Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / skia / b4bbc64adee74f99ddfc30ad4e5043f98380bdbb / . / src / jumper / SkJumper_generated.S
blob: a3ea251a7ca697d6b80a36f52d5d1c9d9703bf3e [file] [log] [blame]
Mike Kleind7e06ae2017-03-29 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 Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]9#if defined(__MACH__)
10 #define HIDDEN .private_extern
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11 #define FUNCTION(name)
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]12 #define BALIGN4 .align 2
13 #define BALIGN16 .align 4
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]14 #define BALIGN32 .align 5
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]15#else
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]16 .section .note.GNU-stack,"",%progbits
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17 #define HIDDEN .hidden
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18 #define FUNCTION(name) .type name,%function
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]19 #define BALIGN4 .balign 4
20 #define BALIGN16 .balign 16
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]21 #define BALIGN32 .balign 32
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22#endif
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23.text
24#if defined(__aarch64__)
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]25BALIGN4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]27HIDDEN _sk_start_pipeline_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]28.globl _sk_start_pipeline_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]29FUNCTION(_sk_start_pipeline_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]30_sk_start_pipeline_aarch64:
31 .long 0xa9bd5bf7 // stp x23, x22, [sp, #-48]!
32 .long 0xa90153f5 // stp x21, x20, [sp, #16]
33 .long 0xa9027bf3 // stp x19, x30, [sp, #32]
34 .long 0xaa0103f4 // mov x20, x1
35 .long 0xf8408697 // ldr x23, [x20], #8
36 .long 0xaa0003f5 // mov x21, x0
37 .long 0xaa0303f3 // mov x19, x3
38 .long 0x910012a8 // add x8, x21, #0x4
39 .long 0xeb13011f // cmp x8, x19
40 .long 0xaa0203f6 // mov x22, x2
41 .long 0x54000069 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast
42 .long 0xaa1503e0 // mov x0, x21
43 .long 0x14000012 // b 78 <sk_start_pipeline_aarch64+0x78>
44 .long 0x6f00e400 // movi v0.2d, #0x0
45 .long 0x6f00e401 // movi v1.2d, #0x0
46 .long 0x6f00e402 // movi v2.2d, #0x0
47 .long 0x6f00e403 // movi v3.2d, #0x0
48 .long 0x6f00e404 // movi v4.2d, #0x0
49 .long 0x6f00e405 // movi v5.2d, #0x0
50 .long 0x6f00e406 // movi v6.2d, #0x0
51 .long 0x6f00e407 // movi v7.2d, #0x0
52 .long 0xaa1503e0 // mov x0, x21
53 .long 0xaa1403e1 // mov x1, x20
54 .long 0xaa1603e2 // mov x2, x22
55 .long 0xd63f02e0 // blr x23
56 .long 0x910012a0 // add x0, x21, #0x4
57 .long 0x910022a8 // add x8, x21, #0x8
58 .long 0xeb13011f // cmp x8, x19
59 .long 0xaa0003f5 // mov x21, x0
60 .long 0x54fffe09 // b.ls 34 <sk_start_pipeline_aarch64+0x34> // b.plast
61 .long 0xa9427bf3 // ldp x19, x30, [sp, #32]
62 .long 0xa94153f5 // ldp x21, x20, [sp, #16]
63 .long 0xa8c35bf7 // ldp x23, x22, [sp], #48
64 .long 0xd65f03c0 // ret
65
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]66HIDDEN _sk_just_return_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]67.globl _sk_just_return_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]68FUNCTION(_sk_just_return_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]69_sk_just_return_aarch64:
70 .long 0xd65f03c0 // ret
71
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]72HIDDEN _sk_seed_shader_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]73.globl _sk_seed_shader_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]74FUNCTION(_sk_seed_shader_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]75_sk_seed_shader_aarch64:
76 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
77 .long 0x3dc00046 // ldr q6, [x2]
78 .long 0x4e040c00 // dup v0.4s, w0
79 .long 0x4f0167e7 // movi v7.4s, #0x3f, lsl #24
80 .long 0x4d40c901 // ld1r {v1.4s}, [x8]
81 .long 0x4e21d800 // scvtf v0.4s, v0.4s
82 .long 0x4e27d400 // fadd v0.4s, v0.4s, v7.4s
83 .long 0x4f03f602 // fmov v2.4s, #1.000000000000000000e+00
84 .long 0x4e21d821 // scvtf v1.4s, v1.4s
85 .long 0x6f00e403 // movi v3.2d, #0x0
86 .long 0x6f00e404 // movi v4.2d, #0x0
87 .long 0x6f00e405 // movi v5.2d, #0x0
88 .long 0x4e26d400 // fadd v0.4s, v0.4s, v6.4s
89 .long 0x6f00e406 // movi v6.2d, #0x0
90 .long 0x4e27d421 // fadd v1.4s, v1.4s, v7.4s
91 .long 0x6f00e407 // movi v7.2d, #0x0
92 .long 0xd61f0060 // br x3
93
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]94HIDDEN _sk_constant_color_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]95.globl _sk_constant_color_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]96FUNCTION(_sk_constant_color_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]97_sk_constant_color_aarch64:
98 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]99 .long 0xaa0803ea // mov x10, x8
100 .long 0x4ddfc940 // ld1r {v0.4s}, [x10], #4
101 .long 0x91002109 // add x9, x8, #0x8
102 .long 0x91003108 // add x8, x8, #0xc
103 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
104 .long 0x4d40c903 // ld1r {v3.4s}, [x8]
105 .long 0x4d40c941 // ld1r {v1.4s}, [x10]
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]106 .long 0xd61f0060 // br x3
107
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]108HIDDEN _sk_clear_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]109.globl _sk_clear_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]110FUNCTION(_sk_clear_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]111_sk_clear_aarch64:
112 .long 0xf8408423 // ldr x3, [x1], #8
113 .long 0x6f00e400 // movi v0.2d, #0x0
114 .long 0x6f00e401 // movi v1.2d, #0x0
115 .long 0x6f00e402 // movi v2.2d, #0x0
116 .long 0x6f00e403 // movi v3.2d, #0x0
117 .long 0xd61f0060 // br x3
118
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]119HIDDEN _sk_srcatop_aarch64
120.globl _sk_srcatop_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]121FUNCTION(_sk_srcatop_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]122_sk_srcatop_aarch64:
123 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
124 .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s
125 .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s
126 .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s
127 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]128 .long 0xf8408423 // ldr x3, [x1], #8
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]129 .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s
130 .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s
131 .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s
132 .long 0x6e27de10 // fmul v16.4s, v16.4s, v7.4s
133 .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s
134 .long 0x4eb01e03 // mov v3.16b, v16.16b
135 .long 0xd61f0060 // br x3
136
137HIDDEN _sk_dstatop_aarch64
138.globl _sk_dstatop_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]139FUNCTION(_sk_dstatop_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]140_sk_dstatop_aarch64:
141 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
142 .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s
143 .long 0xf8408423 // ldr x3, [x1], #8
144 .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
145 .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
146 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
147 .long 0x6e23de10 // fmul v16.4s, v16.4s, v3.4s
148 .long 0x4e23ccf0 // fmla v16.4s, v7.4s, v3.4s
149 .long 0x4e23cc80 // fmla v0.4s, v4.4s, v3.4s
150 .long 0x4e23cca1 // fmla v1.4s, v5.4s, v3.4s
151 .long 0x4e23ccc2 // fmla v2.4s, v6.4s, v3.4s
152 .long 0x4eb01e03 // mov v3.16b, v16.16b
153 .long 0xd61f0060 // br x3
154
155HIDDEN _sk_srcin_aarch64
156.globl _sk_srcin_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]157FUNCTION(_sk_srcin_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]158_sk_srcin_aarch64:
159 .long 0xf8408423 // ldr x3, [x1], #8
160 .long 0x6e27dc00 // fmul v0.4s, v0.4s, v7.4s
161 .long 0x6e27dc21 // fmul v1.4s, v1.4s, v7.4s
162 .long 0x6e27dc42 // fmul v2.4s, v2.4s, v7.4s
163 .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s
164 .long 0xd61f0060 // br x3
165
166HIDDEN _sk_dstin_aarch64
167.globl _sk_dstin_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]168FUNCTION(_sk_dstin_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]169_sk_dstin_aarch64:
170 .long 0xf8408423 // ldr x3, [x1], #8
171 .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s
172 .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s
173 .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s
174 .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s
175 .long 0xd61f0060 // br x3
176
177HIDDEN _sk_srcout_aarch64
178.globl _sk_srcout_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]179FUNCTION(_sk_srcout_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]180_sk_srcout_aarch64:
181 .long 0xf8408423 // ldr x3, [x1], #8
182 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
183 .long 0x4ea7d610 // fsub v16.4s, v16.4s, v7.4s
184 .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
185 .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
186 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
187 .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
188 .long 0xd61f0060 // br x3
189
190HIDDEN _sk_dstout_aarch64
191.globl _sk_dstout_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]192FUNCTION(_sk_dstout_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]193_sk_dstout_aarch64:
194 .long 0xf8408423 // ldr x3, [x1], #8
195 .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00
196 .long 0x4ea3d403 // fsub v3.4s, v0.4s, v3.4s
197 .long 0x6e24dc60 // fmul v0.4s, v3.4s, v4.4s
198 .long 0x6e25dc61 // fmul v1.4s, v3.4s, v5.4s
199 .long 0x6e26dc62 // fmul v2.4s, v3.4s, v6.4s
200 .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]201 .long 0xd61f0060 // br x3
202
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]203HIDDEN _sk_srcover_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]204.globl _sk_srcover_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]205FUNCTION(_sk_srcover_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]206_sk_srcover_aarch64:
207 .long 0xf8408423 // ldr x3, [x1], #8
208 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
209 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
210 .long 0x4e24ce00 // fmla v0.4s, v16.4s, v4.4s
211 .long 0x4e25ce01 // fmla v1.4s, v16.4s, v5.4s
212 .long 0x4e26ce02 // fmla v2.4s, v16.4s, v6.4s
213 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
214 .long 0xd61f0060 // br x3
215
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]216HIDDEN _sk_dstover_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]217.globl _sk_dstover_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]218FUNCTION(_sk_dstover_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]219_sk_dstover_aarch64:
220 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
221 .long 0xf8408423 // ldr x3, [x1], #8
222 .long 0x4ea41c90 // mov v16.16b, v4.16b
223 .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s
224 .long 0x4ea51cb1 // mov v17.16b, v5.16b
225 .long 0x4ea61cd2 // mov v18.16b, v6.16b
226 .long 0x4ea71cf3 // mov v19.16b, v7.16b
227 .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s
228 .long 0x4e21ce91 // fmla v17.4s, v20.4s, v1.4s
229 .long 0x4e22ce92 // fmla v18.4s, v20.4s, v2.4s
230 .long 0x4e23ce93 // fmla v19.4s, v20.4s, v3.4s
231 .long 0x4eb01e00 // mov v0.16b, v16.16b
232 .long 0x4eb11e21 // mov v1.16b, v17.16b
233 .long 0x4eb21e42 // mov v2.16b, v18.16b
234 .long 0x4eb31e63 // mov v3.16b, v19.16b
235 .long 0xd61f0060 // br x3
236
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]237HIDDEN _sk_modulate_aarch64
238.globl _sk_modulate_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]239FUNCTION(_sk_modulate_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]240_sk_modulate_aarch64:
241 .long 0xf8408423 // ldr x3, [x1], #8
242 .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s
243 .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s
244 .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s
245 .long 0x6e27dc63 // fmul v3.4s, v3.4s, v7.4s
246 .long 0xd61f0060 // br x3
247
248HIDDEN _sk_multiply_aarch64
249.globl _sk_multiply_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]250FUNCTION(_sk_multiply_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]251_sk_multiply_aarch64:
252 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
253 .long 0x4ea7d613 // fsub v19.4s, v16.4s, v7.4s
254 .long 0x4ea3d614 // fsub v20.4s, v16.4s, v3.4s
255 .long 0x6e20de70 // fmul v16.4s, v19.4s, v0.4s
256 .long 0x6e21de71 // fmul v17.4s, v19.4s, v1.4s
257 .long 0x6e22de72 // fmul v18.4s, v19.4s, v2.4s
258 .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s
259 .long 0xf8408423 // ldr x3, [x1], #8
260 .long 0x4e34cc90 // fmla v16.4s, v4.4s, v20.4s
261 .long 0x4e34ccb1 // fmla v17.4s, v5.4s, v20.4s
262 .long 0x4e34ccd2 // fmla v18.4s, v6.4s, v20.4s
263 .long 0x4e34ccf3 // fmla v19.4s, v7.4s, v20.4s
264 .long 0x4e20cc90 // fmla v16.4s, v4.4s, v0.4s
265 .long 0x4e21ccb1 // fmla v17.4s, v5.4s, v1.4s
266 .long 0x4e22ccd2 // fmla v18.4s, v6.4s, v2.4s
267 .long 0x4e23ccf3 // fmla v19.4s, v7.4s, v3.4s
268 .long 0x4eb01e00 // mov v0.16b, v16.16b
269 .long 0x4eb11e21 // mov v1.16b, v17.16b
270 .long 0x4eb21e42 // mov v2.16b, v18.16b
271 .long 0x4eb31e63 // mov v3.16b, v19.16b
272 .long 0xd61f0060 // br x3
273
274HIDDEN _sk_plus__aarch64
275.globl _sk_plus__aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]276FUNCTION(_sk_plus__aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]277_sk_plus__aarch64:
278 .long 0xf8408423 // ldr x3, [x1], #8
279 .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s
280 .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s
281 .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s
282 .long 0x4e27d463 // fadd v3.4s, v3.4s, v7.4s
283 .long 0xd61f0060 // br x3
284
285HIDDEN _sk_screen_aarch64
286.globl _sk_screen_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]287FUNCTION(_sk_screen_aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]288_sk_screen_aarch64:
289 .long 0xf8408423 // ldr x3, [x1], #8
290 .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s
291 .long 0x4e25d431 // fadd v17.4s, v1.4s, v5.4s
292 .long 0x4e26d452 // fadd v18.4s, v2.4s, v6.4s
293 .long 0x4e27d473 // fadd v19.4s, v3.4s, v7.4s
294 .long 0x4ea4cc10 // fmls v16.4s, v0.4s, v4.4s
295 .long 0x4ea5cc31 // fmls v17.4s, v1.4s, v5.4s
296 .long 0x4ea6cc52 // fmls v18.4s, v2.4s, v6.4s
297 .long 0x4ea7cc73 // fmls v19.4s, v3.4s, v7.4s
298 .long 0x4eb01e00 // mov v0.16b, v16.16b
299 .long 0x4eb11e21 // mov v1.16b, v17.16b
300 .long 0x4eb21e42 // mov v2.16b, v18.16b
301 .long 0x4eb31e63 // mov v3.16b, v19.16b
302 .long 0xd61f0060 // br x3
303
304HIDDEN _sk_xor__aarch64
305.globl _sk_xor__aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]306FUNCTION(_sk_xor__aarch64)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]307_sk_xor__aarch64:
308 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
309 .long 0xf8408423 // ldr x3, [x1], #8
310 .long 0x4ea7d611 // fsub v17.4s, v16.4s, v7.4s
311 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
312 .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s
313 .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s
314 .long 0x6e22de22 // fmul v2.4s, v17.4s, v2.4s
315 .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s
316 .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s
317 .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s
318 .long 0x4e30ccc2 // fmla v2.4s, v6.4s, v16.4s
319 .long 0x4e30cce3 // fmla v3.4s, v7.4s, v16.4s
320 .long 0xd61f0060 // br x3
321
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]322HIDDEN _sk_darken_aarch64
323.globl _sk_darken_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]324FUNCTION(_sk_darken_aarch64)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]325_sk_darken_aarch64:
326 .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s
327 .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s
328 .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s
329 .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s
330 .long 0x4e31f610 // fmax v16.4s, v16.4s, v17.4s
331 .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s
332 .long 0xf8408423 // ldr x3, [x1], #8
333 .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s
334 .long 0x4e33f652 // fmax v18.4s, v18.4s, v19.4s
335 .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s
336 .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s
337 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
338 .long 0x4e33f631 // fmax v17.4s, v17.4s, v19.4s
339 .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s
340 .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s
341 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
342 .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s
343 .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s
344 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
345 .long 0xd61f0060 // br x3
346
347HIDDEN _sk_lighten_aarch64
348.globl _sk_lighten_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]349FUNCTION(_sk_lighten_aarch64)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]350_sk_lighten_aarch64:
351 .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s
352 .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s
353 .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s
354 .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s
355 .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s
356 .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s
357 .long 0xf8408423 // ldr x3, [x1], #8
358 .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s
359 .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s
360 .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s
361 .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s
362 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
363 .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s
364 .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s
365 .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s
366 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
367 .long 0x4eb2d421 // fsub v1.4s, v1.4s, v18.4s
368 .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s
369 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
370 .long 0xd61f0060 // br x3
371
372HIDDEN _sk_difference_aarch64
373.globl _sk_difference_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]374FUNCTION(_sk_difference_aarch64)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]375_sk_difference_aarch64:
376 .long 0x6e27dc10 // fmul v16.4s, v0.4s, v7.4s
377 .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s
378 .long 0x6e27dc32 // fmul v18.4s, v1.4s, v7.4s
379 .long 0x6e25dc73 // fmul v19.4s, v3.4s, v5.4s
380 .long 0x4eb1f610 // fmin v16.4s, v16.4s, v17.4s
381 .long 0x4eb3f652 // fmin v18.4s, v18.4s, v19.4s
382 .long 0x4e24d400 // fadd v0.4s, v0.4s, v4.4s
383 .long 0x4e30d610 // fadd v16.4s, v16.4s, v16.4s
384 .long 0x6e27dc51 // fmul v17.4s, v2.4s, v7.4s
385 .long 0x6e26dc73 // fmul v19.4s, v3.4s, v6.4s
386 .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s
387 .long 0x4e25d421 // fadd v1.4s, v1.4s, v5.4s
388 .long 0x4e32d650 // fadd v16.4s, v18.4s, v18.4s
389 .long 0xf8408423 // ldr x3, [x1], #8
390 .long 0x4eb3f631 // fmin v17.4s, v17.4s, v19.4s
391 .long 0x4eb0d421 // fsub v1.4s, v1.4s, v16.4s
392 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
393 .long 0x4e26d442 // fadd v2.4s, v2.4s, v6.4s
394 .long 0x4e31d631 // fadd v17.4s, v17.4s, v17.4s
395 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
396 .long 0x4eb1d442 // fsub v2.4s, v2.4s, v17.4s
397 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
398 .long 0xd61f0060 // br x3
399
400HIDDEN _sk_exclusion_aarch64
401.globl _sk_exclusion_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]402FUNCTION(_sk_exclusion_aarch64)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]403_sk_exclusion_aarch64:
404 .long 0x4e24d410 // fadd v16.4s, v0.4s, v4.4s
405 .long 0x6e24dc00 // fmul v0.4s, v0.4s, v4.4s
406 .long 0x4e20d400 // fadd v0.4s, v0.4s, v0.4s
407 .long 0x4ea0d600 // fsub v0.4s, v16.4s, v0.4s
408 .long 0x4e25d430 // fadd v16.4s, v1.4s, v5.4s
409 .long 0x6e25dc21 // fmul v1.4s, v1.4s, v5.4s
410 .long 0x4e21d421 // fadd v1.4s, v1.4s, v1.4s
411 .long 0x4ea1d601 // fsub v1.4s, v16.4s, v1.4s
412 .long 0x4e26d450 // fadd v16.4s, v2.4s, v6.4s
413 .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s
414 .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s
415 .long 0xf8408423 // ldr x3, [x1], #8
416 .long 0x4ea2d602 // fsub v2.4s, v16.4s, v2.4s
417 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
418 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
419 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
420 .long 0xd61f0060 // br x3
421
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]422HIDDEN _sk_colorburn_aarch64
423.globl _sk_colorburn_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]424FUNCTION(_sk_colorburn_aarch64)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]425_sk_colorburn_aarch64:
426 .long 0x4ea4d4f3 // fsub v19.4s, v7.4s, v4.4s
427 .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s
428 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
429 .long 0x6e20fe73 // fdiv v19.4s, v19.4s, v0.4s
430 .long 0x4ea7d634 // fsub v20.4s, v17.4s, v7.4s
431 .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s
432 .long 0x6e20de95 // fmul v21.4s, v20.4s, v0.4s
433 .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s
434 .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s
435 .long 0x4e33cc75 // fmla v21.4s, v3.4s, v19.4s
436 .long 0x4ea5d4f3 // fsub v19.4s, v7.4s, v5.4s
437 .long 0x6e23de73 // fmul v19.4s, v19.4s, v3.4s
438 .long 0x6e21fe73 // fdiv v19.4s, v19.4s, v1.4s
439 .long 0x4ea0d812 // fcmeq v18.4s, v0.4s, #0.0
440 .long 0x4eb3f4f3 // fmin v19.4s, v7.4s, v19.4s
441 .long 0x6e751c12 // bsl v18.16b, v0.16b, v21.16b
442 .long 0x6e21de80 // fmul v0.4s, v20.4s, v1.4s
443 .long 0x4eb3d4f3 // fsub v19.4s, v7.4s, v19.4s
444 .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s
445 .long 0x4e33cc60 // fmla v0.4s, v3.4s, v19.4s
446 .long 0x4ea0d833 // fcmeq v19.4s, v1.4s, #0.0
447 .long 0x6e601c33 // bsl v19.16b, v1.16b, v0.16b
448 .long 0x4ea6d4e0 // fsub v0.4s, v7.4s, v6.4s
449 .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s
450 .long 0x6e22fc00 // fdiv v0.4s, v0.4s, v2.4s
451 .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s
452 .long 0x6e22de81 // fmul v1.4s, v20.4s, v2.4s
453 .long 0x4ea0d4e0 // fsub v0.4s, v7.4s, v0.4s
454 .long 0x4e26d434 // fadd v20.4s, v1.4s, v6.4s
455 .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s
456 .long 0x4ea0d840 // fcmeq v0.4s, v2.4s, #0.0
457 .long 0x4ea3d631 // fsub v17.4s, v17.4s, v3.4s
458 .long 0xf8408423 // ldr x3, [x1], #8
459 .long 0x4e27e490 // fcmeq v16.4s, v4.4s, v7.4s
460 .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b
461 .long 0x4e31cc92 // fmla v18.4s, v4.4s, v17.4s
462 .long 0x4e27e4a1 // fcmeq v1.4s, v5.4s, v7.4s
463 .long 0x4e27e4c2 // fcmeq v2.4s, v6.4s, v7.4s
464 .long 0x4e31ccb3 // fmla v19.4s, v5.4s, v17.4s
465 .long 0x4e31ccc0 // fmla v0.4s, v6.4s, v17.4s
466 .long 0x6e721ed0 // bsl v16.16b, v22.16b, v18.16b
467 .long 0x6e731ea1 // bsl v1.16b, v21.16b, v19.16b
468 .long 0x6e601e82 // bsl v2.16b, v20.16b, v0.16b
469 .long 0x4e27ce23 // fmla v3.4s, v17.4s, v7.4s
470 .long 0x4eb01e00 // mov v0.16b, v16.16b
471 .long 0xd61f0060 // br x3
472
473HIDDEN _sk_colordodge_aarch64
474.globl _sk_colordodge_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]475FUNCTION(_sk_colordodge_aarch64)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]476_sk_colordodge_aarch64:
477 .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
478 .long 0x6e24dc71 // fmul v17.4s, v3.4s, v4.4s
479 .long 0x4ea0d474 // fsub v20.4s, v3.4s, v0.4s
480 .long 0x6e25dc75 // fmul v21.4s, v3.4s, v5.4s
481 .long 0x4ea1d476 // fsub v22.4s, v3.4s, v1.4s
482 .long 0x4ea7d657 // fsub v23.4s, v18.4s, v7.4s
483 .long 0x6e34fe31 // fdiv v17.4s, v17.4s, v20.4s
484 .long 0x6e36feb4 // fdiv v20.4s, v21.4s, v22.4s
485 .long 0x6e20def5 // fmul v21.4s, v23.4s, v0.4s
486 .long 0x4eb1f4f1 // fmin v17.4s, v7.4s, v17.4s
487 .long 0x4e23e413 // fcmeq v19.4s, v0.4s, v3.4s
488 .long 0x4e24d6b6 // fadd v22.4s, v21.4s, v4.4s
489 .long 0x4e31cc75 // fmla v21.4s, v3.4s, v17.4s
490 .long 0x6e751c13 // bsl v19.16b, v0.16b, v21.16b
491 .long 0x6e21dee0 // fmul v0.4s, v23.4s, v1.4s
492 .long 0x4eb4f4f4 // fmin v20.4s, v7.4s, v20.4s
493 .long 0x4e25d415 // fadd v21.4s, v0.4s, v5.4s
494 .long 0x4e34cc60 // fmla v0.4s, v3.4s, v20.4s
495 .long 0x4e23e434 // fcmeq v20.4s, v1.4s, v3.4s
496 .long 0x6e601c34 // bsl v20.16b, v1.16b, v0.16b
497 .long 0x6e26dc60 // fmul v0.4s, v3.4s, v6.4s
498 .long 0x4ea2d461 // fsub v1.4s, v3.4s, v2.4s
499 .long 0x6e21fc00 // fdiv v0.4s, v0.4s, v1.4s
500 .long 0x6e22dee1 // fmul v1.4s, v23.4s, v2.4s
501 .long 0x4ea0f4e0 // fmin v0.4s, v7.4s, v0.4s
502 .long 0x4e26d437 // fadd v23.4s, v1.4s, v6.4s
503 .long 0x4e20cc61 // fmla v1.4s, v3.4s, v0.4s
504 .long 0x4e23e440 // fcmeq v0.4s, v2.4s, v3.4s
505 .long 0x6e611c40 // bsl v0.16b, v2.16b, v1.16b
506 .long 0x4ea3d641 // fsub v1.4s, v18.4s, v3.4s
507 .long 0xf8408423 // ldr x3, [x1], #8
508 .long 0x4ea0d890 // fcmeq v16.4s, v4.4s, #0.0
509 .long 0x4ea0d8b1 // fcmeq v17.4s, v5.4s, #0.0
510 .long 0x4e21cc93 // fmla v19.4s, v4.4s, v1.4s
511 .long 0x4e21ccb4 // fmla v20.4s, v5.4s, v1.4s
512 .long 0x4ea0d8c2 // fcmeq v2.4s, v6.4s, #0.0
513 .long 0x4e21ccc0 // fmla v0.4s, v6.4s, v1.4s
514 .long 0x6e731ed0 // bsl v16.16b, v22.16b, v19.16b
515 .long 0x6e741eb1 // bsl v17.16b, v21.16b, v20.16b
516 .long 0x6e601ee2 // bsl v2.16b, v23.16b, v0.16b
517 .long 0x4e27cc23 // fmla v3.4s, v1.4s, v7.4s
518 .long 0x4eb01e00 // mov v0.16b, v16.16b
519 .long 0x4eb11e21 // mov v1.16b, v17.16b
520 .long 0xd61f0060 // br x3
521
522HIDDEN _sk_hardlight_aarch64
523.globl _sk_hardlight_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]524FUNCTION(_sk_hardlight_aarch64)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]525_sk_hardlight_aarch64:
526 .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s
527 .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s
528 .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s
529 .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s
530 .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s
531 .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s
532 .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s
533 .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s
534 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
535 .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s
536 .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b
537 .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s
538 .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s
539 .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s
540 .long 0x4e21d436 // fadd v22.4s, v1.4s, v1.4s
541 .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s
542 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
543 .long 0x6e36e476 // fcmge v22.4s, v3.4s, v22.4s
544 .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s
545 .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s
546 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
547 .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b
548 .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s
549 .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s
550 .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s
551 .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s
552 .long 0x4e22d454 // fadd v20.4s, v2.4s, v2.4s
553 .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s
554 .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s
555 .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s
556 .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s
557 .long 0xf8408423 // ldr x3, [x1], #8
558 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
559 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
560 .long 0x6e34e474 // fcmge v20.4s, v3.4s, v20.4s
561 .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s
562 .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s
563 .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s
564 .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s
565 .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s
566 .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b
567 .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s
568 .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s
569 .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s
570 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
571 .long 0xd61f0060 // br x3
572
573HIDDEN _sk_overlay_aarch64
574.globl _sk_overlay_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]575FUNCTION(_sk_overlay_aarch64)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]576_sk_overlay_aarch64:
577 .long 0x4ea4d4f4 // fsub v20.4s, v7.4s, v4.4s
578 .long 0x4ea0d475 // fsub v21.4s, v3.4s, v0.4s
579 .long 0x6e34deb4 // fmul v20.4s, v21.4s, v20.4s
580 .long 0x4e24d491 // fadd v17.4s, v4.4s, v4.4s
581 .long 0x6e24dc12 // fmul v18.4s, v0.4s, v4.4s
582 .long 0x6e27dc73 // fmul v19.4s, v3.4s, v7.4s
583 .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s
584 .long 0x6e31e4f1 // fcmge v17.4s, v7.4s, v17.4s
585 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
586 .long 0x4eb4d674 // fsub v20.4s, v19.4s, v20.4s
587 .long 0x6e741e51 // bsl v17.16b, v18.16b, v20.16b
588 .long 0x4ea5d4f2 // fsub v18.4s, v7.4s, v5.4s
589 .long 0x4ea1d474 // fsub v20.4s, v3.4s, v1.4s
590 .long 0x6e32de92 // fmul v18.4s, v20.4s, v18.4s
591 .long 0x4e25d4b6 // fadd v22.4s, v5.4s, v5.4s
592 .long 0x6e25dc35 // fmul v21.4s, v1.4s, v5.4s
593 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
594 .long 0x6e36e4f6 // fcmge v22.4s, v7.4s, v22.4s
595 .long 0x4e35d6b5 // fadd v21.4s, v21.4s, v21.4s
596 .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s
597 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
598 .long 0x6e721eb6 // bsl v22.16b, v21.16b, v18.16b
599 .long 0x4ea6d4f2 // fsub v18.4s, v7.4s, v6.4s
600 .long 0x4ea2d475 // fsub v21.4s, v3.4s, v2.4s
601 .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s
602 .long 0x4ea7d615 // fsub v21.4s, v16.4s, v7.4s
603 .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s
604 .long 0x6e20dea0 // fmul v0.4s, v21.4s, v0.4s
605 .long 0x6e21dea1 // fmul v1.4s, v21.4s, v1.4s
606 .long 0x6e22deb5 // fmul v21.4s, v21.4s, v2.4s
607 .long 0x6e26dc42 // fmul v2.4s, v2.4s, v6.4s
608 .long 0xf8408423 // ldr x3, [x1], #8
609 .long 0x4e32d652 // fadd v18.4s, v18.4s, v18.4s
610 .long 0x4ea3d610 // fsub v16.4s, v16.4s, v3.4s
611 .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s
612 .long 0x4e22d442 // fadd v2.4s, v2.4s, v2.4s
613 .long 0x4eb2d672 // fsub v18.4s, v19.4s, v18.4s
614 .long 0x4e30cc80 // fmla v0.4s, v4.4s, v16.4s
615 .long 0x4e30cca1 // fmla v1.4s, v5.4s, v16.4s
616 .long 0x4e30ccd5 // fmla v21.4s, v6.4s, v16.4s
617 .long 0x6e721c54 // bsl v20.16b, v2.16b, v18.16b
618 .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s
619 .long 0x4e36d421 // fadd v1.4s, v1.4s, v22.4s
620 .long 0x4e34d6a2 // fadd v2.4s, v21.4s, v20.4s
621 .long 0x4e27ce03 // fmla v3.4s, v16.4s, v7.4s
622 .long 0xd61f0060 // br x3
623
624HIDDEN _sk_softlight_aarch64
625.globl _sk_softlight_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]626FUNCTION(_sk_softlight_aarch64)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]627_sk_softlight_aarch64:
628 .long 0x4ea0c8f5 // fcmgt v21.4s, v7.4s, #0.0
629 .long 0x6e27fc96 // fdiv v22.4s, v4.4s, v7.4s
630 .long 0x6e27fcb8 // fdiv v24.4s, v5.4s, v7.4s
631 .long 0x6e27fcd9 // fdiv v25.4s, v6.4s, v7.4s
632 .long 0x4e351ed6 // and v22.16b, v22.16b, v21.16b
633 .long 0x4e351f18 // and v24.16b, v24.16b, v21.16b
634 .long 0x4e351f35 // and v21.16b, v25.16b, v21.16b
635 .long 0x6ea1dad9 // frsqrte v25.4s, v22.4s
636 .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s
637 .long 0x4ebdfedd // frsqrts v29.4s, v22.4s, v29.4s
638 .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s
639 .long 0x4ea1db3d // frecpe v29.4s, v25.4s
640 .long 0x6ea0fada // fneg v26.4s, v22.4s
641 .long 0x6ea1db1b // frsqrte v27.4s, v24.4s
642 .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s
643 .long 0x4e3dcf3a // fmla v26.4s, v25.4s, v29.4s
644 .long 0x6e3bdf7d // fmul v29.4s, v27.4s, v27.4s
645 .long 0x4ebdff1d // frsqrts v29.4s, v24.4s, v29.4s
646 .long 0x6e3ddf7b // fmul v27.4s, v27.4s, v29.4s
647 .long 0x4ea1db7d // frecpe v29.4s, v27.4s
648 .long 0x6ea0fb1c // fneg v28.4s, v24.4s
649 .long 0x6ea1dab9 // frsqrte v25.4s, v21.4s
650 .long 0x4e3dff7b // frecps v27.4s, v27.4s, v29.4s
651 .long 0x4e3dcf7c // fmla v28.4s, v27.4s, v29.4s
652 .long 0x6e39df3d // fmul v29.4s, v25.4s, v25.4s
653 .long 0x4ebdfebd // frsqrts v29.4s, v21.4s, v29.4s
654 .long 0x6e3ddf39 // fmul v25.4s, v25.4s, v29.4s
655 .long 0x4ea1db3d // frecpe v29.4s, v25.4s
656 .long 0x6ea0fabb // fneg v27.4s, v21.4s
657 .long 0x4e3dff39 // frecps v25.4s, v25.4s, v29.4s
658 .long 0x4e3dcf3b // fmla v27.4s, v25.4s, v29.4s
659 .long 0x4e36d6d9 // fadd v25.4s, v22.4s, v22.4s
660 .long 0x4f07f613 // fmov v19.4s, #-1.000000000000000000e+00
661 .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s
662 .long 0x4e24d497 // fadd v23.4s, v4.4s, v4.4s
663 .long 0x4e33d6dd // fadd v29.4s, v22.4s, v19.4s
664 .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s
665 .long 0x4f00f794 // fmov v20.4s, #7.000000000000000000e+00
666 .long 0x6e39dfb9 // fmul v25.4s, v29.4s, v25.4s
667 .long 0x4e37d6f7 // fadd v23.4s, v23.4s, v23.4s
668 .long 0x6e37e4f7 // fcmge v23.4s, v7.4s, v23.4s
669 .long 0x4e36ce99 // fmla v25.4s, v20.4s, v22.4s
670 .long 0x6e7a1f37 // bsl v23.16b, v25.16b, v26.16b
671 .long 0x4e38d719 // fadd v25.4s, v24.4s, v24.4s
672 .long 0x4e39d739 // fadd v25.4s, v25.4s, v25.4s
673 .long 0x4e33d71a // fadd v26.4s, v24.4s, v19.4s
674 .long 0x4e39cf39 // fmla v25.4s, v25.4s, v25.4s
675 .long 0x6e39df59 // fmul v25.4s, v26.4s, v25.4s
676 .long 0x4e25d4ba // fadd v26.4s, v5.4s, v5.4s
677 .long 0x4e3ad75a // fadd v26.4s, v26.4s, v26.4s
678 .long 0x6e3ae4fa // fcmge v26.4s, v7.4s, v26.4s
679 .long 0x4e38ce99 // fmla v25.4s, v20.4s, v24.4s
680 .long 0x6e7c1f3a // bsl v26.16b, v25.16b, v28.16b
681 .long 0x4e35d6bc // fadd v28.4s, v21.4s, v21.4s
682 .long 0x4e3cd79c // fadd v28.4s, v28.4s, v28.4s
683 .long 0x4e33d6b3 // fadd v19.4s, v21.4s, v19.4s
684 .long 0x4e3ccf9c // fmla v28.4s, v28.4s, v28.4s
685 .long 0x6e3cde73 // fmul v19.4s, v19.4s, v28.4s
686 .long 0x4e35ce93 // fmla v19.4s, v20.4s, v21.4s
687 .long 0x4e26d4d4 // fadd v20.4s, v6.4s, v6.4s
688 .long 0x4e34d694 // fadd v20.4s, v20.4s, v20.4s
689 .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
690 .long 0x6e34e4f4 // fcmge v20.4s, v7.4s, v20.4s
691 .long 0x4e20d411 // fadd v17.4s, v0.4s, v0.4s
692 .long 0x6e7b1e74 // bsl v20.16b, v19.16b, v27.16b
693 .long 0x4ea7d65b // fsub v27.4s, v18.4s, v7.4s
694 .long 0x4ea31c70 // mov v16.16b, v3.16b
695 .long 0x4e21d43d // fadd v29.4s, v1.4s, v1.4s
696 .long 0x4e22d45c // fadd v28.4s, v2.4s, v2.4s
697 .long 0x6e20df60 // fmul v0.4s, v27.4s, v0.4s
698 .long 0x6e21df61 // fmul v1.4s, v27.4s, v1.4s
699 .long 0x6e22df62 // fmul v2.4s, v27.4s, v2.4s
700 .long 0x4ea3d63b // fsub v27.4s, v17.4s, v3.4s
701 .long 0x4eb6d656 // fsub v22.4s, v18.4s, v22.4s
702 .long 0x4ea31c79 // mov v25.16b, v3.16b
703 .long 0x4e3bced0 // fmla v16.4s, v22.4s, v27.4s
704 .long 0x4ea3d7b6 // fsub v22.4s, v29.4s, v3.4s
705 .long 0x4eb8d658 // fsub v24.4s, v18.4s, v24.4s
706 .long 0x4ea31c73 // mov v19.16b, v3.16b
707 .long 0x4e36cf19 // fmla v25.4s, v24.4s, v22.4s
708 .long 0x4ea3d798 // fsub v24.4s, v28.4s, v3.4s
709 .long 0x4eb5d655 // fsub v21.4s, v18.4s, v21.4s
710 .long 0x4e38ceb3 // fmla v19.4s, v21.4s, v24.4s
711 .long 0x6e27df7b // fmul v27.4s, v27.4s, v7.4s
712 .long 0x6e27ded6 // fmul v22.4s, v22.4s, v7.4s
713 .long 0x6e27df18 // fmul v24.4s, v24.4s, v7.4s
714 .long 0xf8408423 // ldr x3, [x1], #8
715 .long 0x6e37df77 // fmul v23.4s, v27.4s, v23.4s
716 .long 0x6e3aded6 // fmul v22.4s, v22.4s, v26.4s
717 .long 0x6e34df14 // fmul v20.4s, v24.4s, v20.4s
718 .long 0x4ea3d652 // fsub v18.4s, v18.4s, v3.4s
719 .long 0x6e31e471 // fcmge v17.4s, v3.4s, v17.4s
720 .long 0x6e3de475 // fcmge v21.4s, v3.4s, v29.4s
721 .long 0x6e3ce47c // fcmge v28.4s, v3.4s, v28.4s
722 .long 0x6e24de10 // fmul v16.4s, v16.4s, v4.4s
723 .long 0x6e25df39 // fmul v25.4s, v25.4s, v5.4s
724 .long 0x6e26de73 // fmul v19.4s, v19.4s, v6.4s
725 .long 0x4e23cc97 // fmla v23.4s, v4.4s, v3.4s
726 .long 0x4e23ccb6 // fmla v22.4s, v5.4s, v3.4s
727 .long 0x4e23ccd4 // fmla v20.4s, v6.4s, v3.4s
728 .long 0x4e32cc80 // fmla v0.4s, v4.4s, v18.4s
729 .long 0x4e32cca1 // fmla v1.4s, v5.4s, v18.4s
730 .long 0x4e32ccc2 // fmla v2.4s, v6.4s, v18.4s
731 .long 0x6e771e11 // bsl v17.16b, v16.16b, v23.16b
732 .long 0x6e761f35 // bsl v21.16b, v25.16b, v22.16b
733 .long 0x6e741e7c // bsl v28.16b, v19.16b, v20.16b
734 .long 0x4e31d400 // fadd v0.4s, v0.4s, v17.4s
735 .long 0x4e35d421 // fadd v1.4s, v1.4s, v21.4s
736 .long 0x4e3cd442 // fadd v2.4s, v2.4s, v28.4s
737 .long 0x4e27ce43 // fmla v3.4s, v18.4s, v7.4s
738 .long 0xd61f0060 // br x3
739
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]740HIDDEN _sk_clamp_0_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]741.globl _sk_clamp_0_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]742FUNCTION(_sk_clamp_0_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]743_sk_clamp_0_aarch64:
744 .long 0xf8408423 // ldr x3, [x1], #8
745 .long 0x6f00e410 // movi v16.2d, #0x0
746 .long 0x4e30f400 // fmax v0.4s, v0.4s, v16.4s
747 .long 0x4e30f421 // fmax v1.4s, v1.4s, v16.4s
748 .long 0x4e30f442 // fmax v2.4s, v2.4s, v16.4s
749 .long 0x4e30f463 // fmax v3.4s, v3.4s, v16.4s
750 .long 0xd61f0060 // br x3
751
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]752HIDDEN _sk_clamp_1_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]753.globl _sk_clamp_1_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]754FUNCTION(_sk_clamp_1_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]755_sk_clamp_1_aarch64:
756 .long 0xf8408423 // ldr x3, [x1], #8
757 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
758 .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s
759 .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s
760 .long 0x4eb0f442 // fmin v2.4s, v2.4s, v16.4s
761 .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s
762 .long 0xd61f0060 // br x3
763
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]764HIDDEN _sk_clamp_a_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]765.globl _sk_clamp_a_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]766FUNCTION(_sk_clamp_a_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]767_sk_clamp_a_aarch64:
768 .long 0xf8408423 // ldr x3, [x1], #8
769 .long 0x4f03f610 // fmov v16.4s, #1.000000000000000000e+00
770 .long 0x4eb0f463 // fmin v3.4s, v3.4s, v16.4s
771 .long 0x4ea3f400 // fmin v0.4s, v0.4s, v3.4s
772 .long 0x4ea3f421 // fmin v1.4s, v1.4s, v3.4s
773 .long 0x4ea3f442 // fmin v2.4s, v2.4s, v3.4s
774 .long 0xd61f0060 // br x3
775
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]776HIDDEN _sk_set_rgb_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]777.globl _sk_set_rgb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]778FUNCTION(_sk_set_rgb_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]779_sk_set_rgb_aarch64:
780 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
781 .long 0xaa0803e9 // mov x9, x8
782 .long 0x4ddfc920 // ld1r {v0.4s}, [x9], #4
783 .long 0x91002108 // add x8, x8, #0x8
784 .long 0x4d40c902 // ld1r {v2.4s}, [x8]
785 .long 0x4d40c921 // ld1r {v1.4s}, [x9]
786 .long 0xd61f0060 // br x3
787
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]788HIDDEN _sk_swap_rb_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]789.globl _sk_swap_rb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]790FUNCTION(_sk_swap_rb_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]791_sk_swap_rb_aarch64:
792 .long 0xf8408423 // ldr x3, [x1], #8
793 .long 0x4ea01c10 // mov v16.16b, v0.16b
794 .long 0x4ea21c40 // mov v0.16b, v2.16b
795 .long 0x4eb01e02 // mov v2.16b, v16.16b
796 .long 0xd61f0060 // br x3
797
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]798HIDDEN _sk_swap_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]799.globl _sk_swap_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]800FUNCTION(_sk_swap_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]801_sk_swap_aarch64:
802 .long 0xf8408423 // ldr x3, [x1], #8
803 .long 0x4ea31c70 // mov v16.16b, v3.16b
804 .long 0x4ea21c51 // mov v17.16b, v2.16b
805 .long 0x4ea11c32 // mov v18.16b, v1.16b
806 .long 0x4ea01c13 // mov v19.16b, v0.16b
807 .long 0x4ea41c80 // mov v0.16b, v4.16b
808 .long 0x4ea51ca1 // mov v1.16b, v5.16b
809 .long 0x4ea61cc2 // mov v2.16b, v6.16b
810 .long 0x4ea71ce3 // mov v3.16b, v7.16b
811 .long 0x4eb31e64 // mov v4.16b, v19.16b
812 .long 0x4eb21e45 // mov v5.16b, v18.16b
813 .long 0x4eb11e26 // mov v6.16b, v17.16b
814 .long 0x4eb01e07 // mov v7.16b, v16.16b
815 .long 0xd61f0060 // br x3
816
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]817HIDDEN _sk_move_src_dst_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]818.globl _sk_move_src_dst_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]819FUNCTION(_sk_move_src_dst_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]820_sk_move_src_dst_aarch64:
821 .long 0xf8408423 // ldr x3, [x1], #8
822 .long 0x4ea01c04 // mov v4.16b, v0.16b
823 .long 0x4ea11c25 // mov v5.16b, v1.16b
824 .long 0x4ea21c46 // mov v6.16b, v2.16b
825 .long 0x4ea31c67 // mov v7.16b, v3.16b
826 .long 0xd61f0060 // br x3
827
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]828HIDDEN _sk_move_dst_src_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]829.globl _sk_move_dst_src_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]830FUNCTION(_sk_move_dst_src_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]831_sk_move_dst_src_aarch64:
832 .long 0xf8408423 // ldr x3, [x1], #8
833 .long 0x4ea41c80 // mov v0.16b, v4.16b
834 .long 0x4ea51ca1 // mov v1.16b, v5.16b
835 .long 0x4ea61cc2 // mov v2.16b, v6.16b
836 .long 0x4ea71ce3 // mov v3.16b, v7.16b
837 .long 0xd61f0060 // br x3
838
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]839HIDDEN _sk_premul_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]840.globl _sk_premul_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]841FUNCTION(_sk_premul_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]842_sk_premul_aarch64:
843 .long 0xf8408423 // ldr x3, [x1], #8
844 .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s
845 .long 0x6e23dc21 // fmul v1.4s, v1.4s, v3.4s
846 .long 0x6e23dc42 // fmul v2.4s, v2.4s, v3.4s
847 .long 0xd61f0060 // br x3
848
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]849HIDDEN _sk_unpremul_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]850.globl _sk_unpremul_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]851FUNCTION(_sk_unpremul_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]852_sk_unpremul_aarch64:
853 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
854 .long 0xf8408423 // ldr x3, [x1], #8
855 .long 0x4ea0d870 // fcmeq v16.4s, v3.4s, #0.0
856 .long 0x6e23fe31 // fdiv v17.4s, v17.4s, v3.4s
857 .long 0x4e701e30 // bic v16.16b, v17.16b, v16.16b
858 .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
859 .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
860 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
861 .long 0xd61f0060 // br x3
862
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]863HIDDEN _sk_from_srgb_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]864.globl _sk_from_srgb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]865FUNCTION(_sk_from_srgb_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]866_sk_from_srgb_aarch64:
867 .long 0x52a7d328 // mov w8, #0x3e990000
868 .long 0x72933348 // movk w8, #0x999a
869 .long 0x4e040d10 // dup v16.4s, w8
870 .long 0x52a7e648 // mov w8, #0x3f320000
871 .long 0x7291eb88 // movk w8, #0x8f5c
872 .long 0x4e040d11 // dup v17.4s, w8
873 .long 0x52a76468 // mov w8, #0x3b230000
874 .long 0x729ae148 // movk w8, #0xd70a
875 .long 0x4e040d12 // dup v18.4s, w8
876 .long 0x52a7b3c8 // mov w8, #0x3d9e0000
877 .long 0x72907228 // movk w8, #0x8391
878 .long 0x6e22dc54 // fmul v20.4s, v2.4s, v2.4s
879 .long 0x4eb11e35 // mov v21.16b, v17.16b
880 .long 0x4eb11e37 // mov v23.16b, v17.16b
881 .long 0x4e22ce11 // fmla v17.4s, v16.4s, v2.4s
882 .long 0x4eb21e56 // mov v22.16b, v18.16b
883 .long 0x4eb21e58 // mov v24.16b, v18.16b
884 .long 0x4e34ce32 // fmla v18.4s, v17.4s, v20.4s
885 .long 0x4e040d11 // dup v17.4s, w8
886 .long 0x52a7ac28 // mov w8, #0x3d610000
887 .long 0x6e20dc13 // fmul v19.4s, v0.4s, v0.4s
888 .long 0x7288f5c8 // movk w8, #0x47ae
889 .long 0x4e20ce15 // fmla v21.4s, v16.4s, v0.4s
890 .long 0xf8408423 // ldr x3, [x1], #8
891 .long 0x6e21dc34 // fmul v20.4s, v1.4s, v1.4s
892 .long 0x4e33ceb6 // fmla v22.4s, v21.4s, v19.4s
893 .long 0x4e040d13 // dup v19.4s, w8
894 .long 0x4e21ce17 // fmla v23.4s, v16.4s, v1.4s
895 .long 0x6e31dc15 // fmul v21.4s, v0.4s, v17.4s
896 .long 0x6ea0e660 // fcmgt v0.4s, v19.4s, v0.4s
897 .long 0x6e31dc30 // fmul v16.4s, v1.4s, v17.4s
898 .long 0x6ea1e661 // fcmgt v1.4s, v19.4s, v1.4s
899 .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s
900 .long 0x6ea2e662 // fcmgt v2.4s, v19.4s, v2.4s
901 .long 0x4e34cef8 // fmla v24.4s, v23.4s, v20.4s
902 .long 0x6e761ea0 // bsl v0.16b, v21.16b, v22.16b
903 .long 0x6e781e01 // bsl v1.16b, v16.16b, v24.16b
904 .long 0x6e721e22 // bsl v2.16b, v17.16b, v18.16b
905 .long 0xd61f0060 // br x3
906
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]907HIDDEN _sk_to_srgb_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]908.globl _sk_to_srgb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]909FUNCTION(_sk_to_srgb_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]910_sk_to_srgb_aarch64:
911 .long 0x52a828e8 // mov w8, #0x41470000
912 .long 0x728b8528 // movk w8, #0x5c29
913 .long 0x4e040d12 // dup v18.4s, w8
914 .long 0x52a7e608 // mov w8, #0x3f300000
915 .long 0x728df9c8 // movk w8, #0x6fce
916 .long 0x6ea1d811 // frsqrte v17.4s, v0.4s
917 .long 0x4e040d13 // dup v19.4s, w8
918 .long 0x52b7b948 // mov w8, #0xbdca0000
919 .long 0x728af508 // movk w8, #0x57a8
920 .long 0x6ea1d834 // frsqrte v20.4s, v1.4s
921 .long 0x6e31de36 // fmul v22.4s, v17.4s, v17.4s
922 .long 0x4e040d10 // dup v16.4s, w8
923 .long 0x52a77188 // mov w8, #0x3b8c0000
924 .long 0x6ea1d855 // frsqrte v21.4s, v2.4s
925 .long 0x6e34de98 // fmul v24.4s, v20.4s, v20.4s
926 .long 0x4eb6fc16 // frsqrts v22.4s, v0.4s, v22.4s
927 .long 0x729ce088 // movk w8, #0xe704
928 .long 0x6e35deb9 // fmul v25.4s, v21.4s, v21.4s
929 .long 0x4eb8fc38 // frsqrts v24.4s, v1.4s, v24.4s
930 .long 0x6e36de31 // fmul v17.4s, v17.4s, v22.4s
931 .long 0x4e040d17 // dup v23.4s, w8
932 .long 0x4eb9fc59 // frsqrts v25.4s, v2.4s, v25.4s
933 .long 0x6e38de94 // fmul v20.4s, v20.4s, v24.4s
934 .long 0x4ea1da36 // frecpe v22.4s, v17.4s
935 .long 0x6e32dc1a // fmul v26.4s, v0.4s, v18.4s
936 .long 0x6ea0e6e0 // fcmgt v0.4s, v23.4s, v0.4s
937 .long 0x6e32dc3c // fmul v28.4s, v1.4s, v18.4s
938 .long 0x6ea1e6e1 // fcmgt v1.4s, v23.4s, v1.4s
939 .long 0x6e32dc52 // fmul v18.4s, v2.4s, v18.4s
940 .long 0x6ea2e6e2 // fcmgt v2.4s, v23.4s, v2.4s
941 .long 0x6e39deb5 // fmul v21.4s, v21.4s, v25.4s
942 .long 0x4ea1da97 // frecpe v23.4s, v20.4s
943 .long 0x4e36fe39 // frecps v25.4s, v17.4s, v22.4s
944 .long 0x4ea1dab8 // frecpe v24.4s, v21.4s
945 .long 0x6e39ded6 // fmul v22.4s, v22.4s, v25.4s
946 .long 0x4e37fe99 // frecps v25.4s, v20.4s, v23.4s
947 .long 0x4eb01e1b // mov v27.16b, v16.16b
948 .long 0x6e39def7 // fmul v23.4s, v23.4s, v25.4s
949 .long 0x4e38feb9 // frecps v25.4s, v21.4s, v24.4s
950 .long 0x6e39df18 // fmul v24.4s, v24.4s, v25.4s
951 .long 0x4eb01e19 // mov v25.16b, v16.16b
952 .long 0x4e36ce7b // fmla v27.4s, v19.4s, v22.4s
953 .long 0x6ea1da36 // frsqrte v22.4s, v17.4s
954 .long 0x4e37ce79 // fmla v25.4s, v19.4s, v23.4s
955 .long 0x6ea1da97 // frsqrte v23.4s, v20.4s
956 .long 0x4e38ce70 // fmla v16.4s, v19.4s, v24.4s
957 .long 0x6e36ded8 // fmul v24.4s, v22.4s, v22.4s
958 .long 0x6ea1dab3 // frsqrte v19.4s, v21.4s
959 .long 0x4eb8fe31 // frsqrts v17.4s, v17.4s, v24.4s
960 .long 0x6e37def8 // fmul v24.4s, v23.4s, v23.4s
961 .long 0x4eb8fe94 // frsqrts v20.4s, v20.4s, v24.4s
962 .long 0x6e33de78 // fmul v24.4s, v19.4s, v19.4s
963 .long 0x52a7da48 // mov w8, #0x3ed20000
964 .long 0x4eb8feb5 // frsqrts v21.4s, v21.4s, v24.4s
965 .long 0x7290f848 // movk w8, #0x87c2
966 .long 0x6e31ded1 // fmul v17.4s, v22.4s, v17.4s
967 .long 0x6e34def4 // fmul v20.4s, v23.4s, v20.4s
968 .long 0x6e35de73 // fmul v19.4s, v19.4s, v21.4s
969 .long 0x4e040d15 // dup v21.4s, w8
970 .long 0xf8408423 // ldr x3, [x1], #8
971 .long 0x4e31cebb // fmla v27.4s, v21.4s, v17.4s
972 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
973 .long 0x4e34ceb9 // fmla v25.4s, v21.4s, v20.4s
974 .long 0x4e33ceb0 // fmla v16.4s, v21.4s, v19.4s
975 .long 0x4ebbf633 // fmin v19.4s, v17.4s, v27.4s
976 .long 0x4eb9f634 // fmin v20.4s, v17.4s, v25.4s
977 .long 0x4eb0f630 // fmin v16.4s, v17.4s, v16.4s
978 .long 0x6e731f40 // bsl v0.16b, v26.16b, v19.16b
979 .long 0x6e741f81 // bsl v1.16b, v28.16b, v20.16b
980 .long 0x6e701e42 // bsl v2.16b, v18.16b, v16.16b
981 .long 0xd61f0060 // br x3
982
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]983HIDDEN _sk_rgb_to_hsl_aarch64
984.globl _sk_rgb_to_hsl_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]985FUNCTION(_sk_rgb_to_hsl_aarch64)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]986_sk_rgb_to_hsl_aarch64:
987 .long 0x4e21f410 // fmax v16.4s, v0.4s, v1.4s
988 .long 0x4ea1f411 // fmin v17.4s, v0.4s, v1.4s
989 .long 0x6ea1e454 // fcmgt v20.4s, v2.4s, v1.4s
990 .long 0x4f00f715 // fmov v21.4s, #6.000000000000000000e+00
991 .long 0x4e22f610 // fmax v16.4s, v16.4s, v2.4s
992 .long 0x4ea2f631 // fmin v17.4s, v17.4s, v2.4s
993 .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
994 .long 0x4e341eb4 // and v20.16b, v21.16b, v20.16b
995 .long 0x4eb1d615 // fsub v21.4s, v16.4s, v17.4s
996 .long 0x4ea2d433 // fsub v19.4s, v1.4s, v2.4s
997 .long 0x4ea0d456 // fsub v22.4s, v2.4s, v0.4s
998 .long 0x4f026417 // movi v23.4s, #0x40, lsl #24
999 .long 0x6e35fe42 // fdiv v2.4s, v18.4s, v21.4s
1000 .long 0x4ea1d418 // fsub v24.4s, v0.4s, v1.4s
1001 .long 0x4f00f619 // fmov v25.4s, #4.000000000000000000e+00
1002 .long 0x4f0167fa // movi v26.4s, #0x3f, lsl #24
1003 .long 0x4eb0d6f2 // fsub v18.4s, v23.4s, v16.4s
1004 .long 0x4e36cc57 // fmla v23.4s, v2.4s, v22.4s
1005 .long 0x4e31e616 // fcmeq v22.4s, v16.4s, v17.4s
1006 .long 0x4e20e600 // fcmeq v0.4s, v16.4s, v0.4s
1007 .long 0x4e21e601 // fcmeq v1.4s, v16.4s, v1.4s
1008 .long 0x4e31d610 // fadd v16.4s, v16.4s, v17.4s
1009 .long 0x52a7c548 // mov w8, #0x3e2a0000
1010 .long 0x4e33cc54 // fmla v20.4s, v2.4s, v19.4s
1011 .long 0x4e38cc59 // fmla v25.4s, v2.4s, v24.4s
1012 .long 0x6e3ade02 // fmul v2.4s, v16.4s, v26.4s
1013 .long 0x72955568 // movk w8, #0xaaab
1014 .long 0xf8408423 // ldr x3, [x1], #8
1015 .long 0x4eb1d651 // fsub v17.4s, v18.4s, v17.4s
1016 .long 0x6ebae452 // fcmgt v18.4s, v2.4s, v26.4s
1017 .long 0x6e791ee1 // bsl v1.16b, v23.16b, v25.16b
1018 .long 0x4e040d13 // dup v19.4s, w8
1019 .long 0x6e701e32 // bsl v18.16b, v17.16b, v16.16b
1020 .long 0x6e611e80 // bsl v0.16b, v20.16b, v1.16b
1021 .long 0x6e32fea1 // fdiv v1.4s, v21.4s, v18.4s
1022 .long 0x6e33dc00 // fmul v0.4s, v0.4s, v19.4s
1023 .long 0x4e761c00 // bic v0.16b, v0.16b, v22.16b
1024 .long 0x4e761c21 // bic v1.16b, v1.16b, v22.16b
1025 .long 0xd61f0060 // br x3
1026
1027HIDDEN _sk_hsl_to_rgb_aarch64
1028.globl _sk_hsl_to_rgb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1029FUNCTION(_sk_hsl_to_rgb_aarch64)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]1030_sk_hsl_to_rgb_aarch64:
1031 .long 0x52a7d548 // mov w8, #0x3eaa0000
1032 .long 0x72955568 // movk w8, #0xaaab
1033 .long 0x4e040d17 // dup v23.4s, w8
