400 | [diff] [blame] | 12892 | .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| 12893 | .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 12894 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12895 | .byte 255,224 // jmpq *%rax |
| 12896 | |
| Mike Klein | 3146bb9 | 2017-04-05 14:45:02 -0400 | [diff] [blame^] | 12897 | HIDDEN _sk_load_u16_be_sse2 |
| 12898 | .globl _sk_load_u16_be_sse2 |
| 12899 | _sk_load_u16_be_sse2: |
| 12900 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12901 | .byte 72,139,0 // mov (%rax),%rax |
| 12902 | .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0 |
| 12903 | .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1 |
| 12904 | .byte 102,15,111,208 // movdqa %xmm0,%xmm2 |
| 12905 | .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2 |
| 12906 | .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0 |
| 12907 | .byte 102,15,111,202 // movdqa %xmm2,%xmm1 |
| 12908 | .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1 |
| 12909 | .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2 |
| 12910 | .byte 184,128,0,128,55 // mov $0x37800080,%eax |
| 12911 | .byte 102,68,15,110,192 // movd %eax,%xmm8 |
| 12912 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 12913 | .byte 102,15,111,193 // movdqa %xmm1,%xmm0 |
| 12914 | .byte 102,15,113,240,8 // psllw $0x8,%xmm0 |
| 12915 | .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3 |
| 12916 | .byte 102,15,113,209,8 // psrlw $0x8,%xmm1 |
| 12917 | .byte 102,15,235,200 // por %xmm0,%xmm1 |
| 12918 | .byte 102,69,15,239,201 // pxor %xmm9,%xmm9 |
| 12919 | .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1 |
| 12920 | .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0 |
| 12921 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 12922 | .byte 102,15,111,203 // movdqa %xmm3,%xmm1 |
| 12923 | .byte 102,15,113,241,8 // psllw $0x8,%xmm1 |
| 12924 | .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| 12925 | .byte 102,15,235,217 // por %xmm1,%xmm3 |
| 12926 | .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| 12927 | .byte 15,91,203 // cvtdq2ps %xmm3,%xmm1 |
| 12928 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 12929 | .byte 102,68,15,111,210 // movdqa %xmm2,%xmm10 |
| 12930 | .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| 12931 | .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3 |
| 12932 | .byte 102,15,113,210,8 // psrlw $0x8,%xmm2 |
| 12933 | .byte 102,65,15,235,210 // por %xmm10,%xmm2 |
| 12934 | .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2 |
| 12935 | .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2 |
| 12936 | .byte 65,15,89,208 // mulps %xmm8,%xmm2 |
| 12937 | .byte 102,68,15,111,211 // movdqa %xmm3,%xmm10 |
| 12938 | .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| 12939 | .byte 102,15,113,211,8 // psrlw $0x8,%xmm3 |
| 12940 | .byte 102,65,15,235,218 // por %xmm10,%xmm3 |
| 12941 | .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3 |
| 12942 | .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3 |
| 12943 | .byte 65,15,89,216 // mulps %xmm8,%xmm3 |
| 12944 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12945 | .byte 255,224 // jmpq *%rax |
| 12946 | |
| 12947 | HIDDEN _sk_store_u16_be_sse2 |
| 12948 | .globl _sk_store_u16_be_sse2 |
| 12949 | _sk_store_u16_be_sse2: |
| 12950 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 12951 | .byte 72,139,0 // mov (%rax),%rax |
| 12952 | .byte 185,0,255,127,71 // mov $0x477fff00,%ecx |
| 12953 | .byte 102,68,15,110,201 // movd %ecx,%xmm9 |
| 12954 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 12955 | .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| 12956 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 12957 | .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8 |
| 12958 | .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8 |
| 12959 | .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8 |
| 12960 | .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8 |
| 12961 | .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10 |
| 12962 | .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10 |
| 12963 | .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8 |
| 12964 | .byte 102,69,15,235,194 // por %xmm10,%xmm8 |
| 12965 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 12966 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 12967 | .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10 |
| 12968 | .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10 |
| 12969 | .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10 |
| 12970 | .byte 102,69,15,107,210 // packssdw %xmm10,%xmm10 |
| 12971 | .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11 |
| 12972 | .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11 |
| 12973 | .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10 |
| 12974 | .byte 102,69,15,235,211 // por %xmm11,%xmm10 |
| 12975 | .byte 69,15,40,217 // movaps %xmm9,%xmm11 |
| 12976 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 12977 | .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11 |
| 12978 | .byte 102,65,15,114,243,16 // pslld $0x10,%xmm11 |
| 12979 | .byte 102,65,15,114,227,16 // psrad $0x10,%xmm11 |
| 12980 | .byte 102,69,15,107,219 // packssdw %xmm11,%xmm11 |
| 12981 | .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12 |
| 12982 | .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| 12983 | .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11 |
| 12984 | .byte 102,69,15,235,220 // por %xmm12,%xmm11 |
| 12985 | .byte 68,15,89,203 // mulps %xmm3,%xmm9 |
| 12986 | .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9 |
| 12987 | .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9 |
| 12988 | .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9 |
| 12989 | .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9 |
| 12990 | .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12 |
| 12991 | .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12 |
| 12992 | .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9 |
| 12993 | .byte 102,69,15,235,204 // por %xmm12,%xmm9 |
| 12994 | .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8 |
| 12995 | .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11 |
| 12996 | .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9 |
| 12997 | .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9 |
| 12998 | .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8) |
| 12999 | .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8 |
| 13000 | .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8) |
| 13001 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13002 | .byte 255,224 // jmpq *%rax |
| 13003 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13004 | HIDDEN _sk_store_f32_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13005 | .globl _sk_store_f32_sse2 |
| 13006 | _sk_store_f32_sse2: |
| 13007 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13008 | .byte 72,139,0 // mov (%rax),%rax |
| 13009 | .byte 72,137,249 // mov %rdi,%rcx |
| 13010 | .byte 72,193,225,4 // shl $0x4,%rcx |
| 13011 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 13012 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 13013 | .byte 68,15,20,201 // unpcklps %xmm1,%xmm9 |
| 13014 | .byte 68,15,40,210 // movaps %xmm2,%xmm10 |
| 13015 | .byte 68,15,40,218 // movaps %xmm2,%xmm11 |
| 13016 | .byte 68,15,20,219 // unpcklps %xmm3,%xmm11 |
| 13017 | .byte 68,15,21,193 // unpckhps %xmm1,%xmm8 |
| 13018 | .byte 68,15,21,211 // unpckhps %xmm3,%xmm10 |
| 13019 | .byte 69,15,40,225 // movaps %xmm9,%xmm12 |
| 13020 | .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12 |
| 13021 | .byte 69,15,18,217 // movhlps %xmm9,%xmm11 |
| 13022 | .byte 69,15,40,200 // movaps %xmm8,%xmm9 |
| 13023 | .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9 |
| 13024 | .byte 69,15,18,208 // movhlps %xmm8,%xmm10 |
| 13025 | .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1) |
| 13026 | .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1) |
| 13027 | .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1) |
| 13028 | .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1) |
| 13029 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13030 | .byte 255,224 // jmpq *%rax |
| 13031 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13032 | HIDDEN _sk_clamp_x_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13033 | .globl _sk_clamp_x_sse2 |
| 13034 | _sk_clamp_x_sse2: |
| 13035 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13036 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 13037 | .byte 68,15,95,192 // maxps %xmm0,%xmm8 |
| 13038 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 13039 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13040 | .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0 |
| 13041 | .byte 102,65,15,254,193 // paddd %xmm9,%xmm0 |
| 13042 | .byte 68,15,93,192 // minps %xmm0,%xmm8 |
| 13043 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13044 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 13045 | .byte 255,224 // jmpq *%rax |
| 13046 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13047 | HIDDEN _sk_clamp_y_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13048 | .globl _sk_clamp_y_sse2 |
| 13049 | _sk_clamp_y_sse2: |
| 13050 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13051 | .byte 69,15,87,192 // xorps %xmm8,%xmm8 |
| 13052 | .byte 68,15,95,193 // maxps %xmm1,%xmm8 |
| 13053 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 13054 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13055 | .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1 |
| 13056 | .byte 102,65,15,254,201 // paddd %xmm9,%xmm1 |
| 13057 | .byte 68,15,93,193 // minps %xmm1,%xmm8 |
| 13058 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13059 | .byte 65,15,40,200 // movaps %xmm8,%xmm1 |
| 13060 | .byte 255,224 // jmpq *%rax |
| 13061 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13062 | HIDDEN _sk_repeat_x_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13063 | .globl _sk_repeat_x_sse2 |
| 13064 | _sk_repeat_x_sse2: |
| 13065 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13066 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 13067 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 13068 | .byte 68,15,40,200 // movaps %xmm0,%xmm9 |
| 13069 | .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| 13070 | .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 |
| 13071 | .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| 13072 | .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 |