1034 .long 0x52a7c548 // mov w8, #0x3e2a0000
1035 .long 0x72955568 // movk w8, #0xaaab
1036 .long 0x4e040d13 // dup v19.4s, w8
1037 .long 0x52a7e548 // mov w8, #0x3f2a0000
1038 .long 0x4f03f612 // fmov v18.4s, #1.000000000000000000e+00
1039 .long 0x4f07f616 // fmov v22.4s, #-1.000000000000000000e+00
1040 .long 0x72955568 // movk w8, #0xaaab
1041 .long 0x4e22d435 // fadd v21.4s, v1.4s, v2.4s
1042 .long 0x4e040d1a // dup v26.4s, w8
1043 .long 0x52b7d548 // mov w8, #0xbeaa0000
1044 .long 0x6eb2e41d // fcmgt v29.4s, v0.4s, v18.4s
1045 .long 0x4e36d41e // fadd v30.4s, v0.4s, v22.4s
1046 .long 0x4f0167f1 // movi v17.4s, #0x3f, lsl #24
1047 .long 0x4ea0d830 // fcmeq v16.4s, v1.4s, #0.0
1048 .long 0x4ea0e819 // fcmlt v25.4s, v0.4s, #0.0
1049 .long 0x72955568 // movk w8, #0xaaab
1050 .long 0x4e32d43c // fadd v28.4s, v1.4s, v18.4s
1051 .long 0x4ea2cc35 // fmls v21.4s, v1.4s, v2.4s
1052 .long 0x4e32d401 // fadd v1.4s, v0.4s, v18.4s
1053 .long 0x6e601fdd // bsl v29.16b, v30.16b, v0.16b
1054 .long 0x4e37d417 // fadd v23.4s, v0.4s, v23.4s
1055 .long 0x6ea2e63b // fcmgt v27.4s, v17.4s, v2.4s
1056 .long 0x4e040d1e // dup v30.4s, w8
1057 .long 0x6e22df9c // fmul v28.4s, v28.4s, v2.4s
1058 .long 0x6e7d1c39 // bsl v25.16b, v1.16b, v29.16b
1059 .long 0x6eb2e6e1 // fcmgt v1.4s, v23.4s, v18.4s
1060 .long 0x4e36d6fd // fadd v29.4s, v23.4s, v22.4s
1061 .long 0x4e3ed41e // fadd v30.4s, v0.4s, v30.4s
1062 .long 0x6e751f9b // bsl v27.16b, v28.16b, v21.16b
1063 .long 0x4ea0eaf5 // fcmlt v21.4s, v23.4s, #0.0
1064 .long 0x4e32d6fc // fadd v28.4s, v23.4s, v18.4s
1065 .long 0x6e771fa1 // bsl v1.16b, v29.16b, v23.16b
1066 .long 0x4f026414 // movi v20.4s, #0x40, lsl #24
1067 .long 0x6e611f95 // bsl v21.16b, v28.16b, v1.16b
1068 .long 0x4e32d7c1 // fadd v1.4s, v30.4s, v18.4s
1069 .long 0x6eb2e7d2 // fcmgt v18.4s, v30.4s, v18.4s
1070 .long 0x4e36d7d6 // fadd v22.4s, v30.4s, v22.4s
1071 .long 0x6ea0fb7c // fneg v28.4s, v27.4s
1072 .long 0x4ea0ebdd // fcmlt v29.4s, v30.4s, #0.0
1073 .long 0x6e7e1ed2 // bsl v18.16b, v22.16b, v30.16b
1074 .long 0x4e22ce9c // fmla v28.4s, v20.4s, v2.4s
1075 .long 0x4f00f718 // fmov v24.4s, #6.000000000000000000e+00
1076 .long 0x6e721c3d // bsl v29.16b, v1.16b, v18.16b
1077 .long 0x4ebcd761 // fsub v1.4s, v27.4s, v28.4s
1078 .long 0x4eb5d752 // fsub v18.4s, v26.4s, v21.4s
1079 .long 0x4ebc1f94 // mov v20.16b, v28.16b
1080 .long 0x6e38dc38 // fmul v24.4s, v1.4s, v24.4s
1081 .long 0x4eb9d756 // fsub v22.4s, v26.4s, v25.4s
1082 .long 0x4ebc1f9f // mov v31.16b, v28.16b
1083 .long 0x4e32cf14 // fmla v20.4s, v24.4s, v18.4s
1084 .long 0x4ebc1f81 // mov v1.16b, v28.16b
1085 .long 0x4ebc1f92 // mov v18.16b, v28.16b
1086 .long 0x4e38cc1f // fmla v31.4s, v0.4s, v24.4s
1087 .long 0x4e36cf01 // fmla v1.4s, v24.4s, v22.4s
1088 .long 0x4ebdd740 // fsub v0.4s, v26.4s, v29.4s
1089 .long 0x4e3ecf12 // fmla v18.4s, v24.4s, v30.4s
1090 .long 0x4ebc1f96 // mov v22.16b, v28.16b
1091 .long 0x6eb5e75e // fcmgt v30.4s, v26.4s, v21.4s
1092 .long 0x4e20cf16 // fmla v22.4s, v24.4s, v0.4s
1093 .long 0x6e7c1e9e // bsl v30.16b, v20.16b, v28.16b
1094 .long 0x6eb9e754 // fcmgt v20.4s, v26.4s, v25.4s
1095 .long 0x6ebde75a // fcmgt v26.4s, v26.4s, v29.4s
1096 .long 0x6e7c1c34 // bsl v20.16b, v1.16b, v28.16b
1097 .long 0x6e7c1eda // bsl v26.16b, v22.16b, v28.16b
1098 .long 0x4e37cf1c // fmla v28.4s, v24.4s, v23.4s
1099 .long 0x6eb9e637 // fcmgt v23.4s, v17.4s, v25.4s
1100 .long 0x6eb5e678 // fcmgt v24.4s, v19.4s, v21.4s
1101 .long 0x6eb5e635 // fcmgt v21.4s, v17.4s, v21.4s
1102 .long 0x6ebde631 // fcmgt v17.4s, v17.4s, v29.4s
1103 .long 0xf8408423 // ldr x3, [x1], #8
1104 .long 0x6eb9e676 // fcmgt v22.4s, v19.4s, v25.4s
1105 .long 0x6ebde673 // fcmgt v19.4s, v19.4s, v29.4s
1106 .long 0x6e7a1f71 // bsl v17.16b, v27.16b, v26.16b
1107 .long 0x6e7e1f75 // bsl v21.16b, v27.16b, v30.16b
1108 .long 0x6e741f77 // bsl v23.16b, v27.16b, v20.16b
1109 .long 0x6e711e53 // bsl v19.16b, v18.16b, v17.16b
1110 .long 0x4eb01e00 // mov v0.16b, v16.16b
1111 .long 0x4eb01e01 // mov v1.16b, v16.16b
1112 .long 0x6e751f98 // bsl v24.16b, v28.16b, v21.16b
1113 .long 0x6e771ff6 // bsl v22.16b, v31.16b, v23.16b
1114 .long 0x6e731c50 // bsl v16.16b, v2.16b, v19.16b
1115 .long 0x6e781c40 // bsl v0.16b, v2.16b, v24.16b
1116 .long 0x6e761c41 // bsl v1.16b, v2.16b, v22.16b
1117 .long 0x4eb01e02 // mov v2.16b, v16.16b
1118 .long 0xd61f0060 // br x3
1119
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1120HIDDEN _sk_scale_1_float_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1121.globl _sk_scale_1_float_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1122FUNCTION(_sk_scale_1_float_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1123_sk_scale_1_float_aarch64:
1124 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1125 .long 0xbd400110 // ldr s16, [x8]
1126 .long 0x4f909000 // fmul v0.4s, v0.4s, v16.s[0]
1127 .long 0x4f909021 // fmul v1.4s, v1.4s, v16.s[0]
1128 .long 0x4f909042 // fmul v2.4s, v2.4s, v16.s[0]
1129 .long 0x4f909063 // fmul v3.4s, v3.4s, v16.s[0]
1130 .long 0xd61f0060 // br x3
1131
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1132HIDDEN _sk_scale_u8_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1133.globl _sk_scale_u8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1134FUNCTION(_sk_scale_u8_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1135_sk_scale_u8_aarch64:
1136 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1137 .long 0x52a77009 // mov w9, #0x3b800000
1138 .long 0x72901029 // movk w9, #0x8081
1139 .long 0x4e040d30 // dup v16.4s, w9
1140 .long 0xf9400108 // ldr x8, [x8]
1141 .long 0x8b000108 // add x8, x8, x0
1142 .long 0x39400109 // ldrb w9, [x8]
1143 .long 0x3940050a // ldrb w10, [x8, #1]
1144 .long 0x3940090b // ldrb w11, [x8, #2]
1145 .long 0x39400d08 // ldrb w8, [x8, #3]
1146 .long 0x4e021d31 // mov v17.h[0], w9
1147 .long 0x4e061d51 // mov v17.h[1], w10
1148 .long 0x4e0a1d71 // mov v17.h[2], w11
1149 .long 0x4e0e1d11 // mov v17.h[3], w8
1150 .long 0x2f10a631 // uxtl v17.4s, v17.4h
1151 .long 0x6e21da31 // ucvtf v17.4s, v17.4s
1152 .long 0x6e30de30 // fmul v16.4s, v17.4s, v16.4s
1153 .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
1154 .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
1155 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
1156 .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
1157 .long 0xd61f0060 // br x3
1158
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1159HIDDEN _sk_lerp_1_float_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1160.globl _sk_lerp_1_float_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1161FUNCTION(_sk_lerp_1_float_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1162_sk_lerp_1_float_aarch64:
1163 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1164 .long 0x4ea4d411 // fsub v17.4s, v0.4s, v4.4s
1165 .long 0x4ea41c80 // mov v0.16b, v4.16b
1166 .long 0x4ea5d432 // fsub v18.4s, v1.4s, v5.4s
1167 .long 0xbd400110 // ldr s16, [x8]
1168 .long 0x4ea51ca1 // mov v1.16b, v5.16b
1169 .long 0x4f901220 // fmla v0.4s, v17.4s, v16.s[0]
1170 .long 0x4ea6d451 // fsub v17.4s, v2.4s, v6.4s
1171 .long 0x4f901241 // fmla v1.4s, v18.4s, v16.s[0]
1172 .long 0x4ea61cc2 // mov v2.16b, v6.16b
1173 .long 0x4ea7d472 // fsub v18.4s, v3.4s, v7.4s
1174 .long 0x4ea71ce3 // mov v3.16b, v7.16b
1175 .long 0x4f901222 // fmla v2.4s, v17.4s, v16.s[0]
1176 .long 0x4f901243 // fmla v3.4s, v18.4s, v16.s[0]
1177 .long 0xd61f0060 // br x3
1178
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1179HIDDEN _sk_lerp_u8_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1180.globl _sk_lerp_u8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1181FUNCTION(_sk_lerp_u8_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1182_sk_lerp_u8_aarch64:
1183 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1184 .long 0x52a77009 // mov w9, #0x3b800000
1185 .long 0x72901029 // movk w9, #0x8081
1186 .long 0x4e040d30 // dup v16.4s, w9
1187 .long 0xf9400108 // ldr x8, [x8]
1188 .long 0x4ea4d412 // fsub v18.4s, v0.4s, v4.4s
1189 .long 0x8b000108 // add x8, x8, x0
1190 .long 0x3940010a // ldrb w10, [x8]
1191 .long 0x39400509 // ldrb w9, [x8, #1]
1192 .long 0x3940090b // ldrb w11, [x8, #2]
1193 .long 0x39400d08 // ldrb w8, [x8, #3]
1194 .long 0x4e021d51 // mov v17.h[0], w10
1195 .long 0x4e061d31 // mov v17.h[1], w9
1196 .long 0x4e0a1d71 // mov v17.h[2], w11
1197 .long 0x4e0e1d11 // mov v17.h[3], w8
1198 .long 0x2f10a620 // uxtl v0.4s, v17.4h
1199 .long 0x6e21d800 // ucvtf v0.4s, v0.4s
1200 .long 0x6e30dc10 // fmul v16.4s, v0.4s, v16.4s
1201 .long 0x4ea41c80 // mov v0.16b, v4.16b
1202 .long 0x4ea5d431 // fsub v17.4s, v1.4s, v5.4s
1203 .long 0x4ea51ca1 // mov v1.16b, v5.16b
1204 .long 0x4e32ce00 // fmla v0.4s, v16.4s, v18.4s
1205 .long 0x4ea6d452 // fsub v18.4s, v2.4s, v6.4s
1206 .long 0x4e31ce01 // fmla v1.4s, v16.4s, v17.4s
1207 .long 0x4ea61cc2 // mov v2.16b, v6.16b
1208 .long 0x4ea7d471 // fsub v17.4s, v3.4s, v7.4s
1209 .long 0x4ea71ce3 // mov v3.16b, v7.16b
1210 .long 0x4e32ce02 // fmla v2.4s, v16.4s, v18.4s
1211 .long 0x4e31ce03 // fmla v3.4s, v16.4s, v17.4s
1212 .long 0xd61f0060 // br x3
1213
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1214HIDDEN _sk_lerp_565_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1215.globl _sk_lerp_565_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1216FUNCTION(_sk_lerp_565_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1217_sk_lerp_565_aarch64:
1218 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1219 .long 0xd37ff809 // lsl x9, x0, #1
1220 .long 0x4f072710 // movi v16.4s, #0xf8, lsl #8
1221 .long 0x4ea4d413 // fsub v19.4s, v0.4s, v4.4s
1222 .long 0xf9400108 // ldr x8, [x8]
1223 .long 0xfc696903 // ldr d3, [x8, x9]
1224 .long 0x52a6f088 // mov w8, #0x37840000
1225 .long 0x72842108 // movk w8, #0x2108
1226 .long 0x4e040d11 // dup v17.4s, w8
1227 .long 0x2f10a463 // uxtl v3.4s, v3.4h
1228 .long 0x321b17e8 // orr w8, wzr, #0x7e0
1229 .long 0x4e301c60 // and v0.16b, v3.16b, v16.16b
1230 .long 0x4e040d12 // dup v18.4s, w8
1231 .long 0x52a74048 // mov w8, #0x3a020000
1232 .long 0x4e21d800 // scvtf v0.4s, v0.4s
1233 .long 0x72810428 // movk w8, #0x821
1234 .long 0x6e31dc10 // fmul v16.4s, v0.4s, v17.4s
1235 .long 0x4ea41c80 // mov v0.16b, v4.16b
1236 .long 0x4e33ce00 // fmla v0.4s, v16.4s, v19.4s
1237 .long 0x4f0007f0 // movi v16.4s, #0x1f
1238 .long 0x4e040d11 // dup v17.4s, w8
1239 .long 0x52a7a088 // mov w8, #0x3d040000
1240 .long 0x4e321c72 // and v18.16b, v3.16b, v18.16b
1241 .long 0x72842108 // movk w8, #0x2108
1242 .long 0x4e301c63 // and v3.16b, v3.16b, v16.16b
1243 .long 0x4ea6d450 // fsub v16.4s, v2.4s, v6.4s
1244 .long 0x4e21da42 // scvtf v2.4s, v18.4s
1245 .long 0x6e31dc51 // fmul v17.4s, v2.4s, v17.4s
1246 .long 0x4e040d02 // dup v2.4s, w8
1247 .long 0x4e21d863 // scvtf v3.4s, v3.4s
1248 .long 0x4ea5d433 // fsub v19.4s, v1.4s, v5.4s
1249 .long 0x4ea51ca1 // mov v1.16b, v5.16b
1250 .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s
1251 .long 0x4ea61cc2 // mov v2.16b, v6.16b
1252 .long 0x4e33ce21 // fmla v1.4s, v17.4s, v19.4s
1253 .long 0x4e30cc62 // fmla v2.4s, v3.4s, v16.4s
1254 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
1255 .long 0xd61f0060 // br x3
1256
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]1257HIDDEN _sk_load_tables_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1258.globl _sk_load_tables_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1259FUNCTION(_sk_load_tables_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]1260_sk_load_tables_aarch64:
1261 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1262 .long 0xd37ef409 // lsl x9, x0, #2
1263 .long 0x6f00e620 // movi v0.2d, #0xff000000ff
1264 .long 0x52a7700b // mov w11, #0x3b800000
1265 .long 0xa940310a // ldp x10, x12, [x8]
1266 .long 0x7290102b // movk w11, #0x8081
1267 .long 0x4e040d63 // dup v3.4s, w11
1268 .long 0x3ce96942 // ldr q2, [x10, x9]
1269 .long 0xa9412109 // ldp x9, x8, [x8, #16]
1270 .long 0x4e201c41 // and v1.16b, v2.16b, v0.16b
1271 .long 0x1e26002e // fmov w14, s1
1272 .long 0x6f380450 // ushr v16.4s, v2.4s, #8
1273 .long 0x6f300451 // ushr v17.4s, v2.4s, #16
1274 .long 0x8b2e498e // add x14, x12, w14, uxtw #2
1275 .long 0x0e0c3c2a // mov w10, v1.s[1]
1276 .long 0x0e143c2b // mov w11, v1.s[2]
1277 .long 0x0e1c3c2d // mov w13, v1.s[3]
1278 .long 0x4e201e01 // and v1.16b, v16.16b, v0.16b
1279 .long 0x4e201e30 // and v16.16b, v17.16b, v0.16b
1280 .long 0x0d4081c0 // ld1 {v0.s}[0], [x14]
1281 .long 0x8b2a498a // add x10, x12, w10, uxtw #2
1282 .long 0xbc6b5991 // ldr s17, [x12, w11, uxtw #2]
1283 .long 0xbc6d5992 // ldr s18, [x12, w13, uxtw #2]
1284 .long 0x0e0c3c2b // mov w11, v1.s[1]
1285 .long 0x0e143c2c // mov w12, v1.s[2]
1286 .long 0x0e1c3c2d // mov w13, v1.s[3]
1287 .long 0x1e26002e // fmov w14, s1
1288 .long 0x8b2e492e // add x14, x9, w14, uxtw #2
1289 .long 0xbc6c5933 // ldr s19, [x9, w12, uxtw #2]
1290 .long 0xbc6d5934 // ldr s20, [x9, w13, uxtw #2]
1291 .long 0x8b2b4929 // add x9, x9, w11, uxtw #2
1292 .long 0x1e26020b // fmov w11, s16
1293 .long 0x6f280442 // ushr v2.4s, v2.4s, #24
1294 .long 0x0d409140 // ld1 {v0.s}[1], [x10]
1295 .long 0x4e21d842 // scvtf v2.4s, v2.4s
1296 .long 0x8b2b490a // add x10, x8, w11, uxtw #2
1297 .long 0x0d4081c1 // ld1 {v1.s}[0], [x14]
1298 .long 0x6e23dc43 // fmul v3.4s, v2.4s, v3.4s
1299 .long 0x0d408142 // ld1 {v2.s}[0], [x10]
1300 .long 0x0e0c3e0f // mov w15, v16.s[1]
1301 .long 0x0e143e0c // mov w12, v16.s[2]
1302 .long 0x8b2f490a // add x10, x8, w15, uxtw #2
1303 .long 0x0e1c3e0d // mov w13, v16.s[3]
1304 .long 0xbc6c5910 // ldr s16, [x8, w12, uxtw #2]
1305 .long 0x0d409121 // ld1 {v1.s}[1], [x9]
1306 .long 0x0d409142 // ld1 {v2.s}[1], [x10]
1307 .long 0x6e140620 // mov v0.s[2], v17.s[0]
1308 .long 0xbc6d5911 // ldr s17, [x8, w13, uxtw #2]
1309 .long 0x6e140661 // mov v1.s[2], v19.s[0]
1310 .long 0x6e140602 // mov v2.s[2], v16.s[0]
1311 .long 0x6e1c0640 // mov v0.s[3], v18.s[0]
1312 .long 0x6e1c0681 // mov v1.s[3], v20.s[0]
1313 .long 0x6e1c0622 // mov v2.s[3], v17.s[0]
1314 .long 0xd61f0060 // br x3
1315
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]1316HIDDEN _sk_load_tables_u16_be_aarch64
1317.globl _sk_load_tables_u16_be_aarch64
1318FUNCTION(_sk_load_tables_u16_be_aarch64)
1319_sk_load_tables_u16_be_aarch64:
1320 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1321 .long 0x52a6f00a // mov w10, #0x37800000
1322 .long 0x7280100a // movk w10, #0x80
1323 .long 0x4e040d50 // dup v16.4s, w10
1324 .long 0xa9402d09 // ldp x9, x11, [x8]
1325 .long 0x8b000d29 // add x9, x9, x0, lsl #3
1326 .long 0x0c400520 // ld4 {v0.4h-v3.4h}, [x9]
1327 .long 0xa9412109 // ldp x9, x8, [x8, #16]
1328 .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8
1329 .long 0x2f10a411 // uxtl v17.4s, v0.4h
1330 .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8
1331 .long 0x1e26022e // fmov w14, s17
1332 .long 0x0e0c3e2a // mov w10, v17.s[1]
1333 .long 0x0e143e2c // mov w12, v17.s[2]
1334 .long 0x0e1c3e2d // mov w13, v17.s[3]
1335 .long 0x8b2e496e // add x14, x11, w14, uxtw #2
1336 .long 0x2f10a431 // uxtl v17.4s, v1.4h
1337 .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8
1338 .long 0x2f10a453 // uxtl v19.4s, v2.4h
1339 .long 0x2f180462 // ushr v2.4h, v3.4h, #8
1340 .long 0x0d4081c0 // ld1 {v0.s}[0], [x14]
1341 .long 0x0f185474 // shl v20.4h, v3.4h, #8
1342 .long 0x0ea21e82 // orr v2.8b, v20.8b, v2.8b
1343 .long 0x8b2a496a // add x10, x11, w10, uxtw #2
1344 .long 0x1e26022e // fmov w14, s17
1345 .long 0x1e26026f // fmov w15, s19
1346 .long 0x2f10a442 // uxtl v2.4s, v2.4h
1347 .long 0x8b2e492e // add x14, x9, w14, uxtw #2
1348 .long 0x0d409140 // ld1 {v0.s}[1], [x10]
1349 .long 0x8b2f490a // add x10, x8, w15, uxtw #2
1350 .long 0x6e21d842 // ucvtf v2.4s, v2.4s
1351 .long 0x0d4081c1 // ld1 {v1.s}[0], [x14]
1352 .long 0x6e30dc43 // fmul v3.4s, v2.4s, v16.4s
1353 .long 0x0d408142 // ld1 {v2.s}[0], [x10]
1354 .long 0xbc6c5972 // ldr s18, [x11, w12, uxtw #2]
1355 .long 0xbc6d5975 // ldr s21, [x11, w13, uxtw #2]
1356 .long 0x0e0c3e2b // mov w11, v17.s[1]
1357 .long 0x0e143e2c // mov w12, v17.s[2]
1358 .long 0x0e1c3e2d // mov w13, v17.s[3]
1359 .long 0x8b2b492b // add x11, x9, w11, uxtw #2
1360 .long 0xbc6c5931 // ldr s17, [x9, w12, uxtw #2]
1361 .long 0xbc6d5936 // ldr s22, [x9, w13, uxtw #2]
1362 .long 0x0e0c3e69 // mov w9, v19.s[1]
1363 .long 0x0e143e6c // mov w12, v19.s[2]
1364 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1365 .long 0x0e1c3e6d // mov w13, v19.s[3]
1366 .long 0xbc6c5913 // ldr s19, [x8, w12, uxtw #2]
1367 .long 0x0d409161 // ld1 {v1.s}[1], [x11]
1368 .long 0x0d409122 // ld1 {v2.s}[1], [x9]
1369 .long 0xbc6d5910 // ldr s16, [x8, w13, uxtw #2]
1370 .long 0x6e140640 // mov v0.s[2], v18.s[0]
1371 .long 0x6e140621 // mov v1.s[2], v17.s[0]
1372 .long 0x6e140662 // mov v2.s[2], v19.s[0]
1373 .long 0x6e1c06a0 // mov v0.s[3], v21.s[0]
1374 .long 0x6e1c06c1 // mov v1.s[3], v22.s[0]
1375 .long 0x6e1c0602 // mov v2.s[3], v16.s[0]
1376 .long 0xd61f0060 // br x3
1377
1378HIDDEN _sk_load_tables_rgb_u16_be_aarch64
1379.globl _sk_load_tables_rgb_u16_be_aarch64
1380FUNCTION(_sk_load_tables_rgb_u16_be_aarch64)
1381_sk_load_tables_rgb_u16_be_aarch64:
1382 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1383 .long 0x321f07ea // orr w10, wzr, #0x6
1384 .long 0xa9402d09 // ldp x9, x11, [x8]
1385 .long 0x9b0a2409 // madd x9, x0, x10, x9
1386 .long 0x0c404521 // ld3 {v1.4h-v3.4h}, [x9]
1387 .long 0xa9412109 // ldp x9, x8, [x8, #16]
1388 .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8
1389 .long 0x2f10a420 // uxtl v0.4s, v1.4h
1390 .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8
1391 .long 0x1e26000e // fmov w14, s0
1392 .long 0x0e143c0c // mov w12, v0.s[2]
1393 .long 0x8b2e496e // add x14, x11, w14, uxtw #2
1394 .long 0x2f10a450 // uxtl v16.4s, v2.4h
1395 .long 0x0e0c3c0a // mov w10, v0.s[1]
1396 .long 0x0e1c3c0d // mov w13, v0.s[3]
1397 .long 0xbc6c5971 // ldr s17, [x11, w12, uxtw #2]
1398 .long 0x0d4081c0 // ld1 {v0.s}[0], [x14]
1399 .long 0x0e143e0c // mov w12, v16.s[2]
1400 .long 0x1e26020e // fmov w14, s16
1401 .long 0x2f07b7e3 // bic v3.4h, #0xff, lsl #8
1402 .long 0x8b2a496a // add x10, x11, w10, uxtw #2
1403 .long 0xbc6d5972 // ldr s18, [x11, w13, uxtw #2]
1404 .long 0x0e0c3e0b // mov w11, v16.s[1]
1405 .long 0x0e1c3e0d // mov w13, v16.s[3]
1406 .long 0xbc6c5930 // ldr s16, [x9, w12, uxtw #2]
1407 .long 0x8b2e492c // add x12, x9, w14, uxtw #2
1408 .long 0x2f10a462 // uxtl v2.4s, v3.4h
1409 .long 0xbc6d5923 // ldr s3, [x9, w13, uxtw #2]
1410 .long 0x0d408181 // ld1 {v1.s}[0], [x12]
1411 .long 0x0e143c4c // mov w12, v2.s[2]
1412 .long 0x1e26004d // fmov w13, s2
1413 .long 0xbc6c5913 // ldr s19, [x8, w12, uxtw #2]
1414 .long 0x8b2d490c // add x12, x8, w13, uxtw #2
1415 .long 0x8b2b492b // add x11, x9, w11, uxtw #2
1416 .long 0x0e0c3c49 // mov w9, v2.s[1]
1417 .long 0x0d409140 // ld1 {v0.s}[1], [x10]
1418 .long 0x0e1c3c4a // mov w10, v2.s[3]
1419 .long 0x0d408182 // ld1 {v2.s}[0], [x12]
1420 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1421 .long 0x0d409161 // ld1 {v1.s}[1], [x11]
1422 .long 0x6e140620 // mov v0.s[2], v17.s[0]
1423 .long 0x0d409122 // ld1 {v2.s}[1], [x9]
1424 .long 0xbc6a5911 // ldr s17, [x8, w10, uxtw #2]
1425 .long 0x6e140601 // mov v1.s[2], v16.s[0]
1426 .long 0x6e1c0640 // mov v0.s[3], v18.s[0]
1427 .long 0x6e140662 // mov v2.s[2], v19.s[0]
1428 .long 0x6e1c0461 // mov v1.s[3], v3.s[0]
1429 .long 0x6e1c0622 // mov v2.s[3], v17.s[0]
1430 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
1431 .long 0xd61f0060 // br x3
1432
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]1433HIDDEN _sk_byte_tables_aarch64
1434.globl _sk_byte_tables_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1435FUNCTION(_sk_byte_tables_aarch64)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]1436_sk_byte_tables_aarch64:
1437 .long 0xd10083ff // sub sp, sp, #0x20
1438 .long 0xaa0103e8 // mov x8, x1
1439 .long 0x91002109 // add x9, x8, #0x8
1440 .long 0xa9014ff4 // stp x20, x19, [sp, #16]
1441 .long 0xf90007e9 // str x9, [sp, #8]
1442 .long 0xf8410429 // ldr x9, [x1], #16
1443 .long 0x52a86fea // mov w10, #0x437f0000
1444 .long 0x4e040d51 // dup v17.4s, w10
1445 .long 0x52a7700b // mov w11, #0x3b800000
1446 .long 0xa9405126 // ldp x6, x20, [x9]
1447 .long 0x6e31dc00 // fmul v0.4s, v0.4s, v17.4s
1448 .long 0x7290102b // movk w11, #0x8081
1449 .long 0x6e21a800 // fcvtnu v0.4s, v0.4s
1450 .long 0x4e040d70 // dup v16.4s, w11
1451 .long 0x0e0c3c0a // mov w10, v0.s[1]
1452 .long 0x0e143c0b // mov w11, v0.s[2]
1453 .long 0x0e1c3c0c // mov w12, v0.s[3]
1454 .long 0x1e26000d // fmov w13, s0
1455 .long 0x386d48cd // ldrb w13, [x6, w13, uxtw]
1456 .long 0x386a48ca // ldrb w10, [x6, w10, uxtw]
1457 .long 0x386b48cb // ldrb w11, [x6, w11, uxtw]
1458 .long 0x386c48cc // ldrb w12, [x6, w12, uxtw]
1459 .long 0xa9412526 // ldp x6, x9, [x9, #16]
1460 .long 0x6e31dc42 // fmul v2.4s, v2.4s, v17.4s
1461 .long 0x6e31dc21 // fmul v1.4s, v1.4s, v17.4s
1462 .long 0x6e31dc63 // fmul v3.4s, v3.4s, v17.4s
1463 .long 0x6e21a842 // fcvtnu v2.4s, v2.4s
1464 .long 0x6e21a821 // fcvtnu v1.4s, v1.4s
1465 .long 0x6e21a863 // fcvtnu v3.4s, v3.4s
1466 .long 0x0e0c3c52 // mov w18, v2.s[1]
1467 .long 0x0e143c43 // mov w3, v2.s[2]
1468 .long 0x0e1c3c44 // mov w4, v2.s[3]
1469 .long 0x1e260045 // fmov w5, s2
1470 .long 0x1e260031 // fmov w17, s1
1471 .long 0x386548c5 // ldrb w5, [x6, w5, uxtw]
1472 .long 0x387248d2 // ldrb w18, [x6, w18, uxtw]
1473 .long 0x386348c3 // ldrb w3, [x6, w3, uxtw]
1474 .long 0x386448c4 // ldrb w4, [x6, w4, uxtw]
1475 .long 0x1e260066 // fmov w6, s3
1476 .long 0x0e0c3c2e // mov w14, v1.s[1]
1477 .long 0x0e0c3c67 // mov w7, v3.s[1]
1478 .long 0x38714a91 // ldrb w17, [x20, w17, uxtw]
1479 .long 0x38664926 // ldrb w6, [x9, w6, uxtw]
1480 .long 0x0e143c2f // mov w15, v1.s[2]
1481 .long 0x0e1c3c30 // mov w16, v1.s[3]
1482 .long 0x0e143c73 // mov w19, v3.s[2]
1483 .long 0x386e4a8e // ldrb w14, [x20, w14, uxtw]
1484 .long 0x38674927 // ldrb w7, [x9, w7, uxtw]
1485 .long 0x386f4a8f // ldrb w15, [x20, w15, uxtw]
1486 .long 0x38704a90 // ldrb w16, [x20, w16, uxtw]
1487 .long 0x0e1c3c74 // mov w20, v3.s[3]
1488 .long 0x38734933 // ldrb w19, [x9, w19, uxtw]
1489 .long 0x38744929 // ldrb w9, [x9, w20, uxtw]
1490 .long 0x4e021da0 // mov v0.h[0], w13
1491 .long 0x4e021e21 // mov v1.h[0], w17
1492 .long 0x4e021ca2 // mov v2.h[0], w5
1493 .long 0x4e021cc3 // mov v3.h[0], w6
1494 .long 0x4e061d40 // mov v0.h[1], w10
1495 .long 0x4e061dc1 // mov v1.h[1], w14
1496 .long 0x4e061e42 // mov v2.h[1], w18
1497 .long 0x4e061ce3 // mov v3.h[1], w7
1498 .long 0x4e0a1d60 // mov v0.h[2], w11
1499 .long 0x4e0a1de1 // mov v1.h[2], w15
1500 .long 0x4e0a1c62 // mov v2.h[2], w3
1501 .long 0x4e0a1e63 // mov v3.h[2], w19
1502 .long 0x4e0e1d80 // mov v0.h[3], w12
1503 .long 0x4e0e1e01 // mov v1.h[3], w16
1504 .long 0x4e0e1c82 // mov v2.h[3], w4
1505 .long 0x4e0e1d23 // mov v3.h[3], w9
1506 .long 0xf9400505 // ldr x5, [x8, #8]
1507 .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8
1508 .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8
1509 .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8
1510 .long 0x2f07b7e3 // bic v3.4h, #0xff, lsl #8
1511 .long 0xa9414ff4 // ldp x20, x19, [sp, #16]
1512 .long 0x2f10a400 // uxtl v0.4s, v0.4h
1513 .long 0x2f10a421 // uxtl v1.4s, v1.4h
1514 .long 0x2f10a442 // uxtl v2.4s, v2.4h
1515 .long 0x2f10a463 // uxtl v3.4s, v3.4h
1516 .long 0x6e21d800 // ucvtf v0.4s, v0.4s
1517 .long 0x6e21d821 // ucvtf v1.4s, v1.4s
1518 .long 0x6e21d842 // ucvtf v2.4s, v2.4s
1519 .long 0x6e21d863 // ucvtf v3.4s, v3.4s
1520 .long 0x6e30dc00 // fmul v0.4s, v0.4s, v16.4s
1521 .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s
1522 .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s
1523 .long 0x6e30dc63 // fmul v3.4s, v3.4s, v16.4s
1524 .long 0x910083ff // add sp, sp, #0x20
1525 .long 0xd61f00a0 // br x5
1526
1527HIDDEN _sk_byte_tables_rgb_aarch64
1528.globl _sk_byte_tables_rgb_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]1529FUNCTION(_sk_byte_tables_rgb_aarch64)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]1530_sk_byte_tables_rgb_aarch64:
1531 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1532 .long 0x52a77009 // mov w9, #0x3b800000
1533 .long 0x72901029 // movk w9, #0x8081
1534 .long 0x4e040d30 // dup v16.4s, w9
1535 .long 0xb9401909 // ldr w9, [x8, #24]
1536 .long 0xa9402d0a // ldp x10, x11, [x8]
1537 .long 0xf9400908 // ldr x8, [x8, #16]
1538 .long 0x51000529 // sub w9, w9, #0x1
1539 .long 0x4e040d31 // dup v17.4s, w9
1540 .long 0x4e21da31 // scvtf v17.4s, v17.4s
1541 .long 0x6e21de21 // fmul v1.4s, v17.4s, v1.4s
1542 .long 0x6e20de20 // fmul v0.4s, v17.4s, v0.4s
1543 .long 0x6e22de22 // fmul v2.4s, v17.4s, v2.4s
1544 .long 0x6e21a821 // fcvtnu v1.4s, v1.4s
1545 .long 0x6e21a800 // fcvtnu v0.4s, v0.4s
1546 .long 0x6e21a842 // fcvtnu v2.4s, v2.4s
1547 .long 0x0e0c3c2f // mov w15, v1.s[1]
1548 .long 0x0e143c30 // mov w16, v1.s[2]
1549 .long 0x0e1c3c31 // mov w17, v1.s[3]
1550 .long 0x1e260032 // fmov w18, s1
1551 .long 0x1e26000e // fmov w14, s0
1552 .long 0x38724972 // ldrb w18, [x11, w18, uxtw]
1553 .long 0x386f496f // ldrb w15, [x11, w15, uxtw]
1554 .long 0x38704970 // ldrb w16, [x11, w16, uxtw]
1555 .long 0x3871496b // ldrb w11, [x11, w17, uxtw]
1556 .long 0x1e260051 // fmov w17, s2
1557 .long 0x0e0c3c09 // mov w9, v0.s[1]
1558 .long 0x386e494e // ldrb w14, [x10, w14, uxtw]
1559 .long 0x0e0c3c44 // mov w4, v2.s[1]
1560 .long 0x38714911 // ldrb w17, [x8, w17, uxtw]
1561 .long 0x0e143c0c // mov w12, v0.s[2]
1562 .long 0x0e1c3c0d // mov w13, v0.s[3]
1563 .long 0x0e143c45 // mov w5, v2.s[2]
1564 .long 0x38694949 // ldrb w9, [x10, w9, uxtw]
1565 .long 0x38644904 // ldrb w4, [x8, w4, uxtw]
1566 .long 0x386c494c // ldrb w12, [x10, w12, uxtw]
1567 .long 0x386d494a // ldrb w10, [x10, w13, uxtw]
1568 .long 0x0e1c3c4d // mov w13, v2.s[3]
1569 .long 0x38654905 // ldrb w5, [x8, w5, uxtw]
1570 .long 0x386d4908 // ldrb w8, [x8, w13, uxtw]
1571 .long 0x4e021dc0 // mov v0.h[0], w14
1572 .long 0x4e021e41 // mov v1.h[0], w18
1573 .long 0x4e021e22 // mov v2.h[0], w17
1574 .long 0x4e061d20 // mov v0.h[1], w9
1575 .long 0x4e061de1 // mov v1.h[1], w15
1576 .long 0x4e061c82 // mov v2.h[1], w4
1577 .long 0x4e0a1d80 // mov v0.h[2], w12
1578 .long 0x4e0a1e01 // mov v1.h[2], w16
1579 .long 0x4e0a1ca2 // mov v2.h[2], w5
1580 .long 0x4e0e1d40 // mov v0.h[3], w10
1581 .long 0x4e0e1d61 // mov v1.h[3], w11
1582 .long 0x4e0e1d02 // mov v2.h[3], w8
1583 .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8
1584 .long 0x2f07b7e1 // bic v1.4h, #0xff, lsl #8
1585 .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8
1586 .long 0x2f10a400 // uxtl v0.4s, v0.4h
1587 .long 0x2f10a421 // uxtl v1.4s, v1.4h
1588 .long 0x2f10a442 // uxtl v2.4s, v2.4h
1589 .long 0x6e21d800 // ucvtf v0.4s, v0.4s
1590 .long 0x6e21d821 // ucvtf v1.4s, v1.4s
1591 .long 0x6e21d842 // ucvtf v2.4s, v2.4s
1592 .long 0x6e30dc00 // fmul v0.4s, v0.4s, v16.4s
1593 .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s
1594 .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s
1595 .long 0xd61f0060 // br x3
1596
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]1597HIDDEN _sk_table_r_aarch64
1598.globl _sk_table_r_aarch64
1599FUNCTION(_sk_table_r_aarch64)
1600_sk_table_r_aarch64:
1601 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1602 .long 0xb9400909 // ldr w9, [x8, #8]
1603 .long 0xf9400108 // ldr x8, [x8]
1604 .long 0x51000529 // sub w9, w9, #0x1
1605 .long 0x4e040d30 // dup v16.4s, w9
1606 .long 0x4e21da10 // scvtf v16.4s, v16.4s
1607 .long 0x6e20de00 // fmul v0.4s, v16.4s, v0.4s
1608 .long 0x6e21a810 // fcvtnu v16.4s, v0.4s
1609 .long 0x1e26020b // fmov w11, s16
1610 .long 0x8b2b490b // add x11, x8, w11, uxtw #2
1611 .long 0x0d408160 // ld1 {v0.s}[0], [x11]
1612 .long 0x0e0c3e09 // mov w9, v16.s[1]
1613 .long 0x0e143e0a // mov w10, v16.s[2]
1614 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1615 .long 0x0e1c3e0b // mov w11, v16.s[3]
1616 .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2]
1617 .long 0x0d409120 // ld1 {v0.s}[1], [x9]
1618 .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2]
1619 .long 0x6e140600 // mov v0.s[2], v16.s[0]
1620 .long 0x6e1c0620 // mov v0.s[3], v17.s[0]
1621 .long 0xd61f0060 // br x3
1622
1623HIDDEN _sk_table_g_aarch64
1624.globl _sk_table_g_aarch64
1625FUNCTION(_sk_table_g_aarch64)
1626_sk_table_g_aarch64:
1627 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1628 .long 0xb9400909 // ldr w9, [x8, #8]
1629 .long 0xf9400108 // ldr x8, [x8]
1630 .long 0x51000529 // sub w9, w9, #0x1
1631 .long 0x4e040d30 // dup v16.4s, w9
1632 .long 0x4e21da10 // scvtf v16.4s, v16.4s
1633 .long 0x6e21de01 // fmul v1.4s, v16.4s, v1.4s
1634 .long 0x6e21a830 // fcvtnu v16.4s, v1.4s
1635 .long 0x1e26020b // fmov w11, s16
1636 .long 0x8b2b490b // add x11, x8, w11, uxtw #2
1637 .long 0x0d408161 // ld1 {v1.s}[0], [x11]
1638 .long 0x0e0c3e09 // mov w9, v16.s[1]
1639 .long 0x0e143e0a // mov w10, v16.s[2]
1640 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1641 .long 0x0e1c3e0b // mov w11, v16.s[3]
1642 .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2]
1643 .long 0x0d409121 // ld1 {v1.s}[1], [x9]
1644 .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2]
1645 .long 0x6e140601 // mov v1.s[2], v16.s[0]
1646 .long 0x6e1c0621 // mov v1.s[3], v17.s[0]
1647 .long 0xd61f0060 // br x3
1648
1649HIDDEN _sk_table_b_aarch64
1650.globl _sk_table_b_aarch64
1651FUNCTION(_sk_table_b_aarch64)
1652_sk_table_b_aarch64:
1653 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1654 .long 0xb9400909 // ldr w9, [x8, #8]
1655 .long 0xf9400108 // ldr x8, [x8]
1656 .long 0x51000529 // sub w9, w9, #0x1
1657 .long 0x4e040d30 // dup v16.4s, w9
1658 .long 0x4e21da10 // scvtf v16.4s, v16.4s
1659 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
1660 .long 0x6e21a850 // fcvtnu v16.4s, v2.4s
1661 .long 0x1e26020b // fmov w11, s16
1662 .long 0x8b2b490b // add x11, x8, w11, uxtw #2
1663 .long 0x0d408162 // ld1 {v2.s}[0], [x11]
1664 .long 0x0e0c3e09 // mov w9, v16.s[1]
1665 .long 0x0e143e0a // mov w10, v16.s[2]
1666 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1667 .long 0x0e1c3e0b // mov w11, v16.s[3]
1668 .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2]
1669 .long 0x0d409122 // ld1 {v2.s}[1], [x9]
1670 .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2]
1671 .long 0x6e140602 // mov v2.s[2], v16.s[0]
1672 .long 0x6e1c0622 // mov v2.s[3], v17.s[0]
1673 .long 0xd61f0060 // br x3
1674
1675HIDDEN _sk_table_a_aarch64
1676.globl _sk_table_a_aarch64
1677FUNCTION(_sk_table_a_aarch64)
1678_sk_table_a_aarch64:
1679 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
1680 .long 0xb9400909 // ldr w9, [x8, #8]
1681 .long 0xf9400108 // ldr x8, [x8]
1682 .long 0x51000529 // sub w9, w9, #0x1
1683 .long 0x4e040d30 // dup v16.4s, w9
1684 .long 0x4e21da10 // scvtf v16.4s, v16.4s
1685 .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
1686 .long 0x6e21a870 // fcvtnu v16.4s, v3.4s
1687 .long 0x1e26020b // fmov w11, s16
1688 .long 0x8b2b490b // add x11, x8, w11, uxtw #2
1689 .long 0x0d408163 // ld1 {v3.s}[0], [x11]
1690 .long 0x0e0c3e09 // mov w9, v16.s[1]
1691 .long 0x0e143e0a // mov w10, v16.s[2]
1692 .long 0x8b294909 // add x9, x8, w9, uxtw #2
1693 .long 0x0e1c3e0b // mov w11, v16.s[3]
1694 .long 0xbc6a5910 // ldr s16, [x8, w10, uxtw #2]
1695 .long 0x0d409123 // ld1 {v3.s}[1], [x9]
1696 .long 0xbc6b5911 // ldr s17, [x8, w11, uxtw #2]
1697 .long 0x6e140603 // mov v3.s[2], v16.s[0]
1698 .long 0x6e1c0623 // mov v3.s[3], v17.s[0]
1699 .long 0xd61f0060 // br x3
1700
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1701HIDDEN _sk_parametric_r_aarch64
1702.globl _sk_parametric_r_aarch64
1703FUNCTION(_sk_parametric_r_aarch64)
1704_sk_parametric_r_aarch64:
1705 .long 0xf9400028 // ldr x8, [x1]
1706 .long 0x4f016696 // movi v22.4s, #0x34, lsl #24
1707 .long 0x91004109 // add x9, x8, #0x10
1708 .long 0x9100610a // add x10, x8, #0x18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1709 .long 0x4d40c932 // ld1r {v18.4s}, [x9]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1710 .long 0xaa0803e9 // mov x9, x8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1711 .long 0xbd400d11 // ldr s17, [x8, #12]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1712 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1713 .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1714 .long 0x9100210a // add x10, x8, #0x8
1715 .long 0x4d40c954 // ld1r {v20.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1716 .long 0x4f911010 // fmla v16.4s, v0.4s, v17.s[0]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1717 .long 0xbd400135 // ldr s21, [x9]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1718 .long 0x52b85f09 // mov w9, #0xc2f80000
1719 .long 0x728e6ee9 // movk w9, #0x7377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1720 .long 0x4e040d37 // dup v23.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1721 .long 0x52a7f7e9 // mov w9, #0x3fbf0000
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1722 .long 0x7297eea9 // movk w9, #0xbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1723 .long 0x4f951014 // fmla v20.4s, v0.4s, v21.s[0]
1724 .long 0x6e20e640 // fcmge v0.4s, v18.4s, v0.4s
1725 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1726 .long 0x52a7d689 // mov w9, #0x3eb40000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1727 .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1728 .long 0x72889f29 // movk w9, #0x44f9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1729 .long 0x4e21da95 // scvtf v21.4s, v20.4s
1730 .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1731 .long 0x4e040d34 // dup v20.4s, w9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1732 .long 0x52a7fb89 // mov w9, #0x3fdc0000
1733 .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s
1734 .long 0x729d3469 // movk w9, #0xe9a3
1735 .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24
1736 .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s
1737 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1738 .long 0x52a85e49 // mov w9, #0x42f20000
1739 .long 0x72918a29 // movk w9, #0x8c51
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1740 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1741 .long 0x4e040d34 // dup v20.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1742 .long 0x52a7f7c9 // mov w9, #0x3fbe0000
1743 .long 0x729791a9 // movk w9, #0xbc8d
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1744 .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s
1745 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1746 .long 0x52a81349 // mov w9, #0x409a0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1747 .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1748 .long 0x729ebf09 // movk w9, #0xf5f8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1749 .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s
1750 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1751 .long 0x52a83ba9 // mov w9, #0x41dd0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1752 .long 0x4e219a33 // frintm v19.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1753 .long 0x729a5fc9 // movk w9, #0xd2fe
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1754 .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s
1755 .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s
1756 .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s
1757 .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s
1758 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1759 .long 0x91005108 // add x8, x8, #0x14
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1760 .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s
1761 .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s
1762 .long 0x4d40c914 // ld1r {v20.4s}, [x8]
1763 .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24
1764 .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1765 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
1766 .long 0xf9400423 // ldr x3, [x1, #8]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1767 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1768 .long 0x6f00e412 // movi v18.2d, #0x0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1769 .long 0x6e711e00 // bsl v0.16b, v16.16b, v17.16b
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1770 .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00
1771 .long 0x4e32f400 // fmax v0.4s, v0.4s, v18.4s
1772 .long 0x4eb5f400 // fmin v0.4s, v0.4s, v21.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1773 .long 0x91004021 // add x1, x1, #0x10
1774 .long 0xd61f0060 // br x3
1775
1776HIDDEN _sk_parametric_g_aarch64
1777.globl _sk_parametric_g_aarch64
1778FUNCTION(_sk_parametric_g_aarch64)
1779_sk_parametric_g_aarch64:
1780 .long 0xf9400028 // ldr x8, [x1]
1781 .long 0x4f016696 // movi v22.4s, #0x34, lsl #24
1782 .long 0x91004109 // add x9, x8, #0x10
1783 .long 0x9100610a // add x10, x8, #0x18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1784 .long 0x4d40c932 // ld1r {v18.4s}, [x9]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1785 .long 0xaa0803e9 // mov x9, x8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1786 .long 0xbd400d11 // ldr s17, [x8, #12]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1787 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1788 .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1789 .long 0x9100210a // add x10, x8, #0x8
1790 .long 0x4d40c954 // ld1r {v20.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1791 .long 0x4f911030 // fmla v16.4s, v1.4s, v17.s[0]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1792 .long 0xbd400135 // ldr s21, [x9]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1793 .long 0x52b85f09 // mov w9, #0xc2f80000
1794 .long 0x728e6ee9 // movk w9, #0x7377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1795 .long 0x4e040d37 // dup v23.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1796 .long 0x52a7f7e9 // mov w9, #0x3fbf0000
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1797 .long 0x7297eea9 // movk w9, #0xbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1798 .long 0x4f951034 // fmla v20.4s, v1.4s, v21.s[0]
1799 .long 0x6e21e641 // fcmge v1.4s, v18.4s, v1.4s
1800 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1801 .long 0x52a7d689 // mov w9, #0x3eb40000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1802 .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1803 .long 0x72889f29 // movk w9, #0x44f9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1804 .long 0x4e21da95 // scvtf v21.4s, v20.4s
1805 .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1806 .long 0x4e040d34 // dup v20.4s, w9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1807 .long 0x52a7fb89 // mov w9, #0x3fdc0000
1808 .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s
1809 .long 0x729d3469 // movk w9, #0xe9a3
1810 .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24
1811 .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s
1812 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1813 .long 0x52a85e49 // mov w9, #0x42f20000
1814 .long 0x72918a29 // movk w9, #0x8c51
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1815 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1816 .long 0x4e040d34 // dup v20.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1817 .long 0x52a7f7c9 // mov w9, #0x3fbe0000
1818 .long 0x729791a9 // movk w9, #0xbc8d
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1819 .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s
1820 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1821 .long 0x52a81349 // mov w9, #0x409a0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1822 .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1823 .long 0x729ebf09 // movk w9, #0xf5f8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1824 .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s
1825 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1826 .long 0x52a83ba9 // mov w9, #0x41dd0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1827 .long 0x4e219a33 // frintm v19.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1828 .long 0x729a5fc9 // movk w9, #0xd2fe
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1829 .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s
1830 .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s
1831 .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s
1832 .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s
1833 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1834 .long 0x91005108 // add x8, x8, #0x14
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1835 .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s
1836 .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s
1837 .long 0x4d40c914 // ld1r {v20.4s}, [x8]
1838 .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24
1839 .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1840 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
1841 .long 0xf9400423 // ldr x3, [x1, #8]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1842 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1843 .long 0x6f00e412 // movi v18.2d, #0x0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1844 .long 0x6e711e01 // bsl v1.16b, v16.16b, v17.16b
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1845 .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00
1846 .long 0x4e32f421 // fmax v1.4s, v1.4s, v18.4s
1847 .long 0x4eb5f421 // fmin v1.4s, v1.4s, v21.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1848 .long 0x91004021 // add x1, x1, #0x10
1849 .long 0xd61f0060 // br x3
1850
1851HIDDEN _sk_parametric_b_aarch64
1852.globl _sk_parametric_b_aarch64
1853FUNCTION(_sk_parametric_b_aarch64)
1854_sk_parametric_b_aarch64:
1855 .long 0xf9400028 // ldr x8, [x1]
1856 .long 0x4f016696 // movi v22.4s, #0x34, lsl #24
1857 .long 0x91004109 // add x9, x8, #0x10
1858 .long 0x9100610a // add x10, x8, #0x18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1859 .long 0x4d40c932 // ld1r {v18.4s}, [x9]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1860 .long 0xaa0803e9 // mov x9, x8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1861 .long 0xbd400d11 // ldr s17, [x8, #12]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1862 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1863 .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1864 .long 0x9100210a // add x10, x8, #0x8
1865 .long 0x4d40c954 // ld1r {v20.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1866 .long 0x4f911050 // fmla v16.4s, v2.4s, v17.s[0]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1867 .long 0xbd400135 // ldr s21, [x9]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1868 .long 0x52b85f09 // mov w9, #0xc2f80000
1869 .long 0x728e6ee9 // movk w9, #0x7377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1870 .long 0x4e040d37 // dup v23.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1871 .long 0x52a7f7e9 // mov w9, #0x3fbf0000
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1872 .long 0x7297eea9 // movk w9, #0xbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1873 .long 0x4f951054 // fmla v20.4s, v2.4s, v21.s[0]
1874 .long 0x6e22e642 // fcmge v2.4s, v18.4s, v2.4s
1875 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1876 .long 0x52a7d689 // mov w9, #0x3eb40000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1877 .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1878 .long 0x72889f29 // movk w9, #0x44f9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1879 .long 0x4e21da95 // scvtf v21.4s, v20.4s
1880 .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1881 .long 0x4e040d34 // dup v20.4s, w9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1882 .long 0x52a7fb89 // mov w9, #0x3fdc0000
1883 .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s
1884 .long 0x729d3469 // movk w9, #0xe9a3
1885 .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24
1886 .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s
1887 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1888 .long 0x52a85e49 // mov w9, #0x42f20000
1889 .long 0x72918a29 // movk w9, #0x8c51
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1890 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1891 .long 0x4e040d34 // dup v20.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1892 .long 0x52a7f7c9 // mov w9, #0x3fbe0000
1893 .long 0x729791a9 // movk w9, #0xbc8d
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1894 .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s
1895 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1896 .long 0x52a81349 // mov w9, #0x409a0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1897 .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1898 .long 0x729ebf09 // movk w9, #0xf5f8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1899 .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s
1900 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1901 .long 0x52a83ba9 // mov w9, #0x41dd0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1902 .long 0x4e219a33 // frintm v19.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1903 .long 0x729a5fc9 // movk w9, #0xd2fe
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1904 .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s
1905 .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s
1906 .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s
1907 .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s
1908 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1909 .long 0x91005108 // add x8, x8, #0x14
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1910 .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s
1911 .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s
1912 .long 0x4d40c914 // ld1r {v20.4s}, [x8]
1913 .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24
1914 .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1915 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
1916 .long 0xf9400423 // ldr x3, [x1, #8]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1917 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1918 .long 0x6f00e412 // movi v18.2d, #0x0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1919 .long 0x6e711e02 // bsl v2.16b, v16.16b, v17.16b
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1920 .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00
1921 .long 0x4e32f442 // fmax v2.4s, v2.4s, v18.4s
1922 .long 0x4eb5f442 // fmin v2.4s, v2.4s, v21.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1923 .long 0x91004021 // add x1, x1, #0x10
1924 .long 0xd61f0060 // br x3
1925
1926HIDDEN _sk_parametric_a_aarch64
1927.globl _sk_parametric_a_aarch64
1928FUNCTION(_sk_parametric_a_aarch64)
1929_sk_parametric_a_aarch64:
1930 .long 0xf9400028 // ldr x8, [x1]
1931 .long 0x4f016696 // movi v22.4s, #0x34, lsl #24
1932 .long 0x91004109 // add x9, x8, #0x10
1933 .long 0x9100610a // add x10, x8, #0x18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1934 .long 0x4d40c932 // ld1r {v18.4s}, [x9]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1935 .long 0xaa0803e9 // mov x9, x8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1936 .long 0xbd400d11 // ldr s17, [x8, #12]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1937 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1938 .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1939 .long 0x9100210a // add x10, x8, #0x8
1940 .long 0x4d40c954 // ld1r {v20.4s}, [x10]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1941 .long 0x4f911070 // fmla v16.4s, v3.4s, v17.s[0]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1942 .long 0xbd400135 // ldr s21, [x9]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1943 .long 0x52b85f09 // mov w9, #0xc2f80000
1944 .long 0x728e6ee9 // movk w9, #0x7377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1945 .long 0x4e040d37 // dup v23.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1946 .long 0x52a7f7e9 // mov w9, #0x3fbf0000
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1947 .long 0x7297eea9 // movk w9, #0xbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1948 .long 0x4f951074 // fmla v20.4s, v3.4s, v21.s[0]
1949 .long 0x6e23e643 // fcmge v3.4s, v18.4s, v3.4s
1950 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1951 .long 0x52a7d689 // mov w9, #0x3eb40000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1952 .long 0x4f03d7f1 // movi v17.4s, #0x7f, msl #16
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1953 .long 0x72889f29 // movk w9, #0x44f9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1954 .long 0x4e21da95 // scvtf v21.4s, v20.4s
1955 .long 0x4e311e91 // and v17.16b, v20.16b, v17.16b
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1956 .long 0x4e040d34 // dup v20.4s, w9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1957 .long 0x52a7fb89 // mov w9, #0x3fdc0000
1958 .long 0x4e35ced7 // fmla v23.4s, v22.4s, v21.4s
1959 .long 0x729d3469 // movk w9, #0xe9a3
1960 .long 0x4f0177f1 // orr v17.4s, #0x3f, lsl #24
1961 .long 0x4eb2ce37 // fmls v23.4s, v17.4s, v18.4s
1962 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1963 .long 0x52a85e49 // mov w9, #0x42f20000
1964 .long 0x72918a29 // movk w9, #0x8c51
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1965 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1966 .long 0x4e040d34 // dup v20.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1967 .long 0x52a7f7c9 // mov w9, #0x3fbe0000
1968 .long 0x729791a9 // movk w9, #0xbc8d
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1969 .long 0x6e31fe51 // fdiv v17.4s, v18.4s, v17.4s
1970 .long 0x4e040d32 // dup v18.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1971 .long 0x52a81349 // mov w9, #0x409a0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1972 .long 0x4eb1d6f1 // fsub v17.4s, v23.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1973 .long 0x729ebf09 // movk w9, #0xf5f8
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1974 .long 0x6e31de71 // fmul v17.4s, v19.4s, v17.4s
1975 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1976 .long 0x52a83ba9 // mov w9, #0x41dd0000
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1977 .long 0x4e219a33 // frintm v19.4s, v17.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1978 .long 0x729a5fc9 // movk w9, #0xd2fe
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1979 .long 0x4e34d634 // fadd v20.4s, v17.4s, v20.4s
1980 .long 0x4eb3d631 // fsub v17.4s, v17.4s, v19.4s
1981 .long 0x4eb2ce34 // fmls v20.4s, v17.4s, v18.4s
1982 .long 0x4eb1d6b1 // fsub v17.4s, v21.4s, v17.4s
1983 .long 0x4e040d35 // dup v21.4s, w9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1984 .long 0x91005108 // add x8, x8, #0x14
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1985 .long 0x6e31feb1 // fdiv v17.4s, v21.4s, v17.4s
1986 .long 0x4e31d691 // fadd v17.4s, v20.4s, v17.4s
1987 .long 0x4d40c914 // ld1r {v20.4s}, [x8]
1988 .long 0x4f026573 // movi v19.4s, #0x4b, lsl #24
1989 .long 0x6e33de31 // fmul v17.4s, v17.4s, v19.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1990 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
1991 .long 0xf9400423 // ldr x3, [x1, #8]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1992 .long 0x4e34d631 // fadd v17.4s, v17.4s, v20.4s
1993 .long 0x6f00e412 // movi v18.2d, #0x0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1994 .long 0x6e711e03 // bsl v3.16b, v16.16b, v17.16b
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]1995 .long 0x4f03f615 // fmov v21.4s, #1.000000000000000000e+00
1996 .long 0x4e32f463 // fmax v3.4s, v3.4s, v18.4s
1997 .long 0x4eb5f463 // fmin v3.4s, v3.4s, v21.4s
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]1998 .long 0x91004021 // add x1, x1, #0x10
1999 .long 0xd61f0060 // br x3
2000
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]2001HIDDEN _sk_lab_to_xyz_aarch64
2002.globl _sk_lab_to_xyz_aarch64
2003FUNCTION(_sk_lab_to_xyz_aarch64)
2004_sk_lab_to_xyz_aarch64:
2005 .long 0x52a85908 // mov w8, #0x42c80000
2006 .long 0x4e040d10 // dup v16.4s, w8
2007 .long 0x52a86fe8 // mov w8, #0x437f0000
2008 .long 0x4f066471 // movi v17.4s, #0xc3, lsl #24
2009 .long 0x4e040d13 // dup v19.4s, w8
2010 .long 0x52a781a8 // mov w8, #0x3c0d0000
2011 .long 0x7287b968 // movk w8, #0x3dcb
2012 .long 0x4eb11e34 // mov v20.16b, v17.16b
2013 .long 0x4e21ce74 // fmla v20.4s, v19.4s, v1.4s
2014 .long 0x4e040d01 // dup v1.4s, w8
2015 .long 0x52a76068 // mov w8, #0x3b030000