| 13073 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 13074 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 13075 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13076 | .byte 69,15,84,217 // andps %xmm9,%xmm11 |
| 13077 | .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| 13078 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 13079 | .byte 65,15,92,194 // subps %xmm10,%xmm0 |
| 13080 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 13081 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 13082 | .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| 13083 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13084 | .byte 255,224 // jmpq *%rax |
| 13085 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13086 | HIDDEN _sk_repeat_y_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13087 | .globl _sk_repeat_y_sse2 |
| 13088 | _sk_repeat_y_sse2: |
| 13089 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13090 | .byte 243,68,15,16,0 // movss (%rax),%xmm8 |
| 13091 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 13092 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 13093 | .byte 69,15,94,200 // divps %xmm8,%xmm9 |
| 13094 | .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10 |
| 13095 | .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10 |
| 13096 | .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9 |
| 13097 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 13098 | .byte 102,68,15,110,216 // movd %eax,%xmm11 |
| 13099 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13100 | .byte 69,15,84,217 // andps %xmm9,%xmm11 |
| 13101 | .byte 69,15,92,211 // subps %xmm11,%xmm10 |
| 13102 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 13103 | .byte 65,15,92,202 // subps %xmm10,%xmm1 |
| 13104 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 13105 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 13106 | .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| 13107 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13108 | .byte 255,224 // jmpq *%rax |
| 13109 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13110 | HIDDEN _sk_mirror_x_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13111 | .globl _sk_mirror_x_sse2 |
| 13112 | _sk_mirror_x_sse2: |
| 13113 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13114 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 13115 | .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| 13116 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 13117 | .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| 13118 | .byte 243,69,15,88,201 // addss %xmm9,%xmm9 |
| 13119 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13120 | .byte 68,15,40,208 // movaps %xmm0,%xmm10 |
| 13121 | .byte 69,15,94,209 // divps %xmm9,%xmm10 |
| 13122 | .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 |
| 13123 | .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 |
| 13124 | .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 |
| 13125 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 13126 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 13127 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13128 | .byte 69,15,84,226 // andps %xmm10,%xmm12 |
| 13129 | .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| 13130 | .byte 69,15,92,220 // subps %xmm12,%xmm11 |
| 13131 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 13132 | .byte 65,15,92,195 // subps %xmm11,%xmm0 |
| 13133 | .byte 65,15,92,192 // subps %xmm8,%xmm0 |
| 13134 | .byte 68,15,92,208 // subps %xmm0,%xmm10 |
| 13135 | .byte 65,15,84,194 // andps %xmm10,%xmm0 |
| 13136 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 13137 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 13138 | .byte 65,15,93,193 // minps %xmm9,%xmm0 |
| 13139 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13140 | .byte 255,224 // jmpq *%rax |
| 13141 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13142 | HIDDEN _sk_mirror_y_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13143 | .globl _sk_mirror_y_sse2 |
| 13144 | _sk_mirror_y_sse2: |
| 13145 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13146 | .byte 243,68,15,16,8 // movss (%rax),%xmm9 |
| 13147 | .byte 69,15,40,193 // movaps %xmm9,%xmm8 |
| 13148 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 13149 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 13150 | .byte 243,69,15,88,201 // addss %xmm9,%xmm9 |
| 13151 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13152 | .byte 68,15,40,209 // movaps %xmm1,%xmm10 |
| 13153 | .byte 69,15,94,209 // divps %xmm9,%xmm10 |
| 13154 | .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11 |
| 13155 | .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11 |
| 13156 | .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10 |
| 13157 | .byte 184,0,0,128,63 // mov $0x3f800000,%eax |
| 13158 | .byte 102,68,15,110,224 // movd %eax,%xmm12 |
| 13159 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13160 | .byte 69,15,84,226 // andps %xmm10,%xmm12 |
| 13161 | .byte 69,15,87,210 // xorps %xmm10,%xmm10 |
| 13162 | .byte 69,15,92,220 // subps %xmm12,%xmm11 |
| 13163 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 13164 | .byte 65,15,92,203 // subps %xmm11,%xmm1 |
| 13165 | .byte 65,15,92,200 // subps %xmm8,%xmm1 |
| 13166 | .byte 68,15,92,209 // subps %xmm1,%xmm10 |
| 13167 | .byte 65,15,84,202 // andps %xmm10,%xmm1 |
| 13168 | .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9 |
| 13169 | .byte 102,69,15,254,200 // paddd %xmm8,%xmm9 |
| 13170 | .byte 65,15,93,201 // minps %xmm9,%xmm1 |
| 13171 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13172 | .byte 255,224 // jmpq *%rax |
| 13173 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13174 | HIDDEN _sk_luminance_to_alpha_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13175 | .globl _sk_luminance_to_alpha_sse2 |
| 13176 | _sk_luminance_to_alpha_sse2: |
| 13177 | .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax |
| 13178 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 13179 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 13180 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 13181 | .byte 184,89,23,55,63 // mov $0x3f371759,%eax |
| 13182 | .byte 102,15,110,192 // movd %eax,%xmm0 |
| 13183 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 13184 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 13185 | .byte 15,88,195 // addps %xmm3,%xmm0 |
| 13186 | .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax |
| 13187 | .byte 102,15,110,216 // movd %eax,%xmm3 |
| 13188 | .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3 |
| 13189 | .byte 15,89,218 // mulps %xmm2,%xmm3 |
| 13190 | .byte 15,88,216 // addps %xmm0,%xmm3 |
| 13191 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13192 | .byte 15,87,192 // xorps %xmm0,%xmm0 |
| 13193 | .byte 15,87,201 // xorps %xmm1,%xmm1 |
| 13194 | .byte 15,87,210 // xorps %xmm2,%xmm2 |
| 13195 | .byte 255,224 // jmpq *%rax |
| 13196 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13197 | HIDDEN _sk_matrix_2x3_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13198 | .globl _sk_matrix_2x3_sse2 |
| 13199 | _sk_matrix_2x3_sse2: |
| 13200 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 13201 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 13202 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13203 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 13204 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 13205 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 13206 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 13207 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13208 | .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11 |
| 13209 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13210 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13211 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13212 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 13213 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 13214 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 13215 | .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| 13216 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13217 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 13218 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13219 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13220 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13221 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 13222 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 13223 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13224 | .byte 255,224 // jmpq *%rax |
| 13225 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13226 | HIDDEN _sk_matrix_3x4_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13227 | .globl _sk_matrix_3x4_sse2 |
| 13228 | _sk_matrix_3x4_sse2: |
| 13229 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 13230 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 13231 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13232 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 13233 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 13234 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 13235 | .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10 |
| 13236 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13237 | .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| 13238 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13239 | .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12 |
| 13240 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13241 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 13242 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13243 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13244 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13245 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 13246 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 13247 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 13248 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 13249 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13250 | .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| 13251 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13252 | .