2016 .long 0x72824de8 // movk w8, #0x126f
2017 .long 0x4e22ce71 // fmla v17.4s, v19.4s, v2.4s
2018 .long 0x4e040d02 // dup v2.4s, w8
2019 .long 0x52a77468 // mov w8, #0x3ba30000
2020 .long 0x729ae148 // movk w8, #0xd70a
2021 .long 0x4e040d13 // dup v19.4s, w8
2022 .long 0x52a78228 // mov w8, #0x3c110000
2023 .long 0x4f01f612 // fmov v18.4s, #1.600000000000000000e+01
2024 .long 0x72831848 // movk w8, #0x18c2
2025 .long 0x4e20ce12 // fmla v18.4s, v16.4s, v0.4s
2026 .long 0x4e040d00 // dup v0.4s, w8
2027 .long 0x52b7c1a8 // mov w8, #0xbe0d0000
2028 .long 0x7287b968 // movk w8, #0x3dcb
2029 .long 0x6e21de41 // fmul v1.4s, v18.4s, v1.4s
2030 .long 0x4e040d10 // dup v16.4s, w8
2031 .long 0x52a7c068 // mov w8, #0x3e030000
2032 .long 0x4ea11c32 // mov v18.16b, v1.16b
2033 .long 0x72900a08 // movk w8, #0x8050
2034 .long 0x4eb3ce32 // fmls v18.4s, v17.4s, v19.4s
2035 .long 0x6e21dc31 // fmul v17.4s, v1.4s, v1.4s
2036 .long 0x4ea11c35 // mov v21.16b, v1.16b
2037 .long 0x4e30d433 // fadd v19.4s, v1.4s, v16.4s
2038 .long 0x6e31dc31 // fmul v17.4s, v1.4s, v17.4s
2039 .long 0x4e34cc55 // fmla v21.4s, v2.4s, v20.4s
2040 .long 0x4e040d02 // dup v2.4s, w8
2041 .long 0x6e22de73 // fmul v19.4s, v19.4s, v2.4s
2042 .long 0x6ea0e621 // fcmgt v1.4s, v17.4s, v0.4s
2043 .long 0x6e731e21 // bsl v1.16b, v17.16b, v19.16b
2044 .long 0x6e32de51 // fmul v17.4s, v18.4s, v18.4s
2045 .long 0x4e30d653 // fadd v19.4s, v18.4s, v16.4s
2046 .long 0x6e31de51 // fmul v17.4s, v18.4s, v17.4s
2047 .long 0x52a7eec8 // mov w8, #0x3f760000
2048 .long 0x6e22de72 // fmul v18.4s, v19.4s, v2.4s
2049 .long 0x6ea0e633 // fcmgt v19.4s, v17.4s, v0.4s
2050 .long 0x729ae3e8 // movk w8, #0xd71f
2051 .long 0x6e721e33 // bsl v19.16b, v17.16b, v18.16b
2052 .long 0x6e35deb2 // fmul v18.4s, v21.4s, v21.4s
2053 .long 0x4e040d11 // dup v17.4s, w8
2054 .long 0x52a7ea68 // mov w8, #0x3f530000
2055 .long 0x4e30d6b0 // fadd v16.4s, v21.4s, v16.4s
2056 .long 0x6e32deb2 // fmul v18.4s, v21.4s, v18.4s
2057 .long 0xf8408423 // ldr x3, [x1], #8
2058 .long 0x72881ec8 // movk w8, #0x40f6
2059 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
2060 .long 0x6ea0e640 // fcmgt v0.4s, v18.4s, v0.4s
2061 .long 0x4e040d14 // dup v20.4s, w8
2062 .long 0x6e621e40 // bsl v0.16b, v18.16b, v2.16b
2063 .long 0x6e31dc00 // fmul v0.4s, v0.4s, v17.4s
2064 .long 0x6e34de62 // fmul v2.4s, v19.4s, v20.4s
2065 .long 0xd61f0060 // br x3
2066
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2067HIDDEN _sk_load_a8_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2068.globl _sk_load_a8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2069FUNCTION(_sk_load_a8_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2070_sk_load_a8_aarch64:
2071 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2072 .long 0x52a77009 // mov w9, #0x3b800000
2073 .long 0x72901029 // movk w9, #0x8081
2074 .long 0x4e040d22 // dup v2.4s, w9
2075 .long 0xf9400108 // ldr x8, [x8]
2076 .long 0x6f00e400 // movi v0.2d, #0x0
2077 .long 0x6f00e401 // movi v1.2d, #0x0
2078 .long 0x8b000108 // add x8, x8, x0
2079 .long 0x3940010a // ldrb w10, [x8]
2080 .long 0x3940050b // ldrb w11, [x8, #1]
2081 .long 0x3940090c // ldrb w12, [x8, #2]
2082 .long 0x39400d08 // ldrb w8, [x8, #3]
2083 .long 0x4e021d43 // mov v3.h[0], w10
2084 .long 0x4e061d63 // mov v3.h[1], w11
2085 .long 0x4e0a1d83 // mov v3.h[2], w12
2086 .long 0x4e0e1d03 // mov v3.h[3], w8
2087 .long 0x2f10a463 // uxtl v3.4s, v3.4h
2088 .long 0x6e21d863 // ucvtf v3.4s, v3.4s
2089 .long 0x6e22dc63 // fmul v3.4s, v3.4s, v2.4s
2090 .long 0x6f00e402 // movi v2.2d, #0x0
2091 .long 0xd61f0060 // br x3
2092
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2093HIDDEN _sk_gather_a8_aarch64
2094.globl _sk_gather_a8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2095FUNCTION(_sk_gather_a8_aarch64)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2096_sk_gather_a8_aarch64:
2097 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2098 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2099 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2100 .long 0x91004109 // add x9, x8, #0x10
2101 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2102 .long 0xf9400108 // ldr x8, [x8]
2103 .long 0x52a77009 // mov w9, #0x3b800000
2104 .long 0x72901029 // movk w9, #0x8081
2105 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2106 .long 0x1e26000c // fmov w12, s0
2107 .long 0x4e040d23 // dup v3.4s, w9
2108 .long 0x0e0c3c09 // mov w9, v0.s[1]
2109 .long 0x386c490c // ldrb w12, [x8, w12, uxtw]
2110 .long 0x0e143c0a // mov w10, v0.s[2]
2111 .long 0x38694909 // ldrb w9, [x8, w9, uxtw]
2112 .long 0x0e1c3c0b // mov w11, v0.s[3]
2113 .long 0x386a490a // ldrb w10, [x8, w10, uxtw]
2114 .long 0x386b4908 // ldrb w8, [x8, w11, uxtw]
2115 .long 0x4e021d82 // mov v2.h[0], w12
2116 .long 0x4e061d22 // mov v2.h[1], w9
2117 .long 0x4e0a1d42 // mov v2.h[2], w10
2118 .long 0x4e0e1d02 // mov v2.h[3], w8
2119 .long 0x2f07b7e2 // bic v2.4h, #0xff, lsl #8
2120 .long 0x2f10a442 // uxtl v2.4s, v2.4h
2121 .long 0x6e21d842 // ucvtf v2.4s, v2.4s
2122 .long 0x6f00e400 // movi v0.2d, #0x0
2123 .long 0x6f00e401 // movi v1.2d, #0x0
2124 .long 0x6e23dc43 // fmul v3.4s, v2.4s, v3.4s
2125 .long 0x6f00e402 // movi v2.2d, #0x0
2126 .long 0xd61f0060 // br x3
2127
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2128HIDDEN _sk_store_a8_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2129.globl _sk_store_a8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2130FUNCTION(_sk_store_a8_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2131_sk_store_a8_aarch64:
2132 .long 0xf9400028 // ldr x8, [x1]
2133 .long 0x52a86fe9 // mov w9, #0x437f0000
2134 .long 0x4e040d30 // dup v16.4s, w9
2135 .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s
2136 .long 0xf9400108 // ldr x8, [x8]
2137 .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
2138 .long 0x0e612a10 // xtn v16.4h, v16.4s
2139 .long 0x0e0e3e09 // umov w9, v16.h[3]
2140 .long 0x8b000108 // add x8, x8, x0
2141 .long 0x39000d09 // strb w9, [x8, #3]
2142 .long 0x0e0a3e09 // umov w9, v16.h[2]
2143 .long 0x39000909 // strb w9, [x8, #2]
2144 .long 0x0e063e09 // umov w9, v16.h[1]
2145 .long 0x39000509 // strb w9, [x8, #1]
2146 .long 0x0e023e09 // umov w9, v16.h[0]
2147 .long 0x39000109 // strb w9, [x8]
2148 .long 0xf9400423 // ldr x3, [x1, #8]
2149 .long 0x91004021 // add x1, x1, #0x10
2150 .long 0xd61f0060 // br x3
2151
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2152HIDDEN _sk_load_g8_aarch64
2153.globl _sk_load_g8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2154FUNCTION(_sk_load_g8_aarch64)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2155_sk_load_g8_aarch64:
2156 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2157 .long 0x52a77009 // mov w9, #0x3b800000
2158 .long 0x72901029 // movk w9, #0x8081
2159 .long 0x4e040d20 // dup v0.4s, w9
2160 .long 0xf9400108 // ldr x8, [x8]
2161 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
2162 .long 0x8b000108 // add x8, x8, x0
2163 .long 0x3940010a // ldrb w10, [x8]
2164 .long 0x39400509 // ldrb w9, [x8, #1]
2165 .long 0x3940090b // ldrb w11, [x8, #2]
2166 .long 0x39400d08 // ldrb w8, [x8, #3]
2167 .long 0x4e021d41 // mov v1.h[0], w10
2168 .long 0x4e061d21 // mov v1.h[1], w9
2169 .long 0x4e0a1d61 // mov v1.h[2], w11
2170 .long 0x4e0e1d01 // mov v1.h[3], w8
2171 .long 0x2f10a421 // uxtl v1.4s, v1.4h
2172 .long 0x6e21d821 // ucvtf v1.4s, v1.4s
2173 .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
2174 .long 0x4ea01c01 // mov v1.16b, v0.16b
2175 .long 0x4ea01c02 // mov v2.16b, v0.16b
2176 .long 0xd61f0060 // br x3
2177
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2178HIDDEN _sk_gather_g8_aarch64
2179.globl _sk_gather_g8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2180FUNCTION(_sk_gather_g8_aarch64)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2181_sk_gather_g8_aarch64:
2182 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2183 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2184 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2185 .long 0x91004109 // add x9, x8, #0x10
2186 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2187 .long 0xf9400108 // ldr x8, [x8]
2188 .long 0x52a77009 // mov w9, #0x3b800000
2189 .long 0x72901029 // movk w9, #0x8081
2190 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2191 .long 0x1e26000c // fmov w12, s0
2192 .long 0x4e040d23 // dup v3.4s, w9
2193 .long 0x0e0c3c09 // mov w9, v0.s[1]
2194 .long 0x386c490c // ldrb w12, [x8, w12, uxtw]
2195 .long 0x0e143c0a // mov w10, v0.s[2]
2196 .long 0x38694909 // ldrb w9, [x8, w9, uxtw]
2197 .long 0x0e1c3c0b // mov w11, v0.s[3]
2198 .long 0x386a490a // ldrb w10, [x8, w10, uxtw]
2199 .long 0x386b4908 // ldrb w8, [x8, w11, uxtw]
2200 .long 0x4e021d80 // mov v0.h[0], w12
2201 .long 0x4e061d20 // mov v0.h[1], w9
2202 .long 0x4e0a1d40 // mov v0.h[2], w10
2203 .long 0x4e0e1d00 // mov v0.h[3], w8
2204 .long 0x2f07b7e0 // bic v0.4h, #0xff, lsl #8
2205 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2206 .long 0x6e21d800 // ucvtf v0.4s, v0.4s
2207 .long 0x6e23dc00 // fmul v0.4s, v0.4s, v3.4s
2208 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
2209 .long 0x4ea01c01 // mov v1.16b, v0.16b
2210 .long 0x4ea01c02 // mov v2.16b, v0.16b
2211 .long 0xd61f0060 // br x3
2212
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]2213HIDDEN _sk_gather_i8_aarch64
2214.globl _sk_gather_i8_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2215FUNCTION(_sk_gather_i8_aarch64)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]2216_sk_gather_i8_aarch64:
2217 .long 0xaa0103e8 // mov x8, x1
2218 .long 0xf8408429 // ldr x9, [x1], #8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]2219 .long 0xb4000069 // cbz x9, 1d28 <sk_gather_i8_aarch64+0x14>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]2220 .long 0xaa0903ea // mov x10, x9
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]2221 .long 0x14000003 // b 1d30 <sk_gather_i8_aarch64+0x1c>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]2222 .long 0xf940050a // ldr x10, [x8, #8]
2223 .long 0x91004101 // add x1, x8, #0x10
2224 .long 0xf8410548 // ldr x8, [x10], #16
2225 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2226 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2227 .long 0xf9400529 // ldr x9, [x9, #8]
2228 .long 0x4d40c942 // ld1r {v2.4s}, [x10]
2229 .long 0x6f00e623 // movi v3.2d, #0xff000000ff
2230 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2231 .long 0x1e26000d // fmov w13, s0
2232 .long 0x0e0c3c0a // mov w10, v0.s[1]
2233 .long 0x386d490d // ldrb w13, [x8, w13, uxtw]
2234 .long 0x0e143c0b // mov w11, v0.s[2]
2235 .long 0x386a490a // ldrb w10, [x8, w10, uxtw]
2236 .long 0x0e1c3c0c // mov w12, v0.s[3]
2237 .long 0x386b490b // ldrb w11, [x8, w11, uxtw]
2238 .long 0x386c4908 // ldrb w8, [x8, w12, uxtw]
2239 .long 0x4e021da0 // mov v0.h[0], w13
2240 .long 0x4e061d40 // mov v0.h[1], w10
2241 .long 0x4e0a1d60 // mov v0.h[2], w11
2242 .long 0x4e0e1d00 // mov v0.h[3], w8
2243 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2244 .long 0x4e231c00 // and v0.16b, v0.16b, v3.16b
2245 .long 0x1e26000c // fmov w12, s0
2246 .long 0x8b2c492c // add x12, x9, w12, uxtw #2
2247 .long 0x0e0c3c08 // mov w8, v0.s[1]
2248 .long 0x0e143c0a // mov w10, v0.s[2]
2249 .long 0x0e1c3c0b // mov w11, v0.s[3]
2250 .long 0x0d408180 // ld1 {v0.s}[0], [x12]
2251 .long 0x8b284928 // add x8, x9, w8, uxtw #2
2252 .long 0xb86a592a // ldr w10, [x9, w10, uxtw #2]
2253 .long 0x52a7700c // mov w12, #0x3b800000
2254 .long 0x0d409100 // ld1 {v0.s}[1], [x8]
2255 .long 0xb86b5928 // ldr w8, [x9, w11, uxtw #2]
2256 .long 0x7290102c // movk w12, #0x8081
2257 .long 0xf8408423 // ldr x3, [x1], #8
2258 .long 0x4e141d40 // mov v0.s[2], w10
2259 .long 0x4e1c1d00 // mov v0.s[3], w8
2260 .long 0x4e231c01 // and v1.16b, v0.16b, v3.16b
2261 .long 0x6f380402 // ushr v2.4s, v0.4s, #8
2262 .long 0x6f300411 // ushr v17.4s, v0.4s, #16
2263 .long 0x4e040d90 // dup v16.4s, w12
2264 .long 0x6f280400 // ushr v0.4s, v0.4s, #24
2265 .long 0x4e21d821 // scvtf v1.4s, v1.4s
2266 .long 0x4e231c42 // and v2.16b, v2.16b, v3.16b
2267 .long 0x4e231e23 // and v3.16b, v17.16b, v3.16b
2268 .long 0x4e21d811 // scvtf v17.4s, v0.4s
2269 .long 0x6e30dc20 // fmul v0.4s, v1.4s, v16.4s
2270 .long 0x4e21d841 // scvtf v1.4s, v2.4s
2271 .long 0x4e21d862 // scvtf v2.4s, v3.4s
2272 .long 0x6e30dc21 // fmul v1.4s, v1.4s, v16.4s
2273 .long 0x6e30dc42 // fmul v2.4s, v2.4s, v16.4s
2274 .long 0x6e30de23 // fmul v3.4s, v17.4s, v16.4s
2275 .long 0xd61f0060 // br x3
2276
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2277HIDDEN _sk_load_565_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2278.globl _sk_load_565_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2279FUNCTION(_sk_load_565_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2280_sk_load_565_aarch64:
2281 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2282 .long 0xd37ff809 // lsl x9, x0, #1
2283 .long 0x4f072701 // movi v1.4s, #0xf8, lsl #8
2284 .long 0x4f0007e3 // movi v3.4s, #0x1f
2285 .long 0xf9400108 // ldr x8, [x8]
2286 .long 0xfc696900 // ldr d0, [x8, x9]
2287 .long 0x321b17e8 // orr w8, wzr, #0x7e0
2288 .long 0x4e040d02 // dup v2.4s, w8
2289 .long 0x52a6f088 // mov w8, #0x37840000
2290 .long 0x72842108 // movk w8, #0x2108
2291 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2292 .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b
2293 .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b
2294 .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b
2295 .long 0x4e040d00 // dup v0.4s, w8
2296 .long 0x52a74048 // mov w8, #0x3a020000
2297 .long 0x72810428 // movk w8, #0x821
2298 .long 0x4e21d821 // scvtf v1.4s, v1.4s
2299 .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
2300 .long 0x4e040d01 // dup v1.4s, w8
2301 .long 0x52a7a088 // mov w8, #0x3d040000
2302 .long 0x72842108 // movk w8, #0x2108
2303 .long 0x4e21d842 // scvtf v2.4s, v2.4s
2304 .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s
2305 .long 0x4e040d02 // dup v2.4s, w8
2306 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2307 .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s
2308 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
2309 .long 0xd61f0060 // br x3
2310
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2311HIDDEN _sk_gather_565_aarch64
2312.globl _sk_gather_565_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2313FUNCTION(_sk_gather_565_aarch64)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2314_sk_gather_565_aarch64:
2315 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2316 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2317 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2318 .long 0x91004109 // add x9, x8, #0x10
2319 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2320 .long 0xf9400108 // ldr x8, [x8]
2321 .long 0x321b17e9 // orr w9, wzr, #0x7e0
2322 .long 0x4e040d23 // dup v3.4s, w9
2323 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2324 .long 0x1e26000c // fmov w12, s0
2325 .long 0x0e0c3c09 // mov w9, v0.s[1]
2326 .long 0x8b2c450c // add x12, x8, w12, uxtw #1
2327 .long 0x0e143c0a // mov w10, v0.s[2]
2328 .long 0x0e1c3c0b // mov w11, v0.s[3]
2329 .long 0x0d404180 // ld1 {v0.h}[0], [x12]
2330 .long 0x78695909 // ldrh w9, [x8, w9, uxtw #1]
2331 .long 0x786a590a // ldrh w10, [x8, w10, uxtw #1]
2332 .long 0x786b5908 // ldrh w8, [x8, w11, uxtw #1]
2333 .long 0x4f072701 // movi v1.4s, #0xf8, lsl #8
2334 .long 0x4e061d20 // mov v0.h[1], w9
2335 .long 0x4e0a1d40 // mov v0.h[2], w10
2336 .long 0x4e0e1d00 // mov v0.h[3], w8
2337 .long 0x52a6f08b // mov w11, #0x37840000
2338 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2339 .long 0x7284210b // movk w11, #0x2108
2340 .long 0x52a74049 // mov w9, #0x3a020000
2341 .long 0x4f0007e2 // movi v2.4s, #0x1f
2342 .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b
2343 .long 0x72810429 // movk w9, #0x821
2344 .long 0x52a7a08a // mov w10, #0x3d040000
2345 .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b
2346 .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b
2347 .long 0x4e040d60 // dup v0.4s, w11
2348 .long 0x4e21d821 // scvtf v1.4s, v1.4s
2349 .long 0x7284210a // movk w10, #0x2108
2350 .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
2351 .long 0x4e040d21 // dup v1.4s, w9
2352 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2353 .long 0x6e21dc61 // fmul v1.4s, v3.4s, v1.4s
2354 .long 0x4e040d43 // dup v3.4s, w10
2355 .long 0x4e21d842 // scvtf v2.4s, v2.4s
2356 .long 0x6e23dc42 // fmul v2.4s, v2.4s, v3.4s
2357 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
2358 .long 0xd61f0060 // br x3
2359
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2360HIDDEN _sk_store_565_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2361.globl _sk_store_565_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2362FUNCTION(_sk_store_565_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2363_sk_store_565_aarch64:
2364 .long 0xf9400028 // ldr x8, [x1]
2365 .long 0x52a84f8a // mov w10, #0x427c0000
2366 .long 0x4f01f7f0 // fmov v16.4s, #3.100000000000000000e+01
2367 .long 0x4e040d52 // dup v18.4s, w10
2368 .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s
2369 .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s
2370 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
2371 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2372 .long 0x6e30dc50 // fmul v16.4s, v2.4s, v16.4s
2373 .long 0x4f2b5631 // shl v17.4s, v17.4s, #11
2374 .long 0xf9400108 // ldr x8, [x8]
2375 .long 0x4f255652 // shl v18.4s, v18.4s, #5
2376 .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b
2377 .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
2378 .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b
2379 .long 0xd37ff809 // lsl x9, x0, #1
2380 .long 0x0e612a10 // xtn v16.4h, v16.4s
2381 .long 0xfc296910 // str d16, [x8, x9]
2382 .long 0xf9400423 // ldr x3, [x1, #8]
2383 .long 0x91004021 // add x1, x1, #0x10
2384 .long 0xd61f0060 // br x3
2385
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2386HIDDEN _sk_load_4444_aarch64
2387.globl _sk_load_4444_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2388FUNCTION(_sk_load_4444_aarch64)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2389_sk_load_4444_aarch64:
2390 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2391 .long 0xd37ff809 // lsl x9, x0, #1
2392 .long 0x4f072601 // movi v1.4s, #0xf0, lsl #8
2393 .long 0x4f0025e2 // movi v2.4s, #0xf, lsl #8
2394 .long 0xf9400108 // ldr x8, [x8]
2395 .long 0x4f070603 // movi v3.4s, #0xf0
2396 .long 0x4f0005f0 // movi v16.4s, #0xf
2397 .long 0xfc696900 // ldr d0, [x8, x9]
2398 .long 0x52a6f108 // mov w8, #0x37880000
2399 .long 0x72911128 // movk w8, #0x8889
2400 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2401 .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b
2402 .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b
2403 .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b
2404 .long 0x4e301c10 // and v16.16b, v0.16b, v16.16b
2405 .long 0x4e040d00 // dup v0.4s, w8
2406 .long 0x52a73108 // mov w8, #0x39880000
2407 .long 0x72911128 // movk w8, #0x8889
2408 .long 0x4e21d821 // scvtf v1.4s, v1.4s
2409 .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
2410 .long 0x4e040d01 // dup v1.4s, w8
2411 .long 0x52a77108 // mov w8, #0x3b880000
2412 .long 0x72911128 // movk w8, #0x8889
2413 .long 0x4e21d842 // scvtf v2.4s, v2.4s
2414 .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s
2415 .long 0x4e040d02 // dup v2.4s, w8
2416 .long 0x52a7b108 // mov w8, #0x3d880000
2417 .long 0x72911128 // movk w8, #0x8889
2418 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2419 .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s
2420 .long 0x4e040d03 // dup v3.4s, w8
2421 .long 0x4e21da10 // scvtf v16.4s, v16.4s
2422 .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
2423 .long 0xd61f0060 // br x3
2424
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2425HIDDEN _sk_gather_4444_aarch64
2426.globl _sk_gather_4444_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2427FUNCTION(_sk_gather_4444_aarch64)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]2428_sk_gather_4444_aarch64:
2429 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2430 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2431 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2432 .long 0x4f070603 // movi v3.4s, #0xf0
2433 .long 0x91004109 // add x9, x8, #0x10
2434 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2435 .long 0xf9400108 // ldr x8, [x8]
2436 .long 0x4f0005f0 // movi v16.4s, #0xf
2437 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2438 .long 0x1e26000c // fmov w12, s0
2439 .long 0x0e0c3c09 // mov w9, v0.s[1]
2440 .long 0x8b2c450c // add x12, x8, w12, uxtw #1
2441 .long 0x0e143c0a // mov w10, v0.s[2]
2442 .long 0x0e1c3c0b // mov w11, v0.s[3]
2443 .long 0x0d404180 // ld1 {v0.h}[0], [x12]
2444 .long 0x78695909 // ldrh w9, [x8, w9, uxtw #1]
2445 .long 0x786a590a // ldrh w10, [x8, w10, uxtw #1]
2446 .long 0x786b5908 // ldrh w8, [x8, w11, uxtw #1]
2447 .long 0x4f072601 // movi v1.4s, #0xf0, lsl #8
2448 .long 0x4e061d20 // mov v0.h[1], w9
2449 .long 0x4e0a1d40 // mov v0.h[2], w10
2450 .long 0x4e0e1d00 // mov v0.h[3], w8
2451 .long 0x52a6f10b // mov w11, #0x37880000
2452 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2453 .long 0x7291112b // movk w11, #0x8889
2454 .long 0x4f0025e2 // movi v2.4s, #0xf, lsl #8
2455 .long 0x52a73109 // mov w9, #0x39880000
2456 .long 0x4e211c01 // and v1.16b, v0.16b, v1.16b
2457 .long 0x72911129 // movk w9, #0x8889
2458 .long 0x52a7710a // mov w10, #0x3b880000
2459 .long 0x4e221c02 // and v2.16b, v0.16b, v2.16b
2460 .long 0x4e231c03 // and v3.16b, v0.16b, v3.16b
2461 .long 0x4e301c10 // and v16.16b, v0.16b, v16.16b
2462 .long 0x4e040d60 // dup v0.4s, w11
2463 .long 0x4e21d821 // scvtf v1.4s, v1.4s
2464 .long 0x7291112a // movk w10, #0x8889
2465 .long 0x52a7b108 // mov w8, #0x3d880000
2466 .long 0x6e20dc20 // fmul v0.4s, v1.4s, v0.4s
2467 .long 0x4e040d21 // dup v1.4s, w9
2468 .long 0x4e21d842 // scvtf v2.4s, v2.4s
2469 .long 0x72911128 // movk w8, #0x8889
2470 .long 0x6e21dc41 // fmul v1.4s, v2.4s, v1.4s
2471 .long 0x4e040d42 // dup v2.4s, w10
2472 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2473 .long 0x6e22dc62 // fmul v2.4s, v3.4s, v2.4s
2474 .long 0x4e040d03 // dup v3.4s, w8
2475 .long 0x4e21da10 // scvtf v16.4s, v16.4s
2476 .long 0x6e23de03 // fmul v3.4s, v16.4s, v3.4s
2477 .long 0xd61f0060 // br x3
2478
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2479HIDDEN _sk_store_4444_aarch64
2480.globl _sk_store_4444_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2481FUNCTION(_sk_store_4444_aarch64)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]2482_sk_store_4444_aarch64:
2483 .long 0x4f01f5d0 // fmov v16.4s, #1.500000000000000000e+01
2484 .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s
2485 .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s
2486 .long 0xf9400028 // ldr x8, [x1]
2487 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
2488 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2489 .long 0x4f2c5631 // shl v17.4s, v17.4s, #12
2490 .long 0x4f285652 // shl v18.4s, v18.4s, #8
2491 .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b
2492 .long 0x6e30dc52 // fmul v18.4s, v2.4s, v16.4s
2493 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2494 .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s
2495 .long 0x4f245652 // shl v18.4s, v18.4s, #4
2496 .long 0xf9400108 // ldr x8, [x8]
2497 .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b
2498 .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
2499 .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b
2500 .long 0xd37ff809 // lsl x9, x0, #1
2501 .long 0x0e612a10 // xtn v16.4h, v16.4s
2502 .long 0xfc296910 // str d16, [x8, x9]
2503 .long 0xf9400423 // ldr x3, [x1, #8]
2504 .long 0x91004021 // add x1, x1, #0x10
2505 .long 0xd61f0060 // br x3
2506
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2507HIDDEN _sk_load_8888_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2508.globl _sk_load_8888_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2509FUNCTION(_sk_load_8888_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2510_sk_load_8888_aarch64:
2511 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2512 .long 0xd37ef409 // lsl x9, x0, #2
2513 .long 0x6f00e621 // movi v1.2d, #0xff000000ff
2514 .long 0xf9400108 // ldr x8, [x8]
2515 .long 0x3ce96900 // ldr q0, [x8, x9]
2516 .long 0x52a77008 // mov w8, #0x3b800000
2517 .long 0x72901028 // movk w8, #0x8081
2518 .long 0x4e040d02 // dup v2.4s, w8
2519 .long 0x6f380410 // ushr v16.4s, v0.4s, #8
2520 .long 0x6f300411 // ushr v17.4s, v0.4s, #16
2521 .long 0x4e211c03 // and v3.16b, v0.16b, v1.16b
2522 .long 0x6f280400 // ushr v0.4s, v0.4s, #24
2523 .long 0x4e211e10 // and v16.16b, v16.16b, v1.16b
2524 .long 0x4e211e21 // and v1.16b, v17.16b, v1.16b
2525 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2526 .long 0x4e21d811 // scvtf v17.4s, v0.4s
2527 .long 0x4e21da10 // scvtf v16.4s, v16.4s
2528 .long 0x4e21d832 // scvtf v18.4s, v1.4s
2529 .long 0x6e22dc60 // fmul v0.4s, v3.4s, v2.4s
2530 .long 0x6e22de23 // fmul v3.4s, v17.4s, v2.4s
2531 .long 0x6e22de01 // fmul v1.4s, v16.4s, v2.4s
2532 .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s
2533 .long 0xd61f0060 // br x3
2534
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]2535HIDDEN _sk_gather_8888_aarch64
2536.globl _sk_gather_8888_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2537FUNCTION(_sk_gather_8888_aarch64)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]2538_sk_gather_8888_aarch64:
2539 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2540 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2541 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2542 .long 0x91004109 // add x9, x8, #0x10
2543 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2544 .long 0xf9400108 // ldr x8, [x8]
2545 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2546 .long 0x1e26000c // fmov w12, s0
2547 .long 0x8b2c490c // add x12, x8, w12, uxtw #2
2548 .long 0x0e0c3c09 // mov w9, v0.s[1]
2549 .long 0x0e143c0a // mov w10, v0.s[2]
2550 .long 0x0e1c3c0b // mov w11, v0.s[3]
2551 .long 0x0d408180 // ld1 {v0.s}[0], [x12]
2552 .long 0x8b294909 // add x9, x8, w9, uxtw #2
2553 .long 0xb86a590a // ldr w10, [x8, w10, uxtw #2]
2554 .long 0xb86b5908 // ldr w8, [x8, w11, uxtw #2]
2555 .long 0x0d409120 // ld1 {v0.s}[1], [x9]
2556 .long 0x6f00e621 // movi v1.2d, #0xff000000ff
2557 .long 0x52a77009 // mov w9, #0x3b800000
2558 .long 0x72901029 // movk w9, #0x8081
2559 .long 0x4e141d40 // mov v0.s[2], w10
2560 .long 0x4e1c1d00 // mov v0.s[3], w8
2561 .long 0x6f380410 // ushr v16.4s, v0.4s, #8
2562 .long 0x6f300411 // ushr v17.4s, v0.4s, #16
2563 .long 0x4e211c03 // and v3.16b, v0.16b, v1.16b
2564 .long 0x6f280400 // ushr v0.4s, v0.4s, #24
2565 .long 0x4e211e10 // and v16.16b, v16.16b, v1.16b
2566 .long 0x4e211e21 // and v1.16b, v17.16b, v1.16b
2567 .long 0x4e040d22 // dup v2.4s, w9
2568 .long 0x4e21d863 // scvtf v3.4s, v3.4s
2569 .long 0x4e21d811 // scvtf v17.4s, v0.4s
2570 .long 0x4e21da10 // scvtf v16.4s, v16.4s
2571 .long 0x4e21d832 // scvtf v18.4s, v1.4s
2572 .long 0x6e22dc60 // fmul v0.4s, v3.4s, v2.4s
2573 .long 0x6e22de23 // fmul v3.4s, v17.4s, v2.4s
2574 .long 0x6e22de01 // fmul v1.4s, v16.4s, v2.4s
2575 .long 0x6e22de42 // fmul v2.4s, v18.4s, v2.4s
2576 .long 0xd61f0060 // br x3
2577
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2578HIDDEN _sk_store_8888_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2579.globl _sk_store_8888_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2580FUNCTION(_sk_store_8888_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2581_sk_store_8888_aarch64:
2582 .long 0x52a86fea // mov w10, #0x437f0000
2583 .long 0x4e040d50 // dup v16.4s, w10
2584 .long 0xf9400028 // ldr x8, [x1]
2585 .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s
2586 .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s
2587 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2588 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
2589 .long 0x4f285652 // shl v18.4s, v18.4s, #8
2590 .long 0x4eb11e51 // orr v17.16b, v18.16b, v17.16b
2591 .long 0x6e30dc52 // fmul v18.4s, v2.4s, v16.4s
2592 .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s
2593 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2594 .long 0xf9400108 // ldr x8, [x8]
2595 .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
2596 .long 0x4f305652 // shl v18.4s, v18.4s, #16
2597 .long 0x4eb21e31 // orr v17.16b, v17.16b, v18.16b
2598 .long 0x4f385610 // shl v16.4s, v16.4s, #24
2599 .long 0xd37ef409 // lsl x9, x0, #2
2600 .long 0x4eb01e30 // orr v16.16b, v17.16b, v16.16b
2601 .long 0x3ca96910 // str q16, [x8, x9]
2602 .long 0xf9400423 // ldr x3, [x1, #8]
2603 .long 0x91004021 // add x1, x1, #0x10
2604 .long 0xd61f0060 // br x3
2605
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2606HIDDEN _sk_load_f16_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2607.globl _sk_load_f16_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2608FUNCTION(_sk_load_f16_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2609_sk_load_f16_aarch64:
2610 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2611 .long 0xf9400108 // ldr x8, [x8]
2612 .long 0x8b000d08 // add x8, x8, x0, lsl #3
2613 .long 0x0c400510 // ld4 {v16.4h-v19.4h}, [x8]
2614 .long 0x0e217a00 // fcvtl v0.4s, v16.4h
2615 .long 0x0e217a21 // fcvtl v1.4s, v17.4h
2616 .long 0x0e217a42 // fcvtl v2.4s, v18.4h
2617 .long 0x0e217a63 // fcvtl v3.4s, v19.4h
2618 .long 0xd61f0060 // br x3
2619
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]2620HIDDEN _sk_gather_f16_aarch64
2621.globl _sk_gather_f16_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2622FUNCTION(_sk_gather_f16_aarch64)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]2623_sk_gather_f16_aarch64:
2624 .long 0xa9bf7bfd // stp x29, x30, [sp, #-16]!
2625 .long 0xd100c3e9 // sub x9, sp, #0x30
2626 .long 0x910003fd // mov x29, sp
2627 .long 0x927be93f // and sp, x9, #0xffffffffffffffe0
2628 .long 0xf9400028 // ldr x8, [x1]
2629 .long 0x4ea1b821 // fcvtzs v1.4s, v1.4s
2630 .long 0x4ea1b800 // fcvtzs v0.4s, v0.4s
2631 .long 0x91004109 // add x9, x8, #0x10
2632 .long 0x4d40c922 // ld1r {v2.4s}, [x9]
2633 .long 0xf9400108 // ldr x8, [x8]
2634 .long 0x4ea19440 // mla v0.4s, v2.4s, v1.4s
2635 .long 0x0e143c0a // mov w10, v0.s[2]
2636 .long 0x1e26000c // fmov w12, s0
2637 .long 0x8b2c4d0c // add x12, x8, w12, uxtw #3
2638 .long 0x8b2a4d0a // add x10, x8, w10, uxtw #3
2639 .long 0x0e0c3c09 // mov w9, v0.s[1]
2640 .long 0x0e1c3c0b // mov w11, v0.s[3]
2641 .long 0x0d408540 // ld1 {v0.d}[0], [x10]
2642 .long 0x0d408581 // ld1 {v1.d}[0], [x12]
2643 .long 0x8b294d09 // add x9, x8, w9, uxtw #3
2644 .long 0x8b2b4d08 // add x8, x8, w11, uxtw #3
2645 .long 0x4d408500 // ld1 {v0.d}[1], [x8]
2646 .long 0x4d408521 // ld1 {v1.d}[1], [x9]
2647 .long 0x910003e8 // mov x8, sp
2648 .long 0xad0003e1 // stp q1, q0, [sp]
2649 .long 0x0c400510 // ld4 {v16.4h-v19.4h}, [x8]
2650 .long 0xf9400428 // ldr x8, [x1, #8]
2651 .long 0x91004021 // add x1, x1, #0x10
2652 .long 0x0e217a00 // fcvtl v0.4s, v16.4h
2653 .long 0x0e217a21 // fcvtl v1.4s, v17.4h
2654 .long 0x0e217a42 // fcvtl v2.4s, v18.4h
2655 .long 0x0e217a63 // fcvtl v3.4s, v19.4h
2656 .long 0xd63f0100 // blr x8
2657 .long 0x910003bf // mov sp, x29
2658 .long 0xa8c17bfd // ldp x29, x30, [sp], #16
2659 .long 0xd65f03c0 // ret
2660
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2661HIDDEN _sk_store_f16_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2662.globl _sk_store_f16_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2663FUNCTION(_sk_store_f16_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2664_sk_store_f16_aarch64:
2665 .long 0xf9400028 // ldr x8, [x1]
2666 .long 0x0e216810 // fcvtn v16.4h, v0.4s
2667 .long 0x0e216831 // fcvtn v17.4h, v1.4s
2668 .long 0x0e216852 // fcvtn v18.4h, v2.4s
2669 .long 0xf9400108 // ldr x8, [x8]
2670 .long 0x0e216873 // fcvtn v19.4h, v3.4s
2671 .long 0x8b000d08 // add x8, x8, x0, lsl #3
2672 .long 0x0c000510 // st4 {v16.4h-v19.4h}, [x8]
2673 .long 0xf9400423 // ldr x3, [x1, #8]
2674 .long 0x91004021 // add x1, x1, #0x10
2675 .long 0xd61f0060 // br x3
2676
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]2677HIDDEN _sk_load_u16_be_aarch64
2678.globl _sk_load_u16_be_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2679FUNCTION(_sk_load_u16_be_aarch64)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]2680_sk_load_u16_be_aarch64:
2681 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2682 .long 0xf9400108 // ldr x8, [x8]
2683 .long 0x8b000d08 // add x8, x8, x0, lsl #3
2684 .long 0x0c400500 // ld4 {v0.4h-v3.4h}, [x8]
2685 .long 0x52a6f008 // mov w8, #0x37800000
2686 .long 0x72801008 // movk w8, #0x80
2687 .long 0x0f185410 // shl v16.4h, v0.4h, #8
2688 .long 0x2f180411 // ushr v17.4h, v0.4h, #8
2689 .long 0x0f185432 // shl v18.4h, v1.4h, #8
2690 .long 0x2f180433 // ushr v19.4h, v1.4h, #8
2691 .long 0x0f185454 // shl v20.4h, v2.4h, #8
2692 .long 0x2f180455 // ushr v21.4h, v2.4h, #8
2693 .long 0x0f185476 // shl v22.4h, v3.4h, #8
2694 .long 0x2f180460 // ushr v0.4h, v3.4h, #8
2695 .long 0x0eb11e01 // orr v1.8b, v16.8b, v17.8b
2696 .long 0x0eb31e42 // orr v2.8b, v18.8b, v19.8b
2697 .long 0x0eb51e90 // orr v16.8b, v20.8b, v21.8b
2698 .long 0x0ea01ec0 // orr v0.8b, v22.8b, v0.8b
2699 .long 0x2f10a421 // uxtl v1.4s, v1.4h
2700 .long 0x2f10a442 // uxtl v2.4s, v2.4h
2701 .long 0x2f10a610 // uxtl v16.4s, v16.4h
2702 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2703 .long 0x4e040d03 // dup v3.4s, w8
2704 .long 0x6e21d821 // ucvtf v1.4s, v1.4s
2705 .long 0x6e21d842 // ucvtf v2.4s, v2.4s
2706 .long 0x6e21da10 // ucvtf v16.4s, v16.4s
2707 .long 0x6e21d811 // ucvtf v17.4s, v0.4s
2708 .long 0x6e23dc20 // fmul v0.4s, v1.4s, v3.4s
2709 .long 0x6e23dc41 // fmul v1.4s, v2.4s, v3.4s
2710 .long 0x6e23de02 // fmul v2.4s, v16.4s, v3.4s
2711 .long 0x6e23de23 // fmul v3.4s, v17.4s, v3.4s
2712 .long 0xd61f0060 // br x3
2713
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]2714HIDDEN _sk_load_rgb_u16_be_aarch64
2715.globl _sk_load_rgb_u16_be_aarch64
2716FUNCTION(_sk_load_rgb_u16_be_aarch64)
2717_sk_load_rgb_u16_be_aarch64:
2718 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2719 .long 0x321f07e9 // orr w9, wzr, #0x6
2720 .long 0xf9400108 // ldr x8, [x8]
2721 .long 0x9b092008 // madd x8, x0, x9, x8
2722 .long 0x0c404500 // ld3 {v0.4h-v2.4h}, [x8]
2723 .long 0x52a6f008 // mov w8, #0x37800000
2724 .long 0x72801008 // movk w8, #0x80
2725 .long 0x0f185403 // shl v3.4h, v0.4h, #8
2726 .long 0x2f180410 // ushr v16.4h, v0.4h, #8
2727 .long 0x0f185431 // shl v17.4h, v1.4h, #8
2728 .long 0x2f180432 // ushr v18.4h, v1.4h, #8
2729 .long 0x0f185453 // shl v19.4h, v2.4h, #8
2730 .long 0x2f180440 // ushr v0.4h, v2.4h, #8
2731 .long 0x0eb01c61 // orr v1.8b, v3.8b, v16.8b
2732 .long 0x0eb21e23 // orr v3.8b, v17.8b, v18.8b
2733 .long 0x0ea01e60 // orr v0.8b, v19.8b, v0.8b
2734 .long 0x2f10a421 // uxtl v1.4s, v1.4h
2735 .long 0x2f10a463 // uxtl v3.4s, v3.4h
2736 .long 0x2f10a400 // uxtl v0.4s, v0.4h
2737 .long 0x4e040d02 // dup v2.4s, w8
2738 .long 0x6e21d821 // ucvtf v1.4s, v1.4s
2739 .long 0x6e21d863 // ucvtf v3.4s, v3.4s
2740 .long 0x6e21d810 // ucvtf v16.4s, v0.4s
2741 .long 0x6e22dc20 // fmul v0.4s, v1.4s, v2.4s
2742 .long 0x6e22dc61 // fmul v1.4s, v3.4s, v2.4s
2743 .long 0x6e22de02 // fmul v2.4s, v16.4s, v2.4s
2744 .long 0x4f03f603 // fmov v3.4s, #1.000000000000000000e+00
2745 .long 0xd61f0060 // br x3
2746
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]2747HIDDEN _sk_store_u16_be_aarch64
2748.globl _sk_store_u16_be_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2749FUNCTION(_sk_store_u16_be_aarch64)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]2750_sk_store_u16_be_aarch64:
2751 .long 0x52a8efe9 // mov w9, #0x477f0000
2752 .long 0x729fe009 // movk w9, #0xff00
2753 .long 0x4e040d30 // dup v16.4s, w9
2754 .long 0x6e30dc11 // fmul v17.4s, v0.4s, v16.4s
2755 .long 0xf9400028 // ldr x8, [x1]
2756 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
2757 .long 0x0e612a31 // xtn v17.4h, v17.4s
2758 .long 0x6e30dc32 // fmul v18.4s, v1.4s, v16.4s
2759 .long 0x0f185633 // shl v19.4h, v17.4h, #8
2760 .long 0x2f180631 // ushr v17.4h, v17.4h, #8
2761 .long 0x6e21aa52 // fcvtnu v18.4s, v18.4s
2762 .long 0x0eb11e75 // orr v21.8b, v19.8b, v17.8b
2763 .long 0x6e30dc51 // fmul v17.4s, v2.4s, v16.4s
2764 .long 0x0e612a52 // xtn v18.4h, v18.4s
2765 .long 0x6e30dc70 // fmul v16.4s, v3.4s, v16.4s
2766 .long 0x6e21aa31 // fcvtnu v17.4s, v17.4s
2767 .long 0xf9400108 // ldr x8, [x8]
2768 .long 0x0f185654 // shl v20.4h, v18.4h, #8
2769 .long 0x2f180652 // ushr v18.4h, v18.4h, #8
2770 .long 0x6e21aa10 // fcvtnu v16.4s, v16.4s
2771 .long 0x0e612a31 // xtn v17.4h, v17.4s
2772 .long 0x0eb21e96 // orr v22.8b, v20.8b, v18.8b
2773 .long 0x0e612a10 // xtn v16.4h, v16.4s
2774 .long 0x0f185632 // shl v18.4h, v17.4h, #8
2775 .long 0x2f180631 // ushr v17.4h, v17.4h, #8
2776 .long 0x0eb11e57 // orr v23.8b, v18.8b, v17.8b
2777 .long 0x0f185611 // shl v17.4h, v16.4h, #8
2778 .long 0x2f180610 // ushr v16.4h, v16.4h, #8
2779 .long 0x8b000d08 // add x8, x8, x0, lsl #3
2780 .long 0x0eb01e38 // orr v24.8b, v17.8b, v16.8b
2781 .long 0x0c000515 // st4 {v21.4h-v24.4h}, [x8]
2782 .long 0xf9400423 // ldr x3, [x1, #8]
2783 .long 0x91004021 // add x1, x1, #0x10
2784 .long 0xd61f0060 // br x3
2785
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]2786HIDDEN _sk_load_f32_aarch64
2787.globl _sk_load_f32_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2788FUNCTION(_sk_load_f32_aarch64)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]2789_sk_load_f32_aarch64:
2790 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2791 .long 0xf9400108 // ldr x8, [x8]
2792 .long 0x8b001108 // add x8, x8, x0, lsl #4
2793 .long 0x4c400900 // ld4 {v0.4s-v3.4s}, [x8]
2794 .long 0xd61f0060 // br x3
2795
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2796HIDDEN _sk_store_f32_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2797.globl _sk_store_f32_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2798FUNCTION(_sk_store_f32_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2799_sk_store_f32_aarch64:
2800 .long 0xf9400028 // ldr x8, [x1]
2801 .long 0xf9400108 // ldr x8, [x8]
2802 .long 0x8b001108 // add x8, x8, x0, lsl #4
2803 .long 0x4c000900 // st4 {v0.4s-v3.4s}, [x8]
2804 .long 0xf9400423 // ldr x3, [x1, #8]
2805 .long 0x91004021 // add x1, x1, #0x10
2806 .long 0xd61f0060 // br x3
2807
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2808HIDDEN _sk_clamp_x_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2809.globl _sk_clamp_x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2810FUNCTION(_sk_clamp_x_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2811_sk_clamp_x_aarch64:
2812 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2813 .long 0x6f00e411 // movi v17.2d, #0x0
2814 .long 0x4e20f620 // fmax v0.4s, v17.4s, v0.4s
2815 .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff
2816 .long 0x4d40c910 // ld1r {v16.4s}, [x8]
2817 .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s
2818 .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s
2819 .long 0xd61f0060 // br x3
2820
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2821HIDDEN _sk_clamp_y_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2822.globl _sk_clamp_y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2823FUNCTION(_sk_clamp_y_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2824_sk_clamp_y_aarch64:
2825 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2826 .long 0x6f00e411 // movi v17.2d, #0x0
2827 .long 0x4e21f621 // fmax v1.4s, v17.4s, v1.4s
2828 .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff
2829 .long 0x4d40c910 // ld1r {v16.4s}, [x8]
2830 .long 0x4eb18610 // add v16.4s, v16.4s, v17.4s
2831 .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s
2832 .long 0xd61f0060 // br x3
2833
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2834HIDDEN _sk_repeat_x_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2835.globl _sk_repeat_x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2836FUNCTION(_sk_repeat_x_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2837_sk_repeat_x_aarch64:
2838 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2839 .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff
2840 .long 0xbd400110 // ldr s16, [x8]
2841 .long 0x4e040612 // dup v18.4s, v16.s[0]
2842 .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s
2843 .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s
2844 .long 0x4e219a52 // frintm v18.4s, v18.4s
2845 .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0]
2846 .long 0x4eb1f400 // fmin v0.4s, v0.4s, v17.4s
2847 .long 0xd61f0060 // br x3
2848
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2849HIDDEN _sk_repeat_y_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2850.globl _sk_repeat_y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2851FUNCTION(_sk_repeat_y_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2852_sk_repeat_y_aarch64:
2853 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2854 .long 0x6f07e7f1 // movi v17.2d, #0xffffffffffffffff
2855 .long 0xbd400110 // ldr s16, [x8]
2856 .long 0x4e040612 // dup v18.4s, v16.s[0]
2857 .long 0x4eb18651 // add v17.4s, v18.4s, v17.4s
2858 .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s
2859 .long 0x4e219a52 // frintm v18.4s, v18.4s
2860 .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0]
2861 .long 0x4eb1f421 // fmin v1.4s, v1.4s, v17.4s
2862 .long 0xd61f0060 // br x3
2863
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2864HIDDEN _sk_mirror_x_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2865.globl _sk_mirror_x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2866FUNCTION(_sk_mirror_x_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2867_sk_mirror_x_aarch64:
2868 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2869 .long 0xbd400110 // ldr s16, [x8]
2870 .long 0x4e040611 // dup v17.4s, v16.s[0]
2871 .long 0x1e302a10 // fadd s16, s16, s16
2872 .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s
2873 .long 0x4e040612 // dup v18.4s, v16.s[0]
2874 .long 0x6e32fc12 // fdiv v18.4s, v0.4s, v18.4s
2875 .long 0x4e219a52 // frintm v18.4s, v18.4s
2876 .long 0x4f905240 // fmls v0.4s, v18.4s, v16.s[0]
2877 .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff
2878 .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s
2879 .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s
2880 .long 0x4ea0f800 // fabs v0.4s, v0.4s
2881 .long 0x4eb0f400 // fmin v0.4s, v0.4s, v16.4s
2882 .long 0xd61f0060 // br x3
2883
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2884HIDDEN _sk_mirror_y_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2885.globl _sk_mirror_y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2886FUNCTION(_sk_mirror_y_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2887_sk_mirror_y_aarch64:
2888 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2889 .long 0xbd400110 // ldr s16, [x8]
2890 .long 0x4e040611 // dup v17.4s, v16.s[0]
2891 .long 0x1e302a10 // fadd s16, s16, s16
2892 .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s
2893 .long 0x4e040612 // dup v18.4s, v16.s[0]
2894 .long 0x6e32fc32 // fdiv v18.4s, v1.4s, v18.4s
2895 .long 0x4e219a52 // frintm v18.4s, v18.4s
2896 .long 0x4f905241 // fmls v1.4s, v18.4s, v16.s[0]
2897 .long 0x6f07e7f0 // movi v16.2d, #0xffffffffffffffff
2898 .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s
2899 .long 0x4eb08630 // add v16.4s, v17.4s, v16.4s
2900 .long 0x4ea0f821 // fabs v1.4s, v1.4s
2901 .long 0x4eb0f421 // fmin v1.4s, v1.4s, v16.4s
2902 .long 0xd61f0060 // br x3
2903
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2904HIDDEN _sk_luminance_to_alpha_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2905.globl _sk_luminance_to_alpha_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2906FUNCTION(_sk_luminance_to_alpha_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2907_sk_luminance_to_alpha_aarch64:
2908 .long 0x52a7cb28 // mov w8, #0x3e590000
2909 .long 0x72967a08 // movk w8, #0xb3d0
2910 .long 0x4e040d11 // dup v17.4s, w8
2911 .long 0x52a7e6e8 // mov w8, #0x3f370000
2912 .long 0x7282eb28 // movk w8, #0x1759
2913 .long 0x4ea01c10 // mov v16.16b, v0.16b
2914 .long 0x4e040d00 // dup v0.4s, w8
2915 .long 0x52a7b268 // mov w8, #0x3d930000
2916 .long 0xf8408423 // ldr x3, [x1], #8
2917 .long 0x729bb308 // movk w8, #0xdd98
2918 .long 0x6e20dc23 // fmul v3.4s, v1.4s, v0.4s
2919 .long 0x4e30ce23 // fmla v3.4s, v17.4s, v16.4s
2920 .long 0x4e040d10 // dup v16.4s, w8
2921 .long 0x6f00e400 // movi v0.2d, #0x0
2922 .long 0x6f00e401 // movi v1.2d, #0x0
2923 .long 0x4e22ce03 // fmla v3.4s, v16.4s, v2.4s
2924 .long 0x6f00e402 // movi v2.2d, #0x0
2925 .long 0xd61f0060 // br x3
2926
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2927HIDDEN _sk_matrix_2x3_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2928.globl _sk_matrix_2x3_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2929FUNCTION(_sk_matrix_2x3_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2930_sk_matrix_2x3_aarch64:
2931 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2932 .long 0xaa0803e9 // mov x9, x8
2933 .long 0x9100410a // add x10, x8, #0x10
2934 .long 0x4ddfc932 // ld1r {v18.4s}, [x9], #4
2935 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
2936 .long 0x2d415113 // ldp s19, s20, [x8, #8]
2937 .long 0x9100510a // add x10, x8, #0x14
2938 .long 0x4d40c951 // ld1r {v17.4s}, [x10]
2939 .long 0x4f931030 // fmla v16.4s, v1.4s, v19.s[0]
2940 .long 0xbd400133 // ldr s19, [x9]
2941 .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0]
2942 .long 0x4e20ce50 // fmla v16.4s, v18.4s, v0.4s
2943 .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0]
2944 .long 0x4eb01e00 // mov v0.16b, v16.16b
2945 .long 0x4eb11e21 // mov v1.16b, v17.16b
2946 .long 0xd61f0060 // br x3
2947
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2948HIDDEN _sk_matrix_3x4_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2949.globl _sk_matrix_3x4_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2950FUNCTION(_sk_matrix_3x4_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2951_sk_matrix_3x4_aarch64:
2952 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
2953 .long 0xaa0803e9 // mov x9, x8
2954 .long 0x9100910a // add x10, x8, #0x24
2955 .long 0x4ddfc933 // ld1r {v19.4s}, [x9], #4
2956 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
2957 .long 0x9100a10a // add x10, x8, #0x28
2958 .long 0x4d40c951 // ld1r {v17.4s}, [x10]
2959 .long 0x9100b10a // add x10, x8, #0x2c
2960 .long 0x2d435514 // ldp s20, s21, [x8, #24]
2961 .long 0xbd402116 // ldr s22, [x8, #32]
2962 .long 0x4d40c952 // ld1r {v18.4s}, [x10]
2963 .long 0x4f941050 // fmla v16.4s, v2.4s, v20.s[0]
2964 .long 0x4f951051 // fmla v17.4s, v2.4s, v21.s[0]
2965 .long 0x4f961052 // fmla v18.4s, v2.4s, v22.s[0]
2966 .long 0x2d425502 // ldp s2, s21, [x8, #16]
2967 .long 0x2d415d14 // ldp s20, s23, [x8, #8]
2968 .long 0x4f821031 // fmla v17.4s, v1.4s, v2.s[0]
2969 .long 0xbd400122 // ldr s2, [x9]
2970 .long 0x4f971030 // fmla v16.4s, v1.4s, v23.s[0]
2971 .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0]
2972 .long 0x4e20ce70 // fmla v16.4s, v19.4s, v0.4s
2973 .long 0x4f941012 // fmla v18.4s, v0.4s, v20.s[0]
2974 .long 0x4f821011 // fmla v17.4s, v0.4s, v2.s[0]
2975 .long 0x4eb01e00 // mov v0.16b, v16.16b
2976 .long 0x4eb11e21 // mov v1.16b, v17.16b
2977 .long 0x4eb21e42 // mov v2.16b, v18.16b
2978 .long 0xd61f0060 // br x3
2979
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]2980HIDDEN _sk_matrix_4x5_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2981.globl _sk_matrix_4x5_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]2982FUNCTION(_sk_matrix_4x5_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]2983_sk_matrix_4x5_aarch64:
2984 .long 0xf9400029 // ldr x9, [x1]
2985 .long 0xaa0903e8 // mov x8, x9
2986 .long 0x9101012a // add x10, x9, #0x40
2987 .long 0x4ddfc914 // ld1r {v20.4s}, [x8], #4
2988 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
2989 .long 0x9101112a // add x10, x9, #0x44
2990 .long 0x4d40c951 // ld1r {v17.4s}, [x10]
2991 .long 0x9101212a // add x10, x9, #0x48
2992 .long 0x4d40c952 // ld1r {v18.4s}, [x10]
2993 .long 0x2d465533 // ldp s19, s21, [x9, #48]
2994 .long 0x2d475d36 // ldp s22, s23, [x9, #56]
2995 .long 0x9101312a // add x10, x9, #0x4c
2996 .long 0xf9400423 // ldr x3, [x1, #8]
2997 .long 0x4f931070 // fmla v16.4s, v3.4s, v19.s[0]
2998 .long 0x4d40c953 // ld1r {v19.4s}, [x10]
2999 .long 0x4f951071 // fmla v17.4s, v3.4s, v21.s[0]
3000 .long 0x4f961072 // fmla v18.4s, v3.4s, v22.s[0]
3001 .long 0x2d445935 // ldp s21, s22, [x9, #32]
3002 .long 0x4f971073 // fmla v19.4s, v3.4s, v23.s[0]
3003 .long 0x2d455d23 // ldp s3, s23, [x9, #40]
3004 .long 0x91004021 // add x1, x1, #0x10
3005 .long 0x4f951050 // fmla v16.4s, v2.4s, v21.s[0]
3006 .long 0x4f961051 // fmla v17.4s, v2.4s, v22.s[0]
3007 .long 0x2d425935 // ldp s21, s22, [x9, #16]
3008 .long 0x4f971053 // fmla v19.4s, v2.4s, v23.s[0]
3009 .long 0x4f831052 // fmla v18.4s, v2.4s, v3.s[0]
3010 .long 0x2d410d22 // ldp s2, s3, [x9, #8]
3011 .long 0x4f951030 // fmla v16.4s, v1.4s, v21.s[0]
3012 .long 0x2d435d35 // ldp s21, s23, [x9, #24]
3013 .long 0x4f961031 // fmla v17.4s, v1.4s, v22.s[0]
3014 .long 0xbd400116 // ldr s22, [x8]
3015 .long 0x4e20ce90 // fmla v16.4s, v20.4s, v0.4s
3016 .long 0x4f951032 // fmla v18.4s, v1.4s, v21.s[0]
3017 .long 0x4f971033 // fmla v19.4s, v1.4s, v23.s[0]
3018 .long 0x4f821012 // fmla v18.4s, v0.4s, v2.s[0]
3019 .long 0x4f831013 // fmla v19.4s, v0.4s, v3.s[0]
3020 .long 0x4f961011 // fmla v17.4s, v0.4s, v22.s[0]
3021 .long 0x4eb01e00 // mov v0.16b, v16.16b
3022 .long 0x4eb11e21 // mov v1.16b, v17.16b
3023 .long 0x4eb21e42 // mov v2.16b, v18.16b
3024 .long 0x4eb31e63 // mov v3.16b, v19.16b
3025 .long 0xd61f0060 // br x3
3026
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3027HIDDEN _sk_matrix_perspective_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3028.globl _sk_matrix_perspective_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3029FUNCTION(_sk_matrix_perspective_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3030_sk_matrix_perspective_aarch64:
3031 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
3032 .long 0xaa0803e9 // mov x9, x8
3033 .long 0x9100510a // add x10, x8, #0x14
3034 .long 0x4ddfc930 // ld1r {v16.4s}, [x9], #4
3035 .long 0x4d40c951 // ld1r {v17.4s}, [x10]
3036 .long 0x9100810a // add x10, x8, #0x20
3037 .long 0x4d40c952 // ld1r {v18.4s}, [x10]
3038 .long 0x2d41d113 // ldp s19, s20, [x8, #12]
3039 .long 0x2d435915 // ldp s21, s22, [x8, #24]
3040 .long 0x91002108 // add x8, x8, #0x8
3041 .long 0x4f941031 // fmla v17.4s, v1.4s, v20.s[0]
3042 .long 0x4d40c914 // ld1r {v20.4s}, [x8]
3043 .long 0x4f961032 // fmla v18.4s, v1.4s, v22.s[0]
3044 .long 0xbd400136 // ldr s22, [x9]
3045 .long 0x4f951012 // fmla v18.4s, v0.4s, v21.s[0]
3046 .long 0x4f931011 // fmla v17.4s, v0.4s, v19.s[0]
3047 .long 0x4f961034 // fmla v20.4s, v1.4s, v22.s[0]
3048 .long 0x4ea1da41 // frecpe v1.4s, v18.4s
3049 .long 0x4e21fe52 // frecps v18.4s, v18.4s, v1.4s
3050 .long 0x6e32dc32 // fmul v18.4s, v1.4s, v18.4s
3051 .long 0x4e20ce14 // fmla v20.4s, v16.4s, v0.4s
3052 .long 0x6e32de21 // fmul v1.4s, v17.4s, v18.4s
3053 .long 0x6e32de80 // fmul v0.4s, v20.4s, v18.4s
3054 .long 0xd61f0060 // br x3
3055
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3056HIDDEN _sk_linear_gradient_aarch64
3057.globl _sk_linear_gradient_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3058FUNCTION(_sk_linear_gradient_aarch64)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3059_sk_linear_gradient_aarch64:
3060 .long 0xf9400029 // ldr x9, [x1]
3061 .long 0x91004128 // add x8, x9, #0x10
3062 .long 0x9100512a // add x10, x9, #0x14
3063 .long 0x4d40c910 // ld1r {v16.4s}, [x8]
3064 .long 0x91006128 // add x8, x9, #0x18
3065 .long 0x4d40c941 // ld1r {v1.4s}, [x10]
3066 .long 0x9100712a // add x10, x9, #0x1c
3067 .long 0x4d40c902 // ld1r {v2.4s}, [x8]
3068 .long 0xf9400128 // ldr x8, [x9]
3069 .long 0x4d40c943 // ld1r {v3.4s}, [x10]
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]3070 .long 0xb40006c8 // cbz x8, 28fc <sk_linear_gradient_aarch64+0x100>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3071 .long 0x6dbf23e9 // stp d9, d8, [sp, #-16]!