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| 13253 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13254 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 13255 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13256 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13257 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13258 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 13259 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 13260 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 13261 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13262 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 13263 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13264 | .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| 13265 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13266 | .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| 13267 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 13268 | .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| 13269 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 13270 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 13271 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13272 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 13273 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13274 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13275 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 13276 | .byte 255,224 // jmpq *%rax |
| 13277 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13278 | HIDDEN _sk_matrix_4x5_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13279 | .globl _sk_matrix_4x5_sse2 |
| 13280 | _sk_matrix_4x5_sse2: |
| 13281 | .byte 68,15,40,201 // movaps %xmm1,%xmm9 |
| 13282 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 13283 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13284 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 13285 | .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1 |
| 13286 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 13287 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 13288 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13289 | .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11 |
| 13290 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13291 | .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12 |
| 13292 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13293 | .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13 |
| 13294 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 13295 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 13296 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 13297 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 13298 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13299 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13300 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13301 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 13302 | .byte 65,15,88,194 // addps %xmm10,%xmm0 |
| 13303 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 13304 | .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10 |
| 13305 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13306 | .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11 |
| 13307 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13308 | .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12 |
| 13309 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13310 | .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13 |
| 13311 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 13312 | .byte 68,15,89,227 // mulps %xmm3,%xmm12 |
| 13313 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 13314 | .byte 68,15,89,218 // mulps %xmm2,%xmm11 |
| 13315 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13316 | .byte 69,15,89,209 // mulps %xmm9,%xmm10 |
| 13317 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13318 | .byte 65,15,89,200 // mulps %xmm8,%xmm1 |
| 13319 | .byte 65,15,88,202 // addps %xmm10,%xmm1 |
| 13320 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 13321 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13322 | .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11 |
| 13323 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13324 | .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12 |
| 13325 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13326 | .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13 |
| 13327 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 13328 | .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14 |
| 13329 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 13330 | .byte 68,15,89,235 // mulps %xmm3,%xmm13 |
| 13331 | .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| 13332 | .byte 68,15,89,226 // mulps %xmm2,%xmm12 |
| 13333 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 13334 | .byte 69,15,89,217 // mulps %xmm9,%xmm11 |
| 13335 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13336 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 13337 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13338 | .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11 |
| 13339 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13340 | .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12 |
| 13341 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13342 | .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13 |
| 13343 | .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13 |