3072 .long 0xf9400529 // ldr x9, [x9, #8]
3073 .long 0x6f00e413 // movi v19.2d, #0x0
3074 .long 0x6f00e411 // movi v17.2d, #0x0
3075 .long 0x6f00e412 // movi v18.2d, #0x0
3076 .long 0x91004129 // add x9, x9, #0x10
3077 .long 0x6f00e414 // movi v20.2d, #0x0
3078 .long 0xd100412a // sub x10, x9, #0x10
3079 .long 0x4d40c955 // ld1r {v21.4s}, [x10]
3080 .long 0xd100312b // sub x11, x9, #0xc
3081 .long 0xd100212a // sub x10, x9, #0x8
3082 .long 0x4d40c976 // ld1r {v22.4s}, [x11]
3083 .long 0xd100112b // sub x11, x9, #0x4
3084 .long 0x4d40c957 // ld1r {v23.4s}, [x10]
3085 .long 0xaa0903ea // mov x10, x9
3086 .long 0x4d40c978 // ld1r {v24.4s}, [x11]
3087 .long 0x4ddfc959 // ld1r {v25.4s}, [x10], #4
3088 .long 0x9100412b // add x11, x9, #0x10
3089 .long 0x4ea31c7b // mov v27.16b, v3.16b
3090 .long 0x6ea0e6a3 // fcmgt v3.4s, v21.4s, v0.4s
3091 .long 0x4d40c97a // ld1r {v26.4s}, [x11]
3092 .long 0x4eb41e95 // mov v21.16b, v20.16b
3093 .long 0x4ea31c74 // mov v20.16b, v3.16b
3094 .long 0x9100212b // add x11, x9, #0x8
3095 .long 0x4eb31e69 // mov v9.16b, v19.16b
3096 .long 0x4ea31c73 // mov v19.16b, v3.16b
3097 .long 0x6e771eb4 // bsl v20.16b, v21.16b, v23.16b
3098 .long 0x4d40c975 // ld1r {v21.4s}, [x11]
3099 .long 0x9100312b // add x11, x9, #0xc
3100 .long 0x6e761d33 // bsl v19.16b, v9.16b, v22.16b
3101 .long 0x4d40c976 // ld1r {v22.4s}, [x11]
3102 .long 0x4d40c957 // ld1r {v23.4s}, [x10]
3103 .long 0x4eb21e5c // mov v28.16b, v18.16b
3104 .long 0x4eb11e3d // mov v29.16b, v17.16b
3105 .long 0x4eb01e1e // mov v30.16b, v16.16b
3106 .long 0x4ea11c3f // mov v31.16b, v1.16b
3107 .long 0x4ea21c48 // mov v8.16b, v2.16b
3108 .long 0x4ea31c72 // mov v18.16b, v3.16b
3109 .long 0x4ea31c71 // mov v17.16b, v3.16b
3110 .long 0x4ea31c70 // mov v16.16b, v3.16b
3111 .long 0x4ea31c61 // mov v1.16b, v3.16b
3112 .long 0x4ea31c62 // mov v2.16b, v3.16b
3113 .long 0x6e7a1f63 // bsl v3.16b, v27.16b, v26.16b
3114 .long 0x6e781f92 // bsl v18.16b, v28.16b, v24.16b
3115 .long 0x6e791fb1 // bsl v17.16b, v29.16b, v25.16b
3116 .long 0x6e751fe1 // bsl v1.16b, v31.16b, v21.16b
3117 .long 0x6e761d02 // bsl v2.16b, v8.16b, v22.16b
3118 .long 0xd1000508 // sub x8, x8, #0x1
3119 .long 0x6e771fd0 // bsl v16.16b, v30.16b, v23.16b
3120 .long 0x91009129 // add x9, x9, #0x24
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]3121 .long 0xb5fffaa8 // cbnz x8, 2844 <sk_linear_gradient_aarch64+0x48>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3122 .long 0x6cc123e9 // ldp d9, d8, [sp], #16
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]3123 .long 0x14000005 // b 290c <sk_linear_gradient_aarch64+0x110>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3124 .long 0x6f00e414 // movi v20.2d, #0x0
3125 .long 0x6f00e412 // movi v18.2d, #0x0
3126 .long 0x6f00e411 // movi v17.2d, #0x0
3127 .long 0x6f00e413 // movi v19.2d, #0x0
3128 .long 0xf9400423 // ldr x3, [x1, #8]
3129 .long 0x4e20ce70 // fmla v16.4s, v19.4s, v0.4s
3130 .long 0x4e20ce81 // fmla v1.4s, v20.4s, v0.4s
3131 .long 0x4e20ce42 // fmla v2.4s, v18.4s, v0.4s
3132 .long 0x4e20ce23 // fmla v3.4s, v17.4s, v0.4s
3133 .long 0x91004021 // add x1, x1, #0x10
3134 .long 0x4eb01e00 // mov v0.16b, v16.16b
3135 .long 0xd61f0060 // br x3
3136
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3137HIDDEN _sk_linear_gradient_2stops_aarch64
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3138.globl _sk_linear_gradient_2stops_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3139FUNCTION(_sk_linear_gradient_2stops_aarch64)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3140_sk_linear_gradient_2stops_aarch64:
3141 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]3142 .long 0xaa0803e9 // mov x9, x8
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3143 .long 0x9100410a // add x10, x8, #0x10
3144 .long 0x4ddfc931 // ld1r {v17.4s}, [x9], #4
3145 .long 0x4d40c950 // ld1r {v16.4s}, [x10]
3146 .long 0x9100510a // add x10, x8, #0x14
3147 .long 0x4d40c941 // ld1r {v1.4s}, [x10]
3148 .long 0x9100610a // add x10, x8, #0x18
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]3149 .long 0x4d40c942 // ld1r {v2.4s}, [x10]
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]3150 .long 0x9100710a // add x10, x8, #0x1c
3151 .long 0x2d414d12 // ldp s18, s19, [x8, #8]
3152 .long 0x4d40c943 // ld1r {v3.4s}, [x10]
3153 .long 0x4e20ce30 // fmla v16.4s, v17.4s, v0.4s
3154 .long 0xbd400131 // ldr s17, [x9]
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]3155 .long 0x4f921002 // fmla v2.4s, v0.4s, v18.s[0]
3156 .long 0x4f931003 // fmla v3.4s, v0.4s, v19.s[0]
3157 .long 0x4f911001 // fmla v1.4s, v0.4s, v17.s[0]
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3158 .long 0x4eb01e00 // mov v0.16b, v16.16b
3159 .long 0xd61f0060 // br x3
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3160
3161HIDDEN _sk_save_xy_aarch64
3162.globl _sk_save_xy_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3163FUNCTION(_sk_save_xy_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3164_sk_save_xy_aarch64:
3165 .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24
3166 .long 0xf9400028 // ldr x8, [x1]
3167 .long 0x4e30d411 // fadd v17.4s, v0.4s, v16.4s
3168 .long 0x4e30d430 // fadd v16.4s, v1.4s, v16.4s
3169 .long 0x4e219a32 // frintm v18.4s, v17.4s
3170 .long 0x4eb2d631 // fsub v17.4s, v17.4s, v18.4s
3171 .long 0x4e219a12 // frintm v18.4s, v16.4s
3172 .long 0x4eb2d610 // fsub v16.4s, v16.4s, v18.4s
3173 .long 0x3d800100 // str q0, [x8]
3174 .long 0x3d800901 // str q1, [x8, #32]
3175 .long 0x3d801111 // str q17, [x8, #64]
3176 .long 0x3d801910 // str q16, [x8, #96]
3177 .long 0xf9400423 // ldr x3, [x1, #8]
3178 .long 0x91004021 // add x1, x1, #0x10
3179 .long 0xd61f0060 // br x3
3180
3181HIDDEN _sk_accumulate_aarch64
3182.globl _sk_accumulate_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3183FUNCTION(_sk_accumulate_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3184_sk_accumulate_aarch64:
3185 .long 0xa8c10c28 // ldp x8, x3, [x1], #16
3186 .long 0x3dc02110 // ldr q16, [x8, #128]
3187 .long 0x3dc02911 // ldr q17, [x8, #160]
3188 .long 0x6e31de10 // fmul v16.4s, v16.4s, v17.4s
3189 .long 0x4e30cc04 // fmla v4.4s, v0.4s, v16.4s
3190 .long 0x4e30cc25 // fmla v5.4s, v1.4s, v16.4s
3191 .long 0x4e30cc46 // fmla v6.4s, v2.4s, v16.4s
3192 .long 0x4e30cc67 // fmla v7.4s, v3.4s, v16.4s
3193 .long 0xd61f0060 // br x3
3194
3195HIDDEN _sk_bilinear_nx_aarch64
3196.globl _sk_bilinear_nx_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3197FUNCTION(_sk_bilinear_nx_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3198_sk_bilinear_nx_aarch64:
3199 .long 0xf9400028 // ldr x8, [x1]
3200 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
3201 .long 0x3dc01100 // ldr q0, [x8, #64]
3202 .long 0x3dc00110 // ldr q16, [x8]
3203 .long 0x4ea0d620 // fsub v0.4s, v17.4s, v0.4s
3204 .long 0x3d802100 // str q0, [x8, #128]
3205 .long 0xf9400423 // ldr x3, [x1, #8]
3206 .long 0x4f0567e0 // movi v0.4s, #0xbf, lsl #24
3207 .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s
3208 .long 0x91004021 // add x1, x1, #0x10
3209 .long 0xd61f0060 // br x3
3210
3211HIDDEN _sk_bilinear_px_aarch64
3212.globl _sk_bilinear_px_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3213FUNCTION(_sk_bilinear_px_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3214_sk_bilinear_px_aarch64:
3215 .long 0xf9400028 // ldr x8, [x1]
3216 .long 0x3dc01100 // ldr q0, [x8, #64]
3217 .long 0x3dc00110 // ldr q16, [x8]
3218 .long 0x3d802100 // str q0, [x8, #128]
3219 .long 0xf9400423 // ldr x3, [x1, #8]
3220 .long 0x4f0167e0 // movi v0.4s, #0x3f, lsl #24
3221 .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s
3222 .long 0x91004021 // add x1, x1, #0x10
3223 .long 0xd61f0060 // br x3
3224
3225HIDDEN _sk_bilinear_ny_aarch64
3226.globl _sk_bilinear_ny_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3227FUNCTION(_sk_bilinear_ny_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3228_sk_bilinear_ny_aarch64:
3229 .long 0xf9400028 // ldr x8, [x1]
3230 .long 0x4f03f611 // fmov v17.4s, #1.000000000000000000e+00
3231 .long 0x3dc01901 // ldr q1, [x8, #96]
3232 .long 0x3dc00910 // ldr q16, [x8, #32]
3233 .long 0x4ea1d621 // fsub v1.4s, v17.4s, v1.4s
3234 .long 0x3d802901 // str q1, [x8, #160]
3235 .long 0xf9400423 // ldr x3, [x1, #8]
3236 .long 0x4f0567e1 // movi v1.4s, #0xbf, lsl #24
3237 .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s
3238 .long 0x91004021 // add x1, x1, #0x10
3239 .long 0xd61f0060 // br x3
3240
3241HIDDEN _sk_bilinear_py_aarch64
3242.globl _sk_bilinear_py_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3243FUNCTION(_sk_bilinear_py_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3244_sk_bilinear_py_aarch64:
3245 .long 0xf9400028 // ldr x8, [x1]
3246 .long 0x3dc01901 // ldr q1, [x8, #96]
3247 .long 0x3dc00910 // ldr q16, [x8, #32]
3248 .long 0x3d802901 // str q1, [x8, #160]
3249 .long 0xf9400423 // ldr x3, [x1, #8]
3250 .long 0x4f0167e1 // movi v1.4s, #0x3f, lsl #24
3251 .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s
3252 .long 0x91004021 // add x1, x1, #0x10
3253 .long 0xd61f0060 // br x3
3254
3255HIDDEN _sk_bicubic_n3x_aarch64
3256.globl _sk_bicubic_n3x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3257FUNCTION(_sk_bicubic_n3x_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3258_sk_bicubic_n3x_aarch64:
3259 .long 0xf9400028 // ldr x8, [x1]
3260 .long 0x52a7d8e9 // mov w9, #0x3ec70000
3261 .long 0x72838e49 // movk w9, #0x1c72
3262 .long 0x4e040d30 // dup v16.4s, w9
3263 .long 0x3dc01111 // ldr q17, [x8, #64]
3264 .long 0x52b7d549 // mov w9, #0xbeaa0000
3265 .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00
3266 .long 0x72955569 // movk w9, #0xaaab
3267 .long 0x4e040d32 // dup v18.4s, w9
3268 .long 0x4eb1d400 // fsub v0.4s, v0.4s, v17.4s
3269 .long 0x6e20dc11 // fmul v17.4s, v0.4s, v0.4s
3270 .long 0x4e20ce12 // fmla v18.4s, v16.4s, v0.4s
3271 .long 0x6e32de20 // fmul v0.4s, v17.4s, v18.4s
3272 .long 0x3dc00113 // ldr q19, [x8]
3273 .long 0x3d802100 // str q0, [x8, #128]
3274 .long 0xf9400423 // ldr x3, [x1, #8]
3275 .long 0x4f07f700 // fmov v0.4s, #-1.500000000000000000e+00
3276 .long 0x4e20d660 // fadd v0.4s, v19.4s, v0.4s
3277 .long 0x91004021 // add x1, x1, #0x10
3278 .long 0xd61f0060 // br x3
3279
3280HIDDEN _sk_bicubic_n1x_aarch64
3281.globl _sk_bicubic_n1x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3282FUNCTION(_sk_bicubic_n1x_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3283_sk_bicubic_n1x_aarch64:
3284 .long 0xf9400028 // ldr x8, [x1]
3285 .long 0x52b7f2a9 // mov w9, #0xbf950000
3286 .long 0x4f03f600 // fmov v0.4s, #1.000000000000000000e+00
3287 .long 0x728aaaa9 // movk w9, #0x5555
3288 .long 0x3dc01110 // ldr q16, [x8, #64]
3289 .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00
3290 .long 0x4f0167f2 // movi v18.4s, #0x3f, lsl #24
3291 .long 0x4eb0d400 // fsub v0.4s, v0.4s, v16.4s
3292 .long 0x4e040d30 // dup v16.4s, w9
3293 .long 0x52a7ac69 // mov w9, #0x3d630000
3294 .long 0x7291c729 // movk w9, #0x8e39
3295 .long 0x4e20ce11 // fmla v17.4s, v16.4s, v0.4s
3296 .long 0x4e20ce32 // fmla v18.4s, v17.4s, v0.4s
3297 .long 0x4e040d31 // dup v17.4s, w9
3298 .long 0x4e20ce51 // fmla v17.4s, v18.4s, v0.4s
3299 .long 0x3dc00110 // ldr q16, [x8]
3300 .long 0x3d802111 // str q17, [x8, #128]
3301 .long 0xf9400423 // ldr x3, [x1, #8]
3302 .long 0x4f0567e0 // movi v0.4s, #0xbf, lsl #24
3303 .long 0x4e20d600 // fadd v0.4s, v16.4s, v0.4s
3304 .long 0x91004021 // add x1, x1, #0x10
3305 .long 0xd61f0060 // br x3
3306
3307HIDDEN _sk_bicubic_p1x_aarch64
3308.globl _sk_bicubic_p1x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3309FUNCTION(_sk_bicubic_p1x_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3310_sk_bicubic_p1x_aarch64:
3311 .long 0xf9400028 // ldr x8, [x1]
3312 .long 0x52b7f2a9 // mov w9, #0xbf950000
3313 .long 0x728aaaa9 // movk w9, #0x5555
3314 .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00
3315 .long 0x3dc01112 // ldr q18, [x8, #64]
3316 .long 0x3dc00100 // ldr q0, [x8]
3317 .long 0x4e040d33 // dup v19.4s, w9
3318 .long 0x52a7ac69 // mov w9, #0x3d630000
3319 .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24
3320 .long 0x7291c729 // movk w9, #0x8e39
3321 .long 0x4e32ce71 // fmla v17.4s, v19.4s, v18.4s
3322 .long 0x4e30d400 // fadd v0.4s, v0.4s, v16.4s
3323 .long 0x4e32ce30 // fmla v16.4s, v17.4s, v18.4s
3324 .long 0x4e040d31 // dup v17.4s, w9
3325 .long 0x4e32ce11 // fmla v17.4s, v16.4s, v18.4s
3326 .long 0x3d802111 // str q17, [x8, #128]
3327 .long 0xf9400423 // ldr x3, [x1, #8]
3328 .long 0x91004021 // add x1, x1, #0x10
3329 .long 0xd61f0060 // br x3
3330
3331HIDDEN _sk_bicubic_p3x_aarch64
3332.globl _sk_bicubic_p3x_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3333FUNCTION(_sk_bicubic_p3x_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3334_sk_bicubic_p3x_aarch64:
3335 .long 0xf9400028 // ldr x8, [x1]
3336 .long 0x52a7d8e9 // mov w9, #0x3ec70000
3337 .long 0x72838e49 // movk w9, #0x1c72
3338 .long 0x4e040d20 // dup v0.4s, w9
3339 .long 0x3dc01110 // ldr q16, [x8, #64]
3340 .long 0x52b7d549 // mov w9, #0xbeaa0000
3341 .long 0x72955569 // movk w9, #0xaaab
3342 .long 0x4e040d31 // dup v17.4s, w9
3343 .long 0x6e30de13 // fmul v19.4s, v16.4s, v16.4s
3344 .long 0x4e30cc11 // fmla v17.4s, v0.4s, v16.4s
3345 .long 0x6e31de60 // fmul v0.4s, v19.4s, v17.4s
3346 .long 0x3dc00112 // ldr q18, [x8]
3347 .long 0x3d802100 // str q0, [x8, #128]
3348 .long 0xf9400423 // ldr x3, [x1, #8]
3349 .long 0x4f03f700 // fmov v0.4s, #1.500000000000000000e+00
3350 .long 0x4e20d640 // fadd v0.4s, v18.4s, v0.4s
3351 .long 0x91004021 // add x1, x1, #0x10
3352 .long 0xd61f0060 // br x3
3353
3354HIDDEN _sk_bicubic_n3y_aarch64
3355.globl _sk_bicubic_n3y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3356FUNCTION(_sk_bicubic_n3y_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3357_sk_bicubic_n3y_aarch64:
3358 .long 0xf9400028 // ldr x8, [x1]
3359 .long 0x52a7d8e9 // mov w9, #0x3ec70000
3360 .long 0x72838e49 // movk w9, #0x1c72
3361 .long 0x4e040d30 // dup v16.4s, w9
3362 .long 0x3dc01911 // ldr q17, [x8, #96]
3363 .long 0x52b7d549 // mov w9, #0xbeaa0000
3364 .long 0x4f03f601 // fmov v1.4s, #1.000000000000000000e+00
3365 .long 0x72955569 // movk w9, #0xaaab
3366 .long 0x4e040d32 // dup v18.4s, w9
3367 .long 0x4eb1d421 // fsub v1.4s, v1.4s, v17.4s
3368 .long 0x6e21dc31 // fmul v17.4s, v1.4s, v1.4s
3369 .long 0x4e21ce12 // fmla v18.4s, v16.4s, v1.4s
3370 .long 0x6e32de21 // fmul v1.4s, v17.4s, v18.4s
3371 .long 0x3dc00913 // ldr q19, [x8, #32]
3372 .long 0x3d802901 // str q1, [x8, #160]
3373 .long 0xf9400423 // ldr x3, [x1, #8]
3374 .long 0x4f07f701 // fmov v1.4s, #-1.500000000000000000e+00
3375 .long 0x4e21d661 // fadd v1.4s, v19.4s, v1.4s
3376 .long 0x91004021 // add x1, x1, #0x10
3377 .long 0xd61f0060 // br x3
3378
3379HIDDEN _sk_bicubic_n1y_aarch64
3380.globl _sk_bicubic_n1y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3381FUNCTION(_sk_bicubic_n1y_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3382_sk_bicubic_n1y_aarch64:
3383 .long 0xf9400028 // ldr x8, [x1]
3384 .long 0x52b7f2a9 // mov w9, #0xbf950000
3385 .long 0x4f03f601 // fmov v1.4s, #1.000000000000000000e+00
3386 .long 0x728aaaa9 // movk w9, #0x5555
3387 .long 0x3dc01910 // ldr q16, [x8, #96]
3388 .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00
3389 .long 0x4f0167f2 // movi v18.4s, #0x3f, lsl #24
3390 .long 0x4eb0d421 // fsub v1.4s, v1.4s, v16.4s
3391 .long 0x4e040d30 // dup v16.4s, w9
3392 .long 0x52a7ac69 // mov w9, #0x3d630000
3393 .long 0x7291c729 // movk w9, #0x8e39
3394 .long 0x4e21ce11 // fmla v17.4s, v16.4s, v1.4s
3395 .long 0x4e21ce32 // fmla v18.4s, v17.4s, v1.4s
3396 .long 0x4e040d31 // dup v17.4s, w9
3397 .long 0x4e21ce51 // fmla v17.4s, v18.4s, v1.4s
3398 .long 0x3dc00910 // ldr q16, [x8, #32]
3399 .long 0x3d802911 // str q17, [x8, #160]
3400 .long 0xf9400423 // ldr x3, [x1, #8]
3401 .long 0x4f0567e1 // movi v1.4s, #0xbf, lsl #24
3402 .long 0x4e21d601 // fadd v1.4s, v16.4s, v1.4s
3403 .long 0x91004021 // add x1, x1, #0x10
3404 .long 0xd61f0060 // br x3
3405
3406HIDDEN _sk_bicubic_p1y_aarch64
3407.globl _sk_bicubic_p1y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3408FUNCTION(_sk_bicubic_p1y_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3409_sk_bicubic_p1y_aarch64:
3410 .long 0xf9400028 // ldr x8, [x1]
3411 .long 0x52b7f2a9 // mov w9, #0xbf950000
3412 .long 0x728aaaa9 // movk w9, #0x5555
3413 .long 0x4f03f711 // fmov v17.4s, #1.500000000000000000e+00
3414 .long 0x3dc01912 // ldr q18, [x8, #96]
3415 .long 0x3dc00901 // ldr q1, [x8, #32]
3416 .long 0x4e040d33 // dup v19.4s, w9
3417 .long 0x52a7ac69 // mov w9, #0x3d630000
3418 .long 0x4f0167f0 // movi v16.4s, #0x3f, lsl #24
3419 .long 0x7291c729 // movk w9, #0x8e39
3420 .long 0x4e32ce71 // fmla v17.4s, v19.4s, v18.4s
3421 .long 0x4e30d421 // fadd v1.4s, v1.4s, v16.4s
3422 .long 0x4e32ce30 // fmla v16.4s, v17.4s, v18.4s
3423 .long 0x4e040d31 // dup v17.4s, w9
3424 .long 0x4e32ce11 // fmla v17.4s, v16.4s, v18.4s
3425 .long 0x3d802911 // str q17, [x8, #160]
3426 .long 0xf9400423 // ldr x3, [x1, #8]
3427 .long 0x91004021 // add x1, x1, #0x10
3428 .long 0xd61f0060 // br x3
3429
3430HIDDEN _sk_bicubic_p3y_aarch64
3431.globl _sk_bicubic_p3y_aarch64
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3432FUNCTION(_sk_bicubic_p3y_aarch64)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]3433_sk_bicubic_p3y_aarch64:
3434 .long 0xf9400028 // ldr x8, [x1]
3435 .long 0x52a7d8e9 // mov w9, #0x3ec70000
3436 .long 0x72838e49 // movk w9, #0x1c72
3437 .long 0x4e040d21 // dup v1.4s, w9
3438 .long 0x3dc01910 // ldr q16, [x8, #96]
3439 .long 0x52b7d549 // mov w9, #0xbeaa0000
3440 .long 0x72955569 // movk w9, #0xaaab
3441 .long 0x4e040d31 // dup v17.4s, w9
3442 .long 0x6e30de13 // fmul v19.4s, v16.4s, v16.4s
3443 .long 0x4e30cc31 // fmla v17.4s, v1.4s, v16.4s
3444 .long 0x6e31de61 // fmul v1.4s, v19.4s, v17.4s
3445 .long 0x3dc00912 // ldr q18, [x8, #32]
3446 .long 0x3d802901 // str q1, [x8, #160]
3447 .long 0xf9400423 // ldr x3, [x1, #8]
3448 .long 0x4f03f701 // fmov v1.4s, #1.500000000000000000e+00
3449 .long 0x4e21d641 // fadd v1.4s, v18.4s, v1.4s
3450 .long 0x91004021 // add x1, x1, #0x10
3451 .long 0xd61f0060 // br x3
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3452
3453HIDDEN _sk_callback_aarch64
3454.globl _sk_callback_aarch64
3455FUNCTION(_sk_callback_aarch64)
3456_sk_callback_aarch64:
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]3457 .long 0xd101c3ff // sub sp, sp, #0x70
3458 .long 0xf90023f6 // str x22, [sp, #64]
3459 .long 0xa90553f5 // stp x21, x20, [sp, #80]
3460 .long 0xa9067bf3 // stp x19, x30, [sp, #96]
3461 .long 0xad011fe6 // stp q6, q7, [sp, #32]
3462 .long 0xad0017e4 // stp q4, q5, [sp]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3463 .long 0xaa0103f4 // mov x20, x1
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]3464 .long 0xf9400295 // ldr x21, [x20]
3465 .long 0xaa0003f6 // mov x22, x0
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3466 .long 0x321e03e1 // orr w1, wzr, #0x4
3467 .long 0xaa0203f3 // mov x19, x2
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]3468 .long 0x910022a8 // add x8, x21, #0x8
3469 .long 0x4c000900 // st4 {v0.4s-v3.4s}, [x8]
3470 .long 0xf94002a8 // ldr x8, [x21]
3471 .long 0xaa1503e0 // mov x0, x21
3472 .long 0xd63f0100 // blr x8
3473 .long 0xf94046a8 // ldr x8, [x21, #136]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3474 .long 0xf9400683 // ldr x3, [x20, #8]
3475 .long 0x91004281 // add x1, x20, #0x10
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]3476 .long 0xaa1603e0 // mov x0, x22
3477 .long 0x4c400900 // ld4 {v0.4s-v3.4s}, [x8]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3478 .long 0xaa1303e2 // mov x2, x19
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]3479 .long 0xad4017e4 // ldp q4, q5, [sp]
3480 .long 0xad411fe6 // ldp q6, q7, [sp, #32]
3481 .long 0xa9467bf3 // ldp x19, x30, [sp, #96]
3482 .long 0xa94553f5 // ldp x21, x20, [sp, #80]
3483 .long 0xf94023f6 // ldr x22, [sp, #64]
3484 .long 0x9101c3ff // add sp, sp, #0x70
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]3485 .long 0xd61f0060 // br x3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3486#elif defined(__arm__)
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]3487BALIGN4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3488
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3489HIDDEN _sk_start_pipeline_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3490.globl _sk_start_pipeline_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3491FUNCTION(_sk_start_pipeline_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3492_sk_start_pipeline_vfp4:
3493 .long 0xe92d41f0 // push {r4, r5, r6, r7, r8, lr}
3494 .long 0xe1a04000 // mov r4, r0
3495 .long 0xe2840002 // add r0, r4, #2
3496 .long 0xe1a05003 // mov r5, r3
3497 .long 0xe1a08002 // mov r8, r2
3498 .long 0xe1a07001 // mov r7, r1
3499 .long 0xe1500005 // cmp r0, r5
3500 .long 0x8a000010 // bhi 64 <sk_start_pipeline_vfp4+0x64>
3501 .long 0xe4976004 // ldr r6, [r7], #4
3502 .long 0xf2800010 // vmov.i32 d0, #0
3503 .long 0xe1a00004 // mov r0, r4
3504 .long 0xf2801010 // vmov.i32 d1, #0
3505 .long 0xe1a01007 // mov r1, r7
3506 .long 0xf2802010 // vmov.i32 d2, #0
3507 .long 0xe1a02008 // mov r2, r8
3508 .long 0xf2803010 // vmov.i32 d3, #0
3509 .long 0xf2804010 // vmov.i32 d4, #0
3510 .long 0xf2805010 // vmov.i32 d5, #0
3511 .long 0xf2806010 // vmov.i32 d6, #0
3512 .long 0xf2807010 // vmov.i32 d7, #0
3513 .long 0xe12fff36 // blx r6
3514 .long 0xe2840004 // add r0, r4, #4
3515 .long 0xe2844002 // add r4, r4, #2
3516 .long 0xe1500005 // cmp r0, r5
3517 .long 0x9affffef // bls 24 <sk_start_pipeline_vfp4+0x24>
3518 .long 0xe1a00004 // mov r0, r4
3519 .long 0xe8bd81f0 // pop {r4, r5, r6, r7, r8, pc}
3520
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3521HIDDEN _sk_just_return_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3522.globl _sk_just_return_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3523FUNCTION(_sk_just_return_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3524_sk_just_return_vfp4:
3525 .long 0xe12fff1e // bx lr
3526
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3527HIDDEN _sk_seed_shader_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3528.globl _sk_seed_shader_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3529FUNCTION(_sk_seed_shader_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3530_sk_seed_shader_vfp4:
3531 .long 0xee800b90 // vdup.32 d16, r0
3532 .long 0xe8911008 // ldm r1, {r3, ip}
3533 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
3534 .long 0xf2c3161f // vmov.i32 d17, #1056964608
3535 .long 0xedd23b00 // vldr d19, [r2]
3536 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
3537 .long 0xf2872f10 // vmov.f32 d2, #1
3538 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
3539 .long 0xe2811008 // add r1, r1, #8
3540 .long 0xf2400da1 // vadd.f32 d16, d16, d17
3541 .long 0xf2803010 // vmov.i32 d3, #0
3542 .long 0xf2804010 // vmov.i32 d4, #0
3543 .long 0xf2021da1 // vadd.f32 d1, d18, d17
3544 .long 0xf2000da3 // vadd.f32 d0, d16, d19
3545 .long 0xf2805010 // vmov.i32 d5, #0
3546 .long 0xf2806010 // vmov.i32 d6, #0
3547 .long 0xf2807010 // vmov.i32 d7, #0
3548 .long 0xe12fff1c // bx ip
3549
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3550HIDDEN _sk_constant_color_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3551.globl _sk_constant_color_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3552FUNCTION(_sk_constant_color_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3553_sk_constant_color_vfp4:
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]3554 .long 0xe92d4010 // push {r4, lr}
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3555 .long 0xe8911008 // ldm r1, {r3, ip}
3556 .long 0xe2811008 // add r1, r1, #8
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]3557 .long 0xe283400c // add r4, r3, #12
3558 .long 0xe1a0e003 // mov lr, r3
3559 .long 0xe2833008 // add r3, r3, #8
3560 .long 0xf4ae0c9d // vld1.32 {d0[]}, [lr :32]!
3561 .long 0xf4a43c9f // vld1.32 {d3[]}, [r4 :32]
3562 .long 0xf4a32c9f // vld1.32 {d2[]}, [r3 :32]
3563 .long 0xf4ae1c9f // vld1.32 {d1[]}, [lr :32]
3564 .long 0xe8bd4010 // pop {r4, lr}
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3565 .long 0xe12fff1c // bx ip
3566
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3567HIDDEN _sk_clear_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3568.globl _sk_clear_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3569FUNCTION(_sk_clear_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3570_sk_clear_vfp4:
3571 .long 0xe4913004 // ldr r3, [r1], #4
3572 .long 0xf2800010 // vmov.i32 d0, #0
3573 .long 0xf2801010 // vmov.i32 d1, #0
3574 .long 0xf2802010 // vmov.i32 d2, #0
3575 .long 0xf2803010 // vmov.i32 d3, #0
3576 .long 0xe12fff13 // bx r3
3577
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3578HIDDEN _sk_srcatop_vfp4
3579.globl _sk_srcatop_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3580FUNCTION(_sk_srcatop_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3581_sk_srcatop_vfp4:
3582 .long 0xf2c70f10 // vmov.f32 d16, #1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3583 .long 0xe4913004 // ldr r3, [r1], #4
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3584 .long 0xf2603d83 // vsub.f32 d19, d16, d3
3585 .long 0xf3033d17 // vmul.f32 d3, d3, d7
3586 .long 0xf3430d94 // vmul.f32 d16, d19, d4
3587 .long 0xf3431d95 // vmul.f32 d17, d19, d5
3588 .long 0xf3432d96 // vmul.f32 d18, d19, d6
3589 .long 0xf2400c17 // vfma.f32 d16, d0, d7
3590 .long 0xf2411c17 // vfma.f32 d17, d1, d7
3591 .long 0xf2422c17 // vfma.f32 d18, d2, d7
3592 .long 0xf2033c97 // vfma.f32 d3, d19, d7
3593 .long 0xf22001b0 // vorr d0, d16, d16
3594 .long 0xf22111b1 // vorr d1, d17, d17
3595 .long 0xf22221b2 // vorr d2, d18, d18
3596 .long 0xe12fff13 // bx r3
3597
3598HIDDEN _sk_dstatop_vfp4
3599.globl _sk_dstatop_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3600FUNCTION(_sk_dstatop_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3601_sk_dstatop_vfp4:
3602 .long 0xf2c70f10 // vmov.f32 d16, #1
3603 .long 0xe4913004 // ldr r3, [r1], #4
3604 .long 0xf3431d15 // vmul.f32 d17, d3, d5
3605 .long 0xf2604d87 // vsub.f32 d20, d16, d7
3606 .long 0xf3430d14 // vmul.f32 d16, d3, d4
3607 .long 0xf3432d16 // vmul.f32 d18, d3, d6
3608 .long 0xf3433d17 // vmul.f32 d19, d3, d7
3609 .long 0xf2440c90 // vfma.f32 d16, d20, d0
3610 .long 0xf2441c91 // vfma.f32 d17, d20, d1
3611 .long 0xf2442c92 // vfma.f32 d18, d20, d2
3612 .long 0xf2443c93 // vfma.f32 d19, d20, d3
3613 .long 0xf22001b0 // vorr d0, d16, d16
3614 .long 0xf22111b1 // vorr d1, d17, d17
3615 .long 0xf22221b2 // vorr d2, d18, d18
3616 .long 0xf22331b3 // vorr d3, d19, d19
3617 .long 0xe12fff13 // bx r3
3618
3619HIDDEN _sk_srcin_vfp4
3620.globl _sk_srcin_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3621FUNCTION(_sk_srcin_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3622_sk_srcin_vfp4:
3623 .long 0xf3000d17 // vmul.f32 d0, d0, d7
3624 .long 0xe4913004 // ldr r3, [r1], #4
3625 .long 0xf3011d17 // vmul.f32 d1, d1, d7
3626 .long 0xf3022d17 // vmul.f32 d2, d2, d7
3627 .long 0xf3033d17 // vmul.f32 d3, d3, d7
3628 .long 0xe12fff13 // bx r3
3629
3630HIDDEN _sk_dstin_vfp4
3631.globl _sk_dstin_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3632FUNCTION(_sk_dstin_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3633_sk_dstin_vfp4:
3634 .long 0xf3030d14 // vmul.f32 d0, d3, d4
3635 .long 0xe4913004 // ldr r3, [r1], #4
3636 .long 0xf3031d15 // vmul.f32 d1, d3, d5
3637 .long 0xf3032d16 // vmul.f32 d2, d3, d6
3638 .long 0xf3033d17 // vmul.f32 d3, d3, d7
3639 .long 0xe12fff13 // bx r3
3640
3641HIDDEN _sk_srcout_vfp4
3642.globl _sk_srcout_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3643FUNCTION(_sk_srcout_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3644_sk_srcout_vfp4:
3645 .long 0xf2c70f10 // vmov.f32 d16, #1
3646 .long 0xe4913004 // ldr r3, [r1], #4
3647 .long 0xf2600d87 // vsub.f32 d16, d16, d7
3648 .long 0xf3000d90 // vmul.f32 d0, d16, d0
3649 .long 0xf3001d91 // vmul.f32 d1, d16, d1
3650 .long 0xf3002d92 // vmul.f32 d2, d16, d2
3651 .long 0xf3003d93 // vmul.f32 d3, d16, d3
3652 .long 0xe12fff13 // bx r3
3653
3654HIDDEN _sk_dstout_vfp4
3655.globl _sk_dstout_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3656FUNCTION(_sk_dstout_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3657_sk_dstout_vfp4:
3658 .long 0xf2c70f10 // vmov.f32 d16, #1
3659 .long 0xe4913004 // ldr r3, [r1], #4
3660 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3661 .long 0xf3000d94 // vmul.f32 d0, d16, d4
3662 .long 0xf3001d95 // vmul.f32 d1, d16, d5
3663 .long 0xf3002d96 // vmul.f32 d2, d16, d6
3664 .long 0xf3003d97 // vmul.f32 d3, d16, d7
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3665 .long 0xe12fff13 // bx r3
3666
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3667HIDDEN _sk_srcover_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3668.globl _sk_srcover_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3669FUNCTION(_sk_srcover_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3670_sk_srcover_vfp4:
3671 .long 0xf2c70f10 // vmov.f32 d16, #1
3672 .long 0xe4913004 // ldr r3, [r1], #4
3673 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3674 .long 0xf2040c30 // vfma.f32 d0, d4, d16
3675 .long 0xf2051c30 // vfma.f32 d1, d5, d16
3676 .long 0xf2062c30 // vfma.f32 d2, d6, d16
3677 .long 0xf2073c30 // vfma.f32 d3, d7, d16
3678 .long 0xe12fff13 // bx r3
3679
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]3680HIDDEN _sk_dstover_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3681.globl _sk_dstover_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3682FUNCTION(_sk_dstover_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]3683_sk_dstover_vfp4:
3684 .long 0xf2c70f10 // vmov.f32 d16, #1
3685 .long 0xe4913004 // ldr r3, [r1], #4
3686 .long 0xf2651115 // vorr d17, d5, d5
3687 .long 0xf2604d87 // vsub.f32 d20, d16, d7
3688 .long 0xf2640114 // vorr d16, d4, d4
3689 .long 0xf2662116 // vorr d18, d6, d6
3690 .long 0xf2673117 // vorr d19, d7, d7
3691 .long 0xf2400c34 // vfma.f32 d16, d0, d20
3692 .long 0xf2411c34 // vfma.f32 d17, d1, d20
3693 .long 0xf2422c34 // vfma.f32 d18, d2, d20
3694 .long 0xf2433c34 // vfma.f32 d19, d3, d20
3695 .long 0xf22001b0 // vorr d0, d16, d16
3696 .long 0xf22111b1 // vorr d1, d17, d17
3697 .long 0xf22221b2 // vorr d2, d18, d18
3698 .long 0xf22331b3 // vorr d3, d19, d19
3699 .long 0xe12fff13 // bx r3
3700
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3701HIDDEN _sk_modulate_vfp4
3702.globl _sk_modulate_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3703FUNCTION(_sk_modulate_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3704_sk_modulate_vfp4:
3705 .long 0xf3000d14 // vmul.f32 d0, d0, d4
3706 .long 0xe4913004 // ldr r3, [r1], #4
3707 .long 0xf3011d15 // vmul.f32 d1, d1, d5
3708 .long 0xf3022d16 // vmul.f32 d2, d2, d6
3709 .long 0xf3033d17 // vmul.f32 d3, d3, d7
3710 .long 0xe12fff13 // bx r3
3711
3712HIDDEN _sk_multiply_vfp4
3713.globl _sk_multiply_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3714FUNCTION(_sk_multiply_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3715_sk_multiply_vfp4:
3716 .long 0xf2c70f10 // vmov.f32 d16, #1
3717 .long 0xe4913004 // ldr r3, [r1], #4
3718 .long 0xf2601d87 // vsub.f32 d17, d16, d7
3719 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3720 .long 0xf3412d90 // vmul.f32 d18, d17, d0
3721 .long 0xf3403d94 // vmul.f32 d19, d16, d4
3722 .long 0xf3414d91 // vmul.f32 d20, d17, d1
3723 .long 0xf3405d95 // vmul.f32 d21, d16, d5
3724 .long 0xf3416d92 // vmul.f32 d22, d17, d2
3725 .long 0xf3418d93 // vmul.f32 d24, d17, d3
3726 .long 0xf3407d96 // vmul.f32 d23, d16, d6
3727 .long 0xf3409d97 // vmul.f32 d25, d16, d7
3728 .long 0xf2430da2 // vadd.f32 d16, d19, d18
3729 .long 0xf2451da4 // vadd.f32 d17, d21, d20
3730 .long 0xf2472da6 // vadd.f32 d18, d23, d22
3731 .long 0xf2493da8 // vadd.f32 d19, d25, d24
3732 .long 0xf2400c14 // vfma.f32 d16, d0, d4
3733 .long 0xf2411c15 // vfma.f32 d17, d1, d5
3734 .long 0xf2422c16 // vfma.f32 d18, d2, d6
3735 .long 0xf2433c17 // vfma.f32 d19, d3, d7
3736 .long 0xf22001b0 // vorr d0, d16, d16
3737 .long 0xf22111b1 // vorr d1, d17, d17
3738 .long 0xf22221b2 // vorr d2, d18, d18
3739 .long 0xf22331b3 // vorr d3, d19, d19
3740 .long 0xe12fff13 // bx r3
3741
3742HIDDEN _sk_plus__vfp4
3743.globl _sk_plus__vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3744FUNCTION(_sk_plus__vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3745_sk_plus__vfp4:
3746 .long 0xf2000d04 // vadd.f32 d0, d0, d4
3747 .long 0xe4913004 // ldr r3, [r1], #4
3748 .long 0xf2011d05 // vadd.f32 d1, d1, d5
3749 .long 0xf2022d06 // vadd.f32 d2, d2, d6
3750 .long 0xf2033d07 // vadd.f32 d3, d3, d7
3751 .long 0xe12fff13 // bx r3
3752
3753HIDDEN _sk_screen_vfp4
3754.globl _sk_screen_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3755FUNCTION(_sk_screen_vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3756_sk_screen_vfp4:
3757 .long 0xf2400d04 // vadd.f32 d16, d0, d4
3758 .long 0xe4913004 // ldr r3, [r1], #4
3759 .long 0xf2411d05 // vadd.f32 d17, d1, d5
3760 .long 0xf2422d06 // vadd.f32 d18, d2, d6
3761 .long 0xf2433d07 // vadd.f32 d19, d3, d7
3762 .long 0xf2600c14 // vfms.f32 d16, d0, d4
3763 .long 0xf2611c15 // vfms.f32 d17, d1, d5
3764 .long 0xf2622c16 // vfms.f32 d18, d2, d6
3765 .long 0xf2633c17 // vfms.f32 d19, d3, d7
3766 .long 0xf22001b0 // vorr d0, d16, d16
3767 .long 0xf22111b1 // vorr d1, d17, d17
3768 .long 0xf22221b2 // vorr d2, d18, d18
3769 .long 0xf22331b3 // vorr d3, d19, d19
3770 .long 0xe12fff13 // bx r3
3771
3772HIDDEN _sk_xor__vfp4
3773.globl _sk_xor__vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3774FUNCTION(_sk_xor__vfp4)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]3775_sk_xor__vfp4:
3776 .long 0xf2c70f10 // vmov.f32 d16, #1
3777 .long 0xe4913004 // ldr r3, [r1], #4
3778 .long 0xf2603d83 // vsub.f32 d19, d16, d3
3779 .long 0xf2604d87 // vsub.f32 d20, d16, d7
3780 .long 0xf3430d94 // vmul.f32 d16, d19, d4
3781 .long 0xf3431d95 // vmul.f32 d17, d19, d5
3782 .long 0xf3432d96 // vmul.f32 d18, d19, d6
3783 .long 0xf3433d97 // vmul.f32 d19, d19, d7
3784 .long 0xf2440c90 // vfma.f32 d16, d20, d0
3785 .long 0xf2441c91 // vfma.f32 d17, d20, d1
3786 .long 0xf2442c92 // vfma.f32 d18, d20, d2
3787 .long 0xf2443c93 // vfma.f32 d19, d20, d3
3788 .long 0xf22001b0 // vorr d0, d16, d16
3789 .long 0xf22111b1 // vorr d1, d17, d17
3790 .long 0xf22221b2 // vorr d2, d18, d18
3791 .long 0xf22331b3 // vorr d3, d19, d19
3792 .long 0xe12fff13 // bx r3
3793
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]3794HIDDEN _sk_darken_vfp4
3795.globl _sk_darken_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3796FUNCTION(_sk_darken_vfp4)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]3797_sk_darken_vfp4:
3798 .long 0xf2c70f10 // vmov.f32 d16, #1
3799 .long 0xe4913004 // ldr r3, [r1], #4
3800 .long 0xf3431d14 // vmul.f32 d17, d3, d4
3801 .long 0xf3402d17 // vmul.f32 d18, d0, d7
3802 .long 0xf3433d15 // vmul.f32 d19, d3, d5
3803 .long 0xf3414d17 // vmul.f32 d20, d1, d7
3804 .long 0xf3435d16 // vmul.f32 d21, d3, d6
3805 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3806 .long 0xf3426d17 // vmul.f32 d22, d2, d7
3807 .long 0xf2421fa1 // vmax.f32 d17, d18, d17
3808 .long 0xf2407d04 // vadd.f32 d23, d0, d4
3809 .long 0xf2443fa3 // vmax.f32 d19, d20, d19
3810 .long 0xf2412d05 // vadd.f32 d18, d1, d5
3811 .long 0xf2424d06 // vadd.f32 d20, d2, d6
3812 .long 0xf2465fa5 // vmax.f32 d21, d22, d21
3813 .long 0xf2073c30 // vfma.f32 d3, d7, d16
3814 .long 0xf2270da1 // vsub.f32 d0, d23, d17
3815 .long 0xf2221da3 // vsub.f32 d1, d18, d19
3816 .long 0xf2242da5 // vsub.f32 d2, d20, d21
3817 .long 0xe12fff13 // bx r3
3818
3819HIDDEN _sk_lighten_vfp4
3820.globl _sk_lighten_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3821FUNCTION(_sk_lighten_vfp4)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]3822_sk_lighten_vfp4:
3823 .long 0xf2c70f10 // vmov.f32 d16, #1
3824 .long 0xe4913004 // ldr r3, [r1], #4
3825 .long 0xf3431d14 // vmul.f32 d17, d3, d4
3826 .long 0xf3402d17 // vmul.f32 d18, d0, d7
3827 .long 0xf3433d15 // vmul.f32 d19, d3, d5
3828 .long 0xf3414d17 // vmul.f32 d20, d1, d7
3829 .long 0xf3435d16 // vmul.f32 d21, d3, d6
3830 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3831 .long 0xf3426d17 // vmul.f32 d22, d2, d7
3832 .long 0xf2621fa1 // vmin.f32 d17, d18, d17
3833 .long 0xf2407d04 // vadd.f32 d23, d0, d4
3834 .long 0xf2643fa3 // vmin.f32 d19, d20, d19
3835 .long 0xf2412d05 // vadd.f32 d18, d1, d5
3836 .long 0xf2424d06 // vadd.f32 d20, d2, d6
3837 .long 0xf2665fa5 // vmin.f32 d21, d22, d21
3838 .long 0xf2073c30 // vfma.f32 d3, d7, d16
3839 .long 0xf2270da1 // vsub.f32 d0, d23, d17
3840 .long 0xf2221da3 // vsub.f32 d1, d18, d19
3841 .long 0xf2242da5 // vsub.f32 d2, d20, d21
3842 .long 0xe12fff13 // bx r3
3843
3844HIDDEN _sk_difference_vfp4
3845.globl _sk_difference_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3846FUNCTION(_sk_difference_vfp4)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]3847_sk_difference_vfp4:
3848 .long 0xf3430d14 // vmul.f32 d16, d3, d4
3849 .long 0xe4913004 // ldr r3, [r1], #4
3850 .long 0xf3401d17 // vmul.f32 d17, d0, d7
3851 .long 0xf3432d15 // vmul.f32 d18, d3, d5
3852 .long 0xf3413d17 // vmul.f32 d19, d1, d7
3853 .long 0xf3434d16 // vmul.f32 d20, d3, d6
3854 .long 0xf3425d17 // vmul.f32 d21, d2, d7
3855 .long 0xf2c76f10 // vmov.f32 d22, #1
3856 .long 0xf2610fa0 // vmin.f32 d16, d17, d16
3857 .long 0xf2631fa2 // vmin.f32 d17, d19, d18
3858 .long 0xf2662d83 // vsub.f32 d18, d22, d3
3859 .long 0xf2653fa4 // vmin.f32 d19, d21, d20
3860 .long 0xf2404d04 // vadd.f32 d20, d0, d4
3861 .long 0xf2400da0 // vadd.f32 d16, d16, d16
3862 .long 0xf2073c32 // vfma.f32 d3, d7, d18
3863 .long 0xf2415d05 // vadd.f32 d21, d1, d5
3864 .long 0xf2411da1 // vadd.f32 d17, d17, d17
3865 .long 0xf2426d06 // vadd.f32 d22, d2, d6
3866 .long 0xf2432da3 // vadd.f32 d18, d19, d19
3867 .long 0xf2240da0 // vsub.f32 d0, d20, d16
3868 .long 0xf2251da1 // vsub.f32 d1, d21, d17
3869 .long 0xf2262da2 // vsub.f32 d2, d22, d18
3870 .long 0xe12fff13 // bx r3
3871
3872HIDDEN _sk_exclusion_vfp4
3873.globl _sk_exclusion_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3874FUNCTION(_sk_exclusion_vfp4)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]3875_sk_exclusion_vfp4:
3876 .long 0xf2c70f10 // vmov.f32 d16, #1
3877 .long 0xe4913004 // ldr r3, [r1], #4
3878 .long 0xf3401d14 // vmul.f32 d17, d0, d4
3879 .long 0xf3412d15 // vmul.f32 d18, d1, d5
3880 .long 0xf3423d16 // vmul.f32 d19, d2, d6
3881 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3882 .long 0xf2404d04 // vadd.f32 d20, d0, d4
3883 .long 0xf2411da1 // vadd.f32 d17, d17, d17
3884 .long 0xf2415d05 // vadd.f32 d21, d1, d5
3885 .long 0xf2422da2 // vadd.f32 d18, d18, d18
3886 .long 0xf2426d06 // vadd.f32 d22, d2, d6
3887 .long 0xf2433da3 // vadd.f32 d19, d19, d19
3888 .long 0xf2073c30 // vfma.f32 d3, d7, d16
3889 .long 0xf2240da1 // vsub.f32 d0, d20, d17
3890 .long 0xf2251da2 // vsub.f32 d1, d21, d18
3891 .long 0xf2262da3 // vsub.f32 d2, d22, d19
3892 .long 0xe12fff13 // bx r3
3893
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]3894HIDDEN _sk_colorburn_vfp4
3895.globl _sk_colorburn_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3896FUNCTION(_sk_colorburn_vfp4)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]3897_sk_colorburn_vfp4:
3898 .long 0xed2d8b08 // vpush {d8-d11}
3899 .long 0xf2670d04 // vsub.f32 d16, d7, d4
3900 .long 0xe4913004 // ldr r3, [r1], #4
3901 .long 0xf2671d06 // vsub.f32 d17, d7, d6
3902 .long 0xf2672d05 // vsub.f32 d18, d7, d5
3903 .long 0xf3008d93 // vmul.f32 d8, d16, d3
3904 .long 0xf3019d93 // vmul.f32 d9, d17, d3
3905 .long 0xf302ad93 // vmul.f32 d10, d18, d3
3906 .long 0xf2c71f10 // vmov.f32 d17, #1
3907 .long 0xeec8baa0 // vdiv.f32 s23, s17, s1
3908 .long 0xee88ba00 // vdiv.f32 s22, s16, s0
3909 .long 0xeec98aa2 // vdiv.f32 s17, s19, s5
3910 .long 0xee898a02 // vdiv.f32 s16, s18, s4
3911 .long 0xeeca9aa1 // vdiv.f32 s19, s21, s3
3912 .long 0xee8a9a01 // vdiv.f32 s18, s20, s2
3913 .long 0xf2672f08 // vmin.f32 d18, d7, d8
3914 .long 0xf2673f09 // vmin.f32 d19, d7, d9
3915 .long 0xf2670f0b // vmin.f32 d16, d7, d11
3916 .long 0xf2614d87 // vsub.f32 d20, d17, d7
3917 .long 0xf2672d22 // vsub.f32 d18, d7, d18
3918 .long 0xf2673d23 // vsub.f32 d19, d7, d19
3919 .long 0xf2611d83 // vsub.f32 d17, d17, d3
3920 .long 0xf2670d20 // vsub.f32 d16, d7, d16
3921 .long 0xf3445d90 // vmul.f32 d21, d20, d0
3922 .long 0xf3446d92 // vmul.f32 d22, d20, d2
3923 .long 0xf3422d93 // vmul.f32 d18, d18, d3
3924 .long 0xf3444d91 // vmul.f32 d20, d20, d1
3925 .long 0xf3433d93 // vmul.f32 d19, d19, d3
3926 .long 0xf3400d93 // vmul.f32 d16, d16, d3
3927 .long 0xf3417d95 // vmul.f32 d23, d17, d5
3928 .long 0xf3418d94 // vmul.f32 d24, d17, d4
3929 .long 0xf3419d96 // vmul.f32 d25, d17, d6
3930 .long 0xf2443da3 // vadd.f32 d19, d20, d19
3931 .long 0xf2462da2 // vadd.f32 d18, d22, d18
3932 .long 0xf245ada0 // vadd.f32 d26, d21, d16
3933 .long 0xf247bd81 // vadd.f32 d27, d23, d1
3934 .long 0xf248cd80 // vadd.f32 d28, d24, d0
3935 .long 0xf249dd82 // vadd.f32 d29, d25, d2
3936 .long 0xf2073c31 // vfma.f32 d3, d7, d17
3937 .long 0xf2499da2 // vadd.f32 d25, d25, d18
3938 .long 0xf2473da3 // vadd.f32 d19, d23, d19
3939 .long 0xf3f97501 // vceq.f32 d23, d1, #0
3940 .long 0xf2455d84 // vadd.f32 d21, d21, d4
3941 .long 0xf2444d85 // vadd.f32 d20, d20, d5
3942 .long 0xf2440e07 // vceq.f32 d16, d4, d7
3943 .long 0xf2466d86 // vadd.f32 d22, d22, d6
3944 .long 0xf2451e07 // vceq.f32 d17, d5, d7
3945 .long 0xf2462e07 // vceq.f32 d18, d6, d7
3946 .long 0xf35b71b3 // vbsl d23, d27, d19
3947 .long 0xf3f93500 // vceq.f32 d19, d0, #0
3948 .long 0xf2488daa // vadd.f32 d24, d24, d26
3949 .long 0xf35c31b8 // vbsl d19, d28, d24
3950 .long 0xf3f98502 // vceq.f32 d24, d2, #0
3951 .long 0xf35d81b9 // vbsl d24, d29, d25
3952 .long 0xf35501b3 // vbsl d16, d21, d19
3953 .long 0xf35411b7 // vbsl d17, d20, d23
3954 .long 0xf35621b8 // vbsl d18, d22, d24
3955 .long 0xf22001b0 // vorr d0, d16, d16
3956 .long 0xf22111b1 // vorr d1, d17, d17
3957 .long 0xf22221b2 // vorr d2, d18, d18
3958 .long 0xecbd8b08 // vpop {d8-d11}
3959 .long 0xe12fff13 // bx r3
3960
3961HIDDEN _sk_colordodge_vfp4
3962.globl _sk_colordodge_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]3963FUNCTION(_sk_colordodge_vfp4)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]3964_sk_colordodge_vfp4:
3965 .long 0xed2d8b0e // vpush {d8-d14}
3966 .long 0xf2238d02 // vsub.f32 d8, d3, d2
3967 .long 0xe4913004 // ldr r3, [r1], #4