| 13344 | .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14 |
| 13345 | .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14 |
| 13346 | .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15 |
| 13347 | .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15 |
| 13348 | .byte 68,15,89,243 // mulps %xmm3,%xmm14 |
| 13349 | .byte 69,15,88,247 // addps %xmm15,%xmm14 |
| 13350 | .byte 68,15,89,234 // mulps %xmm2,%xmm13 |
| 13351 | .byte 69,15,88,238 // addps %xmm14,%xmm13 |
| 13352 | .byte 69,15,89,225 // mulps %xmm9,%xmm12 |
| 13353 | .byte 69,15,88,229 // addps %xmm13,%xmm12 |
| 13354 | .byte 69,15,89,216 // mulps %xmm8,%xmm11 |
| 13355 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13356 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13357 | .byte 65,15,40,210 // movaps %xmm10,%xmm2 |
| 13358 | .byte 65,15,40,219 // movaps %xmm11,%xmm3 |
| 13359 | .byte 255,224 // jmpq *%rax |
| 13360 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13361 | HIDDEN _sk_matrix_perspective_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13362 | .globl _sk_matrix_perspective_sse2 |
| 13363 | _sk_matrix_perspective_sse2: |
| 13364 | .byte 68,15,40,192 // movaps %xmm0,%xmm8 |
| 13365 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13366 | .byte 243,15,16,0 // movss (%rax),%xmm0 |
| 13367 | .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9 |
| 13368 | .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0 |
| 13369 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13370 | .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10 |
| 13371 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13372 | .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| 13373 | .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| 13374 | .byte 65,15,89,192 // mulps %xmm8,%xmm0 |
| 13375 | .byte 65,15,88,193 // addps %xmm9,%xmm0 |
| 13376 | .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9 |
| 13377 | .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9 |
| 13378 | .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10 |
| 13379 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13380 | .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11 |
| 13381 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13382 | .byte 68,15,89,209 // mulps %xmm1,%xmm10 |
| 13383 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13384 | .byte 69,15,89,200 // mulps %xmm8,%xmm9 |
| 13385 | .byte 69,15,88,202 // addps %xmm10,%xmm9 |
| 13386 | .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10 |
| 13387 | .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10 |
| 13388 | .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11 |
| 13389 | .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11 |
| 13390 | .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12 |
| 13391 | .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12 |
| 13392 | .byte 68,15,89,217 // mulps %xmm1,%xmm11 |
| 13393 | .byte 69,15,88,220 // addps %xmm12,%xmm11 |
| 13394 | .byte 69,15,89,208 // mulps %xmm8,%xmm10 |
| 13395 | .byte 69,15,88,211 // addps %xmm11,%xmm10 |
| 13396 | .byte 65,15,83,202 // rcpps %xmm10,%xmm1 |
| 13397 | .byte 15,89,193 // mulps %xmm1,%xmm0 |
| 13398 | .byte 68,15,89,201 // mulps %xmm1,%xmm9 |
| 13399 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13400 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 13401 | .byte 255,224 // jmpq *%rax |
| 13402 | |
| Mike Klein | 7c4af54 | 2017-03-29 18:21:14 -0400 | [diff] [blame] | 13403 | HIDDEN _sk_linear_gradient_2stops_sse2 |
| Mike Klein | d7e06ae | 2017-03-29 16:33:06 -0400 | [diff] [blame] | 13404 | .globl _sk_linear_gradient_2stops_sse2 |
| 13405 | _sk_linear_gradient_2stops_sse2: |
| 13406 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13407 | .byte 68,15,16,8 // movups (%rax),%xmm9 |
| 13408 | .byte 15,16,88,16 // movups 0x10(%rax),%xmm3 |
| 13409 | .byte 68,15,40,195 // movaps %xmm3,%xmm8 |
| 13410 | .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8 |
| 13411 | .byte 65,15,40,201 // movaps %xmm9,%xmm1 |
| 13412 | .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1 |
| 13413 | .byte 68,15,89,192 // mulps %xmm0,%xmm8 |
| 13414 | .byte 68,15,88,193 // addps %xmm1,%xmm8 |
| 13415 | .byte 15,40,203 // movaps %xmm3,%xmm1 |
| 13416 | .byte 15,198,201,85 // shufps $0x55,%xmm1,%xmm1 |
| 13417 | .byte 65,15,40,209 // movaps %xmm9,%xmm2 |
| 13418 | .byte 15,198,210,85 // shufps $0x55,%xmm2,%xmm2 |
| 13419 | .byte 15,89,200 // mulps %xmm0,%xmm1 |
| 13420 | .byte 15,88,202 // addps %xmm2,%xmm1 |
| 13421 | .byte 15,40,211 // movaps %xmm3,%xmm2 |
| 13422 | .byte 15,198,210,170 // shufps $0xaa,%xmm2,%xmm2 |
| 13423 | .byte 69,15,40,209 // movaps %xmm9,%xmm10 |
| 13424 | .byte 69,15,198,210,170 // shufps $0xaa,%xmm10,%xmm10 |
| 13425 | .byte 15,89,208 // mulps %xmm0,%xmm2 |
| 13426 | .byte 65,15,88,210 // addps %xmm10,%xmm2 |
| 13427 | .byte 15,198,219,255 // shufps $0xff,%xmm3,%xmm3 |
| 13428 | .byte 69,15,198,201,255 // shufps $0xff,%xmm9,%xmm9 |
| 13429 | .byte 15,89,216 // mulps %xmm0,%xmm3 |
| 13430 | .byte 65,15,88,217 // addps %xmm9,%xmm3 |
| 13431 | .byte 72,173 // lods %ds:(%rsi),%rax |
| 13432 | .byte 65,15,40,192 // movaps %xmm8,%xmm0 |
| 13433 | .byte 255,224 // jmpq *%rax |
| 13434 | #endif |