3968 .long 0xf3039d16 // vmul.f32 d9, d3, d6
3969 .long 0xf223ad01 // vsub.f32 d10, d3, d1
3970 .long 0xf303bd15 // vmul.f32 d11, d3, d5
3971 .long 0xf223cd00 // vsub.f32 d12, d3, d0
3972 .long 0xf303dd14 // vmul.f32 d13, d3, d4
3973 .long 0xeec9eaa8 // vdiv.f32 s29, s19, s17
3974 .long 0xee89ea08 // vdiv.f32 s28, s18, s16
3975 .long 0xeecb8aaa // vdiv.f32 s17, s23, s21
3976 .long 0xeecd9aac // vdiv.f32 s19, s27, s25
3977 .long 0xee8b8a0a // vdiv.f32 s16, s22, s20
3978 .long 0xee8d9a0c // vdiv.f32 s18, s26, s24
3979 .long 0xf2c70f10 // vmov.f32 d16, #1
3980 .long 0xf2672f0e // vmin.f32 d18, d7, d14
3981 .long 0xf2601d87 // vsub.f32 d17, d16, d7
3982 .long 0xf2673f08 // vmin.f32 d19, d7, d8
3983 .long 0xf2674f09 // vmin.f32 d20, d7, d9
3984 .long 0xf2600d83 // vsub.f32 d16, d16, d3
3985 .long 0xf3415d92 // vmul.f32 d21, d17, d2
3986 .long 0xf3422d93 // vmul.f32 d18, d18, d3
3987 .long 0xf3416d91 // vmul.f32 d22, d17, d1
3988 .long 0xf3433d93 // vmul.f32 d19, d19, d3
3989 .long 0xf3411d90 // vmul.f32 d17, d17, d0
3990 .long 0xf3444d93 // vmul.f32 d20, d20, d3
3991 .long 0xf3407d95 // vmul.f32 d23, d16, d5
3992 .long 0xf3408d94 // vmul.f32 d24, d16, d4
3993 .long 0xf3409d96 // vmul.f32 d25, d16, d6
3994 .long 0xf2452da2 // vadd.f32 d18, d21, d18
3995 .long 0xf2463da3 // vadd.f32 d19, d22, d19
3996 .long 0xf2414da4 // vadd.f32 d20, d17, d20
3997 .long 0xf241ae03 // vceq.f32 d26, d1, d3
3998 .long 0xf247bd81 // vadd.f32 d27, d23, d1
3999 .long 0xf3b91505 // vceq.f32 d1, d5, #0
4000 .long 0xf240ce03 // vceq.f32 d28, d0, d3
4001 .long 0xf248dd80 // vadd.f32 d29, d24, d0
4002 .long 0xf3b90504 // vceq.f32 d0, d4, #0
4003 .long 0xf242ee03 // vceq.f32 d30, d2, d3
4004 .long 0xf249fd82 // vadd.f32 d31, d25, d2
4005 .long 0xf3b92506 // vceq.f32 d2, d6, #0
4006 .long 0xf2073c30 // vfma.f32 d3, d7, d16
4007 .long 0xf2410d84 // vadd.f32 d16, d17, d4
4008 .long 0xf2491da2 // vadd.f32 d17, d25, d18
4009 .long 0xf2462d85 // vadd.f32 d18, d22, d5
4010 .long 0xf2455d86 // vadd.f32 d21, d21, d6
4011 .long 0xf2473da3 // vadd.f32 d19, d23, d19
4012 .long 0xf2484da4 // vadd.f32 d20, d24, d20
4013 .long 0xf35fe1b1 // vbsl d30, d31, d17
4014 .long 0xf35ba1b3 // vbsl d26, d27, d19
4015 .long 0xf35dc1b4 // vbsl d28, d29, d20
4016 .long 0xf31001bc // vbsl d0, d16, d28
4017 .long 0xf31211ba // vbsl d1, d18, d26
4018 .long 0xf31521be // vbsl d2, d21, d30
4019 .long 0xecbd8b0e // vpop {d8-d14}
4020 .long 0xe12fff13 // bx r3
4021
4022HIDDEN _sk_hardlight_vfp4
4023.globl _sk_hardlight_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4024FUNCTION(_sk_hardlight_vfp4)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]4025_sk_hardlight_vfp4:
4026 .long 0xf2c71f10 // vmov.f32 d17, #1
4027 .long 0xe4913004 // ldr r3, [r1], #4
4028 .long 0xf2670d04 // vsub.f32 d16, d7, d4
4029 .long 0xf2617d87 // vsub.f32 d23, d17, d7
4030 .long 0xf2611d83 // vsub.f32 d17, d17, d3
4031 .long 0xf2672d05 // vsub.f32 d18, d7, d5
4032 .long 0xf2674d06 // vsub.f32 d20, d7, d6
4033 .long 0xf2633d00 // vsub.f32 d19, d3, d0
4034 .long 0xf2635d01 // vsub.f32 d21, d3, d1
4035 .long 0xf2636d02 // vsub.f32 d22, d3, d2
4036 .long 0xf347bd90 // vmul.f32 d27, d23, d0
4037 .long 0xf341cd94 // vmul.f32 d28, d17, d4
4038 .long 0xf3430db0 // vmul.f32 d16, d19, d16
4039 .long 0xf3463db4 // vmul.f32 d19, d22, d20
4040 .long 0xf3452db2 // vmul.f32 d18, d21, d18
4041 .long 0xf2404d00 // vadd.f32 d20, d0, d0
4042 .long 0xf3405d14 // vmul.f32 d21, d0, d4
4043 .long 0xf2416d01 // vadd.f32 d22, d1, d1
4044 .long 0xf3418d15 // vmul.f32 d24, d1, d5
4045 .long 0xf2429d02 // vadd.f32 d25, d2, d2
4046 .long 0xf342ad16 // vmul.f32 d26, d2, d6
4047 .long 0xf347dd91 // vmul.f32 d29, d23, d1
4048 .long 0xf341fd95 // vmul.f32 d31, d17, d5
4049 .long 0xf24cbdab // vadd.f32 d27, d28, d27
4050 .long 0xf3477d92 // vmul.f32 d23, d23, d2
4051 .long 0xf341cd96 // vmul.f32 d28, d17, d6
4052 .long 0xf2400da0 // vadd.f32 d16, d16, d16
4053 .long 0xf343ed17 // vmul.f32 d30, d3, d7
4054 .long 0xf2422da2 // vadd.f32 d18, d18, d18
4055 .long 0xf2433da3 // vadd.f32 d19, d19, d19
4056 .long 0xf3434e24 // vcge.f32 d20, d3, d20
4057 .long 0xf2455da5 // vadd.f32 d21, d21, d21
4058 .long 0xf3436e26 // vcge.f32 d22, d3, d22
4059 .long 0xf3439e29 // vcge.f32 d25, d3, d25
4060 .long 0xf2488da8 // vadd.f32 d24, d24, d24
4061 .long 0xf24aadaa // vadd.f32 d26, d26, d26
4062 .long 0xf2073c31 // vfma.f32 d3, d7, d17
4063 .long 0xf24fddad // vadd.f32 d29, d31, d29
4064 .long 0xf24c1da7 // vadd.f32 d17, d28, d23
4065 .long 0xf26e0da0 // vsub.f32 d16, d30, d16
4066 .long 0xf26e2da2 // vsub.f32 d18, d30, d18
4067 .long 0xf26e3da3 // vsub.f32 d19, d30, d19
4068 .long 0xf35541b0 // vbsl d20, d21, d16
4069 .long 0xf35861b2 // vbsl d22, d24, d18
4070 .long 0xf35a91b3 // vbsl d25, d26, d19
4071 .long 0xf20b0da4 // vadd.f32 d0, d27, d20
4072 .long 0xf20d1da6 // vadd.f32 d1, d29, d22
4073 .long 0xf2012da9 // vadd.f32 d2, d17, d25
4074 .long 0xe12fff13 // bx r3
4075
4076HIDDEN _sk_overlay_vfp4
4077.globl _sk_overlay_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4078FUNCTION(_sk_overlay_vfp4)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]4079_sk_overlay_vfp4:
4080 .long 0xf2c71f10 // vmov.f32 d17, #1
4081 .long 0xe4913004 // ldr r3, [r1], #4
4082 .long 0xf2670d04 // vsub.f32 d16, d7, d4
4083 .long 0xf2617d87 // vsub.f32 d23, d17, d7
4084 .long 0xf2611d83 // vsub.f32 d17, d17, d3
4085 .long 0xf2672d05 // vsub.f32 d18, d7, d5
4086 .long 0xf2674d06 // vsub.f32 d20, d7, d6
4087 .long 0xf2633d00 // vsub.f32 d19, d3, d0
4088 .long 0xf2635d01 // vsub.f32 d21, d3, d1
4089 .long 0xf2636d02 // vsub.f32 d22, d3, d2
4090 .long 0xf347bd90 // vmul.f32 d27, d23, d0
4091 .long 0xf341cd94 // vmul.f32 d28, d17, d4
4092 .long 0xf3430db0 // vmul.f32 d16, d19, d16
4093 .long 0xf3463db4 // vmul.f32 d19, d22, d20
4094 .long 0xf3452db2 // vmul.f32 d18, d21, d18
4095 .long 0xf2444d04 // vadd.f32 d20, d4, d4
4096 .long 0xf3405d14 // vmul.f32 d21, d0, d4
4097 .long 0xf2456d05 // vadd.f32 d22, d5, d5
4098 .long 0xf3418d15 // vmul.f32 d24, d1, d5
4099 .long 0xf2469d06 // vadd.f32 d25, d6, d6
4100 .long 0xf342ad16 // vmul.f32 d26, d2, d6
4101 .long 0xf347dd91 // vmul.f32 d29, d23, d1
4102 .long 0xf341fd95 // vmul.f32 d31, d17, d5
4103 .long 0xf24cbdab // vadd.f32 d27, d28, d27
4104 .long 0xf3477d92 // vmul.f32 d23, d23, d2
4105 .long 0xf341cd96 // vmul.f32 d28, d17, d6
4106 .long 0xf343ed17 // vmul.f32 d30, d3, d7
4107 .long 0xf2400da0 // vadd.f32 d16, d16, d16
4108 .long 0xf2422da2 // vadd.f32 d18, d18, d18
4109 .long 0xf2433da3 // vadd.f32 d19, d19, d19
4110 .long 0xf3474e24 // vcge.f32 d20, d7, d20
4111 .long 0xf2455da5 // vadd.f32 d21, d21, d21
4112 .long 0xf3476e26 // vcge.f32 d22, d7, d22
4113 .long 0xf2488da8 // vadd.f32 d24, d24, d24
4114 .long 0xf3479e29 // vcge.f32 d25, d7, d25
4115 .long 0xf24aadaa // vadd.f32 d26, d26, d26
4116 .long 0xf2073c31 // vfma.f32 d3, d7, d17
4117 .long 0xf24fddad // vadd.f32 d29, d31, d29
4118 .long 0xf24c1da7 // vadd.f32 d17, d28, d23
4119 .long 0xf26e0da0 // vsub.f32 d16, d30, d16
4120 .long 0xf26e2da2 // vsub.f32 d18, d30, d18
4121 .long 0xf26e3da3 // vsub.f32 d19, d30, d19
4122 .long 0xf35541b0 // vbsl d20, d21, d16
4123 .long 0xf35861b2 // vbsl d22, d24, d18
4124 .long 0xf35a91b3 // vbsl d25, d26, d19
4125 .long 0xf20b0da4 // vadd.f32 d0, d27, d20
4126 .long 0xf20d1da6 // vadd.f32 d1, d29, d22
4127 .long 0xf2012da9 // vadd.f32 d2, d17, d25
4128 .long 0xe12fff13 // bx r3
4129
4130HIDDEN _sk_softlight_vfp4
4131.globl _sk_softlight_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4132FUNCTION(_sk_softlight_vfp4)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]4133_sk_softlight_vfp4:
4134 .long 0xed2d8b06 // vpush {d8-d10}
4135 .long 0xeec58aa7 // vdiv.f32 s17, s11, s15
4136 .long 0xf3f90407 // vcgt.f32 d16, d7, #0
4137 .long 0xe4913004 // ldr r3, [r1], #4
4138 .long 0xeec49aa7 // vdiv.f32 s19, s9, s15
4139 .long 0xeec6aaa7 // vdiv.f32 s21, s13, s15
4140 .long 0xee858a07 // vdiv.f32 s16, s10, s14
4141 .long 0xee849a07 // vdiv.f32 s18, s8, s14
4142 .long 0xee86aa07 // vdiv.f32 s20, s12, s14
4143 .long 0xf26021b0 // vorr d18, d16, d16
4144 .long 0xf2c01010 // vmov.i32 d17, #0
4145 .long 0xf3582131 // vbsl d18, d8, d17
4146 .long 0xf26031b0 // vorr d19, d16, d16
4147 .long 0xf3fb45a2 // vrsqrte.f32 d20, d18
4148 .long 0xf3593131 // vbsl d19, d9, d17
4149 .long 0xf35a0131 // vbsl d16, d10, d17
4150 .long 0xf3fb15a3 // vrsqrte.f32 d17, d19
4151 .long 0xf3fb55a0 // vrsqrte.f32 d21, d16
4152 .long 0xf3446db4 // vmul.f32 d22, d20, d20
4153 .long 0xf243ada3 // vadd.f32 d26, d19, d19
4154 .long 0xf240bda0 // vadd.f32 d27, d16, d16
4155 .long 0xf3417db1 // vmul.f32 d23, d17, d17
4156 .long 0xf3458db5 // vmul.f32 d24, d21, d21
4157 .long 0xf2626fb6 // vrsqrts.f32 d22, d18, d22
4158 .long 0xf2429da2 // vadd.f32 d25, d18, d18
4159 .long 0xf2637fb7 // vrsqrts.f32 d23, d19, d23
4160 .long 0xf2608fb8 // vrsqrts.f32 d24, d16, d24
4161 .long 0xf2818f1c // vmov.f32 d8, #7
4162 .long 0xf2499da9 // vadd.f32 d25, d25, d25
4163 .long 0xf3444db6 // vmul.f32 d20, d20, d22
4164 .long 0xf24a6daa // vadd.f32 d22, d26, d26
4165 .long 0xf24badab // vadd.f32 d26, d27, d27
4166 .long 0xf3411db7 // vmul.f32 d17, d17, d23
4167 .long 0xf3455db8 // vmul.f32 d21, d21, d24
4168 .long 0xf3fb7524 // vrecpe.f32 d23, d20
4169 .long 0xf3498db9 // vmul.f32 d24, d25, d25
4170 .long 0xf3fbd521 // vrecpe.f32 d29, d17
4171 .long 0xf34aedba // vmul.f32 d30, d26, d26
4172 .long 0xf3fbf525 // vrecpe.f32 d31, d21
4173 .long 0xf2444fb7 // vrecps.f32 d20, d20, d23
4174 .long 0xf346cdb6 // vmul.f32 d28, d22, d22
4175 .long 0xf2411fbd // vrecps.f32 d17, d17, d29
4176 .long 0xf3c7bf10 // vmov.f32 d27, #-1
4177 .long 0xf2455fbf // vrecps.f32 d21, d21, d31
4178 .long 0xf24aadae // vadd.f32 d26, d26, d30
4179 .long 0xf2498da8 // vadd.f32 d24, d25, d24
4180 .long 0xf2429dab // vadd.f32 d25, d18, d27
4181 .long 0xf2466dac // vadd.f32 d22, d22, d28
4182 .long 0xf243cdab // vadd.f32 d28, d19, d27
4183 .long 0xf240bdab // vadd.f32 d27, d16, d27
4184 .long 0xf3474db4 // vmul.f32 d20, d23, d20
4185 .long 0xf2c7ef10 // vmov.f32 d30, #1
4186 .long 0xf34d1db1 // vmul.f32 d17, d29, d17
4187 .long 0xf34badba // vmul.f32 d26, d27, d26
4188 .long 0xf242bd02 // vadd.f32 d27, d2, d2
4189 .long 0xf26edda0 // vsub.f32 d29, d30, d16
4190 .long 0xf3498db8 // vmul.f32 d24, d25, d24
4191 .long 0xf3429d98 // vmul.f32 d25, d18, d8
4192 .long 0xf34f5db5 // vmul.f32 d21, d31, d21
4193 .long 0xf26efda2 // vsub.f32 d31, d30, d18
4194 .long 0xf2642da2 // vsub.f32 d18, d20, d18
4195 .long 0xf26b4d83 // vsub.f32 d20, d27, d3
4196 .long 0xf2498da8 // vadd.f32 d24, d25, d24
4197 .long 0xf34c6db6 // vmul.f32 d22, d28, d22
4198 .long 0xf3437d98 // vmul.f32 d23, d19, d8
4199 .long 0xf3449dbd // vmul.f32 d25, d20, d29
4200 .long 0xf245dd05 // vadd.f32 d29, d5, d5
4201 .long 0xf340cd98 // vmul.f32 d28, d16, d8
4202 .long 0xf2476da6 // vadd.f32 d22, d23, d22
4203 .long 0xf2611da3 // vsub.f32 d17, d17, d19
4204 .long 0xf24dddad // vadd.f32 d29, d29, d29
4205 .long 0xf24c7daa // vadd.f32 d23, d28, d26
4206 .long 0xf2650da0 // vsub.f32 d16, d21, d16
4207 .long 0xf26e3da3 // vsub.f32 d19, d30, d19
4208 .long 0xf347de2d // vcge.f32 d29, d7, d29
4209 .long 0xf241ad01 // vadd.f32 d26, d1, d1
4210 .long 0xf3444d97 // vmul.f32 d20, d20, d7
4211 .long 0xf358d1b2 // vbsl d29, d24, d18
4212 .long 0xf2448d04 // vadd.f32 d24, d4, d4
4213 .long 0xf2462d06 // vadd.f32 d18, d6, d6
4214 .long 0xf26a5d83 // vsub.f32 d21, d26, d3
4215 .long 0xf2488da8 // vadd.f32 d24, d24, d24
4216 .long 0xf2422da2 // vadd.f32 d18, d18, d18
4217 .long 0xf345cdbf // vmul.f32 d28, d21, d31
4218 .long 0xf3455d97 // vmul.f32 d21, d21, d7
4219 .long 0xf3478e28 // vcge.f32 d24, d7, d24
4220 .long 0xf3472e22 // vcge.f32 d18, d7, d18
4221 .long 0xf343fd14 // vmul.f32 d31, d3, d4
4222 .long 0xf3455dbd // vmul.f32 d21, d21, d29
4223 .long 0xf35681b1 // vbsl d24, d22, d17
4224 .long 0xf2401d00 // vadd.f32 d17, d0, d0
4225 .long 0xf35721b0 // vbsl d18, d23, d16
4226 .long 0xf24c0d83 // vadd.f32 d16, d28, d3
4227 .long 0xf2496d83 // vadd.f32 d22, d25, d3
4228 .long 0xf2617d83 // vsub.f32 d23, d17, d3
4229 .long 0xf3442db2 // vmul.f32 d18, d20, d18
4230 .long 0xf3434e2a // vcge.f32 d20, d3, d26
4231 .long 0xf343ae2b // vcge.f32 d26, d3, d27
4232 .long 0xf3473db3 // vmul.f32 d19, d23, d19
4233 .long 0xf3477d97 // vmul.f32 d23, d23, d7
4234 .long 0xf3431e21 // vcge.f32 d17, d3, d17
4235 .long 0xf3400d95 // vmul.f32 d16, d16, d5
4236 .long 0xf2433d83 // vadd.f32 d19, d19, d3
4237 .long 0xf3477db8 // vmul.f32 d23, d23, d24
4238 .long 0xf26e8d87 // vsub.f32 d24, d30, d7
4239 .long 0xf26eed83 // vsub.f32 d30, d30, d3
4240 .long 0xf3433d94 // vmul.f32 d19, d19, d4
4241 .long 0xf24f7da7 // vadd.f32 d23, d31, d23
4242 .long 0xf3489d91 // vmul.f32 d25, d24, d1
4243 .long 0xf348cd90 // vmul.f32 d28, d24, d0
4244 .long 0xf34edd94 // vmul.f32 d29, d30, d4
4245 .long 0xf34ebd95 // vmul.f32 d27, d30, d5
4246 .long 0xf3488d92 // vmul.f32 d24, d24, d2
4247 .long 0xf34efd96 // vmul.f32 d31, d30, d6
4248 .long 0xf24dcdac // vadd.f32 d28, d29, d28
4249 .long 0xf343dd15 // vmul.f32 d29, d3, d5
4250 .long 0xf24b9da9 // vadd.f32 d25, d27, d25
4251 .long 0xf343bd16 // vmul.f32 d27, d3, d6
4252 .long 0xf3466d96 // vmul.f32 d22, d22, d6
4253 .long 0xf24f8da8 // vadd.f32 d24, d31, d24
4254 .long 0xf24d5da5 // vadd.f32 d21, d29, d21
4255 .long 0xf24b2da2 // vadd.f32 d18, d27, d18
4256 .long 0xf35311b7 // vbsl d17, d19, d23
4257 .long 0xf35041b5 // vbsl d20, d16, d21
4258 .long 0xf356a1b2 // vbsl d26, d22, d18
4259 .long 0xf2073c3e // vfma.f32 d3, d7, d30
4260 .long 0xf20c0da1 // vadd.f32 d0, d28, d17
4261 .long 0xf2091da4 // vadd.f32 d1, d25, d20
4262 .long 0xf2082daa // vadd.f32 d2, d24, d26
4263 .long 0xecbd8b06 // vpop {d8-d10}
4264 .long 0xe12fff13 // bx r3
4265
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4266HIDDEN _sk_clamp_0_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4267.globl _sk_clamp_0_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4268FUNCTION(_sk_clamp_0_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4269_sk_clamp_0_vfp4:
4270 .long 0xf2c00010 // vmov.i32 d16, #0
4271 .long 0xe4913004 // ldr r3, [r1], #4
4272 .long 0xf2000f20 // vmax.f32 d0, d0, d16
4273 .long 0xf2011f20 // vmax.f32 d1, d1, d16
4274 .long 0xf2022f20 // vmax.f32 d2, d2, d16
4275 .long 0xf2033f20 // vmax.f32 d3, d3, d16
4276 .long 0xe12fff13 // bx r3
4277
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4278HIDDEN _sk_clamp_1_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4279.globl _sk_clamp_1_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4280FUNCTION(_sk_clamp_1_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4281_sk_clamp_1_vfp4:
4282 .long 0xf2c70f10 // vmov.f32 d16, #1
4283 .long 0xe4913004 // ldr r3, [r1], #4
4284 .long 0xf2200f20 // vmin.f32 d0, d0, d16
4285 .long 0xf2211f20 // vmin.f32 d1, d1, d16
4286 .long 0xf2222f20 // vmin.f32 d2, d2, d16
4287 .long 0xf2233f20 // vmin.f32 d3, d3, d16
4288 .long 0xe12fff13 // bx r3
4289
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4290HIDDEN _sk_clamp_a_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4291.globl _sk_clamp_a_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4292FUNCTION(_sk_clamp_a_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4293_sk_clamp_a_vfp4:
4294 .long 0xf2c70f10 // vmov.f32 d16, #1
4295 .long 0xe4913004 // ldr r3, [r1], #4
4296 .long 0xf2233f20 // vmin.f32 d3, d3, d16
4297 .long 0xf2200f03 // vmin.f32 d0, d0, d3
4298 .long 0xf2211f03 // vmin.f32 d1, d1, d3
4299 .long 0xf2222f03 // vmin.f32 d2, d2, d3
4300 .long 0xe12fff13 // bx r3
4301
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4302HIDDEN _sk_set_rgb_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4303.globl _sk_set_rgb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4304FUNCTION(_sk_set_rgb_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4305_sk_set_rgb_vfp4:
4306 .long 0xe92d4800 // push {fp, lr}
4307 .long 0xe8911008 // ldm r1, {r3, ip}
4308 .long 0xe2811008 // add r1, r1, #8
4309 .long 0xe283e008 // add lr, r3, #8
4310 .long 0xf4a30c9d // vld1.32 {d0[]}, [r3 :32]!
4311 .long 0xf4ae2c9f // vld1.32 {d2[]}, [lr :32]
4312 .long 0xf4a31c9f // vld1.32 {d1[]}, [r3 :32]
4313 .long 0xe8bd4800 // pop {fp, lr}
4314 .long 0xe12fff1c // bx ip
4315
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4316HIDDEN _sk_swap_rb_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4317.globl _sk_swap_rb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4318FUNCTION(_sk_swap_rb_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4319_sk_swap_rb_vfp4:
4320 .long 0xeef00b40 // vmov.f64 d16, d0
4321 .long 0xe4913004 // ldr r3, [r1], #4
4322 .long 0xeeb00b42 // vmov.f64 d0, d2
4323 .long 0xeeb02b60 // vmov.f64 d2, d16
4324 .long 0xe12fff13 // bx r3
4325
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4326HIDDEN _sk_swap_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4327.globl _sk_swap_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4328FUNCTION(_sk_swap_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4329_sk_swap_vfp4:
4330 .long 0xeef00b43 // vmov.f64 d16, d3
4331 .long 0xe4913004 // ldr r3, [r1], #4
4332 .long 0xeef01b42 // vmov.f64 d17, d2
4333 .long 0xeef02b41 // vmov.f64 d18, d1
4334 .long 0xeef03b40 // vmov.f64 d19, d0
4335 .long 0xeeb00b44 // vmov.f64 d0, d4
4336 .long 0xeeb01b45 // vmov.f64 d1, d5
4337 .long 0xeeb02b46 // vmov.f64 d2, d6
4338 .long 0xeeb03b47 // vmov.f64 d3, d7
4339 .long 0xeeb04b63 // vmov.f64 d4, d19
4340 .long 0xeeb05b62 // vmov.f64 d5, d18
4341 .long 0xeeb06b61 // vmov.f64 d6, d17
4342 .long 0xeeb07b60 // vmov.f64 d7, d16
4343 .long 0xe12fff13 // bx r3
4344
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4345HIDDEN _sk_move_src_dst_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4346.globl _sk_move_src_dst_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4347FUNCTION(_sk_move_src_dst_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4348_sk_move_src_dst_vfp4:
4349 .long 0xeeb04b40 // vmov.f64 d4, d0
4350 .long 0xe4913004 // ldr r3, [r1], #4
4351 .long 0xeeb05b41 // vmov.f64 d5, d1
4352 .long 0xeeb06b42 // vmov.f64 d6, d2
4353 .long 0xeeb07b43 // vmov.f64 d7, d3
4354 .long 0xe12fff13 // bx r3
4355
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4356HIDDEN _sk_move_dst_src_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4357.globl _sk_move_dst_src_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4358FUNCTION(_sk_move_dst_src_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4359_sk_move_dst_src_vfp4:
4360 .long 0xeeb00b44 // vmov.f64 d0, d4
4361 .long 0xe4913004 // ldr r3, [r1], #4
4362 .long 0xeeb01b45 // vmov.f64 d1, d5
4363 .long 0xeeb02b46 // vmov.f64 d2, d6
4364 .long 0xeeb03b47 // vmov.f64 d3, d7
4365 .long 0xe12fff13 // bx r3
4366
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4367HIDDEN _sk_premul_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4368.globl _sk_premul_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4369FUNCTION(_sk_premul_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4370_sk_premul_vfp4:
4371 .long 0xf3000d13 // vmul.f32 d0, d0, d3
4372 .long 0xe4913004 // ldr r3, [r1], #4
4373 .long 0xf3011d13 // vmul.f32 d1, d1, d3
4374 .long 0xf3022d13 // vmul.f32 d2, d2, d3
4375 .long 0xe12fff13 // bx r3
4376
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4377HIDDEN _sk_unpremul_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4378.globl _sk_unpremul_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4379FUNCTION(_sk_unpremul_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4380_sk_unpremul_vfp4:
4381 .long 0xed2d8b04 // vpush {d8-d9}
4382 .long 0xeeb78a00 // vmov.f32 s16, #112
4383 .long 0xf3f91503 // vceq.f32 d17, d3, #0
4384 .long 0xf2c00010 // vmov.i32 d16, #0
4385 .long 0xe4913004 // ldr r3, [r1], #4
4386 .long 0xeec89a23 // vdiv.f32 s19, s16, s7
4387 .long 0xee889a03 // vdiv.f32 s18, s16, s6
4388 .long 0xf3501199 // vbsl d17, d16, d9
4389 .long 0xf3010d90 // vmul.f32 d0, d17, d0
4390 .long 0xf3011d91 // vmul.f32 d1, d17, d1
4391 .long 0xf3012d92 // vmul.f32 d2, d17, d2
4392 .long 0xecbd8b04 // vpop {d8-d9}
4393 .long 0xe12fff13 // bx r3
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]4394 .long 0xe320f000 // nop {0}
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4395
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4396HIDDEN _sk_from_srgb_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4397.globl _sk_from_srgb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4398FUNCTION(_sk_from_srgb_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4399_sk_from_srgb_vfp4:
4400 .long 0xeddf3b20 // vldr d19, [pc, #128]
4401 .long 0xf3408d10 // vmul.f32 d24, d0, d0
4402 .long 0xeddf0b1c // vldr d16, [pc, #112]
4403 .long 0xf26341b3 // vorr d20, d19, d19
4404 .long 0xf26351b3 // vorr d21, d19, d19
4405 .long 0xeddf9b1f // vldr d25, [pc, #124]
4406 .long 0xf2404c30 // vfma.f32 d20, d0, d16
4407 .long 0xeddf2b1b // vldr d18, [pc, #108]
4408 .long 0xf2415c30 // vfma.f32 d21, d1, d16
4409 .long 0xeddfcb1d // vldr d28, [pc, #116]
4410 .long 0xf2423c30 // vfma.f32 d19, d2, d16
4411 .long 0xe4913004 // ldr r3, [r1], #4
4412 .long 0xf3426d12 // vmul.f32 d22, d2, d2
4413 .long 0xf3417d11 // vmul.f32 d23, d1, d1
4414 .long 0xf3620e80 // vcgt.f32 d16, d18, d0
4415 .long 0xf3621e81 // vcgt.f32 d17, d18, d1
4416 .long 0xf341ad39 // vmul.f32 d26, d1, d25
4417 .long 0xf342bd39 // vmul.f32 d27, d2, d25
4418 .long 0xf3622e82 // vcgt.f32 d18, d18, d2
4419 .long 0xf3409d39 // vmul.f32 d25, d0, d25
4420 .long 0xf26cd1bc // vorr d29, d28, d28
4421 .long 0xf248dcb4 // vfma.f32 d29, d24, d20
4422 .long 0xf26c41bc // vorr d20, d28, d28
4423 .long 0xf2474cb5 // vfma.f32 d20, d23, d21
4424 .long 0xf246ccb3 // vfma.f32 d28, d22, d19
4425 .long 0xf35901bd // vbsl d16, d25, d29
4426 .long 0xf35a11b4 // vbsl d17, d26, d20
4427 .long 0xf35b21bc // vbsl d18, d27, d28
4428 .long 0xf22001b0 // vorr d0, d16, d16
4429 .long 0xf22111b1 // vorr d1, d17, d17
4430 .long 0xf22221b2 // vorr d2, d18, d18
4431 .long 0xe12fff13 // bx r3
4432 .long 0x3e99999a // .word 0x3e99999a
4433 .long 0x3e99999a // .word 0x3e99999a
4434 .long 0x3f328f5c // .word 0x3f328f5c
4435 .long 0x3f328f5c // .word 0x3f328f5c
4436 .long 0x3d6147ae // .word 0x3d6147ae
4437 .long 0x3d6147ae // .word 0x3d6147ae
4438 .long 0x3d9e8391 // .word 0x3d9e8391
4439 .long 0x3d9e8391 // .word 0x3d9e8391
4440 .long 0x3b23d70a // .word 0x3b23d70a
4441 .long 0x3b23d70a // .word 0x3b23d70a
4442
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4443HIDDEN _sk_to_srgb_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4444.globl _sk_to_srgb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4445FUNCTION(_sk_to_srgb_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4446_sk_to_srgb_vfp4:
4447 .long 0xf3fb0582 // vrsqrte.f32 d16, d2
4448 .long 0xe4913004 // ldr r3, [r1], #4
4449 .long 0xf3fb1581 // vrsqrte.f32 d17, d1
4450 .long 0xf3fb2580 // vrsqrte.f32 d18, d0
4451 .long 0xf3403db0 // vmul.f32 d19, d16, d16
4452 .long 0xf3414db1 // vmul.f32 d20, d17, d17
4453 .long 0xf3425db2 // vmul.f32 d21, d18, d18
4454 .long 0xf2623f33 // vrsqrts.f32 d19, d2, d19
4455 .long 0xf2614f34 // vrsqrts.f32 d20, d1, d20
4456 .long 0xf2605f35 // vrsqrts.f32 d21, d0, d21
4457 .long 0xf3400db3 // vmul.f32 d16, d16, d19
4458 .long 0xf3411db4 // vmul.f32 d17, d17, d20
4459 .long 0xf3422db5 // vmul.f32 d18, d18, d21
4460 .long 0xf3fb3520 // vrecpe.f32 d19, d16
4461 .long 0xf3fb4521 // vrecpe.f32 d20, d17
4462 .long 0xf3fb6522 // vrecpe.f32 d22, d18
4463 .long 0xf3fb55a0 // vrsqrte.f32 d21, d16
4464 .long 0xf3fb75a1 // vrsqrte.f32 d23, d17
4465 .long 0xf3fb85a2 // vrsqrte.f32 d24, d18
4466 .long 0xf2409fb3 // vrecps.f32 d25, d16, d19
4467 .long 0xf241afb4 // vrecps.f32 d26, d17, d20
4468 .long 0xf242bfb6 // vrecps.f32 d27, d18, d22
4469 .long 0xf345cdb5 // vmul.f32 d28, d21, d21
4470 .long 0xf347ddb7 // vmul.f32 d29, d23, d23
4471 .long 0xf348edb8 // vmul.f32 d30, d24, d24
4472 .long 0xf2600fbc // vrsqrts.f32 d16, d16, d28
4473 .long 0xf2611fbd // vrsqrts.f32 d17, d17, d29
4474 .long 0xf2622fbe // vrsqrts.f32 d18, d18, d30
4475 .long 0xf3433db9 // vmul.f32 d19, d19, d25
4476 .long 0xeddf9b21 // vldr d25, [pc, #132]
4477 .long 0xf3444dba // vmul.f32 d20, d20, d26
4478 .long 0xeddfab21 // vldr d26, [pc, #132]
4479 .long 0xf3466dbb // vmul.f32 d22, d22, d27
4480 .long 0xf26ab1ba // vorr d27, d26, d26
4481 .long 0xf243bcb9 // vfma.f32 d27, d19, d25
4482 .long 0xf26a31ba // vorr d19, d26, d26
4483 .long 0xf2443cb9 // vfma.f32 d19, d20, d25
4484 .long 0xeddf4b1d // vldr d20, [pc, #116]
4485 .long 0xf246acb9 // vfma.f32 d26, d22, d25
4486 .long 0xf3450db0 // vmul.f32 d16, d21, d16
4487 .long 0xeddf5b1c // vldr d21, [pc, #112]
4488 .long 0xf3471db1 // vmul.f32 d17, d23, d17
4489 .long 0xf3482db2 // vmul.f32 d18, d24, d18
4490 .long 0xf3406d35 // vmul.f32 d22, d0, d21
4491 .long 0xf240bcb4 // vfma.f32 d27, d16, d20
4492 .long 0xf2413cb4 // vfma.f32 d19, d17, d20
4493 .long 0xf242acb4 // vfma.f32 d26, d18, d20
4494 .long 0xeddf2b17 // vldr d18, [pc, #92]
4495 .long 0xf3417d35 // vmul.f32 d23, d1, d21
4496 .long 0xf3620e80 // vcgt.f32 d16, d18, d0
4497 .long 0xf3621e81 // vcgt.f32 d17, d18, d1
4498 .long 0xf3622e82 // vcgt.f32 d18, d18, d2
4499 .long 0xf3425d35 // vmul.f32 d21, d2, d21
4500 .long 0xf2c74f10 // vmov.f32 d20, #1
4501 .long 0xf2648faa // vmin.f32 d24, d20, d26
4502 .long 0xf2643fa3 // vmin.f32 d19, d20, d19
4503 .long 0xf2644fab // vmin.f32 d20, d20, d27
4504 .long 0xf35601b8 // vbsl d16, d22, d24
4505 .long 0xf35711b3 // vbsl d17, d23, d19
4506 .long 0xf35521b4 // vbsl d18, d21, d20
4507 .long 0xf22001b0 // vorr d0, d16, d16
4508 .long 0xf22111b1 // vorr d1, d17, d17
4509 .long 0xf22221b2 // vorr d2, d18, d18
4510 .long 0xe12fff13 // bx r3
4511 .long 0x3f306fce // .word 0x3f306fce
4512 .long 0x3f306fce // .word 0x3f306fce
4513 .long 0xbdca57a8 // .word 0xbdca57a8
4514 .long 0xbdca57a8 // .word 0xbdca57a8
4515 .long 0x3ed287c2 // .word 0x3ed287c2
4516 .long 0x3ed287c2 // .word 0x3ed287c2
4517 .long 0x41475c29 // .word 0x41475c29
4518 .long 0x41475c29 // .word 0x41475c29
4519 .long 0x3b8ce704 // .word 0x3b8ce704
4520 .long 0x3b8ce704 // .word 0x3b8ce704
4521
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]4522HIDDEN _sk_rgb_to_hsl_vfp4
4523.globl _sk_rgb_to_hsl_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4524FUNCTION(_sk_rgb_to_hsl_vfp4)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]4525_sk_rgb_to_hsl_vfp4:
4526 .long 0xed2d8b08 // vpush {d8-d11}
4527 .long 0xf2401f01 // vmax.f32 d17, d0, d1
4528 .long 0xeddf9b2c // vldr d25, [pc, #176]
4529 .long 0xf2600f01 // vmin.f32 d16, d0, d1
4530 .long 0xe4913004 // ldr r3, [r1], #4
4531 .long 0xeeb78a00 // vmov.f32 s16, #112
4532 .long 0xf2c3461f // vmov.i32 d20, #1056964608
4533 .long 0xf2411f82 // vmax.f32 d17, d17, d2
4534 .long 0xf2602f82 // vmin.f32 d18, d16, d2
4535 .long 0xf2c45610 // vmov.i32 d21, #1073741824
4536 .long 0xf2607d01 // vsub.f32 d23, d0, d1
4537 .long 0xf2656da1 // vsub.f32 d22, d21, d17
4538 .long 0xf221ada2 // vsub.f32 d10, d17, d18
4539 .long 0xf2413da2 // vadd.f32 d19, d17, d18
4540 .long 0xf2c08010 // vmov.i32 d24, #0
4541 .long 0xf2666da2 // vsub.f32 d22, d22, d18
4542 .long 0xf241ae80 // vceq.f32 d26, d17, d0
4543 .long 0xeec8ba2a // vdiv.f32 s23, s16, s21
4544 .long 0xf3430db4 // vmul.f32 d16, d19, d20
4545 .long 0xee88ba0a // vdiv.f32 s22, s16, s20
4546 .long 0xf3209ea4 // vcgt.f32 d9, d16, d20
4547 .long 0xf2614d02 // vsub.f32 d20, d1, d2
4548 .long 0xf3477d9b // vmul.f32 d23, d23, d11
4549 .long 0xf31691b3 // vbsl d9, d22, d19
4550 .long 0xf2623d00 // vsub.f32 d19, d2, d0
4551 .long 0xf3626e01 // vcgt.f32 d22, d2, d1
4552 .long 0xeeca8aa9 // vdiv.f32 s17, s21, s19
4553 .long 0xee8a8a09 // vdiv.f32 s16, s20, s18
4554 .long 0xf3433d9b // vmul.f32 d19, d19, d11
4555 .long 0xf3444d9b // vmul.f32 d20, d20, d11
4556 .long 0xf35961b8 // vbsl d22, d25, d24
4557 .long 0xf2419e81 // vceq.f32 d25, d17, d1
4558 .long 0xf2011ea2 // vceq.f32 d1, d17, d18
4559 .long 0xf2433da5 // vadd.f32 d19, d19, d21
4560 .long 0xf2c15f10 // vmov.f32 d21, #4
4561 .long 0xf2464da4 // vadd.f32 d20, d22, d20
4562 .long 0xf2471da5 // vadd.f32 d17, d23, d21
4563 .long 0xf35391b1 // vbsl d25, d19, d17
4564 .long 0xeddf1b0a // vldr d17, [pc, #40]
4565 .long 0xf2612111 // vorr d18, d1, d1
4566 .long 0xf354a1b9 // vbsl d26, d20, d25
4567 .long 0xf35821ba // vbsl d18, d24, d26
4568 .long 0xf3181198 // vbsl d1, d24, d8
4569 .long 0xf3020db1 // vmul.f32 d0, d18, d17
4570 .long 0xf22021b0 // vorr d2, d16, d16
4571 .long 0xecbd8b08 // vpop {d8-d11}
4572 .long 0xe12fff13 // bx r3
4573 .long 0xe320f000 // nop {0}
4574 .long 0x40c00000 // .word 0x40c00000
4575 .long 0x40c00000 // .word 0x40c00000
4576 .long 0x3e2aaaab // .word 0x3e2aaaab
4577 .long 0x3e2aaaab // .word 0x3e2aaaab
4578
4579HIDDEN _sk_hsl_to_rgb_vfp4
4580.globl _sk_hsl_to_rgb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4581FUNCTION(_sk_hsl_to_rgb_vfp4)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]4582_sk_hsl_to_rgb_vfp4:
4583 .long 0xf2c72f10 // vmov.f32 d18, #1
4584 .long 0xeddf0b4f // vldr d16, [pc, #316]
4585 .long 0xf2c3161f // vmov.i32 d17, #1056964608
4586 .long 0xeddf9b4f // vldr d25, [pc, #316]
4587 .long 0xf2415d22 // vadd.f32 d21, d1, d18
4588 .long 0xe4913004 // ldr r3, [r1], #4
4589 .long 0xf3414d12 // vmul.f32 d20, d1, d2
4590 .long 0xf2416d02 // vadd.f32 d22, d1, d2
4591 .long 0xf2407d20 // vadd.f32 d23, d0, d16
4592 .long 0xf3610e82 // vcgt.f32 d16, d17, d2
4593 .long 0xf3455d92 // vmul.f32 d21, d21, d2
4594 .long 0xf2664da4 // vsub.f32 d20, d22, d20
4595 .long 0xf2426d02 // vadd.f32 d22, d2, d2
4596 .long 0xf3c73f10 // vmov.f32 d19, #-1
4597 .long 0xf35501b4 // vbsl d16, d21, d20
4598 .long 0xf2409d29 // vadd.f32 d25, d0, d25
4599 .long 0xf2408d23 // vadd.f32 d24, d0, d19
4600 .long 0xf3f9e629 // vclt.f32 d30, d25, #0
4601 .long 0xf360ae22 // vcgt.f32 d26, d0, d18
4602 .long 0xf247cda3 // vadd.f32 d28, d23, d19
4603 .long 0xf367dea2 // vcgt.f32 d29, d23, d18
4604 .long 0xf240bd22 // vadd.f32 d27, d0, d18
4605 .long 0xf2666da0 // vsub.f32 d22, d22, d16
4606 .long 0xf2474da2 // vadd.f32 d20, d23, d18
4607 .long 0xf358a190 // vbsl d26, d24, d0
4608 .long 0xf3f98600 // vclt.f32 d24, d0, #0
4609 .long 0xf3695ea2 // vcgt.f32 d21, d25, d18
4610 .long 0xf2493da3 // vadd.f32 d19, d25, d19
4611 .long 0xf35b81ba // vbsl d24, d27, d26
4612 .long 0xf3f9a627 // vclt.f32 d26, d23, #0
4613 .long 0xf35cd1b7 // vbsl d29, d28, d23
4614 .long 0xeddfcb35 // vldr d28, [pc, #212]
4615 .long 0xf2492da2 // vadd.f32 d18, d25, d18
4616 .long 0xf260bda6 // vsub.f32 d27, d16, d22
4617 .long 0xf354a1bd // vbsl d26, d20, d29
4618 .long 0xf2c14f18 // vmov.f32 d20, #6
4619 .long 0xf35351b9 // vbsl d21, d19, d25
4620 .long 0xf26cddaa // vsub.f32 d29, d28, d26
4621 .long 0xf352e1b5 // vbsl d30, d18, d21
4622 .long 0xf34b2db4 // vmul.f32 d18, d27, d20
4623 .long 0xf26c3da8 // vsub.f32 d19, d28, d24
4624 .long 0xf26c4dae // vsub.f32 d20, d28, d30
4625 .long 0xf36cbeaa // vcgt.f32 d27, d28, d26
4626 .long 0xf3425dbd // vmul.f32 d21, d18, d29
4627 .long 0xf3477db2 // vmul.f32 d23, d23, d18
4628 .long 0xf3423db3 // vmul.f32 d19, d18, d19
4629 .long 0xf3444db2 // vmul.f32 d20, d20, d18
4630 .long 0xf2465da5 // vadd.f32 d21, d22, d21
4631 .long 0xf342dd90 // vmul.f32 d29, d18, d0
4632 .long 0xf3210eaa // vcgt.f32 d0, d17, d26
4633 .long 0xf3492db2 // vmul.f32 d18, d25, d18
4634 .long 0xf355b1b6 // vbsl d27, d21, d22
4635 .long 0xeddf5b22 // vldr d21, [pc, #136]
4636 .long 0xf36cfea8 // vcgt.f32 d31, d28, d24
4637 .long 0xf2463da3 // vadd.f32 d19, d22, d19
4638 .long 0xf36cceae // vcgt.f32 d28, d28, d30
4639 .long 0xf2464da4 // vadd.f32 d20, d22, d20
4640 .long 0xf365aeaa // vcgt.f32 d26, d21, d26
4641 .long 0xf2467da7 // vadd.f32 d23, d22, d23
4642 .long 0xf3619ea8 // vcgt.f32 d25, d17, d24
4643 .long 0xf3611eae // vcgt.f32 d17, d17, d30
4644 .long 0xf31001bb // vbsl d0, d16, d27
4645 .long 0xf353f1b6 // vbsl d31, d19, d22
4646 .long 0xf354c1b6 // vbsl d28, d20, d22
4647 .long 0xf357a190 // vbsl d26, d23, d0
4648 .long 0xf3b90501 // vceq.f32 d0, d1, #0
4649 .long 0xf3658ea8 // vcgt.f32 d24, d21, d24
4650 .long 0xf246ddad // vadd.f32 d29, d22, d29
4651 .long 0xf3653eae // vcgt.f32 d19, d21, d30
4652 .long 0xf2462da2 // vadd.f32 d18, d22, d18
4653 .long 0xf35091bf // vbsl d25, d16, d31
4654 .long 0xf35011bc // vbsl d17, d16, d28
4655 .long 0xf2600110 // vorr d16, d0, d0
4656 .long 0xf2201110 // vorr d1, d0, d0
4657 .long 0xf352013a // vbsl d16, d2, d26
4658 .long 0xf35d81b9 // vbsl d24, d29, d25
4659 .long 0xf35231b1 // vbsl d19, d18, d17
4660 .long 0xf3121138 // vbsl d1, d2, d24
4661 .long 0xf3120133 // vbsl d0, d2, d19
4662 .long 0xf22021b0 // vorr d2, d16, d16
4663 .long 0xe12fff13 // bx r3
4664 .long 0xe320f000 // nop {0}
4665 .long 0xbeaaaaab // .word 0xbeaaaaab
4666 .long 0xbeaaaaab // .word 0xbeaaaaab
4667 .long 0x3eaaaaab // .word 0x3eaaaaab
4668 .long 0x3eaaaaab // .word 0x3eaaaaab
4669 .long 0x3f2aaaab // .word 0x3f2aaaab
4670 .long 0x3f2aaaab // .word 0x3f2aaaab
4671 .long 0x3e2aaaab // .word 0x3e2aaaab
4672 .long 0x3e2aaaab // .word 0x3e2aaaab
Mike Klein7125ac62017-04-05 15:27:22 -0400[diff] [blame]4673
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4674HIDDEN _sk_scale_1_float_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4675.globl _sk_scale_1_float_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4676FUNCTION(_sk_scale_1_float_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4677_sk_scale_1_float_vfp4:
4678 .long 0xe8911008 // ldm r1, {r3, ip}
4679 .long 0xe2811008 // add r1, r1, #8
4680 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
4681 .long 0xf3000d90 // vmul.f32 d0, d16, d0
4682 .long 0xf3001d91 // vmul.f32 d1, d16, d1
4683 .long 0xf3002d92 // vmul.f32 d2, d16, d2
4684 .long 0xf3003d93 // vmul.f32 d3, d16, d3
4685 .long 0xe12fff1c // bx ip
4686
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4687HIDDEN _sk_scale_u8_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4688.globl _sk_scale_u8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4689FUNCTION(_sk_scale_u8_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4690_sk_scale_u8_vfp4:
4691 .long 0xe24dd004 // sub sp, sp, #4
4692 .long 0xe8911008 // ldm r1, {r3, ip}
4693 .long 0xe2811008 // add r1, r1, #8
4694 .long 0xe5933000 // ldr r3, [r3]
4695 .long 0xe0833000 // add r3, r3, r0
4696 .long 0xe1d330b0 // ldrh r3, [r3]
4697 .long 0xe1cd30b0 // strh r3, [sp]
4698 .long 0xe1a0300d // mov r3, sp
4699 .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16]
4700 .long 0xf3c80a30 // vmovl.u8 q8, d16
4701 .long 0xf3d00a30 // vmovl.u16 q8, d16
4702 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
4703 .long 0xeddf1b06 // vldr d17, [pc, #24]
4704 .long 0xf3400db1 // vmul.f32 d16, d16, d17
4705 .long 0xf3000d90 // vmul.f32 d0, d16, d0
4706 .long 0xf3001d91 // vmul.f32 d1, d16, d1
4707 .long 0xf3002d92 // vmul.f32 d2, d16, d2
4708 .long 0xf3003d93 // vmul.f32 d3, d16, d3
4709 .long 0xe28dd004 // add sp, sp, #4
4710 .long 0xe12fff1c // bx ip
4711 .long 0x3b808081 // .word 0x3b808081
4712 .long 0x3b808081 // .word 0x3b808081
4713
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4714HIDDEN _sk_lerp_1_float_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4715.globl _sk_lerp_1_float_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4716FUNCTION(_sk_lerp_1_float_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4717_sk_lerp_1_float_vfp4:
4718 .long 0xe8911008 // ldm r1, {r3, ip}
4719 .long 0xf2600d04 // vsub.f32 d16, d0, d4
4720 .long 0xf2611d05 // vsub.f32 d17, d1, d5
4721 .long 0xf2622d06 // vsub.f32 d18, d2, d6
4722 .long 0xe2811008 // add r1, r1, #8
4723 .long 0xf2633d07 // vsub.f32 d19, d3, d7
4724 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
4725 .long 0xf2240114 // vorr d0, d4, d4
4726 .long 0xf2251115 // vorr d1, d5, d5
4727 .long 0xf2262116 // vorr d2, d6, d6
4728 .long 0xf2273117 // vorr d3, d7, d7
4729 .long 0xf2000cb4 // vfma.f32 d0, d16, d20
4730 .long 0xf2011cb4 // vfma.f32 d1, d17, d20
4731 .long 0xf2022cb4 // vfma.f32 d2, d18, d20
4732 .long 0xf2033cb4 // vfma.f32 d3, d19, d20
4733 .long 0xe12fff1c // bx ip
4734
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4735HIDDEN _sk_lerp_u8_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4736.globl _sk_lerp_u8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4737FUNCTION(_sk_lerp_u8_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4738_sk_lerp_u8_vfp4:
4739 .long 0xe24dd004 // sub sp, sp, #4
4740 .long 0xe8911008 // ldm r1, {r3, ip}
4741 .long 0xf2602d04 // vsub.f32 d18, d0, d4
4742 .long 0xf2623d06 // vsub.f32 d19, d2, d6
4743 .long 0xf2634d07 // vsub.f32 d20, d3, d7
4744 .long 0xe2811008 // add r1, r1, #8
4745 .long 0xe5933000 // ldr r3, [r3]
4746 .long 0xf2240114 // vorr d0, d4, d4
4747 .long 0xf2262116 // vorr d2, d6, d6
4748 .long 0xe0833000 // add r3, r3, r0
4749 .long 0xf2273117 // vorr d3, d7, d7
4750 .long 0xe1d330b0 // ldrh r3, [r3]
4751 .long 0xe1cd30b0 // strh r3, [sp]
4752 .long 0xe1a0300d // mov r3, sp
4753 .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16]
4754 .long 0xf3c80a30 // vmovl.u8 q8, d16
4755 .long 0xf3d00a30 // vmovl.u16 q8, d16
4756 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
4757 .long 0xeddf1b08 // vldr d17, [pc, #32]
4758 .long 0xf3400db1 // vmul.f32 d16, d16, d17
4759 .long 0xf2611d05 // vsub.f32 d17, d1, d5
4760 .long 0xf2251115 // vorr d1, d5, d5
4761 .long 0xf2020cb0 // vfma.f32 d0, d18, d16
4762 .long 0xf2011cb0 // vfma.f32 d1, d17, d16
4763 .long 0xf2032cb0 // vfma.f32 d2, d19, d16
4764 .long 0xf2043cb0 // vfma.f32 d3, d20, d16
4765 .long 0xe28dd004 // add sp, sp, #4
4766 .long 0xe12fff1c // bx ip
4767 .long 0x3b808081 // .word 0x3b808081
4768 .long 0x3b808081 // .word 0x3b808081
4769
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4770HIDDEN _sk_lerp_565_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4771.globl _sk_lerp_565_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4772FUNCTION(_sk_lerp_565_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4773_sk_lerp_565_vfp4:
4774 .long 0xe24dd004 // sub sp, sp, #4
4775 .long 0xe8911008 // ldm r1, {r3, ip}
4776 .long 0xf3c72218 // vmov.i32 d18, #63488
4777 .long 0xf2c1101f // vmov.i32 d17, #31
4778 .long 0xf2603d04 // vsub.f32 d19, d0, d4
4779 .long 0xe2811008 // add r1, r1, #8
4780 .long 0xe5933000 // ldr r3, [r3]
4781 .long 0xf2616d05 // vsub.f32 d22, d1, d5
4782 .long 0xf2240114 // vorr d0, d4, d4
4783 .long 0xf2251115 // vorr d1, d5, d5
4784 .long 0xe7933080 // ldr r3, [r3, r0, lsl #1]
4785 .long 0xf2873f10 // vmov.f32 d3, #1
4786 .long 0xe58d3000 // str r3, [sp]
4787 .long 0xe1a0300d // mov r3, sp
4788 .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
4789 .long 0xe3a03e7e // mov r3, #2016
4790 .long 0xf3d04a30 // vmovl.u16 q10, d16
4791 .long 0xee803b90 // vdup.32 d16, r3
4792 .long 0xf24421b2 // vand d18, d20, d18
4793 .long 0xf24411b1 // vand d17, d20, d17
4794 .long 0xeddf5b12 // vldr d21, [pc, #72]
4795 .long 0xf24401b0 // vand d16, d20, d16
4796 .long 0xeddf4b0e // vldr d20, [pc, #56]
4797 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
4798 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
4799 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
4800 .long 0xf3422db4 // vmul.f32 d18, d18, d20
4801 .long 0xeddf4b0d // vldr d20, [pc, #52]
4802 .long 0xf3400db5 // vmul.f32 d16, d16, d21
4803 .long 0xf2625d06 // vsub.f32 d21, d2, d6
4804 .long 0xf3411db4 // vmul.f32 d17, d17, d20
4805 .long 0xf2262116 // vorr d2, d6, d6
4806 .long 0xf2030cb2 // vfma.f32 d0, d19, d18
4807 .long 0xf2061cb0 // vfma.f32 d1, d22, d16
4808 .long 0xf2052cb1 // vfma.f32 d2, d21, d17
4809 .long 0xe28dd004 // add sp, sp, #4
4810 .long 0xe12fff1c // bx ip
4811 .long 0xe320f000 // nop {0}
4812 .long 0x37842108 // .word 0x37842108
4813 .long 0x37842108 // .word 0x37842108
4814 .long 0x3a020821 // .word 0x3a020821
4815 .long 0x3a020821 // .word 0x3a020821
4816 .long 0x3d042108 // .word 0x3d042108
4817 .long 0x3d042108 // .word 0x3d042108
4818
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]4819HIDDEN _sk_load_tables_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4820.globl _sk_load_tables_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4821FUNCTION(_sk_load_tables_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]4822_sk_load_tables_vfp4:
4823 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
4824 .long 0xe8911008 // ldm r1, {r3, ip}
4825 .long 0xf3c7001f // vmov.i32 d16, #255
4826 .long 0xe2811008 // add r1, r1, #8
4827 .long 0xe593e000 // ldr lr, [r3]
4828 .long 0xe99300b0 // ldmib r3, {r4, r5, r7}
4829 .long 0xe08e3100 // add r3, lr, r0, lsl #2
4830 .long 0xedd31b00 // vldr d17, [r3]
4831 .long 0xf24121b0 // vand d18, d17, d16
4832 .long 0xf3f83031 // vshr.u32 d19, d17, #8
4833 .long 0xee323b90 // vmov.32 r3, d18[1]
4834 .long 0xee126b90 // vmov.32 r6, d18[0]
4835 .long 0xf3f02031 // vshr.u32 d18, d17, #16
4836 .long 0xf24221b0 // vand d18, d18, d16
4837 .long 0xf24301b0 // vand d16, d19, d16
4838 .long 0xe0843103 // add r3, r4, r3, lsl #2
4839 .long 0xedd30a00 // vldr s1, [r3]
4840 .long 0xe0843106 // add r3, r4, r6, lsl #2
4841 .long 0xee326b90 // vmov.32 r6, d18[1]
4842 .long 0xed930a00 // vldr s0, [r3]
4843 .long 0xee303b90 // vmov.32 r3, d16[1]
4844 .long 0xee104b90 // vmov.32 r4, d16[0]
4845 .long 0xf3e80031 // vshr.u32 d16, d17, #24
4846 .long 0xeddf1b0d // vldr d17, [pc, #52]
4847 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
4848 .long 0xf3003db1 // vmul.f32 d3, d16, d17
4849 .long 0xe087e106 // add lr, r7, r6, lsl #2
4850 .long 0xee126b90 // vmov.32 r6, d18[0]
4851 .long 0xe0853103 // add r3, r5, r3, lsl #2
4852 .long 0xedde2a00 // vldr s5, [lr]
4853 .long 0xedd31a00 // vldr s3, [r3]
4854 .long 0xe0853104 // add r3, r5, r4, lsl #2
4855 .long 0xed931a00 // vldr s2, [r3]
4856 .long 0xe0873106 // add r3, r7, r6, lsl #2
4857 .long 0xed932a00 // vldr s4, [r3]
4858 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
4859 .long 0xe12fff1c // bx ip
4860 .long 0xe320f000 // nop {0}
4861 .long 0x3b808081 // .word 0x3b808081
4862 .long 0x3b808081 // .word 0x3b808081
4863
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]4864HIDDEN _sk_load_tables_u16_be_vfp4
4865.globl _sk_load_tables_u16_be_vfp4
4866FUNCTION(_sk_load_tables_u16_be_vfp4)
4867_sk_load_tables_u16_be_vfp4:
4868 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
4869 .long 0xe8911008 // ldm r1, {r3, ip}
4870 .long 0xf3c7801f // vmov.i32 d24, #255
4871 .long 0xe2811008 // add r1, r1, #8
4872 .long 0xe5936000 // ldr r6, [r3]
4873 .long 0xe9934030 // ldmib r3, {r4, r5, lr}
4874 .long 0xe0863180 // add r3, r6, r0, lsl #3
4875 .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]!
4876 .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3]
4877 .long 0xee903bb0 // vmov.u16 r3, d16[0]
4878 .long 0xee926bb0 // vmov.u16 r6, d18[0]
4879 .long 0xee043b90 // vmov.32 d20[0], r3
4880 .long 0xee913bb0 // vmov.u16 r3, d17[0]
4881 .long 0xee056b90 // vmov.32 d21[0], r6
4882 .long 0xee906bf0 // vmov.u16 r6, d16[1]
4883 .long 0xee063b90 // vmov.32 d22[0], r3
4884 .long 0xee933bb0 // vmov.u16 r3, d19[0]
4885 .long 0xee246b90 // vmov.32 d20[1], r6
4886 .long 0xee926bf0 // vmov.u16 r6, d18[1]
4887 .long 0xf24441b8 // vand d20, d20, d24
4888 .long 0xee073b90 // vmov.32 d23[0], r3
4889 .long 0xee913bf0 // vmov.u16 r3, d17[1]
4890 .long 0xee256b90 // vmov.32 d21[1], r6
4891 .long 0xee936bf0 // vmov.u16 r6, d19[1]
4892 .long 0xf24501b8 // vand d16, d21, d24
4893 .long 0xee107b90 // vmov.32 r7, d16[0]
4894 .long 0xee263b90 // vmov.32 d22[1], r3
4895 .long 0xee343b90 // vmov.32 r3, d20[1]
4896 .long 0xf24611b8 // vand d17, d22, d24
4897 .long 0xee276b90 // vmov.32 d23[1], r6
4898 .long 0xee146b90 // vmov.32 r6, d20[0]
4899 .long 0xf2e82537 // vshl.s32 d18, d23, #8
4900 .long 0xe0843103 // add r3, r4, r3, lsl #2
4901 .long 0xedd30a00 // vldr s1, [r3]
4902 .long 0xe0843106 // add r3, r4, r6, lsl #2
4903 .long 0xee304b90 // vmov.32 r4, d16[1]
4904 .long 0xee116b90 // vmov.32 r6, d17[0]
4905 .long 0xf3c70c1f // vmov.i32 d16, #65535
4906 .long 0xed930a00 // vldr s0, [r3]
4907 .long 0xee313b90 // vmov.32 r3, d17[1]
4908 .long 0xf24711b0 // vand d17, d23, d16
4909 .long 0xf3f81031 // vshr.u32 d17, d17, #8
4910 .long 0xf26211b1 // vorr d17, d18, d17
4911 .long 0xf24101b0 // vand d16, d17, d16
4912 .long 0xeddf1b0c // vldr d17, [pc, #48]
4913 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
4914 .long 0xf3003db1 // vmul.f32 d3, d16, d17
4915 .long 0xe08e4104 // add r4, lr, r4, lsl #2
4916 .long 0xedd42a00 // vldr s5, [r4]
4917 .long 0xe0853103 // add r3, r5, r3, lsl #2
4918 .long 0xedd31a00 // vldr s3, [r3]
4919 .long 0xe0853106 // add r3, r5, r6, lsl #2
4920 .long 0xed931a00 // vldr s2, [r3]
4921 .long 0xe08e3107 // add r3, lr, r7, lsl #2
4922 .long 0xed932a00 // vldr s4, [r3]
4923 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
4924 .long 0xe12fff1c // bx ip
4925 .long 0xe320f000 // nop {0}
4926 .long 0x37800080 // .word 0x37800080
4927 .long 0x37800080 // .word 0x37800080
4928
4929HIDDEN _sk_load_tables_rgb_u16_be_vfp4
4930.globl _sk_load_tables_rgb_u16_be_vfp4
4931FUNCTION(_sk_load_tables_rgb_u16_be_vfp4)
4932_sk_load_tables_rgb_u16_be_vfp4:
4933 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
4934 .long 0xe8911008 // ldm r1, {r3, ip}
4935 .long 0xf3c7601f // vmov.i32 d22, #255
4936 .long 0xe2811008 // add r1, r1, #8
4937 .long 0xf2873f10 // vmov.f32 d3, #1
4938 .long 0xe5936000 // ldr r6, [r3]
4939 .long 0xe9934030 // ldmib r3, {r4, r5, lr}
4940 .long 0xe0803080 // add r3, r0, r0, lsl #1
4941 .long 0xe0863083 // add r3, r6, r3, lsl #1
4942 .long 0xf4e3060d // vld3.16 {d16[0],d17[0],d18[0]}, [r3]!
4943 .long 0xf4e3064f // vld3.16 {d16[1],d17[1],d18[1]}, [r3]
4944 .long 0xee903bb0 // vmov.u16 r3, d16[0]
4945 .long 0xee926bb0 // vmov.u16 r6, d18[0]
4946 .long 0xee043b90 // vmov.32 d20[0], r3
4947 .long 0xee903bf0 // vmov.u16 r3, d16[1]
4948 .long 0xee056b90 // vmov.32 d21[0], r6
4949 .long 0xee916bb0 // vmov.u16 r6, d17[0]
4950 .long 0xee243b90 // vmov.32 d20[1], r3
4951 .long 0xee923bf0 // vmov.u16 r3, d18[1]
4952 .long 0xf24441b6 // vand d20, d20, d22
4953 .long 0xee076b90 // vmov.32 d23[0], r6
4954 .long 0xee916bf0 // vmov.u16 r6, d17[1]
4955 .long 0xee253b90 // vmov.32 d21[1], r3
4956 .long 0xee343b90 // vmov.32 r3, d20[1]
4957 .long 0xf24501b6 // vand d16, d21, d22
4958 .long 0xee276b90 // vmov.32 d23[1], r6
4959 .long 0xee146b90 // vmov.32 r6, d20[0]
4960 .long 0xf24711b6 // vand d17, d23, d22
4961 .long 0xee107b90 // vmov.32 r7, d16[0]
4962 .long 0xe0843103 // add r3, r4, r3, lsl #2
4963 .long 0xedd30a00 // vldr s1, [r3]
4964 .long 0xe0843106 // add r3, r4, r6, lsl #2
4965 .long 0xee304b90 // vmov.32 r4, d16[1]
4966 .long 0xee116b90 // vmov.32 r6, d17[0]
4967 .long 0xed930a00 // vldr s0, [r3]
4968 .long 0xee313b90 // vmov.32 r3, d17[1]
4969 .long 0xe08e4104 // add r4, lr, r4, lsl #2
4970 .long 0xedd42a00 // vldr s5, [r4]
4971 .long 0xe0853103 // add r3, r5, r3, lsl #2
4972 .long 0xedd31a00 // vldr s3, [r3]
4973 .long 0xe0853106 // add r3, r5, r6, lsl #2
4974 .long 0xed931a00 // vldr s2, [r3]
4975 .long 0xe08e3107 // add r3, lr, r7, lsl #2
4976 .long 0xed932a00 // vldr s4, [r3]
4977 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
4978 .long 0xe12fff1c // bx ip
4979
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]4980HIDDEN _sk_byte_tables_vfp4
4981.globl _sk_byte_tables_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]4982FUNCTION(_sk_byte_tables_vfp4)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]4983_sk_byte_tables_vfp4:
4984 .long 0xe92d4bf0 // push {r4, r5, r6, r7, r8, r9, fp, lr}
4985 .long 0xeddf0b37 // vldr d16, [pc, #220]
4986 .long 0xf2c3261f // vmov.i32 d18, #1056964608
4987 .long 0xf2c3361f // vmov.i32 d19, #1056964608
4988 .long 0xe8911010 // ldm r1, {r4, ip}
4989 .long 0xf2422c30 // vfma.f32 d18, d2, d16
4990 .long 0xf2413c30 // vfma.f32 d19, d1, d16
4991 .long 0xe2811008 // add r1, r1, #8
4992 .long 0xf2c3461f // vmov.i32 d20, #1056964608
4993 .long 0xe89402e0 // ldm r4, {r5, r6, r7, r9}
4994 .long 0xf2404c30 // vfma.f32 d20, d0, d16
4995 .long 0xf2c3161f // vmov.i32 d17, #1056964608
4996 .long 0xf2431c30 // vfma.f32 d17, d3, d16
4997 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
4998 .long 0xf3fb37a3 // vcvt.u32.f32 d19, d19
4999 .long 0xf3fb47a4 // vcvt.u32.f32 d20, d20
5000 .long 0xee123b90 // vmov.32 r3, d18[0]
5001 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5002 .long 0xee13eb90 // vmov.32 lr, d19[0]
5003 .long 0xee144b90 // vmov.32 r4, d20[0]
5004 .long 0xe7d78003 // ldrb r8, [r7, r3]
5005 .long 0xe7d6300e // ldrb r3, [r6, lr]
5006 .long 0xee053b90 // vmov.32 d21[0], r3
5007 .long 0xe7d53004 // ldrb r3, [r5, r4]
5008 .long 0xee344b90 // vmov.32 r4, d20[1]
5009 .long 0xee013b90 // vmov.32 d17[0], r3
5010 .long 0xee103b90 // vmov.32 r3, d16[0]
5011 .long 0xee048b90 // vmov.32 d20[0], r8
5012 .long 0xe7d5e004 // ldrb lr, [r5, r4]
5013 .long 0xee334b90 // vmov.32 r4, d19[1]
5014 .long 0xee325b90 // vmov.32 r5, d18[1]
5015 .long 0xf3c7201f // vmov.i32 d18, #255
5016 .long 0xe7d93003 // ldrb r3, [r9, r3]
5017 .long 0xee21eb90 // vmov.32 d17[1], lr
5018 .long 0xf24111b2 // vand d17, d17, d18
5019 .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17
5020 .long 0xe7d64004 // ldrb r4, [r6, r4]
5021 .long 0xee306b90 // vmov.32 r6, d16[1]
5022 .long 0xee003b90 // vmov.32 d16[0], r3
5023 .long 0xee254b90 // vmov.32 d21[1], r4
5024 .long 0xe7d74005 // ldrb r4, [r7, r5]
5025 .long 0xf24531b2 // vand d19, d21, d18
5026 .long 0xee244b90 // vmov.32 d20[1], r4
5027 .long 0xf24441b2 // vand d20, d20, d18
5028 .long 0xf3fb46a4 // vcvt.f32.u32 d20, d20
5029 .long 0xe7d93006 // ldrb r3, [r9, r6]
5030 .long 0xee203b90 // vmov.32 d16[1], r3
5031 .long 0xf24001b2 // vand d16, d16, d18
5032 .long 0xf3fb26a3 // vcvt.f32.u32 d18, d19
5033 .long 0xeddf3b09 // vldr d19, [pc, #36]
5034 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5035 .long 0xf3010db3 // vmul.f32 d0, d17, d19
5036 .long 0xf3042db3 // vmul.f32 d2, d20, d19
5037 .long 0xf3021db3 // vmul.f32 d1, d18, d19
5038 .long 0xf3003db3 // vmul.f32 d3, d16, d19
5039 .long 0xe8bd4bf0 // pop {r4, r5, r6, r7, r8, r9, fp, lr}
5040 .long 0xe12fff1c // bx ip
5041 .long 0xe320f000 // nop {0}
5042 .long 0x437f0000 // .word 0x437f0000
5043 .long 0x437f0000 // .word 0x437f0000
5044 .long 0x3b808081 // .word 0x3b808081
5045 .long 0x3b808081 // .word 0x3b808081
5046
5047HIDDEN _sk_byte_tables_rgb_vfp4
5048.globl _sk_byte_tables_rgb_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5049FUNCTION(_sk_byte_tables_rgb_vfp4)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]5050_sk_byte_tables_rgb_vfp4:
5051 .long 0xe92d41f0 // push {r4, r5, r6, r7, r8, lr}
5052 .long 0xe8911008 // ldm r1, {r3, ip}
5053 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5054 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5055 .long 0xf3c7301f // vmov.i32 d19, #255
5056 .long 0xe2811008 // add r1, r1, #8
5057 .long 0xe593e000 // ldr lr, [r3]
5058 .long 0xe9930110 // ldmib r3, {r4, r8}
5059 .long 0xe593300c // ldr r3, [r3, #12]
5060 .long 0xe2433001 // sub r3, r3, #1
5061 .long 0xee803b90 // vdup.32 d16, r3
5062 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5063 .long 0xf2402c30 // vfma.f32 d18, d0, d16
5064 .long 0xf2411c30 // vfma.f32 d17, d1, d16
5065 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5066 .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
5067 .long 0xee123b90 // vmov.32 r3, d18[0]
5068 .long 0xee326b90 // vmov.32 r6, d18[1]
5069 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5070 .long 0xf2422c30 // vfma.f32 d18, d2, d16
5071 .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18
5072 .long 0xee107b90 // vmov.32 r7, d16[0]
5073 .long 0xee305b90 // vmov.32 r5, d16[1]
5074 .long 0xe7de3003 // ldrb r3, [lr, r3]
5075 .long 0xe7dee006 // ldrb lr, [lr, r6]
5076 .long 0xee116b90 // vmov.32 r6, d17[0]
5077 .long 0xee023b90 // vmov.32 d18[0], r3
5078 .long 0xee313b90 // vmov.32 r3, d17[1]
5079 .long 0xee22eb90 // vmov.32 d18[1], lr
5080 .long 0xf24221b3 // vand d18, d18, d19
5081 .long 0xf3fb26a2 // vcvt.f32.u32 d18, d18
5082 .long 0xe7d87007 // ldrb r7, [r8, r7]
5083 .long 0xee017b90 // vmov.32 d17[0], r7
5084 .long 0xe7d46006 // ldrb r6, [r4, r6]
5085 .long 0xe7d43003 // ldrb r3, [r4, r3]
5086 .long 0xee006b90 // vmov.32 d16[0], r6
5087 .long 0xe7d84005 // ldrb r4, [r8, r5]
5088 .long 0xee203b90 // vmov.32 d16[1], r3
5089 .long 0xee214b90 // vmov.32 d17[1], r4
5090 .long 0xf24001b3 // vand d16, d16, d19
5091 .long 0xf24111b3 // vand d17, d17, d19
5092 .long 0xeddf3b07 // vldr d19, [pc, #28]
5093 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5094 .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17
5095 .long 0xf3020db3 // vmul.f32 d0, d18, d19
5096 .long 0xf3001db3 // vmul.f32 d1, d16, d19
5097 .long 0xf3012db3 // vmul.f32 d2, d17, d19
5098 .long 0xe8bd41f0 // pop {r4, r5, r6, r7, r8, lr}
5099 .long 0xe12fff1c // bx ip
5100 .long 0xe320f000 // nop {0}
5101 .long 0x3b808081 // .word 0x3b808081
5102 .long 0x3b808081 // .word 0x3b808081
5103
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]5104HIDDEN _sk_table_r_vfp4
5105.globl _sk_table_r_vfp4
5106FUNCTION(_sk_table_r_vfp4)
5107_sk_table_r_vfp4:
5108 .long 0xe92d4010 // push {r4, lr}
5109 .long 0xe8911008 // ldm r1, {r3, ip}
5110 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5111 .long 0xe2811008 // add r1, r1, #8
5112 .long 0xe593e000 // ldr lr, [r3]
5113 .long 0xe5933004 // ldr r3, [r3, #4]
5114 .long 0xe2433001 // sub r3, r3, #1
5115 .long 0xee803b90 // vdup.32 d16, r3
5116 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5117 .long 0xf2401c30 // vfma.f32 d17, d0, d16
5118 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5119 .long 0xee303b90 // vmov.32 r3, d16[1]
5120 .long 0xee104b90 // vmov.32 r4, d16[0]
5121 .long 0xe08e3103 // add r3, lr, r3, lsl #2
5122 .long 0xedd30a00 // vldr s1, [r3]
5123 .long 0xe08e3104 // add r3, lr, r4, lsl #2
5124 .long 0xed930a00 // vldr s0, [r3]
5125 .long 0xe8bd4010 // pop {r4, lr}
5126 .long 0xe12fff1c // bx ip
5127
5128HIDDEN _sk_table_g_vfp4
5129.globl _sk_table_g_vfp4
5130FUNCTION(_sk_table_g_vfp4)
5131_sk_table_g_vfp4:
5132 .long 0xe92d4010 // push {r4, lr}
5133 .long 0xe8911008 // ldm r1, {r3, ip}
5134 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5135 .long 0xe2811008 // add r1, r1, #8
5136 .long 0xe593e000 // ldr lr, [r3]
5137 .long 0xe5933004 // ldr r3, [r3, #4]
5138 .long 0xe2433001 // sub r3, r3, #1
5139 .long 0xee803b90 // vdup.32 d16, r3
5140 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5141 .long 0xf2411c30 // vfma.f32 d17, d1, d16
5142 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5143 .long 0xee303b90 // vmov.32 r3, d16[1]
5144 .long 0xee104b90 // vmov.32 r4, d16[0]
5145 .long 0xe08e3103 // add r3, lr, r3, lsl #2
5146 .long 0xedd31a00 // vldr s3, [r3]
5147 .long 0xe08e3104 // add r3, lr, r4, lsl #2
5148 .long 0xed931a00 // vldr s2, [r3]
5149 .long 0xe8bd4010 // pop {r4, lr}
5150 .long 0xe12fff1c // bx ip
5151
5152HIDDEN _sk_table_b_vfp4
5153.globl _sk_table_b_vfp4
5154FUNCTION(_sk_table_b_vfp4)
5155_sk_table_b_vfp4:
5156 .long 0xe92d4010 // push {r4, lr}
5157 .long 0xe8911008 // ldm r1, {r3, ip}
5158 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5159 .long 0xe2811008 // add r1, r1, #8
5160 .long 0xe593e000 // ldr lr, [r3]
5161 .long 0xe5933004 // ldr r3, [r3, #4]
5162 .long 0xe2433001 // sub r3, r3, #1
5163 .long 0xee803b90 // vdup.32 d16, r3
5164 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5165 .long 0xf2421c30 // vfma.f32 d17, d2, d16
5166 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5167 .long 0xee303b90 // vmov.32 r3, d16[1]
5168 .long 0xee104b90 // vmov.32 r4, d16[0]
5169 .long 0xe08e3103 // add r3, lr, r3, lsl #2
5170 .long 0xedd32a00 // vldr s5, [r3]
5171 .long 0xe08e3104 // add r3, lr, r4, lsl #2
5172 .long 0xed932a00 // vldr s4, [r3]
5173 .long 0xe8bd4010 // pop {r4, lr}
5174 .long 0xe12fff1c // bx ip
5175
5176HIDDEN _sk_table_a_vfp4
5177.globl _sk_table_a_vfp4
5178FUNCTION(_sk_table_a_vfp4)
5179_sk_table_a_vfp4:
5180 .long 0xe92d4010 // push {r4, lr}
5181 .long 0xe8911008 // ldm r1, {r3, ip}
5182 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5183 .long 0xe2811008 // add r1, r1, #8
5184 .long 0xe593e000 // ldr lr, [r3]
5185 .long 0xe5933004 // ldr r3, [r3, #4]
5186 .long 0xe2433001 // sub r3, r3, #1
5187 .long 0xee803b90 // vdup.32 d16, r3
5188 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5189 .long 0xf2431c30 // vfma.f32 d17, d3, d16
5190 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5191 .long 0xee303b90 // vmov.32 r3, d16[1]
5192 .long 0xee104b90 // vmov.32 r4, d16[0]
5193 .long 0xe08e3103 // add r3, lr, r3, lsl #2
5194 .long 0xedd33a00 // vldr s7, [r3]
5195 .long 0xe08e3104 // add r3, lr, r4, lsl #2
5196 .long 0xed933a00 // vldr s6, [r3]
5197 .long 0xe8bd4010 // pop {r4, lr}
5198 .long 0xe12fff1c // bx ip
5199
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5200HIDDEN _sk_parametric_r_vfp4
5201.globl _sk_parametric_r_vfp4
5202FUNCTION(_sk_parametric_r_vfp4)
5203_sk_parametric_r_vfp4:
5204 .long 0xe92d4800 // push {fp, lr}
5205 .long 0xed2d8b06 // vpush {d8-d10}
5206 .long 0xe591e000 // ldr lr, [r1]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5207 .long 0xeddf3b41 // vldr d19, [pc, #260]
5208 .long 0xed9f8a4e // vldr s16, [pc, #312]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5209 .long 0xe1a0300e // mov r3, lr
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5210 .long 0xeddf4b42 // vldr d20, [pc, #264]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5211 .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]!
5212 .long 0xe591c004 // ldr ip, [r1, #4]
5213 .long 0xe2811008 // add r1, r1, #8
5214 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5215 .long 0xe28e3008 // add r3, lr, #8
5216 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5217 .long 0xe28e300c // add r3, lr, #12
5218 .long 0xf2412c90 // vfma.f32 d18, d17, d0
5219 .long 0xf2c71d1f // vmov.i32 d17, #8388607
5220 .long 0xf24211b1 // vand d17, d18, d17
5221 .long 0xf2c3171f // vorr.i32 d17, #1056964608
5222 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5223 .long 0xf2019da3 // vadd.f32 d9, d17, d19
5224 .long 0xf2c33614 // vmov.i32 d19, #872415232
5225 .long 0xf3422db3 // vmul.f32 d18, d18, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5226 .long 0xeddf3b30 // vldr d19, [pc, #192]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5227 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5228 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5229 .long 0xf3411db3 // vmul.f32 d17, d17, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5230 .long 0xed9f8a39 // vldr s16, [pc, #228]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5231 .long 0xf2422da4 // vadd.f32 d18, d18, d20
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5232 .long 0xeddf4b2e // vldr d20, [pc, #184]
5233 .long 0xf2c03010 // vmov.i32 d19, #0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5234 .long 0xf2621da1 // vsub.f32 d17, d18, d17
5235 .long 0xf2611d8a // vsub.f32 d17, d17, d10
5236 .long 0xf3400db1 // vmul.f32 d16, d16, d17
5237 .long 0xf3fb1720 // vcvt.s32.f32 d17, d16
5238 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5239 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16
5240 .long 0xf35421b3 // vbsl d18, d20, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5241 .long 0xeddf4b2b // vldr d20, [pc, #172]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5242 .long 0xf2611da2 // vsub.f32 d17, d17, d18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5243 .long 0xeddf2b25 // vldr d18, [pc, #148]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5244 .long 0xf2601da1 // vsub.f32 d17, d16, d17
5245 .long 0xf2400da4 // vadd.f32 d16, d16, d20
5246 .long 0xf2229da1 // vsub.f32 d9, d18, d17
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5247 .long 0xeddf2b23 // vldr d18, [pc, #140]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5248 .long 0xf3411db2 // vmul.f32 d17, d17, d18
5249 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5250 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5251 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
5252 .long 0xf2600da1 // vsub.f32 d16, d16, d17
5253 .long 0xf2c4161b // vmov.i32 d17, #1258291200
5254 .long 0xf2400d8a // vadd.f32 d16, d16, d10
5255 .long 0xf2402cb1 // vfma.f32 d18, d16, d17
5256 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5257 .long 0xe28e3018 // add r3, lr, #24
5258 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5259 .long 0xe28e3010 // add r3, lr, #16
5260 .long 0xf2401c90 // vfma.f32 d17, d16, d0
5261 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5262 .long 0xe28e3014 // add r3, lr, #20
5263 .long 0xf3400e80 // vcge.f32 d16, d16, d0
5264 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
5265 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5266 .long 0xf2442da2 // vadd.f32 d18, d20, d18
5267 .long 0xf35101b2 // vbsl d16, d17, d18
5268 .long 0xf2c71f10 // vmov.f32 d17, #1
5269 .long 0xf2400fa3 // vmax.f32 d16, d16, d19
5270 .long 0xf2200fa1 // vmin.f32 d0, d16, d17
5271 .long 0xecbd8b06 // vpop {d8-d10}
5272 .long 0xe8bd4800 // pop {fp, lr}
5273 .long 0xe12fff1c // bx ip
5274 .long 0x3eb444f9 // .word 0x3eb444f9
5275 .long 0x3eb444f9 // .word 0x3eb444f9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5276 .long 0x3fbfbf75 // .word 0x3fbfbf75
5277 .long 0x3fbfbf75 // .word 0x3fbfbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5278 .long 0xc2f87377 // .word 0xc2f87377
5279 .long 0xc2f87377 // .word 0xc2f87377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5280 .long 0x3f800000 // .word 0x3f800000
5281 .long 0x3f800000 // .word 0x3f800000
5282 .long 0x409af5f8 // .word 0x409af5f8
5283 .long 0x409af5f8 // .word 0x409af5f8
5284 .long 0x3fbebc8d // .word 0x3fbebc8d
5285 .long 0x3fbebc8d // .word 0x3fbebc8d
5286 .long 0x42f28c51 // .word 0x42f28c51
5287 .long 0x42f28c51 // .word 0x42f28c51
5288 .long 0x3fdce9a3 // .word 0x3fdce9a3
5289 .long 0x41ddd2fe // .word 0x41ddd2fe
5290
5291HIDDEN _sk_parametric_g_vfp4
5292.globl _sk_parametric_g_vfp4
5293FUNCTION(_sk_parametric_g_vfp4)
5294_sk_parametric_g_vfp4:
5295 .long 0xe92d4800 // push {fp, lr}
5296 .long 0xed2d8b06 // vpush {d8-d10}
5297 .long 0xe591e000 // ldr lr, [r1]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5298 .long 0xeddf3b41 // vldr d19, [pc, #260]
5299 .long 0xed9f8a4e // vldr s16, [pc, #312]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5300 .long 0xe1a0300e // mov r3, lr
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5301 .long 0xeddf4b42 // vldr d20, [pc, #264]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5302 .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]!
5303 .long 0xe591c004 // ldr ip, [r1, #4]
5304 .long 0xe2811008 // add r1, r1, #8
5305 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5306 .long 0xe28e3008 // add r3, lr, #8
5307 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5308 .long 0xe28e300c // add r3, lr, #12
5309 .long 0xf2412c91 // vfma.f32 d18, d17, d1
5310 .long 0xf2c71d1f // vmov.i32 d17, #8388607
5311 .long 0xf24211b1 // vand d17, d18, d17
5312 .long 0xf2c3171f // vorr.i32 d17, #1056964608
5313 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5314 .long 0xf2019da3 // vadd.f32 d9, d17, d19
5315 .long 0xf2c33614 // vmov.i32 d19, #872415232
5316 .long 0xf3422db3 // vmul.f32 d18, d18, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5317 .long 0xeddf3b30 // vldr d19, [pc, #192]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5318 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5319 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5320 .long 0xf3411db3 // vmul.f32 d17, d17, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5321 .long 0xed9f8a39 // vldr s16, [pc, #228]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5322 .long 0xf2422da4 // vadd.f32 d18, d18, d20
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5323 .long 0xeddf4b2e // vldr d20, [pc, #184]
5324 .long 0xf2c03010 // vmov.i32 d19, #0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5325 .long 0xf2621da1 // vsub.f32 d17, d18, d17
5326 .long 0xf2611d8a // vsub.f32 d17, d17, d10
5327 .long 0xf3400db1 // vmul.f32 d16, d16, d17
5328 .long 0xf3fb1720 // vcvt.s32.f32 d17, d16
5329 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5330 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16
5331 .long 0xf35421b3 // vbsl d18, d20, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5332 .long 0xeddf4b2b // vldr d20, [pc, #172]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5333 .long 0xf2611da2 // vsub.f32 d17, d17, d18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5334 .long 0xeddf2b25 // vldr d18, [pc, #148]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5335 .long 0xf2601da1 // vsub.f32 d17, d16, d17
5336 .long 0xf2400da4 // vadd.f32 d16, d16, d20
5337 .long 0xf2229da1 // vsub.f32 d9, d18, d17
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5338 .long 0xeddf2b23 // vldr d18, [pc, #140]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5339 .long 0xf3411db2 // vmul.f32 d17, d17, d18
5340 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5341 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5342 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
5343 .long 0xf2600da1 // vsub.f32 d16, d16, d17
5344 .long 0xf2c4161b // vmov.i32 d17, #1258291200
5345 .long 0xf2400d8a // vadd.f32 d16, d16, d10
5346 .long 0xf2402cb1 // vfma.f32 d18, d16, d17
5347 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5348 .long 0xe28e3018 // add r3, lr, #24
5349 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5350 .long 0xe28e3010 // add r3, lr, #16
5351 .long 0xf2401c91 // vfma.f32 d17, d16, d1
5352 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5353 .long 0xe28e3014 // add r3, lr, #20
5354 .long 0xf3400e81 // vcge.f32 d16, d16, d1
5355 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
5356 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5357 .long 0xf2442da2 // vadd.f32 d18, d20, d18
5358 .long 0xf35101b2 // vbsl d16, d17, d18
5359 .long 0xf2c71f10 // vmov.f32 d17, #1
5360 .long 0xf2400fa3 // vmax.f32 d16, d16, d19
5361 .long 0xf2201fa1 // vmin.f32 d1, d16, d17
5362 .long 0xecbd8b06 // vpop {d8-d10}
5363 .long 0xe8bd4800 // pop {fp, lr}
5364 .long 0xe12fff1c // bx ip
5365 .long 0x3eb444f9 // .word 0x3eb444f9
5366 .long 0x3eb444f9 // .word 0x3eb444f9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5367 .long 0x3fbfbf75 // .word 0x3fbfbf75
5368 .long 0x3fbfbf75 // .word 0x3fbfbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5369 .long 0xc2f87377 // .word 0xc2f87377
5370 .long 0xc2f87377 // .word 0xc2f87377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5371 .long 0x3f800000 // .word 0x3f800000
5372 .long 0x3f800000 // .word 0x3f800000
5373 .long 0x409af5f8 // .word 0x409af5f8
5374 .long 0x409af5f8 // .word 0x409af5f8
5375 .long 0x3fbebc8d // .word 0x3fbebc8d
5376 .long 0x3fbebc8d // .word 0x3fbebc8d
5377 .long 0x42f28c51 // .word 0x42f28c51
5378 .long 0x42f28c51 // .word 0x42f28c51
5379 .long 0x3fdce9a3 // .word 0x3fdce9a3
5380 .long 0x41ddd2fe // .word 0x41ddd2fe
5381
5382HIDDEN _sk_parametric_b_vfp4
5383.globl _sk_parametric_b_vfp4
5384FUNCTION(_sk_parametric_b_vfp4)
5385_sk_parametric_b_vfp4:
5386 .long 0xe92d4800 // push {fp, lr}
5387 .long 0xed2d8b06 // vpush {d8-d10}
5388 .long 0xe591e000 // ldr lr, [r1]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5389 .long 0xeddf3b41 // vldr d19, [pc, #260]
5390 .long 0xed9f8a4e // vldr s16, [pc, #312]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5391 .long 0xe1a0300e // mov r3, lr
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5392 .long 0xeddf4b42 // vldr d20, [pc, #264]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5393 .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]!
5394 .long 0xe591c004 // ldr ip, [r1, #4]
5395 .long 0xe2811008 // add r1, r1, #8
5396 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5397 .long 0xe28e3008 // add r3, lr, #8
5398 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5399 .long 0xe28e300c // add r3, lr, #12
5400 .long 0xf2412c92 // vfma.f32 d18, d17, d2
5401 .long 0xf2c71d1f // vmov.i32 d17, #8388607
5402 .long 0xf24211b1 // vand d17, d18, d17
5403 .long 0xf2c3171f // vorr.i32 d17, #1056964608
5404 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5405 .long 0xf2019da3 // vadd.f32 d9, d17, d19
5406 .long 0xf2c33614 // vmov.i32 d19, #872415232
5407 .long 0xf3422db3 // vmul.f32 d18, d18, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5408 .long 0xeddf3b30 // vldr d19, [pc, #192]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5409 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5410 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5411 .long 0xf3411db3 // vmul.f32 d17, d17, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5412 .long 0xed9f8a39 // vldr s16, [pc, #228]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5413 .long 0xf2422da4 // vadd.f32 d18, d18, d20
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5414 .long 0xeddf4b2e // vldr d20, [pc, #184]
5415 .long 0xf2c03010 // vmov.i32 d19, #0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5416 .long 0xf2621da1 // vsub.f32 d17, d18, d17
5417 .long 0xf2611d8a // vsub.f32 d17, d17, d10
5418 .long 0xf3400db1 // vmul.f32 d16, d16, d17
5419 .long 0xf3fb1720 // vcvt.s32.f32 d17, d16
5420 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5421 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16
5422 .long 0xf35421b3 // vbsl d18, d20, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5423 .long 0xeddf4b2b // vldr d20, [pc, #172]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5424 .long 0xf2611da2 // vsub.f32 d17, d17, d18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5425 .long 0xeddf2b25 // vldr d18, [pc, #148]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5426 .long 0xf2601da1 // vsub.f32 d17, d16, d17
5427 .long 0xf2400da4 // vadd.f32 d16, d16, d20
5428 .long 0xf2229da1 // vsub.f32 d9, d18, d17
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5429 .long 0xeddf2b23 // vldr d18, [pc, #140]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5430 .long 0xf3411db2 // vmul.f32 d17, d17, d18
5431 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5432 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5433 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
5434 .long 0xf2600da1 // vsub.f32 d16, d16, d17
5435 .long 0xf2c4161b // vmov.i32 d17, #1258291200
5436 .long 0xf2400d8a // vadd.f32 d16, d16, d10
5437 .long 0xf2402cb1 // vfma.f32 d18, d16, d17
5438 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5439 .long 0xe28e3018 // add r3, lr, #24
5440 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5441 .long 0xe28e3010 // add r3, lr, #16
5442 .long 0xf2401c92 // vfma.f32 d17, d16, d2
5443 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5444 .long 0xe28e3014 // add r3, lr, #20
5445 .long 0xf3400e82 // vcge.f32 d16, d16, d2
5446 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
5447 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5448 .long 0xf2442da2 // vadd.f32 d18, d20, d18
5449 .long 0xf35101b2 // vbsl d16, d17, d18
5450 .long 0xf2c71f10 // vmov.f32 d17, #1
5451 .long 0xf2400fa3 // vmax.f32 d16, d16, d19
5452 .long 0xf2202fa1 // vmin.f32 d2, d16, d17
5453 .long 0xecbd8b06 // vpop {d8-d10}
5454 .long 0xe8bd4800 // pop {fp, lr}
5455 .long 0xe12fff1c // bx ip
5456 .long 0x3eb444f9 // .word 0x3eb444f9
5457 .long 0x3eb444f9 // .word 0x3eb444f9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5458 .long 0x3fbfbf75 // .word 0x3fbfbf75
5459 .long 0x3fbfbf75 // .word 0x3fbfbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5460 .long 0xc2f87377 // .word 0xc2f87377
5461 .long 0xc2f87377 // .word 0xc2f87377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5462 .long 0x3f800000 // .word 0x3f800000
5463 .long 0x3f800000 // .word 0x3f800000
5464 .long 0x409af5f8 // .word 0x409af5f8
5465 .long 0x409af5f8 // .word 0x409af5f8
5466 .long 0x3fbebc8d // .word 0x3fbebc8d
5467 .long 0x3fbebc8d // .word 0x3fbebc8d
5468 .long 0x42f28c51 // .word 0x42f28c51
5469 .long 0x42f28c51 // .word 0x42f28c51
5470 .long 0x3fdce9a3 // .word 0x3fdce9a3
5471 .long 0x41ddd2fe // .word 0x41ddd2fe
5472
5473HIDDEN _sk_parametric_a_vfp4
5474.globl _sk_parametric_a_vfp4
5475FUNCTION(_sk_parametric_a_vfp4)
5476_sk_parametric_a_vfp4:
5477 .long 0xe92d4800 // push {fp, lr}
5478 .long 0xed2d8b06 // vpush {d8-d10}
5479 .long 0xe591e000 // ldr lr, [r1]
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5480 .long 0xeddf3b41 // vldr d19, [pc, #260]
5481 .long 0xed9f8a4e // vldr s16, [pc, #312]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5482 .long 0xe1a0300e // mov r3, lr
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5483 .long 0xeddf4b42 // vldr d20, [pc, #264]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5484 .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]!
5485 .long 0xe591c004 // ldr ip, [r1, #4]
5486 .long 0xe2811008 // add r1, r1, #8
5487 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5488 .long 0xe28e3008 // add r3, lr, #8
5489 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5490 .long 0xe28e300c // add r3, lr, #12
5491 .long 0xf2412c93 // vfma.f32 d18, d17, d3
5492 .long 0xf2c71d1f // vmov.i32 d17, #8388607
5493 .long 0xf24211b1 // vand d17, d18, d17
5494 .long 0xf2c3171f // vorr.i32 d17, #1056964608
5495 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5496 .long 0xf2019da3 // vadd.f32 d9, d17, d19
5497 .long 0xf2c33614 // vmov.i32 d19, #872415232
5498 .long 0xf3422db3 // vmul.f32 d18, d18, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5499 .long 0xeddf3b30 // vldr d19, [pc, #192]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5500 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5501 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5502 .long 0xf3411db3 // vmul.f32 d17, d17, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5503 .long 0xed9f8a39 // vldr s16, [pc, #228]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5504 .long 0xf2422da4 // vadd.f32 d18, d18, d20
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5505 .long 0xeddf4b2e // vldr d20, [pc, #184]
5506 .long 0xf2c03010 // vmov.i32 d19, #0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5507 .long 0xf2621da1 // vsub.f32 d17, d18, d17
5508 .long 0xf2611d8a // vsub.f32 d17, d17, d10
5509 .long 0xf3400db1 // vmul.f32 d16, d16, d17
5510 .long 0xf3fb1720 // vcvt.s32.f32 d17, d16
5511 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5512 .long 0xf3612ea0 // vcgt.f32 d18, d17, d16
5513 .long 0xf35421b3 // vbsl d18, d20, d19
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5514 .long 0xeddf4b2b // vldr d20, [pc, #172]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5515 .long 0xf2611da2 // vsub.f32 d17, d17, d18
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5516 .long 0xeddf2b25 // vldr d18, [pc, #148]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5517 .long 0xf2601da1 // vsub.f32 d17, d16, d17
5518 .long 0xf2400da4 // vadd.f32 d16, d16, d20
5519 .long 0xf2229da1 // vsub.f32 d9, d18, d17
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5520 .long 0xeddf2b23 // vldr d18, [pc, #140]
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5521 .long 0xf3411db2 // vmul.f32 d17, d17, d18
5522 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5523 .long 0xeec8aa29 // vdiv.f32 s21, s16, s19
5524 .long 0xee88aa09 // vdiv.f32 s20, s16, s18
5525 .long 0xf2600da1 // vsub.f32 d16, d16, d17
5526 .long 0xf2c4161b // vmov.i32 d17, #1258291200
5527 .long 0xf2400d8a // vadd.f32 d16, d16, d10
5528 .long 0xf2402cb1 // vfma.f32 d18, d16, d17
5529 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5530 .long 0xe28e3018 // add r3, lr, #24
5531 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
5532 .long 0xe28e3010 // add r3, lr, #16
5533 .long 0xf2401c93 // vfma.f32 d17, d16, d3
5534 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
5535 .long 0xe28e3014 // add r3, lr, #20
5536 .long 0xf3400e83 // vcge.f32 d16, d16, d3
5537 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
5538 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5539 .long 0xf2442da2 // vadd.f32 d18, d20, d18
5540 .long 0xf35101b2 // vbsl d16, d17, d18
5541 .long 0xf2c71f10 // vmov.f32 d17, #1
5542 .long 0xf2400fa3 // vmax.f32 d16, d16, d19
5543 .long 0xf2203fa1 // vmin.f32 d3, d16, d17
5544 .long 0xecbd8b06 // vpop {d8-d10}
5545 .long 0xe8bd4800 // pop {fp, lr}
5546 .long 0xe12fff1c // bx ip
5547 .long 0x3eb444f9 // .word 0x3eb444f9
5548 .long 0x3eb444f9 // .word 0x3eb444f9
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5549 .long 0x3fbfbf75 // .word 0x3fbfbf75
5550 .long 0x3fbfbf75 // .word 0x3fbfbf75
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]5551 .long 0xc2f87377 // .word 0xc2f87377
5552 .long 0xc2f87377 // .word 0xc2f87377
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]5553 .long 0x3f800000 // .word 0x3f800000
5554 .long 0x3f800000 // .word 0x3f800000
5555 .long 0x409af5f8 // .word 0x409af5f8
5556 .long 0x409af5f8 // .word 0x409af5f8
5557 .long 0x3fbebc8d // .word 0x3fbebc8d
5558 .long 0x3fbebc8d // .word 0x3fbebc8d
5559 .long 0x42f28c51 // .word 0x42f28c51
5560 .long 0x42f28c51 // .word 0x42f28c51
5561 .long 0x3fdce9a3 // .word 0x3fdce9a3
5562 .long 0x41ddd2fe // .word 0x41ddd2fe
5563
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]5564HIDDEN _sk_lab_to_xyz_vfp4
5565.globl _sk_lab_to_xyz_vfp4
5566FUNCTION(_sk_lab_to_xyz_vfp4)
5567_sk_lab_to_xyz_vfp4:
5568 .long 0xeddf1b2e // vldr d17, [pc, #184]
5569 .long 0xf3c43613 // vmov.i32 d19, #-1023410176
5570 .long 0xeddf0b2a // vldr d16, [pc, #168]
5571 .long 0xf2c34f10 // vmov.f32 d20, #16
5572 .long 0xf3401d31 // vmul.f32 d17, d0, d17
5573 .long 0xeddf5b2f // vldr d21, [pc, #188]
5574 .long 0xf3412d30 // vmul.f32 d18, d1, d16
5575 .long 0xeddf6b2f // vldr d22, [pc, #188]
5576 .long 0xf3420d30 // vmul.f32 d16, d2, d16
5577 .long 0xeddf7b2f // vldr d23, [pc, #188]
5578 .long 0xeddf8b30 // vldr d24, [pc, #192]
5579 .long 0xf2411da4 // vadd.f32 d17, d17, d20
5580 .long 0xeddf4b24 // vldr d20, [pc, #144]
5581 .long 0xf2422da3 // vadd.f32 d18, d18, d19
5582 .long 0xe4913004 // ldr r3, [r1], #4
5583 .long 0xf2400da3 // vadd.f32 d16, d16, d19
5584 .long 0xeddf3b22 // vldr d19, [pc, #136]
5585 .long 0xf3411db3 // vmul.f32 d17, d17, d19
5586 .long 0xf3422db4 // vmul.f32 d18, d18, d20
5587 .long 0xf3400db5 // vmul.f32 d16, d16, d21
5588 .long 0xf2412da2 // vadd.f32 d18, d17, d18
5589 .long 0xf2610da0 // vsub.f32 d16, d17, d16
5590 .long 0xf3415db1 // vmul.f32 d21, d17, d17
5591 .long 0xf3423db2 // vmul.f32 d19, d18, d18
5592 .long 0xf3404db0 // vmul.f32 d20, d16, d16
5593 .long 0xf3415db5 // vmul.f32 d21, d17, d21
5594 .long 0xf2411da6 // vadd.f32 d17, d17, d22
5595 .long 0xf3423db3 // vmul.f32 d19, d18, d19
5596 .long 0xf3404db4 // vmul.f32 d20, d16, d20
5597 .long 0xf2400da6 // vadd.f32 d16, d16, d22
5598 .long 0xf2422da6 // vadd.f32 d18, d18, d22
5599 .long 0xeddf6b1d // vldr d22, [pc, #116]
5600 .long 0xf3639ea7 // vcgt.f32 d25, d19, d23
5601 .long 0xf364aea7 // vcgt.f32 d26, d20, d23
5602 .long 0xf3400db8 // vmul.f32 d16, d16, d24
5603 .long 0xf3422db8 // vmul.f32 d18, d18, d24
5604 .long 0xf35391b2 // vbsl d25, d19, d18
5605 .long 0xeddf2b19 // vldr d18, [pc, #100]
5606 .long 0xf354a1b0 // vbsl d26, d20, d16
5607 .long 0xf3251ea7 // vcgt.f32 d1, d21, d23
5608 .long 0xf3090db6 // vmul.f32 d0, d25, d22
5609 .long 0xf30a2db2 // vmul.f32 d2, d26, d18
5610 .long 0xf3410db8 // vmul.f32 d16, d17, d24
5611 .long 0xf31511b0 // vbsl d1, d21, d16
5612 .long 0xe12fff13 // bx r3
5613 .long 0xe320f000 // nop {0}
5614 .long 0x437f0000 // .word 0x437f0000
5615 .long 0x437f0000 // .word 0x437f0000
5616 .long 0x42c80000 // .word 0x42c80000
5617 .long 0x42c80000 // .word 0x42c80000
5618 .long 0x3b03126f // .word 0x3b03126f
5619 .long 0x3b03126f // .word 0x3b03126f
5620 .long 0x3c0d3dcb // .word 0x3c0d3dcb
5621 .long 0x3c0d3dcb // .word 0x3c0d3dcb
5622 .long 0x3ba3d70a // .word 0x3ba3d70a
5623 .long 0x3ba3d70a // .word 0x3ba3d70a
5624 .long 0xbe0d3dcb // .word 0xbe0d3dcb
5625 .long 0xbe0d3dcb // .word 0xbe0d3dcb
5626 .long 0x3c1118c2 // .word 0x3c1118c2
5627 .long 0x3c1118c2 // .word 0x3c1118c2
5628 .long 0x3e038050 // .word 0x3e038050
5629 .long 0x3e038050 // .word 0x3e038050
5630 .long 0x3f76d71f // .word 0x3f76d71f
5631 .long 0x3f76d71f // .word 0x3f76d71f
5632 .long 0x3f5340f6 // .word 0x3f5340f6
5633 .long 0x3f5340f6 // .word 0x3f5340f6
5634
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]5635HIDDEN _sk_load_a8_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5636.globl _sk_load_a8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5637FUNCTION(_sk_load_a8_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5638_sk_load_a8_vfp4:
5639 .long 0xe24dd004 // sub sp, sp, #4
5640 .long 0xe8911008 // ldm r1, {r3, ip}
5641 .long 0xe2811008 // add r1, r1, #8
5642 .long 0xf2800010 // vmov.i32 d0, #0
5643 .long 0xf2801010 // vmov.i32 d1, #0
5644 .long 0xe5933000 // ldr r3, [r3]
5645 .long 0xf2802010 // vmov.i32 d2, #0
5646 .long 0xe0833000 // add r3, r3, r0
5647 .long 0xe1d330b0 // ldrh r3, [r3]
5648 .long 0xe1cd30b0 // strh r3, [sp]
5649 .long 0xe1a0300d // mov r3, sp
5650 .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16]
5651 .long 0xf3c80a30 // vmovl.u8 q8, d16
5652 .long 0xf3d00a30 // vmovl.u16 q8, d16
5653 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5654 .long 0xeddf1b03 // vldr d17, [pc, #12]
5655 .long 0xf3003db1 // vmul.f32 d3, d16, d17
5656 .long 0xe28dd004 // add sp, sp, #4
5657 .long 0xe12fff1c // bx ip
5658 .long 0xe320f000 // nop {0}
5659 .long 0x3b808081 // .word 0x3b808081
5660 .long 0x3b808081 // .word 0x3b808081
5661
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5662HIDDEN _sk_gather_a8_vfp4
5663.globl _sk_gather_a8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5664FUNCTION(_sk_gather_a8_vfp4)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5665_sk_gather_a8_vfp4:
5666 .long 0xe92d4010 // push {r4, lr}
5667 .long 0xe8911008 // ldm r1, {r3, ip}
5668 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
5669 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
5670 .long 0xe2811008 // add r1, r1, #8
5671 .long 0xf2800010 // vmov.i32 d0, #0
5672 .long 0xe493e008 // ldr lr, [r3], #8
5673 .long 0xf2801010 // vmov.i32 d1, #0
5674 .long 0xf2802010 // vmov.i32 d2, #0
5675 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5676 .long 0xf26219a0 // vmla.i32 d17, d18, d16
5677 .long 0xee113b90 // vmov.32 r3, d17[0]
5678 .long 0xee314b90 // vmov.32 r4, d17[1]
5679 .long 0xf3c7101f // vmov.i32 d17, #255
5680 .long 0xe7de3003 // ldrb r3, [lr, r3]
5681 .long 0xe7de4004 // ldrb r4, [lr, r4]
5682 .long 0xee003b90 // vmov.32 d16[0], r3
5683 .long 0xee204b90 // vmov.32 d16[1], r4
5684 .long 0xf24001b1 // vand d16, d16, d17
5685 .long 0xeddf1b03 // vldr d17, [pc, #12]
5686 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5687 .long 0xf3003db1 // vmul.f32 d3, d16, d17
5688 .long 0xe8bd4010 // pop {r4, lr}
5689 .long 0xe12fff1c // bx ip
5690 .long 0x3b808081 // .word 0x3b808081
5691 .long 0x3b808081 // .word 0x3b808081
5692
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]5693HIDDEN _sk_store_a8_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5694.globl _sk_store_a8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5695FUNCTION(_sk_store_a8_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5696_sk_store_a8_vfp4:
5697 .long 0xe92d4800 // push {fp, lr}
5698 .long 0xeddf0b0d // vldr d16, [pc, #52]
5699 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5700 .long 0xf2431c30 // vfma.f32 d17, d3, d16
5701 .long 0xe5913000 // ldr r3, [r1]
5702 .long 0xe5933000 // ldr r3, [r3]
5703 .long 0xf3fb07a1 // vcvt.u32.f32 d16, d17
5704 .long 0xee10eb90 // vmov.32 lr, d16[0]
5705 .long 0xee30cb90 // vmov.32 ip, d16[1]
5706 .long 0xe7e3e000 // strb lr, [r3, r0]!
5707 .long 0xe5c3c001 // strb ip, [r3, #1]
5708 .long 0xe2813008 // add r3, r1, #8
5709 .long 0xe591c004 // ldr ip, [r1, #4]
5710 .long 0xe1a01003 // mov r1, r3
5711 .long 0xe8bd4800 // pop {fp, lr}
5712 .long 0xe12fff1c // bx ip
5713 .long 0x437f0000 // .word 0x437f0000
5714 .long 0x437f0000 // .word 0x437f0000
5715
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]5716HIDDEN _sk_load_g8_vfp4
5717.globl _sk_load_g8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5718FUNCTION(_sk_load_g8_vfp4)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]5719_sk_load_g8_vfp4:
5720 .long 0xe24dd004 // sub sp, sp, #4
5721 .long 0xe8911008 // ldm r1, {r3, ip}
5722 .long 0xe2811008 // add r1, r1, #8
5723 .long 0xf2873f10 // vmov.f32 d3, #1
5724 .long 0xe5933000 // ldr r3, [r3]
5725 .long 0xe0833000 // add r3, r3, r0
5726 .long 0xe1d330b0 // ldrh r3, [r3]
5727 .long 0xe1cd30b0 // strh r3, [sp]
5728 .long 0xe1a0300d // mov r3, sp
5729 .long 0xf4e3041f // vld1.16 {d16[0]}, [r3 :16]
5730 .long 0xf3c80a30 // vmovl.u8 q8, d16
5731 .long 0xf3d00a30 // vmovl.u16 q8, d16
5732 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5733 .long 0xeddf1b05 // vldr d17, [pc, #20]
5734 .long 0xf3000db1 // vmul.f32 d0, d16, d17
5735 .long 0xf2201110 // vorr d1, d0, d0
5736 .long 0xf2202110 // vorr d2, d0, d0
5737 .long 0xe28dd004 // add sp, sp, #4
5738 .long 0xe12fff1c // bx ip
5739 .long 0xe320f000 // nop {0}
5740 .long 0x3b808081 // .word 0x3b808081
5741 .long 0x3b808081 // .word 0x3b808081
5742
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5743HIDDEN _sk_gather_g8_vfp4
5744.globl _sk_gather_g8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5745FUNCTION(_sk_gather_g8_vfp4)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5746_sk_gather_g8_vfp4:
5747 .long 0xe92d4010 // push {r4, lr}
5748 .long 0xe8911008 // ldm r1, {r3, ip}
5749 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
5750 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
5751 .long 0xe2811008 // add r1, r1, #8
5752 .long 0xf2873f10 // vmov.f32 d3, #1
5753 .long 0xe493e008 // ldr lr, [r3], #8
5754 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5755 .long 0xf26219a0 // vmla.i32 d17, d18, d16
5756 .long 0xee113b90 // vmov.32 r3, d17[0]
5757 .long 0xee314b90 // vmov.32 r4, d17[1]
5758 .long 0xf3c7101f // vmov.i32 d17, #255
5759 .long 0xe7de3003 // ldrb r3, [lr, r3]
5760 .long 0xe7de4004 // ldrb r4, [lr, r4]
5761 .long 0xee003b90 // vmov.32 d16[0], r3
5762 .long 0xee204b90 // vmov.32 d16[1], r4
5763 .long 0xf24001b1 // vand d16, d16, d17
5764 .long 0xeddf1b05 // vldr d17, [pc, #20]
5765 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
5766 .long 0xf3000db1 // vmul.f32 d0, d16, d17
5767 .long 0xf2201110 // vorr d1, d0, d0
5768 .long 0xf2202110 // vorr d2, d0, d0
5769 .long 0xe8bd4010 // pop {r4, lr}
5770 .long 0xe12fff1c // bx ip
5771 .long 0x3b808081 // .word 0x3b808081
5772 .long 0x3b808081 // .word 0x3b808081
5773
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]5774HIDDEN _sk_gather_i8_vfp4
5775.globl _sk_gather_i8_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5776FUNCTION(_sk_gather_i8_vfp4)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]5777_sk_gather_i8_vfp4:
5778 .long 0xe92d4010 // push {r4, lr}
5779 .long 0xe1a0e001 // mov lr, r1
5780 .long 0xe491c004 // ldr ip, [r1], #4
5781 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
5782 .long 0xe35c0000 // cmp ip, #0
5783 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
5784 .long 0xe1a0300c // mov r3, ip
5785 .long 0x028e1008 // addeq r1, lr, #8
5786 .long 0x059e3004 // ldreq r3, [lr, #4]
5787 .long 0xe493e008 // ldr lr, [r3], #8
5788 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5789 .long 0xf26219a0 // vmla.i32 d17, d18, d16
5790 .long 0xee113b90 // vmov.32 r3, d17[0]
5791 .long 0xee314b90 // vmov.32 r4, d17[1]
5792 .long 0xf3c7101f // vmov.i32 d17, #255
5793 .long 0xe7de3003 // ldrb r3, [lr, r3]
5794 .long 0xe7de4004 // ldrb r4, [lr, r4]
5795 .long 0xee003b90 // vmov.32 d16[0], r3
5796 .long 0xee204b90 // vmov.32 d16[1], r4
5797 .long 0xe59c4004 // ldr r4, [ip, #4]
5798 .long 0xf24001b1 // vand d16, d16, d17
5799 .long 0xee103b90 // vmov.32 r3, d16[0]
5800 .long 0xee30eb90 // vmov.32 lr, d16[1]
5801 .long 0xe0843103 // add r3, r4, r3, lsl #2
5802 .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
5803 .long 0xe084310e // add r3, r4, lr, lsl #2
5804 .long 0xf4e308bf // vld1.32 {d16[1]}, [r3 :32]
5805 .long 0xf24021b1 // vand d18, d16, d17
5806 .long 0xf3f83030 // vshr.u32 d19, d16, #8
5807 .long 0xf3e84030 // vshr.u32 d20, d16, #24
5808 .long 0xe4913004 // ldr r3, [r1], #4
5809 .long 0xf3f00030 // vshr.u32 d16, d16, #16
5810 .long 0xf24331b1 // vand d19, d19, d17
5811 .long 0xf24001b1 // vand d16, d16, d17
5812 .long 0xeddf1b0a // vldr d17, [pc, #40]
5813 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5814 .long 0xf3fb4624 // vcvt.f32.s32 d20, d20
5815 .long 0xf3fb3623 // vcvt.f32.s32 d19, d19
5816 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5817 .long 0xf3020db1 // vmul.f32 d0, d18, d17
5818 .long 0xf3043db1 // vmul.f32 d3, d20, d17
5819 .long 0xf3031db1 // vmul.f32 d1, d19, d17
5820 .long 0xf3002db1 // vmul.f32 d2, d16, d17
5821 .long 0xe8bd4010 // pop {r4, lr}
5822 .long 0xe12fff13 // bx r3
5823 .long 0xe320f000 // nop {0}
5824 .long 0x3b808081 // .word 0x3b808081
5825 .long 0x3b808081 // .word 0x3b808081
5826
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]5827HIDDEN _sk_load_565_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5828.globl _sk_load_565_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5829FUNCTION(_sk_load_565_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5830_sk_load_565_vfp4:
5831 .long 0xe24dd004 // sub sp, sp, #4
5832 .long 0xe8911008 // ldm r1, {r3, ip}
5833 .long 0xf2c1101f // vmov.i32 d17, #31
5834 .long 0xf3c72218 // vmov.i32 d18, #63488
5835 .long 0xeddf3b16 // vldr d19, [pc, #88]
5836 .long 0xe2811008 // add r1, r1, #8
5837 .long 0xe5933000 // ldr r3, [r3]
5838 .long 0xf2873f10 // vmov.f32 d3, #1
5839 .long 0xe7933080 // ldr r3, [r3, r0, lsl #1]
5840 .long 0xe58d3000 // str r3, [sp]
5841 .long 0xe1a0300d // mov r3, sp
5842 .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
5843 .long 0xe3a03e7e // mov r3, #2016
5844 .long 0xf3d04a30 // vmovl.u16 q10, d16
5845 .long 0xee803b90 // vdup.32 d16, r3
5846 .long 0xf24411b1 // vand d17, d20, d17
5847 .long 0xeddf5b0e // vldr d21, [pc, #56]
5848 .long 0xf24421b2 // vand d18, d20, d18
5849 .long 0xf24401b0 // vand d16, d20, d16
5850 .long 0xeddf4b09 // vldr d20, [pc, #36]
5851 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5852 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5853 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5854 .long 0xf3020db3 // vmul.f32 d0, d18, d19
5855 .long 0xf3001db4 // vmul.f32 d1, d16, d20
5856 .long 0xf3012db5 // vmul.f32 d2, d17, d21
5857 .long 0xe28dd004 // add sp, sp, #4
5858 .long 0xe12fff1c // bx ip
5859 .long 0x37842108 // .word 0x37842108
5860 .long 0x37842108 // .word 0x37842108
5861 .long 0x3a020821 // .word 0x3a020821
5862 .long 0x3a020821 // .word 0x3a020821
5863 .long 0x3d042108 // .word 0x3d042108
5864 .long 0x3d042108 // .word 0x3d042108
5865
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5866HIDDEN _sk_gather_565_vfp4
5867.globl _sk_gather_565_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5868FUNCTION(_sk_gather_565_vfp4)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5869_sk_gather_565_vfp4:
5870 .long 0xe92d4010 // push {r4, lr}
5871 .long 0xe8911008 // ldm r1, {r3, ip}
5872 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
5873 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
5874 .long 0xeddf4b20 // vldr d20, [pc, #128]
5875 .long 0xe2811008 // add r1, r1, #8
5876 .long 0xe493e008 // ldr lr, [r3], #8
5877 .long 0xf2873f10 // vmov.f32 d3, #1
5878 .long 0xeddf5b1e // vldr d21, [pc, #120]
5879 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
5880 .long 0xf26219a0 // vmla.i32 d17, d18, d16
5881 .long 0xf2c1201f // vmov.i32 d18, #31
5882 .long 0xee113b90 // vmov.32 r3, d17[0]
5883 .long 0xee314b90 // vmov.32 r4, d17[1]
5884 .long 0xf3c71218 // vmov.i32 d17, #63488
5885 .long 0xe08e3083 // add r3, lr, r3, lsl #1
5886 .long 0xe08e4084 // add r4, lr, r4, lsl #1
5887 .long 0xe1d330b0 // ldrh r3, [r3]
5888 .long 0xe1d440b0 // ldrh r4, [r4]
5889 .long 0xee003b90 // vmov.32 d16[0], r3
5890 .long 0xe3a03e7e // mov r3, #2016
5891 .long 0xee833b90 // vdup.32 d19, r3
5892 .long 0xee204b90 // vmov.32 d16[1], r4
5893 .long 0xf24011b1 // vand d17, d16, d17
5894 .long 0xf24031b3 // vand d19, d16, d19
5895 .long 0xf24001b2 // vand d16, d16, d18
5896 .long 0xf3fb2623 // vcvt.f32.s32 d18, d19
5897 .long 0xeddf3b07 // vldr d19, [pc, #28]
5898 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5899 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5900 .long 0xf3021db4 // vmul.f32 d1, d18, d20
5901 .long 0xf3010db3 // vmul.f32 d0, d17, d19
5902 .long 0xf3002db5 // vmul.f32 d2, d16, d21
5903 .long 0xe8bd4010 // pop {r4, lr}
5904 .long 0xe12fff1c // bx ip
5905 .long 0xe320f000 // nop {0}
5906 .long 0x37842108 // .word 0x37842108
5907 .long 0x37842108 // .word 0x37842108
5908 .long 0x3a020821 // .word 0x3a020821
5909 .long 0x3a020821 // .word 0x3a020821
5910 .long 0x3d042108 // .word 0x3d042108
5911 .long 0x3d042108 // .word 0x3d042108
5912
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]5913HIDDEN _sk_store_565_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5914.globl _sk_store_565_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5915FUNCTION(_sk_store_565_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]5916_sk_store_565_vfp4:
5917 .long 0xf2c30f1f // vmov.f32 d16, #31
5918 .long 0xeddf1b15 // vldr d17, [pc, #84]
5919 .long 0xf2c3361f // vmov.i32 d19, #1056964608
5920 .long 0xe5913000 // ldr r3, [r1]
5921 .long 0xf2413c31 // vfma.f32 d19, d1, d17
5922 .long 0xf2c3161f // vmov.i32 d17, #1056964608
5923 .long 0xf2401c30 // vfma.f32 d17, d0, d16
5924 .long 0xe5933000 // ldr r3, [r3]
5925 .long 0xf2c3261f // vmov.i32 d18, #1056964608
5926 .long 0xf2422c30 // vfma.f32 d18, d2, d16
5927 .long 0xe0833080 // add r3, r3, r0, lsl #1
5928 .long 0xf3fb07a3 // vcvt.u32.f32 d16, d19
5929 .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
5930 .long 0xf3fb27a2 // vcvt.u32.f32 d18, d18
5931 .long 0xf2e50530 // vshl.s32 d16, d16, #5
5932 .long 0xf2eb1531 // vshl.s32 d17, d17, #11
5933 .long 0xf26001b1 // vorr d16, d16, d17
5934 .long 0xf26001b2 // vorr d16, d16, d18
5935 .long 0xf3f60121 // vuzp.16 d16, d17
5936 .long 0xf4c3080f // vst1.32 {d16[0]}, [r3]
5937 .long 0xe2813008 // add r3, r1, #8
5938 .long 0xe591c004 // ldr ip, [r1, #4]
5939 .long 0xe1a01003 // mov r1, r3
5940 .long 0xe12fff1c // bx ip
5941 .long 0x427c0000 // .word 0x427c0000
5942 .long 0x427c0000 // .word 0x427c0000
5943
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]5944HIDDEN _sk_load_4444_vfp4
5945.globl _sk_load_4444_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5946FUNCTION(_sk_load_4444_vfp4)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]5947_sk_load_4444_vfp4:
5948 .long 0xe24dd004 // sub sp, sp, #4
5949 .long 0xe8911008 // ldm r1, {r3, ip}
5950 .long 0xf3c71210 // vmov.i32 d17, #61440
5951 .long 0xf3c74010 // vmov.i32 d20, #240
5952 .long 0xf2c0501f // vmov.i32 d21, #15
5953 .long 0xeddf6b1d // vldr d22, [pc, #116]
5954 .long 0xe5933000 // ldr r3, [r3]
5955 .long 0xe2811008 // add r1, r1, #8
5956 .long 0xe7933080 // ldr r3, [r3, r0, lsl #1]
5957 .long 0xe58d3000 // str r3, [sp]
5958 .long 0xe1a0300d // mov r3, sp
5959 .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
5960 .long 0xf3d02a30 // vmovl.u16 q9, d16
5961 .long 0xf2c0021f // vmov.i32 d16, #3840
5962 .long 0xf24211b1 // vand d17, d18, d17
5963 .long 0xf24201b0 // vand d16, d18, d16
5964 .long 0xf24241b4 // vand d20, d18, d20
5965 .long 0xf24221b5 // vand d18, d18, d21
5966 .long 0xeddf3b0c // vldr d19, [pc, #48]
5967 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
5968 .long 0xeddf5b0c // vldr d21, [pc, #48]
5969 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
5970 .long 0xf3fb4624 // vcvt.f32.s32 d20, d20
5971 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
5972 .long 0xf3010db3 // vmul.f32 d0, d17, d19
5973 .long 0xeddf1b0b // vldr d17, [pc, #44]
5974 .long 0xf3001db5 // vmul.f32 d1, d16, d21
5975 .long 0xf3042db6 // vmul.f32 d2, d20, d22
5976 .long 0xf3023db1 // vmul.f32 d3, d18, d17
5977 .long 0xe28dd004 // add sp, sp, #4
5978 .long 0xe12fff1c // bx ip
5979 .long 0xe320f000 // nop {0}
5980 .long 0x37888889 // .word 0x37888889
5981 .long 0x37888889 // .word 0x37888889
5982 .long 0x39888889 // .word 0x39888889
5983 .long 0x39888889 // .word 0x39888889
5984 .long 0x3b888889 // .word 0x3b888889
5985 .long 0x3b888889 // .word 0x3b888889
5986 .long 0x3d888889 // .word 0x3d888889
5987 .long 0x3d888889 // .word 0x3d888889
5988
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5989HIDDEN _sk_gather_4444_vfp4
5990.globl _sk_gather_4444_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]5991FUNCTION(_sk_gather_4444_vfp4)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]5992_sk_gather_4444_vfp4:
5993 .long 0xe92d4010 // push {r4, lr}
5994 .long 0xe8911008 // ldm r1, {r3, ip}
5995 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
5996 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
5997 .long 0xf3c73010 // vmov.i32 d19, #240
5998 .long 0xeddf5b21 // vldr d21, [pc, #132]
5999 .long 0xe493e008 // ldr lr, [r3], #8
6000 .long 0xf2c0401f // vmov.i32 d20, #15
6001 .long 0xeddf6b20 // vldr d22, [pc, #128]
6002 .long 0xe2811008 // add r1, r1, #8
6003 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6004 .long 0xf26219a0 // vmla.i32 d17, d18, d16
6005 .long 0xf2c0221f // vmov.i32 d18, #3840
6006 .long 0xee113b90 // vmov.32 r3, d17[0]
6007 .long 0xee314b90 // vmov.32 r4, d17[1]
6008 .long 0xf3c71210 // vmov.i32 d17, #61440
6009 .long 0xe08e3083 // add r3, lr, r3, lsl #1
6010 .long 0xe08e4084 // add r4, lr, r4, lsl #1
6011 .long 0xe1d330b0 // ldrh r3, [r3]
6012 .long 0xe1d440b0 // ldrh r4, [r4]
6013 .long 0xee003b90 // vmov.32 d16[0], r3
6014 .long 0xee204b90 // vmov.32 d16[1], r4
6015 .long 0xf24011b1 // vand d17, d16, d17
6016 .long 0xf24021b2 // vand d18, d16, d18
6017 .long 0xf24031b3 // vand d19, d16, d19
6018 .long 0xf24001b4 // vand d16, d16, d20
6019 .long 0xeddf4b0a // vldr d20, [pc, #40]
6020 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
6021 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
6022 .long 0xf3fb3623 // vcvt.f32.s32 d19, d19
6023 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
6024 .long 0xf3010db4 // vmul.f32 d0, d17, d20
6025 .long 0xeddf1b0a // vldr d17, [pc, #40]
6026 .long 0xf3021db5 // vmul.f32 d1, d18, d21
6027 .long 0xf3032db6 // vmul.f32 d2, d19, d22
6028 .long 0xf3003db1 // vmul.f32 d3, d16, d17
6029 .long 0xe8bd4010 // pop {r4, lr}
6030 .long 0xe12fff1c // bx ip
6031 .long 0x37888889 // .word 0x37888889
6032 .long 0x37888889 // .word 0x37888889
6033 .long 0x39888889 // .word 0x39888889
6034 .long 0x39888889 // .word 0x39888889
6035 .long 0x3b888889 // .word 0x3b888889
6036 .long 0x3b888889 // .word 0x3b888889
6037 .long 0x3d888889 // .word 0x3d888889
6038 .long 0x3d888889 // .word 0x3d888889
6039
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]6040HIDDEN _sk_store_4444_vfp4
6041.globl _sk_store_4444_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6042FUNCTION(_sk_store_4444_vfp4)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]6043_sk_store_4444_vfp4:
6044 .long 0xf2c20f1e // vmov.f32 d16, #15
6045 .long 0xe5913000 // ldr r3, [r1]
6046 .long 0xf2c3261f // vmov.i32 d18, #1056964608
6047 .long 0xf2c3361f // vmov.i32 d19, #1056964608
6048 .long 0xf2402c30 // vfma.f32 d18, d0, d16
6049 .long 0xe5933000 // ldr r3, [r3]
6050 .long 0xf2413c30 // vfma.f32 d19, d1, d16
6051 .long 0xf2c3461f // vmov.i32 d20, #1056964608
6052 .long 0xe0833080 // add r3, r3, r0, lsl #1
6053 .long 0xf2424c30 // vfma.f32 d20, d2, d16
6054 .long 0xf2c3161f // vmov.i32 d17, #1056964608
6055 .long 0xf2431c30 // vfma.f32 d17, d3, d16
6056 .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18
6057 .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19
6058 .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20
6059 .long 0xf2ec0530 // vshl.s32 d16, d16, #12
6060 .long 0xf2e82532 // vshl.s32 d18, d18, #8
6061 .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
6062 .long 0xf2e43533 // vshl.s32 d19, d19, #4
6063 .long 0xf26201b0 // vorr d16, d18, d16
6064 .long 0xf26001b3 // vorr d16, d16, d19
6065 .long 0xf26001b1 // vorr d16, d16, d17
6066 .long 0xf3f60121 // vuzp.16 d16, d17
6067 .long 0xf4c3080f // vst1.32 {d16[0]}, [r3]
6068 .long 0xe2813008 // add r3, r1, #8
6069 .long 0xe591c004 // ldr ip, [r1, #4]
6070 .long 0xe1a01003 // mov r1, r3
6071 .long 0xe12fff1c // bx ip
6072
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6073HIDDEN _sk_load_8888_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6074.globl _sk_load_8888_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6075FUNCTION(_sk_load_8888_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6076_sk_load_8888_vfp4:
6077 .long 0xe8911008 // ldm r1, {r3, ip}
6078 .long 0xf3c7001f // vmov.i32 d16, #255
6079 .long 0xe2811008 // add r1, r1, #8
6080 .long 0xe5933000 // ldr r3, [r3]
6081 .long 0xe0833100 // add r3, r3, r0, lsl #2
6082 .long 0xedd31b00 // vldr d17, [r3]
6083 .long 0xf24121b0 // vand d18, d17, d16
6084 .long 0xf3f83031 // vshr.u32 d19, d17, #8
6085 .long 0xf3e84031 // vshr.u32 d20, d17, #24
6086 .long 0xf3f01031 // vshr.u32 d17, d17, #16
6087 .long 0xf24331b0 // vand d19, d19, d16
6088 .long 0xf24101b0 // vand d16, d17, d16
6089 .long 0xeddf1b08 // vldr d17, [pc, #32]
6090 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
6091 .long 0xf3fb4624 // vcvt.f32.s32 d20, d20
6092 .long 0xf3fb3623 // vcvt.f32.s32 d19, d19
6093 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
6094 .long 0xf3020db1 // vmul.f32 d0, d18, d17
6095 .long 0xf3043db1 // vmul.f32 d3, d20, d17
6096 .long 0xf3031db1 // vmul.f32 d1, d19, d17
6097 .long 0xf3002db1 // vmul.f32 d2, d16, d17
6098 .long 0xe12fff1c // bx ip
6099 .long 0x3b808081 // .word 0x3b808081
6100 .long 0x3b808081 // .word 0x3b808081
6101
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]6102HIDDEN _sk_gather_8888_vfp4
6103.globl _sk_gather_8888_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6104FUNCTION(_sk_gather_8888_vfp4)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]6105_sk_gather_8888_vfp4:
6106 .long 0xe92d4010 // push {r4, lr}
6107 .long 0xe8911008 // ldm r1, {r3, ip}
6108 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
6109 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
6110 .long 0xe2811008 // add r1, r1, #8
6111 .long 0xe493e008 // ldr lr, [r3], #8
6112 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6113 .long 0xf26219a0 // vmla.i32 d17, d18, d16
6114 .long 0xee113b90 // vmov.32 r3, d17[0]
6115 .long 0xee314b90 // vmov.32 r4, d17[1]
6116 .long 0xf3c7101f // vmov.i32 d17, #255
6117 .long 0xe08e3103 // add r3, lr, r3, lsl #2
6118 .long 0xf4e3083f // vld1.32 {d16[0]}, [r3 :32]
6119 .long 0xe08e3104 // add r3, lr, r4, lsl #2
6120 .long 0xf4e308bf // vld1.32 {d16[1]}, [r3 :32]
6121 .long 0xf24021b1 // vand d18, d16, d17
6122 .long 0xf3f83030 // vshr.u32 d19, d16, #8
6123 .long 0xf3e84030 // vshr.u32 d20, d16, #24
6124 .long 0xf3f00030 // vshr.u32 d16, d16, #16
6125 .long 0xf24331b1 // vand d19, d19, d17
6126 .long 0xf24001b1 // vand d16, d16, d17
6127 .long 0xeddf1b09 // vldr d17, [pc, #36]
6128 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
6129 .long 0xf3fb4624 // vcvt.f32.s32 d20, d20
6130 .long 0xf3fb3623 // vcvt.f32.s32 d19, d19
6131 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
6132 .long 0xf3020db1 // vmul.f32 d0, d18, d17
6133 .long 0xf3043db1 // vmul.f32 d3, d20, d17
6134 .long 0xf3031db1 // vmul.f32 d1, d19, d17
6135 .long 0xf3002db1 // vmul.f32 d2, d16, d17
6136 .long 0xe8bd4010 // pop {r4, lr}
6137 .long 0xe12fff1c // bx ip
6138 .long 0x3b808081 // .word 0x3b808081
6139 .long 0x3b808081 // .word 0x3b808081
6140
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6141HIDDEN _sk_store_8888_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6142.globl _sk_store_8888_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6143FUNCTION(_sk_store_8888_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6144_sk_store_8888_vfp4:
6145 .long 0xeddf0b1a // vldr d16, [pc, #104]
6146 .long 0xf2c3261f // vmov.i32 d18, #1056964608
6147 .long 0xf2412c30 // vfma.f32 d18, d1, d16
6148 .long 0xe5913000 // ldr r3, [r1]
6149 .long 0xf2c3361f // vmov.i32 d19, #1056964608
6150 .long 0xf2c3161f // vmov.i32 d17, #1056964608
6151 .long 0xf2423c30 // vfma.f32 d19, d2, d16
6152 .long 0xe5933000 // ldr r3, [r3]
6153 .long 0xf2c3461f // vmov.i32 d20, #1056964608
6154 .long 0xf2401c30 // vfma.f32 d17, d0, d16
6155 .long 0xe0833100 // add r3, r3, r0, lsl #2
6156 .long 0xf2434c30 // vfma.f32 d20, d3, d16
6157 .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18
6158 .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19
6159 .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
6160 .long 0xf3fb37a4 // vcvt.u32.f32 d19, d20
6161 .long 0xf2e80530 // vshl.s32 d16, d16, #8
6162 .long 0xf2f02532 // vshl.s32 d18, d18, #16
6163 .long 0xf26001b1 // vorr d16, d16, d17
6164 .long 0xf2f81533 // vshl.s32 d17, d19, #24
6165 .long 0xf26001b2 // vorr d16, d16, d18
6166 .long 0xf26001b1 // vorr d16, d16, d17
6167 .long 0xedc30b00 // vstr d16, [r3]
6168 .long 0xe2813008 // add r3, r1, #8
6169 .long 0xe591c004 // ldr ip, [r1, #4]
6170 .long 0xe1a01003 // mov r1, r3
6171 .long 0xe12fff1c // bx ip
6172 .long 0xe320f000 // nop {0}
6173 .long 0x437f0000 // .word 0x437f0000
6174 .long 0x437f0000 // .word 0x437f0000
6175
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6176HIDDEN _sk_load_f16_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6177.globl _sk_load_f16_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6178FUNCTION(_sk_load_f16_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6179_sk_load_f16_vfp4:
6180 .long 0xe8911008 // ldm r1, {r3, ip}
6181 .long 0xe2811008 // add r1, r1, #8
6182 .long 0xe5933000 // ldr r3, [r3]
6183 .long 0xe0833180 // add r3, r3, r0, lsl #3
Mike Klein114e6b32017-04-03 22:21:15 -0400[diff] [blame]6184 .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]!
6185 .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3]
6186 .long 0xf3b60720 // vcvt.f32.f16 q0, d16
6187 .long 0xf3b62722 // vcvt.f32.f16 q1, d18
6188 .long 0xf3f64721 // vcvt.f32.f16 q10, d17
6189 .long 0xf3f60723 // vcvt.f32.f16 q8, d19
6190 .long 0xf22411b4 // vorr d1, d20, d20
6191 .long 0xf22031b0 // vorr d3, d16, d16
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6192 .long 0xe12fff1c // bx ip
6193
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]6194HIDDEN _sk_gather_f16_vfp4
6195.globl _sk_gather_f16_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6196FUNCTION(_sk_gather_f16_vfp4)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]6197_sk_gather_f16_vfp4:
6198 .long 0xe92d4c10 // push {r4, sl, fp, lr}
6199 .long 0xe28db008 // add fp, sp, #8
6200 .long 0xe24dd010 // sub sp, sp, #16
6201 .long 0xe7c3d01f // bfc sp, #0, #4
6202 .long 0xe5913000 // ldr r3, [r1]
6203 .long 0xf3fb0701 // vcvt.s32.f32 d16, d1
6204 .long 0xf3fb1700 // vcvt.s32.f32 d17, d0
6205 .long 0xe493c008 // ldr ip, [r3], #8
6206 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6207 .long 0xf26219a0 // vmla.i32 d17, d18, d16
6208 .long 0xee113b90 // vmov.32 r3, d17[0]
6209 .long 0xee31eb90 // vmov.32 lr, d17[1]
6210 .long 0xe08c4183 // add r4, ip, r3, lsl #3
6211 .long 0xe08c318e // add r3, ip, lr, lsl #3
6212 .long 0xedd31b00 // vldr d17, [r3]
6213 .long 0xe1a0300d // mov r3, sp
6214 .long 0xedd40b00 // vldr d16, [r4]
6215 .long 0xf4430aef // vst1.64 {d16-d17}, [r3 :128]
6216 .long 0xf4e3071f // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3 :64]
6217 .long 0xe3833008 // orr r3, r3, #8
6218 .long 0xf4e3075f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3 :64]
6219 .long 0xf3b60720 // vcvt.f32.f16 q0, d16
6220 .long 0xf3b62722 // vcvt.f32.f16 q1, d18
6221 .long 0xe5913004 // ldr r3, [r1, #4]
6222 .long 0xe2811008 // add r1, r1, #8
6223 .long 0xf3f64721 // vcvt.f32.f16 q10, d17
6224 .long 0xf3f60723 // vcvt.f32.f16 q8, d19
6225 .long 0xf22411b4 // vorr d1, d20, d20
6226 .long 0xf22031b0 // vorr d3, d16, d16
6227 .long 0xe12fff33 // blx r3
6228 .long 0xe24bd008 // sub sp, fp, #8
6229 .long 0xe8bd8c10 // pop {r4, sl, fp, pc}
6230
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6231HIDDEN _sk_store_f16_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6232.globl _sk_store_f16_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6233FUNCTION(_sk_store_f16_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6234_sk_store_f16_vfp4:
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]6235 .long 0xf2630113 // vorr d16, d3, d3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6236 .long 0xe5913000 // ldr r3, [r1]
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]6237 .long 0xf2612111 // vorr d18, d1, d1
6238 .long 0xf3f67620 // vcvt.f16.f32 d23, q8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6239 .long 0xe5933000 // ldr r3, [r3]
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]6240 .long 0xf3f66602 // vcvt.f16.f32 d22, q1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6241 .long 0xe0833180 // add r3, r3, r0, lsl #3
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]6242 .long 0xf3f65622 // vcvt.f16.f32 d21, q9
6243 .long 0xf3f64600 // vcvt.f16.f32 d20, q0
6244 .long 0xf22211b2 // vorr d1, d18, d18
6245 .long 0xf22031b0 // vorr d3, d16, d16
6246 .long 0xf4c3470d // vst4.16 {d20[0],d21[0],d22[0],d23[0]}, [r3]!
6247 .long 0xf4c3474f // vst4.16 {d20[1],d21[1],d22[1],d23[1]}, [r3]
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6248 .long 0xe2813008 // add r3, r1, #8
6249 .long 0xe591c004 // ldr ip, [r1, #4]
6250 .long 0xe1a01003 // mov r1, r3
6251 .long 0xe12fff1c // bx ip
6252
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]6253HIDDEN _sk_load_u16_be_vfp4
6254.globl _sk_load_u16_be_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6255FUNCTION(_sk_load_u16_be_vfp4)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]6256_sk_load_u16_be_vfp4:
6257 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
6258 .long 0xe8911008 // ldm r1, {r3, ip}
6259 .long 0xe2811008 // add r1, r1, #8
6260 .long 0xe5933000 // ldr r3, [r3]
6261 .long 0xe0833180 // add r3, r3, r0, lsl #3
6262 .long 0xf4e3070d // vld4.16 {d16[0],d17[0],d18[0],d19[0]}, [r3]!
6263 .long 0xf4e3074f // vld4.16 {d16[1],d17[1],d18[1],d19[1]}, [r3]
6264 .long 0xee903bb0 // vmov.u16 r3, d16[0]
6265 .long 0xee92ebb0 // vmov.u16 lr, d18[0]
6266 .long 0xee914bb0 // vmov.u16 r4, d17[0]
6267 .long 0xee937bb0 // vmov.u16 r7, d19[0]
6268 .long 0xee905bf0 // vmov.u16 r5, d16[1]
6269 .long 0xee926bf0 // vmov.u16 r6, d18[1]
6270 .long 0xee043b90 // vmov.32 d20[0], r3
6271 .long 0xee05eb90 // vmov.32 d21[0], lr
6272 .long 0xee93ebf0 // vmov.u16 lr, d19[1]
6273 .long 0xee913bf0 // vmov.u16 r3, d17[1]
6274 .long 0xf3c71c1f // vmov.i32 d17, #65535
6275 .long 0xee004b90 // vmov.32 d16[0], r4
6276 .long 0xee027b90 // vmov.32 d18[0], r7
6277 .long 0xee245b90 // vmov.32 d20[1], r5
6278 .long 0xf24431b1 // vand d19, d20, d17
6279 .long 0xee256b90 // vmov.32 d21[1], r6
6280 .long 0xf2e84534 // vshl.s32 d20, d20, #8
6281 .long 0xf24561b1 // vand d22, d21, d17
6282 .long 0xf3f83033 // vshr.u32 d19, d19, #8
6283 .long 0xf2e85535 // vshl.s32 d21, d21, #8
6284 .long 0xf26431b3 // vorr d19, d20, d19
6285 .long 0xf3f86036 // vshr.u32 d22, d22, #8
6286 .long 0xf24331b1 // vand d19, d19, d17
6287 .long 0xf26551b6 // vorr d21, d21, d22
6288 .long 0xf3fb36a3 // vcvt.f32.u32 d19, d19
6289 .long 0xee22eb90 // vmov.32 d18[1], lr
6290 .long 0xee203b90 // vmov.32 d16[1], r3
6291 .long 0xf24281b1 // vand d24, d18, d17
6292 .long 0xf2e82532 // vshl.s32 d18, d18, #8
6293 .long 0xf24071b1 // vand d23, d16, d17
6294 .long 0xf3f84038 // vshr.u32 d20, d24, #8
6295 .long 0xf2e80530 // vshl.s32 d16, d16, #8
6296 .long 0xf3f87037 // vshr.u32 d23, d23, #8
6297 .long 0xf26221b4 // vorr d18, d18, d20
6298 .long 0xf26001b7 // vorr d16, d16, d23
6299 .long 0xf24541b1 // vand d20, d21, d17
6300 .long 0xf24001b1 // vand d16, d16, d17
6301 .long 0xf24211b1 // vand d17, d18, d17
6302 .long 0xeddf2b09 // vldr d18, [pc, #36]
6303 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
6304 .long 0xf3fb46a4 // vcvt.f32.u32 d20, d20
6305 .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17
6306 .long 0xf3030db2 // vmul.f32 d0, d19, d18
6307 .long 0xf3001db2 // vmul.f32 d1, d16, d18
6308 .long 0xf3042db2 // vmul.f32 d2, d20, d18
6309 .long 0xf3013db2 // vmul.f32 d3, d17, d18
6310 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
6311 .long 0xe12fff1c // bx ip
6312 .long 0xe320f000 // nop {0}
6313 .long 0x37800080 // .word 0x37800080
6314 .long 0x37800080 // .word 0x37800080
6315
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]6316HIDDEN _sk_load_rgb_u16_be_vfp4
6317.globl _sk_load_rgb_u16_be_vfp4
6318FUNCTION(_sk_load_rgb_u16_be_vfp4)
6319_sk_load_rgb_u16_be_vfp4:
6320 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
6321 .long 0xe8911008 // ldm r1, {r3, ip}
6322 .long 0xe2811008 // add r1, r1, #8
6323 .long 0xf2873f10 // vmov.f32 d3, #1
6324 .long 0xe593e000 // ldr lr, [r3]
6325 .long 0xe0803080 // add r3, r0, r0, lsl #1
6326 .long 0xe08e3083 // add r3, lr, r3, lsl #1
6327 .long 0xf4e3060d // vld3.16 {d16[0],d17[0],d18[0]}, [r3]!
6328 .long 0xf4e3064f // vld3.16 {d16[1],d17[1],d18[1]}, [r3]
6329 .long 0xee90ebb0 // vmov.u16 lr, d16[0]
6330 .long 0xee913bb0 // vmov.u16 r3, d17[0]
6331 .long 0xee924bb0 // vmov.u16 r4, d18[0]
6332 .long 0xee927bf0 // vmov.u16 r7, d18[1]
6333 .long 0xf3c73c1f // vmov.i32 d19, #65535
6334 .long 0xee905bf0 // vmov.u16 r5, d16[1]
6335 .long 0xee916bf0 // vmov.u16 r6, d17[1]
6336 .long 0xee00eb90 // vmov.32 d16[0], lr
6337 .long 0xee013b90 // vmov.32 d17[0], r3
6338 .long 0xee024b90 // vmov.32 d18[0], r4
6339 .long 0xee205b90 // vmov.32 d16[1], r5
6340 .long 0xee216b90 // vmov.32 d17[1], r6
6341 .long 0xf24041b3 // vand d20, d16, d19
6342 .long 0xf2e80530 // vshl.s32 d16, d16, #8
6343 .long 0xee227b90 // vmov.32 d18[1], r7
6344 .long 0xf24151b3 // vand d21, d17, d19
6345 .long 0xf3f84034 // vshr.u32 d20, d20, #8
6346 .long 0xf24261b3 // vand d22, d18, d19
6347 .long 0xf2e81531 // vshl.s32 d17, d17, #8
6348 .long 0xf3f85035 // vshr.u32 d21, d21, #8
6349 .long 0xf2e82532 // vshl.s32 d18, d18, #8
6350 .long 0xf3f86036 // vshr.u32 d22, d22, #8
6351 .long 0xf26001b4 // vorr d16, d16, d20
6352 .long 0xf26111b5 // vorr d17, d17, d21
6353 .long 0xf26221b6 // vorr d18, d18, d22
6354 .long 0xf24001b3 // vand d16, d16, d19
6355 .long 0xf24111b3 // vand d17, d17, d19
6356 .long 0xf24221b3 // vand d18, d18, d19
6357 .long 0xeddf3b07 // vldr d19, [pc, #28]
6358 .long 0xf3fb06a0 // vcvt.f32.u32 d16, d16
6359 .long 0xf3fb16a1 // vcvt.f32.u32 d17, d17
6360 .long 0xf3fb26a2 // vcvt.f32.u32 d18, d18
6361 .long 0xf3000db3 // vmul.f32 d0, d16, d19
6362 .long 0xf3011db3 // vmul.f32 d1, d17, d19
6363 .long 0xf3022db3 // vmul.f32 d2, d18, d19
6364 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
6365 .long 0xe12fff1c // bx ip
6366 .long 0x37800080 // .word 0x37800080
6367 .long 0x37800080 // .word 0x37800080
6368
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]6369HIDDEN _sk_store_u16_be_vfp4
6370.globl _sk_store_u16_be_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6371FUNCTION(_sk_store_u16_be_vfp4)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]6372_sk_store_u16_be_vfp4:
6373 .long 0xeddf0b2a // vldr d16, [pc, #168]
6374 .long 0xf2c3261f // vmov.i32 d18, #1056964608
6375 .long 0xf2c3361f // vmov.i32 d19, #1056964608
6376 .long 0xe5913000 // ldr r3, [r1]
6377 .long 0xf2432c30 // vfma.f32 d18, d3, d16
6378 .long 0xf2c3461f // vmov.i32 d20, #1056964608
6379 .long 0xf2423c30 // vfma.f32 d19, d2, d16
6380 .long 0xf2c3161f // vmov.i32 d17, #1056964608
6381 .long 0xf2414c30 // vfma.f32 d20, d1, d16
6382 .long 0xf2401c30 // vfma.f32 d17, d0, d16
6383 .long 0xf3fb07a2 // vcvt.u32.f32 d16, d18
6384 .long 0xf3fb27a3 // vcvt.u32.f32 d18, d19
6385 .long 0xf3c73c1f // vmov.i32 d19, #65535
6386 .long 0xf3fb47a4 // vcvt.u32.f32 d20, d20
6387 .long 0xf3fb17a1 // vcvt.u32.f32 d17, d17
6388 .long 0xf24051b3 // vand d21, d16, d19
6389 .long 0xf24261b3 // vand d22, d18, d19
6390 .long 0xf24471b3 // vand d23, d20, d19
6391 .long 0xf24131b3 // vand d19, d17, d19
6392 .long 0xf2e80530 // vshl.s32 d16, d16, #8
6393 .long 0xf3f85035 // vshr.u32 d21, d21, #8
6394 .long 0xf2e82532 // vshl.s32 d18, d18, #8
6395 .long 0xf3f86036 // vshr.u32 d22, d22, #8
6396 .long 0xf260b1b5 // vorr d27, d16, d21
6397 .long 0xf2e84534 // vshl.s32 d20, d20, #8
6398 .long 0xf3f87037 // vshr.u32 d23, d23, #8
6399 .long 0xf262a1b6 // vorr d26, d18, d22
6400 .long 0xf2e81531 // vshl.s32 d17, d17, #8
6401 .long 0xf3f83033 // vshr.u32 d19, d19, #8
6402 .long 0xf26491b7 // vorr d25, d20, d23
6403 .long 0xf26181b3 // vorr d24, d17, d19
6404 .long 0xf3f6b120 // vuzp.16 d27, d16
6405 .long 0xe5933000 // ldr r3, [r3]
6406 .long 0xf3f6a120 // vuzp.16 d26, d16
6407 .long 0xe0833180 // add r3, r3, r0, lsl #3
6408 .long 0xf3f69120 // vuzp.16 d25, d16
6409 .long 0xf3f68120 // vuzp.16 d24, d16
6410 .long 0xf4c3870d // vst4.16 {d24[0],d25[0],d26[0],d27[0]}, [r3]!
6411 .long 0xf4c3874f // vst4.16 {d24[1],d25[1],d26[1],d27[1]}, [r3]
6412 .long 0xe2813008 // add r3, r1, #8
6413 .long 0xe591c004 // ldr ip, [r1, #4]
6414 .long 0xe1a01003 // mov r1, r3
6415 .long 0xe12fff1c // bx ip
6416 .long 0xe320f000 // nop {0}
6417 .long 0x477fff00 // .word 0x477fff00
6418 .long 0x477fff00 // .word 0x477fff00
6419
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]6420HIDDEN _sk_load_f32_vfp4
6421.globl _sk_load_f32_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6422FUNCTION(_sk_load_f32_vfp4)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]6423_sk_load_f32_vfp4:
6424 .long 0xe8911008 // ldm r1, {r3, ip}
6425 .long 0xe2811008 // add r1, r1, #8
6426 .long 0xe5933000 // ldr r3, [r3]
6427 .long 0xe0833200 // add r3, r3, r0, lsl #4
6428 .long 0xf423008f // vld4.32 {d0-d3}, [r3]
6429 .long 0xe12fff1c // bx ip
6430
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6431HIDDEN _sk_store_f32_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6432.globl _sk_store_f32_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6433FUNCTION(_sk_store_f32_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6434_sk_store_f32_vfp4:
6435 .long 0xe5913000 // ldr r3, [r1]
6436 .long 0xe5933000 // ldr r3, [r3]
6437 .long 0xe0833200 // add r3, r3, r0, lsl #4
6438 .long 0xf403008f // vst4.32 {d0-d3}, [r3]
6439 .long 0xe2813008 // add r3, r1, #8
6440 .long 0xe591c004 // ldr ip, [r1, #4]
6441 .long 0xe1a01003 // mov r1, r3
6442 .long 0xe12fff1c // bx ip
6443
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6444HIDDEN _sk_clamp_x_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6445.globl _sk_clamp_x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6446FUNCTION(_sk_clamp_x_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6447_sk_clamp_x_vfp4:
6448 .long 0xe8911008 // ldm r1, {r3, ip}
6449 .long 0xf2c00010 // vmov.i32 d16, #0
6450 .long 0xf3c71e1f // vmov.i8 d17, #255
6451 .long 0xf2400f80 // vmax.f32 d16, d16, d0
6452 .long 0xe2811008 // add r1, r1, #8
6453 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6454 .long 0xf26218a1 // vadd.i32 d17, d18, d17
6455 .long 0xf2200fa1 // vmin.f32 d0, d16, d17
6456 .long 0xe12fff1c // bx ip
6457
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6458HIDDEN _sk_clamp_y_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6459.globl _sk_clamp_y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6460FUNCTION(_sk_clamp_y_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6461_sk_clamp_y_vfp4:
6462 .long 0xe8911008 // ldm r1, {r3, ip}
6463 .long 0xf2c00010 // vmov.i32 d16, #0
6464 .long 0xf3c71e1f // vmov.i8 d17, #255
6465 .long 0xf2400f81 // vmax.f32 d16, d16, d1
6466 .long 0xe2811008 // add r1, r1, #8
6467 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6468 .long 0xf26218a1 // vadd.i32 d17, d18, d17
6469 .long 0xf2201fa1 // vmin.f32 d1, d16, d17
6470 .long 0xe12fff1c // bx ip
6471
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6472HIDDEN _sk_repeat_x_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6473.globl _sk_repeat_x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6474FUNCTION(_sk_repeat_x_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6475_sk_repeat_x_vfp4:
6476 .long 0xed2d8b04 // vpush {d8-d9}
6477 .long 0xe8911008 // ldm r1, {r3, ip}
6478 .long 0xf2c02010 // vmov.i32 d18, #0
6479 .long 0xe2811008 // add r1, r1, #8
6480 .long 0xeddf3b10 // vldr d19, [pc, #64]
6481 .long 0xed938a00 // vldr s16, [r3]
6482 .long 0xeec09a88 // vdiv.f32 s19, s1, s16
6483 .long 0xee809a08 // vdiv.f32 s18, s0, s16
6484 .long 0xf3fb0709 // vcvt.s32.f32 d16, d9
6485 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
6486 .long 0xf3601e89 // vcgt.f32 d17, d16, d9
6487 .long 0xf35311b2 // vbsl d17, d19, d18
6488 .long 0xf3f42c08 // vdup.32 d18, d8[0]
6489 .long 0xf2600da1 // vsub.f32 d16, d16, d17
6490 .long 0xf3c71e1f // vmov.i8 d17, #255
6491 .long 0xf26218a1 // vadd.i32 d17, d18, d17
6492 .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0]
6493 .long 0xf2600d20 // vsub.f32 d16, d0, d16
6494 .long 0xf2200fa1 // vmin.f32 d0, d16, d17
6495 .long 0xecbd8b04 // vpop {d8-d9}
6496 .long 0xe12fff1c // bx ip
6497 .long 0xe320f000 // nop {0}
6498 .long 0x3f800000 // .word 0x3f800000
6499 .long 0x3f800000 // .word 0x3f800000
6500
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6501HIDDEN _sk_repeat_y_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6502.globl _sk_repeat_y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6503FUNCTION(_sk_repeat_y_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6504_sk_repeat_y_vfp4:
6505 .long 0xed2d8b04 // vpush {d8-d9}
6506 .long 0xe8911008 // ldm r1, {r3, ip}
6507 .long 0xf2c02010 // vmov.i32 d18, #0
6508 .long 0xe2811008 // add r1, r1, #8
6509 .long 0xeddf3b10 // vldr d19, [pc, #64]
6510 .long 0xed938a00 // vldr s16, [r3]
6511 .long 0xeec19a88 // vdiv.f32 s19, s3, s16
6512 .long 0xee819a08 // vdiv.f32 s18, s2, s16
6513 .long 0xf3fb0709 // vcvt.s32.f32 d16, d9
6514 .long 0xf3fb0620 // vcvt.f32.s32 d16, d16
6515 .long 0xf3601e89 // vcgt.f32 d17, d16, d9
6516 .long 0xf35311b2 // vbsl d17, d19, d18
6517 .long 0xf3f42c08 // vdup.32 d18, d8[0]
6518 .long 0xf2600da1 // vsub.f32 d16, d16, d17
6519 .long 0xf3c71e1f // vmov.i8 d17, #255
6520 .long 0xf26218a1 // vadd.i32 d17, d18, d17
6521 .long 0xf2e009c8 // vmul.f32 d16, d16, d8[0]
6522 .long 0xf2610d20 // vsub.f32 d16, d1, d16
6523 .long 0xf2201fa1 // vmin.f32 d1, d16, d17
6524 .long 0xecbd8b04 // vpop {d8-d9}
6525 .long 0xe12fff1c // bx ip
6526 .long 0xe320f000 // nop {0}
6527 .long 0x3f800000 // .word 0x3f800000
6528 .long 0x3f800000 // .word 0x3f800000
6529
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6530HIDDEN _sk_mirror_x_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6531.globl _sk_mirror_x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6532FUNCTION(_sk_mirror_x_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6533_sk_mirror_x_vfp4:
6534 .long 0xed2d8b04 // vpush {d8-d9}
6535 .long 0xe8911008 // ldm r1, {r3, ip}
6536 .long 0xf2c03010 // vmov.i32 d19, #0
6537 .long 0xe2811008 // add r1, r1, #8
6538 .long 0xeddf4b14 // vldr d20, [pc, #80]
6539 .long 0xed938a00 // vldr s16, [r3]
6540 .long 0xee389a08 // vadd.f32 s18, s16, s16
6541 .long 0xf3f40c08 // vdup.32 d16, d8[0]
6542 .long 0xf2200d20 // vsub.f32 d0, d0, d16
6543 .long 0xeec08a89 // vdiv.f32 s17, s1, s18
6544 .long 0xee808a09 // vdiv.f32 s16, s0, s18
6545 .long 0xf3fb1708 // vcvt.s32.f32 d17, d8
6546 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
6547 .long 0xf3612e88 // vcgt.f32 d18, d17, d8
6548 .long 0xf35421b3 // vbsl d18, d20, d19
6549 .long 0xf2611da2 // vsub.f32 d17, d17, d18
6550 .long 0xf3c72e1f // vmov.i8 d18, #255
6551 .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0]
6552 .long 0xf2601d21 // vsub.f32 d17, d0, d17
6553 .long 0xf2611da0 // vsub.f32 d17, d17, d16
6554 .long 0xf26008a2 // vadd.i32 d16, d16, d18
6555 .long 0xf3f91721 // vabs.f32 d17, d17
6556 .long 0xf2210fa0 // vmin.f32 d0, d17, d16
6557 .long 0xecbd8b04 // vpop {d8-d9}
6558 .long 0xe12fff1c // bx ip
6559 .long 0xe320f000 // nop {0}
6560 .long 0x3f800000 // .word 0x3f800000
6561 .long 0x3f800000 // .word 0x3f800000
6562
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6563HIDDEN _sk_mirror_y_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6564.globl _sk_mirror_y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6565FUNCTION(_sk_mirror_y_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6566_sk_mirror_y_vfp4:
6567 .long 0xed2d8b04 // vpush {d8-d9}
6568 .long 0xe8911008 // ldm r1, {r3, ip}
6569 .long 0xf2c03010 // vmov.i32 d19, #0
6570 .long 0xe2811008 // add r1, r1, #8
6571 .long 0xeddf4b14 // vldr d20, [pc, #80]
6572 .long 0xed938a00 // vldr s16, [r3]
6573 .long 0xee389a08 // vadd.f32 s18, s16, s16
6574 .long 0xf3f40c08 // vdup.32 d16, d8[0]
6575 .long 0xf2211d20 // vsub.f32 d1, d1, d16
6576 .long 0xeec18a89 // vdiv.f32 s17, s3, s18
6577 .long 0xee818a09 // vdiv.f32 s16, s2, s18
6578 .long 0xf3fb1708 // vcvt.s32.f32 d17, d8
6579 .long 0xf3fb1621 // vcvt.f32.s32 d17, d17
6580 .long 0xf3612e88 // vcgt.f32 d18, d17, d8
6581 .long 0xf35421b3 // vbsl d18, d20, d19
6582 .long 0xf2611da2 // vsub.f32 d17, d17, d18
6583 .long 0xf3c72e1f // vmov.i8 d18, #255
6584 .long 0xf2e119c9 // vmul.f32 d17, d17, d9[0]
6585 .long 0xf2611d21 // vsub.f32 d17, d1, d17
6586 .long 0xf2611da0 // vsub.f32 d17, d17, d16
6587 .long 0xf26008a2 // vadd.i32 d16, d16, d18
6588 .long 0xf3f91721 // vabs.f32 d17, d17
6589 .long 0xf2211fa0 // vmin.f32 d1, d17, d16
6590 .long 0xecbd8b04 // vpop {d8-d9}
6591 .long 0xe12fff1c // bx ip
6592 .long 0xe320f000 // nop {0}
6593 .long 0x3f800000 // .word 0x3f800000
6594 .long 0x3f800000 // .word 0x3f800000
6595
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6596HIDDEN _sk_luminance_to_alpha_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6597.globl _sk_luminance_to_alpha_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6598FUNCTION(_sk_luminance_to_alpha_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6599_sk_luminance_to_alpha_vfp4:
6600 .long 0xeddf0b0a // vldr d16, [pc, #40]
6601 .long 0xeddf1b0b // vldr d17, [pc, #44]
6602 .long 0xf3410d30 // vmul.f32 d16, d1, d16
6603 .long 0xe4913004 // ldr r3, [r1], #4
6604 .long 0xf3401d31 // vmul.f32 d17, d0, d17
6605 .long 0xf2800010 // vmov.i32 d0, #0
6606 .long 0xf2801010 // vmov.i32 d1, #0
6607 .long 0xf2013da0 // vadd.f32 d3, d17, d16
6608 .long 0xeddf0b06 // vldr d16, [pc, #24]
6609 .long 0xf2023c30 // vfma.f32 d3, d2, d16
6610 .long 0xf2802010 // vmov.i32 d2, #0
6611 .long 0xe12fff13 // bx r3
6612 .long 0x3f371759 // .word 0x3f371759
6613 .long 0x3f371759 // .word 0x3f371759
6614 .long 0x3e59b3d0 // .word 0x3e59b3d0
6615 .long 0x3e59b3d0 // .word 0x3e59b3d0
6616 .long 0x3d93dd98 // .word 0x3d93dd98
6617 .long 0x3d93dd98 // .word 0x3d93dd98
6618
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6619HIDDEN _sk_matrix_2x3_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6620.globl _sk_matrix_2x3_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6621FUNCTION(_sk_matrix_2x3_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6622_sk_matrix_2x3_vfp4:
6623 .long 0xe92d4800 // push {fp, lr}
6624 .long 0xe591e000 // ldr lr, [r1]
6625 .long 0xe591c004 // ldr ip, [r1, #4]
6626 .long 0xe2811008 // add r1, r1, #8
6627 .long 0xe28e300c // add r3, lr, #12
6628 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6629 .long 0xe28e3008 // add r3, lr, #8
6630 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
6631 .long 0xe28e3010 // add r3, lr, #16
6632 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
6633 .long 0xe28e3014 // add r3, lr, #20
6634 .long 0xf2410c31 // vfma.f32 d16, d1, d17
6635 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
6636 .long 0xf2411c32 // vfma.f32 d17, d1, d18
6637 .long 0xf4ee2c9d // vld1.32 {d18[]}, [lr :32]!
6638 .long 0xf4ee3c9f // vld1.32 {d19[]}, [lr :32]
6639 .long 0xf2400c32 // vfma.f32 d16, d0, d18
6640 .long 0xf2401c33 // vfma.f32 d17, d0, d19
6641 .long 0xf22001b0 // vorr d0, d16, d16
6642 .long 0xf22111b1 // vorr d1, d17, d17
6643 .long 0xe8bd4800 // pop {fp, lr}
6644 .long 0xe12fff1c // bx ip
6645
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6646HIDDEN _sk_matrix_3x4_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6647.globl _sk_matrix_3x4_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6648FUNCTION(_sk_matrix_3x4_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6649_sk_matrix_3x4_vfp4:
6650 .long 0xe92d4800 // push {fp, lr}
6651 .long 0xe591e000 // ldr lr, [r1]
6652 .long 0xe591c004 // ldr ip, [r1, #4]
6653 .long 0xe2811008 // add r1, r1, #8
6654 .long 0xe28e3020 // add r3, lr, #32
6655 .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32]
6656 .long 0xe28e302c // add r3, lr, #44
6657 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
6658 .long 0xe28e301c // add r3, lr, #28
6659 .long 0xf2420c33 // vfma.f32 d16, d2, d19
6660 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
6661 .long 0xe28e3018 // add r3, lr, #24
6662 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6663 .long 0xe28e3024 // add r3, lr, #36
6664 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
6665 .long 0xe28e3028 // add r3, lr, #40
6666 .long 0xf2421c32 // vfma.f32 d17, d2, d18
6667 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6668 .long 0xe28e3010 // add r3, lr, #16
6669 .long 0xf2422c34 // vfma.f32 d18, d2, d20
6670 .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32]
6671 .long 0xe28e300c // add r3, lr, #12
6672 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
6673 .long 0xe28e3014 // add r3, lr, #20
6674 .long 0xf2411c34 // vfma.f32 d17, d1, d20
6675 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
6676 .long 0xf2410c34 // vfma.f32 d16, d1, d20
6677 .long 0xe28e3008 // add r3, lr, #8
6678 .long 0xf2412c33 // vfma.f32 d18, d1, d19
6679 .long 0xf4ee3c9d // vld1.32 {d19[]}, [lr :32]!
6680 .long 0xf4ee4c9f // vld1.32 {d20[]}, [lr :32]
6681 .long 0xf2401c33 // vfma.f32 d17, d0, d19
6682 .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32]
6683 .long 0xf2400c33 // vfma.f32 d16, d0, d19
6684 .long 0xf2402c34 // vfma.f32 d18, d0, d20
6685 .long 0xf22101b1 // vorr d0, d17, d17
6686 .long 0xf22021b0 // vorr d2, d16, d16
6687 .long 0xf22211b2 // vorr d1, d18, d18
6688 .long 0xe8bd4800 // pop {fp, lr}
6689 .long 0xe12fff1c // bx ip
6690
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6691HIDDEN _sk_matrix_4x5_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6692.globl _sk_matrix_4x5_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6693FUNCTION(_sk_matrix_4x5_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6694_sk_matrix_4x5_vfp4:
6695 .long 0xe92d4010 // push {r4, lr}
6696 .long 0xe8911008 // ldm r1, {r3, ip}
6697 .long 0xf2620112 // vorr d16, d2, d2
6698 .long 0xe2811008 // add r1, r1, #8
6699 .long 0xe2834014 // add r4, r3, #20
6700 .long 0xe1a0e003 // mov lr, r3
6701 .long 0xf4e45c9f // vld1.32 {d21[]}, [r4 :32]
6702 .long 0xe2834028 // add r4, r3, #40
6703 .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32]
6704 .long 0xe2834038 // add r4, r3, #56
6705 .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32]
6706 .long 0xe2834048 // add r4, r3, #72
6707 .long 0xf4a42c9f // vld1.32 {d2[]}, [r4 :32]
6708 .long 0xe2834034 // add r4, r3, #52
6709 .long 0xf2032c37 // vfma.f32 d2, d3, d23
6710 .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32]
6711 .long 0xe2834044 // add r4, r3, #68
6712 .long 0xf4e41c9f // vld1.32 {d17[]}, [r4 :32]
6713 .long 0xe2834030 // add r4, r3, #48
6714 .long 0xf2431c38 // vfma.f32 d17, d3, d24
6715 .long 0xf4e49c9f // vld1.32 {d25[]}, [r4 :32]
6716 .long 0xe283403c // add r4, r3, #60
6717 .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32]
6718 .long 0xe283404c // add r4, r3, #76
6719 .long 0xf2002cb6 // vfma.f32 d2, d16, d22
6720 .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32]
6721 .long 0xe2834040 // add r4, r3, #64
6722 .long 0xf2432c33 // vfma.f32 d18, d3, d19
6723 .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32]
6724 .long 0xe2834020 // add r4, r3, #32
6725 .long 0xf2433c39 // vfma.f32 d19, d3, d25
6726 .long 0xf4e47c9f // vld1.32 {d23[]}, [r4 :32]
6727 .long 0xe283402c // add r4, r3, #44
6728 .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32]
6729 .long 0xe2834024 // add r4, r3, #36
6730 .long 0xf2402cb8 // vfma.f32 d18, d16, d24
6731 .long 0xf4e48c9f // vld1.32 {d24[]}, [r4 :32]
6732 .long 0xf2401cb8 // vfma.f32 d17, d16, d24
6733 .long 0xe2834010 // add r4, r3, #16
6734 .long 0xf2403cb7 // vfma.f32 d19, d16, d23
6735 .long 0xf4ee4c9d // vld1.32 {d20[]}, [lr :32]!
6736 .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32]
6737 .long 0xe283401c // add r4, r3, #28
6738 .long 0xf4e46c9f // vld1.32 {d22[]}, [r4 :32]
6739 .long 0xe2834018 // add r4, r3, #24
6740 .long 0xf2412c36 // vfma.f32 d18, d1, d22
6741 .long 0xf2411c35 // vfma.f32 d17, d1, d21
6742 .long 0xf4ee5c9f // vld1.32 {d21[]}, [lr :32]
6743 .long 0xf2413c30 // vfma.f32 d19, d1, d16
6744 .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32]
6745 .long 0xe2834008 // add r4, r3, #8
6746 .long 0xe283300c // add r3, r3, #12
6747 .long 0xf2012c30 // vfma.f32 d2, d1, d16
6748 .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32]
6749 .long 0xf2401c35 // vfma.f32 d17, d0, d21
6750 .long 0xf2403c34 // vfma.f32 d19, d0, d20
6751 .long 0xf4e34c9f // vld1.32 {d20[]}, [r3 :32]
6752 .long 0xf2402c34 // vfma.f32 d18, d0, d20
6753 .long 0xf2002c30 // vfma.f32 d2, d0, d16
6754 .long 0xf22111b1 // vorr d1, d17, d17
6755 .long 0xf22301b3 // vorr d0, d19, d19
6756 .long 0xf22231b2 // vorr d3, d18, d18
6757 .long 0xe8bd4010 // pop {r4, lr}
6758 .long 0xe12fff1c // bx ip
6759
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6760HIDDEN _sk_matrix_perspective_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6761.globl _sk_matrix_perspective_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6762FUNCTION(_sk_matrix_perspective_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6763_sk_matrix_perspective_vfp4:
6764 .long 0xe92d4010 // push {r4, lr}
6765 .long 0xe591e000 // ldr lr, [r1]
6766 .long 0xe591c004 // ldr ip, [r1, #4]
6767 .long 0xe2811008 // add r1, r1, #8
6768 .long 0xe28e301c // add r3, lr, #28
6769 .long 0xe28e4010 // add r4, lr, #16
6770 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
6771 .long 0xe28e3020 // add r3, lr, #32
6772 .long 0xf4e31c9f // vld1.32 {d17[]}, [r3 :32]
6773 .long 0xe28e3018 // add r3, lr, #24
6774 .long 0xf2411c30 // vfma.f32 d17, d1, d16
6775 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
6776 .long 0xe1a0300e // mov r3, lr
6777 .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32]
6778 .long 0xe28e4008 // add r4, lr, #8
6779 .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32]
6780 .long 0xf2401c30 // vfma.f32 d17, d0, d16
6781 .long 0xf4e30c9d // vld1.32 {d16[]}, [r3 :32]!
6782 .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32]
6783 .long 0xe28e3014 // add r3, lr, #20
6784 .long 0xf2413c35 // vfma.f32 d19, d1, d21
6785 .long 0xf4e35c9f // vld1.32 {d21[]}, [r3 :32]
6786 .long 0xe28e300c // add r3, lr, #12
6787 .long 0xf2415c32 // vfma.f32 d21, d1, d18
6788 .long 0xf4e32c9f // vld1.32 {d18[]}, [r3 :32]
6789 .long 0xf3fb4521 // vrecpe.f32 d20, d17
6790 .long 0xf2403c30 // vfma.f32 d19, d0, d16
6791 .long 0xf2411fb4 // vrecps.f32 d17, d17, d20
6792 .long 0xf2405c32 // vfma.f32 d21, d0, d18
6793 .long 0xf3440db1 // vmul.f32 d16, d20, d17
6794 .long 0xf3030db0 // vmul.f32 d0, d19, d16
6795 .long 0xf3051db0 // vmul.f32 d1, d21, d16
6796 .long 0xe8bd4010 // pop {r4, lr}
6797 .long 0xe12fff1c // bx ip
6798
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6799HIDDEN _sk_linear_gradient_vfp4
6800.globl _sk_linear_gradient_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6801FUNCTION(_sk_linear_gradient_vfp4)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6802_sk_linear_gradient_vfp4:
6803 .long 0xe92d4010 // push {r4, lr}
6804 .long 0xe591e000 // ldr lr, [r1]
6805 .long 0xe28e3014 // add r3, lr, #20
6806 .long 0xe1a0400e // mov r4, lr
6807 .long 0xf4a33c9f // vld1.32 {d3[]}, [r3 :32]
6808 .long 0xe28e3010 // add r3, lr, #16
6809 .long 0xf4a32c9f // vld1.32 {d2[]}, [r3 :32]
6810 .long 0xe28e3008 // add r3, lr, #8
6811 .long 0xf4e30c9f // vld1.32 {d16[]}, [r3 :32]
6812 .long 0xe494c00c // ldr ip, [r4], #12
6813 .long 0xf4a41c9f // vld1.32 {d1[]}, [r4 :32]
6814 .long 0xe35c0000 // cmp ip, #0
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]6815 .long 0x0a000036 // beq 2d70 <sk_linear_gradient_vfp4+0x110>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6816 .long 0xe59e3004 // ldr r3, [lr, #4]
6817 .long 0xf2c01010 // vmov.i32 d17, #0
6818 .long 0xf2c07010 // vmov.i32 d23, #0
6819 .long 0xf2c08010 // vmov.i32 d24, #0
6820 .long 0xe2833020 // add r3, r3, #32
6821 .long 0xf2c06010 // vmov.i32 d22, #0
6822 .long 0xe2434018 // sub r4, r3, #24
6823 .long 0xf4e33c9f // vld1.32 {d19[]}, [r3 :32]
6824 .long 0xe25cc001 // subs ip, ip, #1
6825 .long 0xf4e4dc9f // vld1.32 {d29[]}, [r4 :32]
6826 .long 0xe2434014 // sub r4, r3, #20
6827 .long 0xf4e45c9f // vld1.32 {d21[]}, [r4 :32]
6828 .long 0xe243400c // sub r4, r3, #12
6829 .long 0xf4e44c9f // vld1.32 {d20[]}, [r4 :32]
6830 .long 0xe2434020 // sub r4, r3, #32
6831 .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32]
6832 .long 0xe2434004 // sub r4, r3, #4
6833 .long 0xf3622e80 // vcgt.f32 d18, d18, d0
6834 .long 0xf4e4bc9f // vld1.32 {d27[]}, [r4 :32]
6835 .long 0xe2434008 // sub r4, r3, #8
6836 .long 0xf4e4cc9f // vld1.32 {d28[]}, [r4 :32]
6837 .long 0xe2434010 // sub r4, r3, #16
6838 .long 0xf262a1b2 // vorr d26, d18, d18
6839 .long 0xf4e4ec9f // vld1.32 {d30[]}, [r4 :32]
6840 .long 0xe243401c // sub r4, r3, #28
6841 .long 0xf352a13b // vbsl d26, d2, d27
6842 .long 0xe2833024 // add r3, r3, #36
6843 .long 0xf262b1b2 // vorr d27, d18, d18
6844 .long 0xf26291b2 // vorr d25, d18, d18
6845 .long 0xf351b13c // vbsl d27, d1, d28
6846 .long 0xf262c1b2 // vorr d28, d18, d18
6847 .long 0xf3539133 // vbsl d25, d3, d19
6848 .long 0xf350c1b4 // vbsl d28, d16, d20
6849 .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32]
6850 .long 0xf26241b2 // vorr d20, d18, d18
6851 .long 0xf26231b2 // vorr d19, d18, d18
6852 .long 0xf35841b5 // vbsl d20, d24, d21
6853 .long 0xf26251b2 // vorr d21, d18, d18
6854 .long 0xf35121b0 // vbsl d18, d17, d16
6855 .long 0xf35731be // vbsl d19, d23, d30
6856 .long 0xf35651bd // vbsl d21, d22, d29
6857 .long 0xf26211b2 // vorr d17, d18, d18
6858 .long 0xf22931b9 // vorr d3, d25, d25
6859 .long 0xf22a21ba // vorr d2, d26, d26
6860 .long 0xf22b11bb // vorr d1, d27, d27
6861 .long 0xf26c01bc // vorr d16, d28, d28
6862 .long 0xf26371b3 // vorr d23, d19, d19
6863 .long 0xf26481b4 // vorr d24, d20, d20
6864 .long 0xf26561b5 // vorr d22, d21, d21
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]6865 .long 0x1affffd3 // bne 2cac <sk_linear_gradient_vfp4+0x4c>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6866 .long 0xf26c01bc // vorr d16, d28, d28
6867 .long 0xf22b11bb // vorr d1, d27, d27
6868 .long 0xf22a21ba // vorr d2, d26, d26
6869 .long 0xf22931b9 // vorr d3, d25, d25
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]6870 .long 0xea000003 // b 2d80 <sk_linear_gradient_vfp4+0x120>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6871 .long 0xf2c05010 // vmov.i32 d21, #0
6872 .long 0xf2c04010 // vmov.i32 d20, #0
6873 .long 0xf2c03010 // vmov.i32 d19, #0
6874 .long 0xf2c02010 // vmov.i32 d18, #0
6875 .long 0xf2400c32 // vfma.f32 d16, d0, d18
6876 .long 0xe5913004 // ldr r3, [r1, #4]
6877 .long 0xf2001c35 // vfma.f32 d1, d0, d21
6878 .long 0xe2811008 // add r1, r1, #8
6879 .long 0xf2002c34 // vfma.f32 d2, d0, d20
6880 .long 0xf2003c33 // vfma.f32 d3, d0, d19
6881 .long 0xf22001b0 // vorr d0, d16, d16
6882 .long 0xe8bd4010 // pop {r4, lr}
6883 .long 0xe12fff13 // bx r3
6884
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]6885HIDDEN _sk_linear_gradient_2stops_vfp4
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6886.globl _sk_linear_gradient_2stops_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6887FUNCTION(_sk_linear_gradient_2stops_vfp4)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6888_sk_linear_gradient_2stops_vfp4:
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6889 .long 0xe92d4010 // push {r4, lr}
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6890 .long 0xe8911008 // ldm r1, {r3, ip}
6891 .long 0xe2811008 // add r1, r1, #8
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6892 .long 0xe283400c // add r4, r3, #12
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6893 .long 0xe1a0e003 // mov lr, r3
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6894 .long 0xf4e42c9f // vld1.32 {d18[]}, [r4 :32]
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6895 .long 0xe2834008 // add r4, r3, #8
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6896 .long 0xf4e43c9f // vld1.32 {d19[]}, [r4 :32]
6897 .long 0xe2834018 // add r4, r3, #24
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6898 .long 0xf4a42c9f // vld1.32 {d2[]}, [r4 :32]
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]6899 .long 0xe2834010 // add r4, r3, #16
6900 .long 0xf2002c33 // vfma.f32 d2, d0, d19
6901 .long 0xf4e40c9f // vld1.32 {d16[]}, [r4 :32]
6902 .long 0xe283401c // add r4, r3, #28
6903 .long 0xe2833014 // add r3, r3, #20
6904 .long 0xf4ee1c9d // vld1.32 {d17[]}, [lr :32]!
6905 .long 0xf2400c31 // vfma.f32 d16, d0, d17
6906 .long 0xf4a43c9f // vld1.32 {d3[]}, [r4 :32]
6907 .long 0xf2003c32 // vfma.f32 d3, d0, d18
6908 .long 0xf4ee1c9f // vld1.32 {d17[]}, [lr :32]
6909 .long 0xf4a31c9f // vld1.32 {d1[]}, [r3 :32]
6910 .long 0xf2001c31 // vfma.f32 d1, d0, d17
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6911 .long 0xf22001b0 // vorr d0, d16, d16
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]6912 .long 0xe8bd4010 // pop {r4, lr}
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]6913 .long 0xe12fff1c // bx ip
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6914
6915HIDDEN _sk_save_xy_vfp4
6916.globl _sk_save_xy_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6917FUNCTION(_sk_save_xy_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6918_sk_save_xy_vfp4:
6919 .long 0xf2c3061f // vmov.i32 d16, #1056964608
6920 .long 0xeddf7b17 // vldr d23, [pc, #92]
6921 .long 0xf2c06010 // vmov.i32 d22, #0
6922 .long 0xe5913000 // ldr r3, [r1]
6923 .long 0xf2401d20 // vadd.f32 d17, d0, d16
6924 .long 0xf2410d20 // vadd.f32 d16, d1, d16
6925 .long 0xed830b00 // vstr d0, [r3]
6926 .long 0xed831b08 // vstr d1, [r3, #32]
6927 .long 0xf3fb2721 // vcvt.s32.f32 d18, d17
6928 .long 0xf3fb3720 // vcvt.s32.f32 d19, d16
6929 .long 0xf3fb2622 // vcvt.f32.s32 d18, d18
6930 .long 0xf3fb3623 // vcvt.f32.s32 d19, d19
6931 .long 0xf3624ea1 // vcgt.f32 d20, d18, d17
6932 .long 0xf3635ea0 // vcgt.f32 d21, d19, d16
6933 .long 0xf35741b6 // vbsl d20, d23, d22
6934 .long 0xf35751b6 // vbsl d21, d23, d22
6935 .long 0xf2622da4 // vsub.f32 d18, d18, d20
6936 .long 0xf2633da5 // vsub.f32 d19, d19, d21
6937 .long 0xf2611da2 // vsub.f32 d17, d17, d18
6938 .long 0xf2600da3 // vsub.f32 d16, d16, d19
6939 .long 0xedc31b10 // vstr d17, [r3, #64]
6940 .long 0xedc30b18 // vstr d16, [r3, #96]
6941 .long 0xe2813008 // add r3, r1, #8
6942 .long 0xe591c004 // ldr ip, [r1, #4]
6943 .long 0xe1a01003 // mov r1, r3
6944 .long 0xe12fff1c // bx ip
6945 .long 0x3f800000 // .word 0x3f800000
6946 .long 0x3f800000 // .word 0x3f800000
6947
6948HIDDEN _sk_accumulate_vfp4
6949.globl _sk_accumulate_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6950FUNCTION(_sk_accumulate_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6951_sk_accumulate_vfp4:
6952 .long 0xe8911008 // ldm r1, {r3, ip}
6953 .long 0xe2811008 // add r1, r1, #8
6954 .long 0xedd31b28 // vldr d17, [r3, #160]
6955 .long 0xedd30b20 // vldr d16, [r3, #128]
6956 .long 0xf3400db1 // vmul.f32 d16, d16, d17
6957 .long 0xf2004c90 // vfma.f32 d4, d16, d0
6958 .long 0xf2005c91 // vfma.f32 d5, d16, d1
6959 .long 0xf2006c92 // vfma.f32 d6, d16, d2
6960 .long 0xf2007c93 // vfma.f32 d7, d16, d3
6961 .long 0xe12fff1c // bx ip
6962
6963HIDDEN _sk_bilinear_nx_vfp4
6964.globl _sk_bilinear_nx_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6965FUNCTION(_sk_bilinear_nx_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6966_sk_bilinear_nx_vfp4:
6967 .long 0xe5913000 // ldr r3, [r1]
6968 .long 0xf2c70f10 // vmov.f32 d16, #1
6969 .long 0xedd32b10 // vldr d18, [r3, #64]
6970 .long 0xf2600da2 // vsub.f32 d16, d16, d18
6971 .long 0xedd31b00 // vldr d17, [r3]
6972 .long 0xf3c3261f // vmov.i32 d18, #-1090519040
6973 .long 0xf2010da2 // vadd.f32 d0, d17, d18
6974 .long 0xedc30b20 // vstr d16, [r3, #128]
6975 .long 0xe2813008 // add r3, r1, #8
6976 .long 0xe591c004 // ldr ip, [r1, #4]
6977 .long 0xe1a01003 // mov r1, r3
6978 .long 0xe12fff1c // bx ip
6979
6980HIDDEN _sk_bilinear_px_vfp4
6981.globl _sk_bilinear_px_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6982FUNCTION(_sk_bilinear_px_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6983_sk_bilinear_px_vfp4:
6984 .long 0xe5913000 // ldr r3, [r1]
6985 .long 0xf2c3061f // vmov.i32 d16, #1056964608
6986 .long 0xedd31b00 // vldr d17, [r3]
6987 .long 0xedd32b10 // vldr d18, [r3, #64]
6988 .long 0xf2010da0 // vadd.f32 d0, d17, d16
6989 .long 0xedc32b20 // vstr d18, [r3, #128]
6990 .long 0xe2813008 // add r3, r1, #8
6991 .long 0xe591c004 // ldr ip, [r1, #4]
6992 .long 0xe1a01003 // mov r1, r3
6993 .long 0xe12fff1c // bx ip
6994
6995HIDDEN _sk_bilinear_ny_vfp4
6996.globl _sk_bilinear_ny_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]6997FUNCTION(_sk_bilinear_ny_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]6998_sk_bilinear_ny_vfp4:
6999 .long 0xe5913000 // ldr r3, [r1]
7000 .long 0xf2c70f10 // vmov.f32 d16, #1
7001 .long 0xedd32b18 // vldr d18, [r3, #96]
7002 .long 0xf2600da2 // vsub.f32 d16, d16, d18
7003 .long 0xedd31b08 // vldr d17, [r3, #32]
7004 .long 0xf3c3261f // vmov.i32 d18, #-1090519040
7005 .long 0xf2011da2 // vadd.f32 d1, d17, d18
7006 .long 0xedc30b28 // vstr d16, [r3, #160]
7007 .long 0xe2813008 // add r3, r1, #8
7008 .long 0xe591c004 // ldr ip, [r1, #4]
7009 .long 0xe1a01003 // mov r1, r3
7010 .long 0xe12fff1c // bx ip
7011
7012HIDDEN _sk_bilinear_py_vfp4
7013.globl _sk_bilinear_py_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7014FUNCTION(_sk_bilinear_py_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7015_sk_bilinear_py_vfp4:
7016 .long 0xe5913000 // ldr r3, [r1]
7017 .long 0xf2c3061f // vmov.i32 d16, #1056964608
7018 .long 0xedd31b08 // vldr d17, [r3, #32]
7019 .long 0xedd32b18 // vldr d18, [r3, #96]
7020 .long 0xf2011da0 // vadd.f32 d1, d17, d16
7021 .long 0xedc32b28 // vstr d18, [r3, #160]
7022 .long 0xe2813008 // add r3, r1, #8
7023 .long 0xe591c004 // ldr ip, [r1, #4]
7024 .long 0xe1a01003 // mov r1, r3
7025 .long 0xe12fff1c // bx ip
7026
7027HIDDEN _sk_bicubic_n3x_vfp4
7028.globl _sk_bicubic_n3x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7029FUNCTION(_sk_bicubic_n3x_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7030_sk_bicubic_n3x_vfp4:
7031 .long 0xe5913000 // ldr r3, [r1]
7032 .long 0xf2c70f10 // vmov.f32 d16, #1
7033 .long 0xeddf3b10 // vldr d19, [pc, #64]
7034 .long 0xedd32b10 // vldr d18, [r3, #64]
7035 .long 0xf2600da2 // vsub.f32 d16, d16, d18
7036 .long 0xeddf2b0b // vldr d18, [pc, #44]
7037 .long 0xedd31b00 // vldr d17, [r3]
7038 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7039 .long 0xf3400db0 // vmul.f32 d16, d16, d16
7040 .long 0xf3c72f18 // vmov.f32 d18, #-1.5
7041 .long 0xf2010da2 // vadd.f32 d0, d17, d18
7042 .long 0xf3400db3 // vmul.f32 d16, d16, d19
7043 .long 0xedc30b20 // vstr d16, [r3, #128]
7044 .long 0xe2813008 // add r3, r1, #8
7045 .long 0xe591c004 // ldr ip, [r1, #4]
7046 .long 0xe1a01003 // mov r1, r3
7047 .long 0xe12fff1c // bx ip
7048 .long 0xe320f000 // nop {0}
7049 .long 0x3ec71c72 // .word 0x3ec71c72
7050 .long 0x3ec71c72 // .word 0x3ec71c72
7051 .long 0xbeaaaaab // .word 0xbeaaaaab
7052 .long 0xbeaaaaab // .word 0xbeaaaaab
7053
7054HIDDEN _sk_bicubic_n1x_vfp4
7055.globl _sk_bicubic_n1x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7056FUNCTION(_sk_bicubic_n1x_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7057_sk_bicubic_n1x_vfp4:
7058 .long 0xe5913000 // ldr r3, [r1]
7059 .long 0xf2c70f10 // vmov.f32 d16, #1
7060 .long 0xf2c73f18 // vmov.f32 d19, #1.5
7061 .long 0xedd32b10 // vldr d18, [r3, #64]
7062 .long 0xf2600da2 // vsub.f32 d16, d16, d18
7063 .long 0xeddf2b0d // vldr d18, [pc, #52]
7064 .long 0xedd31b00 // vldr d17, [r3]
7065 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7066 .long 0xf2c3261f // vmov.i32 d18, #1056964608
7067 .long 0xf2402cb3 // vfma.f32 d18, d16, d19
7068 .long 0xeddf3b0a // vldr d19, [pc, #40]
7069 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7070 .long 0xf3c3061f // vmov.i32 d16, #-1090519040
7071 .long 0xf2010da0 // vadd.f32 d0, d17, d16
7072 .long 0xedc33b20 // vstr d19, [r3, #128]
7073 .long 0xe2813008 // add r3, r1, #8
7074 .long 0xe591c004 // ldr ip, [r1, #4]
7075 .long 0xe1a01003 // mov r1, r3
7076 .long 0xe12fff1c // bx ip
7077 .long 0xe320f000 // nop {0}
7078 .long 0xbf955555 // .word 0xbf955555
7079 .long 0xbf955555 // .word 0xbf955555
7080 .long 0x3d638e39 // .word 0x3d638e39
7081 .long 0x3d638e39 // .word 0x3d638e39
7082
7083HIDDEN _sk_bicubic_p1x_vfp4
7084.globl _sk_bicubic_p1x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7085FUNCTION(_sk_bicubic_p1x_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7086_sk_bicubic_p1x_vfp4:
7087 .long 0xe5913000 // ldr r3, [r1]
7088 .long 0xf2c71f18 // vmov.f32 d17, #1.5
7089 .long 0xeddf0b0c // vldr d16, [pc, #48]
7090 .long 0xedd33b10 // vldr d19, [r3, #64]
7091 .long 0xf2431cb0 // vfma.f32 d17, d19, d16
7092 .long 0xedd32b00 // vldr d18, [r3]
7093 .long 0xf2c3061f // vmov.i32 d16, #1056964608
7094 .long 0xf2020da0 // vadd.f32 d0, d18, d16
7095 .long 0xf2430cb1 // vfma.f32 d16, d19, d17
7096 .long 0xeddf1b07 // vldr d17, [pc, #28]
7097 .long 0xf2431cb0 // vfma.f32 d17, d19, d16
7098 .long 0xedc31b20 // vstr d17, [r3, #128]
7099 .long 0xe2813008 // add r3, r1, #8
7100 .long 0xe591c004 // ldr ip, [r1, #4]
7101 .long 0xe1a01003 // mov r1, r3
7102 .long 0xe12fff1c // bx ip
7103 .long 0xbf955555 // .word 0xbf955555
7104 .long 0xbf955555 // .word 0xbf955555
7105 .long 0x3d638e39 // .word 0x3d638e39
7106 .long 0x3d638e39 // .word 0x3d638e39
7107
7108HIDDEN _sk_bicubic_p3x_vfp4
7109.globl _sk_bicubic_p3x_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7110FUNCTION(_sk_bicubic_p3x_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7111_sk_bicubic_p3x_vfp4:
7112 .long 0xe5913000 // ldr r3, [r1]
7113 .long 0xeddf0b0d // vldr d16, [pc, #52]
7114 .long 0xeddf3b0e // vldr d19, [pc, #56]
7115 .long 0xedd32b10 // vldr d18, [r3, #64]
7116 .long 0xf2423cb0 // vfma.f32 d19, d18, d16
7117 .long 0xedd31b00 // vldr d17, [r3]
7118 .long 0xf3420db2 // vmul.f32 d16, d18, d18
7119 .long 0xf2c72f18 // vmov.f32 d18, #1.5
7120 .long 0xf2010da2 // vadd.f32 d0, d17, d18
7121 .long 0xf3400db3 // vmul.f32 d16, d16, d19
7122 .long 0xedc30b20 // vstr d16, [r3, #128]
7123 .long 0xe2813008 // add r3, r1, #8
7124 .long 0xe591c004 // ldr ip, [r1, #4]
7125 .long 0xe1a01003 // mov r1, r3
7126 .long 0xe12fff1c // bx ip
7127 .long 0xe320f000 // nop {0}
7128 .long 0x3ec71c72 // .word 0x3ec71c72
7129 .long 0x3ec71c72 // .word 0x3ec71c72
7130 .long 0xbeaaaaab // .word 0xbeaaaaab
7131 .long 0xbeaaaaab // .word 0xbeaaaaab
7132
7133HIDDEN _sk_bicubic_n3y_vfp4
7134.globl _sk_bicubic_n3y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7135FUNCTION(_sk_bicubic_n3y_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7136_sk_bicubic_n3y_vfp4:
7137 .long 0xe5913000 // ldr r3, [r1]
7138 .long 0xf2c70f10 // vmov.f32 d16, #1
7139 .long 0xeddf3b10 // vldr d19, [pc, #64]
7140 .long 0xedd32b18 // vldr d18, [r3, #96]
7141 .long 0xf2600da2 // vsub.f32 d16, d16, d18
7142 .long 0xeddf2b0b // vldr d18, [pc, #44]
7143 .long 0xedd31b08 // vldr d17, [r3, #32]
7144 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7145 .long 0xf3400db0 // vmul.f32 d16, d16, d16
7146 .long 0xf3c72f18 // vmov.f32 d18, #-1.5
7147 .long 0xf2011da2 // vadd.f32 d1, d17, d18
7148 .long 0xf3400db3 // vmul.f32 d16, d16, d19
7149 .long 0xedc30b28 // vstr d16, [r3, #160]
7150 .long 0xe2813008 // add r3, r1, #8
7151 .long 0xe591c004 // ldr ip, [r1, #4]
7152 .long 0xe1a01003 // mov r1, r3
7153 .long 0xe12fff1c // bx ip
7154 .long 0xe320f000 // nop {0}
7155 .long 0x3ec71c72 // .word 0x3ec71c72
7156 .long 0x3ec71c72 // .word 0x3ec71c72
7157 .long 0xbeaaaaab // .word 0xbeaaaaab
7158 .long 0xbeaaaaab // .word 0xbeaaaaab
7159
7160HIDDEN _sk_bicubic_n1y_vfp4
7161.globl _sk_bicubic_n1y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7162FUNCTION(_sk_bicubic_n1y_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7163_sk_bicubic_n1y_vfp4:
7164 .long 0xe5913000 // ldr r3, [r1]
7165 .long 0xf2c70f10 // vmov.f32 d16, #1
7166 .long 0xf2c73f18 // vmov.f32 d19, #1.5
7167 .long 0xedd32b18 // vldr d18, [r3, #96]
7168 .long 0xf2600da2 // vsub.f32 d16, d16, d18
7169 .long 0xeddf2b0d // vldr d18, [pc, #52]
7170 .long 0xedd31b08 // vldr d17, [r3, #32]
7171 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7172 .long 0xf2c3261f // vmov.i32 d18, #1056964608
7173 .long 0xf2402cb3 // vfma.f32 d18, d16, d19
7174 .long 0xeddf3b0a // vldr d19, [pc, #40]
7175 .long 0xf2403cb2 // vfma.f32 d19, d16, d18
7176 .long 0xf3c3061f // vmov.i32 d16, #-1090519040
7177 .long 0xf2011da0 // vadd.f32 d1, d17, d16
7178 .long 0xedc33b28 // vstr d19, [r3, #160]
7179 .long 0xe2813008 // add r3, r1, #8
7180 .long 0xe591c004 // ldr ip, [r1, #4]
7181 .long 0xe1a01003 // mov r1, r3
7182 .long 0xe12fff1c // bx ip
7183 .long 0xe320f000 // nop {0}
7184 .long 0xbf955555 // .word 0xbf955555
7185 .long 0xbf955555 // .word 0xbf955555
7186 .long 0x3d638e39 // .word 0x3d638e39
7187 .long 0x3d638e39 // .word 0x3d638e39
7188
7189HIDDEN _sk_bicubic_p1y_vfp4
7190.globl _sk_bicubic_p1y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7191FUNCTION(_sk_bicubic_p1y_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7192_sk_bicubic_p1y_vfp4:
7193 .long 0xe5913000 // ldr r3, [r1]
7194 .long 0xf2c71f18 // vmov.f32 d17, #1.5
7195 .long 0xeddf0b0c // vldr d16, [pc, #48]
7196 .long 0xedd33b18 // vldr d19, [r3, #96]
7197 .long 0xf2431cb0 // vfma.f32 d17, d19, d16
7198 .long 0xedd32b08 // vldr d18, [r3, #32]
7199 .long 0xf2c3061f // vmov.i32 d16, #1056964608
7200 .long 0xf2021da0 // vadd.f32 d1, d18, d16
7201 .long 0xf2430cb1 // vfma.f32 d16, d19, d17
7202 .long 0xeddf1b07 // vldr d17, [pc, #28]
7203 .long 0xf2431cb0 // vfma.f32 d17, d19, d16
7204 .long 0xedc31b28 // vstr d17, [r3, #160]
7205 .long 0xe2813008 // add r3, r1, #8
7206 .long 0xe591c004 // ldr ip, [r1, #4]
7207 .long 0xe1a01003 // mov r1, r3
7208 .long 0xe12fff1c // bx ip
7209 .long 0xbf955555 // .word 0xbf955555
7210 .long 0xbf955555 // .word 0xbf955555
7211 .long 0x3d638e39 // .word 0x3d638e39
7212 .long 0x3d638e39 // .word 0x3d638e39
7213
7214HIDDEN _sk_bicubic_p3y_vfp4
7215.globl _sk_bicubic_p3y_vfp4
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7216FUNCTION(_sk_bicubic_p3y_vfp4)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]7217_sk_bicubic_p3y_vfp4:
7218 .long 0xe5913000 // ldr r3, [r1]
7219 .long 0xeddf0b0d // vldr d16, [pc, #52]
7220 .long 0xeddf3b0e // vldr d19, [pc, #56]
7221 .long 0xedd32b18 // vldr d18, [r3, #96]
7222 .long 0xf2423cb0 // vfma.f32 d19, d18, d16
7223 .long 0xedd31b08 // vldr d17, [r3, #32]
7224 .long 0xf3420db2 // vmul.f32 d16, d18, d18
7225 .long 0xf2c72f18 // vmov.f32 d18, #1.5
7226 .long 0xf2011da2 // vadd.f32 d1, d17, d18
7227 .long 0xf3400db3 // vmul.f32 d16, d16, d19
7228 .long 0xedc30b28 // vstr d16, [r3, #160]
7229 .long 0xe2813008 // add r3, r1, #8
7230 .long 0xe591c004 // ldr ip, [r1, #4]
7231 .long 0xe1a01003 // mov r1, r3
7232 .long 0xe12fff1c // bx ip
7233 .long 0xe320f000 // nop {0}
7234 .long 0x3ec71c72 // .word 0x3ec71c72
7235 .long 0x3ec71c72 // .word 0x3ec71c72
7236 .long 0xbeaaaaab // .word 0xbeaaaaab
7237 .long 0xbeaaaaab // .word 0xbeaaaaab
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7238
7239HIDDEN _sk_callback_vfp4
7240.globl _sk_callback_vfp4
7241FUNCTION(_sk_callback_vfp4)
7242_sk_callback_vfp4:
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7243 .long 0xe92d48f0 // push {r4, r5, r6, r7, fp, lr}
7244 .long 0xed2d8b08 // vpush {d8-d11}
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7245 .long 0xe1a05001 // mov r5, r1
7246 .long 0xe1a06000 // mov r6, r0
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7247 .long 0xe5957000 // ldr r7, [r5]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7248 .long 0xe1a04002 // mov r4, r2
7249 .long 0xe3a01002 // mov r1, #2
7250 .long 0xeeb08b47 // vmov.f64 d8, d7
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7251 .long 0xe2870004 // add r0, r7, #4
7252 .long 0xf400008f // vst4.32 {d0-d3}, [r0]
7253 .long 0xe1a00007 // mov r0, r7
7254 .long 0xe5972000 // ldr r2, [r7]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7255 .long 0xeeb09b46 // vmov.f64 d9, d6
7256 .long 0xeeb0ab45 // vmov.f64 d10, d5
7257 .long 0xeeb0bb44 // vmov.f64 d11, d4
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7258 .long 0xe12fff32 // blx r2
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7259 .long 0xe5970084 // ldr r0, [r7, #132]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7260 .long 0xe2851008 // add r1, r5, #8
7261 .long 0xe5953004 // ldr r3, [r5, #4]
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7262 .long 0xe1a02004 // mov r2, r4
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7263 .long 0xeeb04b4b // vmov.f64 d4, d11
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7264 .long 0xf420008f // vld4.32 {d0-d3}, [r0]
7265 .long 0xe1a00006 // mov r0, r6
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7266 .long 0xeeb05b4a // vmov.f64 d5, d10
7267 .long 0xeeb06b49 // vmov.f64 d6, d9
7268 .long 0xeeb07b48 // vmov.f64 d7, d8
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]7269 .long 0xecbd8b08 // vpop {d8-d11}
7270 .long 0xe8bd48f0 // pop {r4, r5, r6, r7, fp, lr}
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]7271 .long 0xe12fff13 // bx r3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7272#elif defined(__x86_64__)
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]7273BALIGN32
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7274
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7275HIDDEN _sk_start_pipeline_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7276.globl _sk_start_pipeline_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7277FUNCTION(_sk_start_pipeline_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7278_sk_start_pipeline_hsw:
7279 .byte 65,87 // push %r15
7280 .byte 65,86 // push %r14
7281 .byte 65,85 // push %r13
7282 .byte 65,84 // push %r12
7283 .byte 83 // push %rbx
7284 .byte 73,137,205 // mov %rcx,%r13
7285 .byte 73,137,214 // mov %rdx,%r14
7286 .byte 72,137,251 // mov %rdi,%rbx
7287 .byte 72,173 // lods %ds:(%rsi),%rax
7288 .byte 73,137,199 // mov %rax,%r15
7289 .byte 73,137,244 // mov %rsi,%r12
7290 .byte 72,141,67,8 // lea 0x8(%rbx),%rax
7291 .byte 76,57,232 // cmp %r13,%rax
7292 .byte 118,5 // jbe 28 <_sk_start_pipeline_hsw+0x28>
7293 .byte 72,137,223 // mov %rbx,%rdi
7294 .byte 235,65 // jmp 69 <_sk_start_pipeline_hsw+0x69>
7295 .byte 185,0,0,0,0 // mov $0x0,%ecx
7296 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
7297 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
7298 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
7299 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
7300 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
7301 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
7302 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
7303 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
7304 .byte 72,137,223 // mov %rbx,%rdi
7305 .byte 76,137,230 // mov %r12,%rsi
7306 .byte 76,137,242 // mov %r14,%rdx
7307 .byte 65,255,215 // callq *%r15
7308 .byte 72,141,123,8 // lea 0x8(%rbx),%rdi
7309 .byte 72,131,195,16 // add $0x10,%rbx
7310 .byte 76,57,235 // cmp %r13,%rbx
7311 .byte 72,137,251 // mov %rdi,%rbx
7312 .byte 118,191 // jbe 28 <_sk_start_pipeline_hsw+0x28>
7313 .byte 76,137,233 // mov %r13,%rcx
7314 .byte 72,41,249 // sub %rdi,%rcx
7315 .byte 116,41 // je 9a <_sk_start_pipeline_hsw+0x9a>
7316 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
7317 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
7318 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
7319 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
7320 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
7321 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
7322 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
7323 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
7324 .byte 76,137,230 // mov %r12,%rsi
7325 .byte 76,137,242 // mov %r14,%rdx
7326 .byte 65,255,215 // callq *%r15
7327 .byte 76,137,232 // mov %r13,%rax
7328 .byte 91 // pop %rbx
7329 .byte 65,92 // pop %r12
7330 .byte 65,93 // pop %r13
7331 .byte 65,94 // pop %r14
7332 .byte 65,95 // pop %r15
7333 .byte 197,248,119 // vzeroupper
7334 .byte 195 // retq
7335
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7336HIDDEN _sk_just_return_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7337.globl _sk_just_return_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7338FUNCTION(_sk_just_return_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7339_sk_just_return_hsw:
7340 .byte 195 // retq
7341
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7342HIDDEN _sk_seed_shader_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7343.globl _sk_seed_shader_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7344FUNCTION(_sk_seed_shader_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7345_sk_seed_shader_hsw:
7346 .byte 72,173 // lods %ds:(%rsi),%rax
7347 .byte 197,249,110,199 // vmovd %edi,%xmm0
7348 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
7349 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]7350 .byte 196,226,125,24,13,61,64,0,0 // vbroadcastss 0x403d(%rip),%ymm1 # 4100 <_sk_callback_hsw+0x126>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7351 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
7352 .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0
7353 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2
7354 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
7355 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7356 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]7357 .byte 196,226,125,24,21,33,64,0,0 // vbroadcastss 0x4021(%rip),%ymm2 # 4104 <_sk_callback_hsw+0x12a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7358 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
7359 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
7360 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
7361 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
7362 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
7363 .byte 255,224 // jmpq *%rax
7364
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7365HIDDEN _sk_constant_color_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7366.globl _sk_constant_color_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7367FUNCTION(_sk_constant_color_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7368_sk_constant_color_hsw:
7369 .byte 72,173 // lods %ds:(%rsi),%rax
7370 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
7371 .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1
7372 .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2
7373 .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3
7374 .byte 72,173 // lods %ds:(%rsi),%rax
7375 .byte 255,224 // jmpq *%rax
7376
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7377HIDDEN _sk_clear_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7378.globl _sk_clear_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7379FUNCTION(_sk_clear_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7380_sk_clear_hsw:
7381 .byte 72,173 // lods %ds:(%rsi),%rax
7382 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
7383 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
7384 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
7385 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
7386 .byte 255,224 // jmpq *%rax
7387
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7388HIDDEN _sk_srcatop_hsw
7389.globl _sk_srcatop_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7390FUNCTION(_sk_srcatop_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7391_sk_srcatop_hsw:
7392 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7393 .byte 197,121,110,192 // vmovd %eax,%xmm8
7394 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7395 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7396 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9
7397 .byte 196,194,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm0
7398 .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9
7399 .byte 196,194,69,168,201 // vfmadd213ps %ymm9,%ymm7,%ymm1
7400 .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9
7401 .byte 196,194,69,168,209 // vfmadd213ps %ymm9,%ymm7,%ymm2
7402 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
7403 .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3
7404 .byte 72,173 // lods %ds:(%rsi),%rax
7405 .byte 255,224 // jmpq *%rax
7406
7407HIDDEN _sk_dstatop_hsw
7408.globl _sk_dstatop_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7409FUNCTION(_sk_dstatop_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7410_sk_dstatop_hsw:
7411 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7412 .byte 197,121,110,192 // vmovd %eax,%xmm8
7413 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7414 .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
7415 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
7416 .byte 196,226,101,184,196 // vfmadd231ps %ymm4,%ymm3,%ymm0
7417 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
7418 .byte 196,226,101,184,205 // vfmadd231ps %ymm5,%ymm3,%ymm1
7419 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
7420 .byte 196,226,101,184,214 // vfmadd231ps %ymm6,%ymm3,%ymm2
7421 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
7422 .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3
7423 .byte 72,173 // lods %ds:(%rsi),%rax
7424 .byte 255,224 // jmpq *%rax
7425
7426HIDDEN _sk_srcin_hsw
7427.globl _sk_srcin_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7428FUNCTION(_sk_srcin_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7429_sk_srcin_hsw:
7430 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
7431 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
7432 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
7433 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
7434 .byte 72,173 // lods %ds:(%rsi),%rax
7435 .byte 255,224 // jmpq *%rax
7436
7437HIDDEN _sk_dstin_hsw
7438.globl _sk_dstin_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7439FUNCTION(_sk_dstin_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7440_sk_dstin_hsw:
7441 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
7442 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
7443 .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2
7444 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
7445 .byte 72,173 // lods %ds:(%rsi),%rax
7446 .byte 255,224 // jmpq *%rax
7447
7448HIDDEN _sk_srcout_hsw
7449.globl _sk_srcout_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7450FUNCTION(_sk_srcout_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7451_sk_srcout_hsw:
7452 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7453 .byte 197,121,110,192 // vmovd %eax,%xmm8
7454 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7455 .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
7456 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
7457 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
7458 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
7459 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
7460 .byte 72,173 // lods %ds:(%rsi),%rax
7461 .byte 255,224 // jmpq *%rax
7462
7463HIDDEN _sk_dstout_hsw
7464.globl _sk_dstout_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7465FUNCTION(_sk_dstout_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7466_sk_dstout_hsw:
7467 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7468 .byte 197,249,110,192 // vmovd %eax,%xmm0
7469 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
7470 .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3
7471 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
7472 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
7473 .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2
7474 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7475 .byte 72,173 // lods %ds:(%rsi),%rax
7476 .byte 255,224 // jmpq *%rax
7477
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7478HIDDEN _sk_srcover_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7479.globl _sk_srcover_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7480FUNCTION(_sk_srcover_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7481_sk_srcover_hsw:
7482 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7483 .byte 197,121,110,192 // vmovd %eax,%xmm8
7484 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7485 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7486 .byte 196,194,93,184,192 // vfmadd231ps %ymm8,%ymm4,%ymm0
7487 .byte 196,194,85,184,200 // vfmadd231ps %ymm8,%ymm5,%ymm1
7488 .byte 196,194,77,184,208 // vfmadd231ps %ymm8,%ymm6,%ymm2
7489 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7490 .byte 72,173 // lods %ds:(%rsi),%rax
7491 .byte 255,224 // jmpq *%rax
7492
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]7493HIDDEN _sk_dstover_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7494.globl _sk_dstover_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7495FUNCTION(_sk_dstover_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]7496_sk_dstover_hsw:
7497 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7498 .byte 197,121,110,192 // vmovd %eax,%xmm8
7499 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7500 .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
7501 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
7502 .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1
7503 .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2
7504 .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3
7505 .byte 72,173 // lods %ds:(%rsi),%rax
7506 .byte 255,224 // jmpq *%rax
7507
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7508HIDDEN _sk_modulate_hsw
7509.globl _sk_modulate_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7510FUNCTION(_sk_modulate_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7511_sk_modulate_hsw:
7512 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
7513 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
7514 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
7515 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
7516 .byte 72,173 // lods %ds:(%rsi),%rax
7517 .byte 255,224 // jmpq *%rax
7518
7519HIDDEN _sk_multiply_hsw
7520.globl _sk_multiply_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7521FUNCTION(_sk_multiply_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7522_sk_multiply_hsw:
7523 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7524 .byte 197,121,110,192 // vmovd %eax,%xmm8
7525 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7526 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
7527 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7528 .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10
7529 .byte 196,98,53,184,208 // vfmadd231ps %ymm0,%ymm9,%ymm10
7530 .byte 196,194,93,168,194 // vfmadd213ps %ymm10,%ymm4,%ymm0
7531 .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10
7532 .byte 196,98,61,184,213 // vfmadd231ps %ymm5,%ymm8,%ymm10
7533 .byte 196,194,85,168,202 // vfmadd213ps %ymm10,%ymm5,%ymm1
7534 .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10
7535 .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10
7536 .byte 196,194,77,168,210 // vfmadd213ps %ymm10,%ymm6,%ymm2
7537 .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9
7538 .byte 196,66,69,168,193 // vfmadd213ps %ymm9,%ymm7,%ymm8
7539 .byte 196,194,69,168,216 // vfmadd213ps %ymm8,%ymm7,%ymm3
7540 .byte 72,173 // lods %ds:(%rsi),%rax
7541 .byte 255,224 // jmpq *%rax
7542
7543HIDDEN _sk_plus__hsw
7544.globl _sk_plus__hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7545FUNCTION(_sk_plus__hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7546_sk_plus__hsw:
7547 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
7548 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
7549 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
7550 .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
7551 .byte 72,173 // lods %ds:(%rsi),%rax
7552 .byte 255,224 // jmpq *%rax
7553
7554HIDDEN _sk_screen_hsw
7555.globl _sk_screen_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7556FUNCTION(_sk_screen_hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7557_sk_screen_hsw:
7558 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
7559 .byte 196,194,93,172,192 // vfnmadd213ps %ymm8,%ymm4,%ymm0
7560 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
7561 .byte 196,194,85,172,200 // vfnmadd213ps %ymm8,%ymm5,%ymm1
7562 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
7563 .byte 196,194,77,172,208 // vfnmadd213ps %ymm8,%ymm6,%ymm2
7564 .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8
7565 .byte 196,194,69,172,216 // vfnmadd213ps %ymm8,%ymm7,%ymm3
7566 .byte 72,173 // lods %ds:(%rsi),%rax
7567 .byte 255,224 // jmpq *%rax
7568
7569HIDDEN _sk_xor__hsw
7570.globl _sk_xor__hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7571FUNCTION(_sk_xor__hsw)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]7572_sk_xor__hsw:
7573 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7574 .byte 197,121,110,192 // vmovd %eax,%xmm8
7575 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7576 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
7577 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7578 .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10
7579 .byte 196,194,53,168,194 // vfmadd213ps %ymm10,%ymm9,%ymm0
7580 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
7581 .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1
7582 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
7583 .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2
7584 .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3
7585 .byte 196,98,69,168,195 // vfmadd213ps %ymm3,%ymm7,%ymm8
7586 .byte 72,173 // lods %ds:(%rsi),%rax
7587 .byte 197,124,41,195 // vmovaps %ymm8,%ymm3
7588 .byte 255,224 // jmpq *%rax
7589
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]7590HIDDEN _sk_darken_hsw
7591.globl _sk_darken_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7592FUNCTION(_sk_darken_hsw)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]7593_sk_darken_hsw:
7594 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
7595 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
7596 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
7597 .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0
7598 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
7599 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
7600 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
7601 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
7602 .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1
7603 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
7604 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
7605 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
7606 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
7607 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2
7608 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
7609 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7610 .byte 197,121,110,192 // vmovd %eax,%xmm8
7611 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7612 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7613 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7614 .byte 72,173 // lods %ds:(%rsi),%rax
7615 .byte 255,224 // jmpq *%rax
7616
7617HIDDEN _sk_lighten_hsw
7618.globl _sk_lighten_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7619FUNCTION(_sk_lighten_hsw)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]7620_sk_lighten_hsw:
7621 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
7622 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
7623 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
7624 .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0
7625 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
7626 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
7627 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
7628 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
7629 .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1
7630 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
7631 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
7632 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
7633 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
7634 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
7635 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
7636 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7637 .byte 197,121,110,192 // vmovd %eax,%xmm8
7638 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7639 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7640 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7641 .byte 72,173 // lods %ds:(%rsi),%rax
7642 .byte 255,224 // jmpq *%rax
7643
7644HIDDEN _sk_difference_hsw
7645.globl _sk_difference_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7646FUNCTION(_sk_difference_hsw)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]7647_sk_difference_hsw:
7648 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
7649 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
7650 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
7651 .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0
7652 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
7653 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
7654 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
7655 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
7656 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
7657 .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1
7658 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
7659 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
7660 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
7661 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
7662 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
7663 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
7664 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
7665 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
7666 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7667 .byte 197,121,110,192 // vmovd %eax,%xmm8
7668 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7669 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7670 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7671 .byte 72,173 // lods %ds:(%rsi),%rax
7672 .byte 255,224 // jmpq *%rax
7673
7674HIDDEN _sk_exclusion_hsw
7675.globl _sk_exclusion_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7676FUNCTION(_sk_exclusion_hsw)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]7677_sk_exclusion_hsw:
7678 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
7679 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
7680 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
7681 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
7682 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
7683 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
7684 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
7685 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
7686 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
7687 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
7688 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
7689 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
7690 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7691 .byte 197,121,110,192 // vmovd %eax,%xmm8
7692 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7693 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7694 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7695 .byte 72,173 // lods %ds:(%rsi),%rax
7696 .byte 255,224 // jmpq *%rax
7697
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7698HIDDEN _sk_colorburn_hsw
7699.globl _sk_colorburn_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7700FUNCTION(_sk_colorburn_hsw)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7701_sk_colorburn_hsw:
7702 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7703 .byte 197,121,110,192 // vmovd %eax,%xmm8
7704 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7705 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
7706 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11
7707 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
7708 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7709 .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12
7710 .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13
7711 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
7712 .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13
7713 .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13
7714 .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13
7715 .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13
7716 .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13
7717 .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12
7718 .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0
7719 .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0
7720 .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12
7721 .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11
7722 .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0
7723 .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11
7724 .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12
7725 .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13
7726 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
7727 .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13
7728 .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13
7729 .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13
7730 .byte 196,66,101,168,235 // vfmadd213ps %ymm11,%ymm3,%ymm13
7731 .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13
7732 .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12
7733 .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1
7734 .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1
7735 .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12
7736 .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11
7737 .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1
7738 .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9
7739 .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10
7740 .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11
7741 .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12
7742 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
7743 .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
7744 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
7745 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
7746 .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12
7747 .byte 196,66,101,168,225 // vfmadd213ps %ymm9,%ymm3,%ymm12
7748 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
7749 .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2
7750 .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10
7751 .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9
7752 .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2
7753 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7754 .byte 72,173 // lods %ds:(%rsi),%rax
7755 .byte 255,224 // jmpq *%rax
7756
7757HIDDEN _sk_colordodge_hsw
7758.globl _sk_colordodge_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7759FUNCTION(_sk_colordodge_hsw)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7760_sk_colordodge_hsw:
7761 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
7762 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7763 .byte 197,121,110,200 // vmovd %eax,%xmm9
7764 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
7765 .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10
7766 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
7767 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9
7768 .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12
7769 .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13
7770 .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12
7771 .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13
7772 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
7773 .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12
7774 .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12
7775 .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13
7776 .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0
7777 .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0
7778 .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12
7779 .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11
7780 .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0
7781 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
7782 .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12
7783 .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13
7784 .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12
7785 .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13
7786 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
7787 .byte 196,66,101,168,227 // vfmadd213ps %ymm11,%ymm3,%ymm12
7788 .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12
7789 .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13
7790 .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1
7791 .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1
7792 .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12
7793 .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11
7794 .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1
7795 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
7796 .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11
7797 .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12
7798 .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11
7799 .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12
7800 .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11
7801 .byte 196,66,101,168,218 // vfmadd213ps %ymm10,%ymm3,%ymm11
7802 .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11
7803 .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12
7804 .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2
7805 .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2
7806 .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8
7807 .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10
7808 .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2
7809 .byte 196,194,69,184,217 // vfmadd231ps %ymm9,%ymm7,%ymm3
7810 .byte 72,173 // lods %ds:(%rsi),%rax
7811 .byte 255,224 // jmpq *%rax
7812
7813HIDDEN _sk_hardlight_hsw
7814.globl _sk_hardlight_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7815FUNCTION(_sk_hardlight_hsw)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7816_sk_hardlight_hsw:
7817 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7818 .byte 197,121,110,192 // vmovd %eax,%xmm8
7819 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7820 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
7821 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7822 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
7823 .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11
7824 .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9
7825 .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12
7826 .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9
7827 .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13
7828 .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9
7829 .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14
7830 .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0
7831 .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0
7832 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
7833 .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0
7834 .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0
7835 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
7836 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
7837 .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11
7838 .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12
7839 .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12
7840 .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13
7841 .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13
7842 .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14
7843 .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1
7844 .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
7845 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
7846 .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1
7847 .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1
7848 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
7849 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
7850 .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10
7851 .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11
7852 .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11
7853 .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12
7854 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
7855 .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13
7856 .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2
7857 .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2
7858 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
7859 .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2
7860 .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2
7861 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
7862 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7863 .byte 72,173 // lods %ds:(%rsi),%rax
7864 .byte 255,224 // jmpq *%rax
7865
7866HIDDEN _sk_overlay_hsw
7867.globl _sk_overlay_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7868FUNCTION(_sk_overlay_hsw)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7869_sk_overlay_hsw:
7870 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7871 .byte 197,121,110,192 // vmovd %eax,%xmm8
7872 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7873 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
7874 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
7875 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
7876 .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11
7877 .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9
7878 .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12
7879 .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9
7880 .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13
7881 .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9
7882 .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14
7883 .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0
7884 .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0
7885 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
7886 .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0
7887 .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0
7888 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
7889 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
7890 .byte 196,98,61,184,221 // vfmadd231ps %ymm5,%ymm8,%ymm11
7891 .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12
7892 .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12
7893 .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13
7894 .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13
7895 .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14
7896 .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1
7897 .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
7898 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
7899 .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1
7900 .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1
7901 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
7902 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
7903 .byte 196,98,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm10
7904 .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11
7905 .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11
7906 .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12
7907 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
7908 .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13
7909 .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2
7910 .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2
7911 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
7912 .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2
7913 .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2
7914 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
7915 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
7916 .byte 72,173 // lods %ds:(%rsi),%rax
7917 .byte 255,224 // jmpq *%rax
7918
7919HIDDEN _sk_softlight_hsw
7920.globl _sk_softlight_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]7921FUNCTION(_sk_softlight_hsw)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7922_sk_softlight_hsw:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]7923 .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7924 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
7925 .byte 197,44,194,223,1 // vcmpltps %ymm7,%ymm10,%ymm11
7926 .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8
7927 .byte 196,67,45,74,224,176 // vblendvps %ymm11,%ymm8,%ymm10,%ymm12
7928 .byte 196,65,28,88,196 // vaddps %ymm12,%ymm12,%ymm8
7929 .byte 196,65,60,88,232 // vaddps %ymm8,%ymm8,%ymm13
7930 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
7931 .byte 197,121,110,192 // vmovd %eax,%xmm8
7932 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
7933 .byte 196,66,21,168,237 // vfmadd213ps %ymm13,%ymm13,%ymm13
7934 .byte 196,65,28,92,240 // vsubps %ymm8,%ymm12,%ymm14
7935 .byte 184,0,0,224,64 // mov $0x40e00000,%eax
7936 .byte 197,121,110,200 // vmovd %eax,%xmm9
7937 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
7938 .byte 196,65,28,89,249 // vmulps %ymm9,%ymm12,%ymm15
7939 .byte 196,66,21,184,254 // vfmadd231ps %ymm14,%ymm13,%ymm15
7940 .byte 196,65,124,82,236 // vrsqrtps %ymm12,%ymm13
7941 .byte 196,65,124,83,237 // vrcpps %ymm13,%ymm13
7942 .byte 196,65,20,92,236 // vsubps %ymm12,%ymm13,%ymm13
7943 .byte 197,92,88,244 // vaddps %ymm4,%ymm4,%ymm14
7944 .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14
7945 .byte 197,12,194,247,2 // vcmpleps %ymm7,%ymm14,%ymm14
7946 .byte 196,67,21,74,239,224 // vblendvps %ymm14,%ymm15,%ymm13,%ymm13
7947 .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14
7948 .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15
7949 .byte 196,65,60,92,228 // vsubps %ymm12,%ymm8,%ymm12
7950 .byte 196,98,5,168,227 // vfmadd213ps %ymm3,%ymm15,%ymm12
7951 .byte 197,28,89,228 // vmulps %ymm4,%ymm12,%ymm12
7952 .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15
7953 .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13
7954 .byte 196,98,101,184,236 // vfmadd231ps %ymm4,%ymm3,%ymm13
7955 .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14
7956 .byte 196,195,21,74,212,224 // vblendvps %ymm14,%ymm12,%ymm13,%ymm2
7957 .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13
7958 .byte 196,67,45,74,237,176 // vblendvps %ymm11,%ymm13,%ymm10,%ymm13
7959 .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14
7960 .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14
7961 .byte 196,66,13,168,246 // vfmadd213ps %ymm14,%ymm14,%ymm14
7962 .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15
7963 .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14
7964 .byte 196,66,53,184,245 // vfmadd231ps %ymm13,%ymm9,%ymm14
7965 .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15
7966 .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15
7967 .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15
7968 .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12
7969 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
7970 .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12
7971 .byte 196,67,5,74,230,192 // vblendvps %ymm12,%ymm14,%ymm15,%ymm12
7972 .byte 197,116,88,241 // vaddps %ymm1,%ymm1,%ymm14
7973 .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13
7974 .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15
7975 .byte 196,98,5,168,235 // vfmadd213ps %ymm3,%ymm15,%ymm13
7976 .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15
7977 .byte 196,65,4,89,228 // vmulps %ymm12,%ymm15,%ymm12
7978 .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13
7979 .byte 196,98,101,184,229 // vfmadd231ps %ymm5,%ymm3,%ymm12
7980 .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14
7981 .byte 196,67,29,74,237,224 // vblendvps %ymm14,%ymm13,%ymm12,%ymm13
7982 .byte 197,76,94,231 // vdivps %ymm7,%ymm6,%ymm12
7983 .byte 196,67,45,74,212,176 // vblendvps %ymm11,%ymm12,%ymm10,%ymm10
7984 .byte 196,65,44,88,218 // vaddps %ymm10,%ymm10,%ymm11
7985 .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11
7986 .byte 196,66,37,168,219 // vfmadd213ps %ymm11,%ymm11,%ymm11
7987 .byte 196,65,44,92,224 // vsubps %ymm8,%ymm10,%ymm12
7988 .byte 196,65,28,89,219 // vmulps %ymm11,%ymm12,%ymm11
7989 .byte 196,66,45,168,203 // vfmadd213ps %ymm11,%ymm10,%ymm9
7990 .byte 196,65,124,82,218 // vrsqrtps %ymm10,%ymm11
7991 .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11
7992 .byte 196,65,36,92,218 // vsubps %ymm10,%ymm11,%ymm11
7993 .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12
7994 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
7995 .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12
7996 .byte 196,67,37,74,201,192 // vblendvps %ymm12,%ymm9,%ymm11,%ymm9
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]7997 .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]7998 .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11
7999 .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12
8000 .byte 196,65,60,92,210 // vsubps %ymm10,%ymm8,%ymm10
8001 .byte 196,98,29,168,211 // vfmadd213ps %ymm3,%ymm12,%ymm10
8002 .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12
8003 .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9
8004 .byte 197,44,89,214 // vmulps %ymm6,%ymm10,%ymm10
8005 .byte 196,98,101,184,206 // vfmadd231ps %ymm6,%ymm3,%ymm9
8006 .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11
8007 .byte 196,67,53,74,202,176 // vblendvps %ymm11,%ymm10,%ymm9,%ymm9
8008 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
8009 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
8010 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
8011 .byte 196,98,45,184,216 // vfmadd231ps %ymm0,%ymm10,%ymm11
8012 .byte 196,193,108,88,195 // vaddps %ymm11,%ymm2,%ymm0
8013 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
8014 .byte 196,226,61,184,205 // vfmadd231ps %ymm5,%ymm8,%ymm1
8015 .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1
8016 .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2
8017 .byte 196,226,61,184,214 // vfmadd231ps %ymm6,%ymm8,%ymm2
8018 .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2
8019 .byte 196,194,69,184,216 // vfmadd231ps %ymm8,%ymm7,%ymm3
8020 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]8021 .byte 255,224 // jmpq *%rax
8022
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8023HIDDEN _sk_clamp_0_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8024.globl _sk_clamp_0_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8025FUNCTION(_sk_clamp_0_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8026_sk_clamp_0_hsw:
8027 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
8028 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
8029 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1
8030 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
8031 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3
8032 .byte 72,173 // lods %ds:(%rsi),%rax
8033 .byte 255,224 // jmpq *%rax
8034
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8035HIDDEN _sk_clamp_1_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8036.globl _sk_clamp_1_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8037FUNCTION(_sk_clamp_1_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8038_sk_clamp_1_hsw:
8039 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8040 .byte 197,121,110,192 // vmovd %eax,%xmm8
8041 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8042 .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
8043 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
8044 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
8045 .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
8046 .byte 72,173 // lods %ds:(%rsi),%rax
8047 .byte 255,224 // jmpq *%rax
8048
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8049HIDDEN _sk_clamp_a_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8050.globl _sk_clamp_a_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8051FUNCTION(_sk_clamp_a_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8052_sk_clamp_a_hsw:
8053 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8054 .byte 197,121,110,192 // vmovd %eax,%xmm8
8055 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8056 .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
8057 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0
8058 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1
8059 .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2
8060 .byte 72,173 // lods %ds:(%rsi),%rax
8061 .byte 255,224 // jmpq *%rax
8062
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8063HIDDEN _sk_set_rgb_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8064.globl _sk_set_rgb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8065FUNCTION(_sk_set_rgb_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8066_sk_set_rgb_hsw:
8067 .byte 72,173 // lods %ds:(%rsi),%rax
8068 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
8069 .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1
8070 .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2
8071 .byte 72,173 // lods %ds:(%rsi),%rax
8072 .byte 255,224 // jmpq *%rax
8073
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8074HIDDEN _sk_swap_rb_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8075.globl _sk_swap_rb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8076FUNCTION(_sk_swap_rb_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8077_sk_swap_rb_hsw:
8078 .byte 197,124,40,192 // vmovaps %ymm0,%ymm8
8079 .byte 72,173 // lods %ds:(%rsi),%rax
8080 .byte 197,252,40,194 // vmovaps %ymm2,%ymm0
8081 .byte 197,124,41,194 // vmovaps %ymm8,%ymm2
8082 .byte 255,224 // jmpq *%rax
8083
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8084HIDDEN _sk_swap_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8085.globl _sk_swap_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8086FUNCTION(_sk_swap_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8087_sk_swap_hsw:
8088 .byte 197,124,40,195 // vmovaps %ymm3,%ymm8
8089 .byte 197,124,40,202 // vmovaps %ymm2,%ymm9
8090 .byte 197,124,40,209 // vmovaps %ymm1,%ymm10
8091 .byte 197,124,40,216 // vmovaps %ymm0,%ymm11
8092 .byte 72,173 // lods %ds:(%rsi),%rax
8093 .byte 197,252,40,196 // vmovaps %ymm4,%ymm0
8094 .byte 197,252,40,205 // vmovaps %ymm5,%ymm1
8095 .byte 197,252,40,214 // vmovaps %ymm6,%ymm2
8096 .byte 197,252,40,223 // vmovaps %ymm7,%ymm3
8097 .byte 197,124,41,220 // vmovaps %ymm11,%ymm4
8098 .byte 197,124,41,213 // vmovaps %ymm10,%ymm5
8099 .byte 197,124,41,206 // vmovaps %ymm9,%ymm6
8100 .byte 197,124,41,199 // vmovaps %ymm8,%ymm7
8101 .byte 255,224 // jmpq *%rax
8102
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8103HIDDEN _sk_move_src_dst_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8104.globl _sk_move_src_dst_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8105FUNCTION(_sk_move_src_dst_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8106_sk_move_src_dst_hsw:
8107 .byte 72,173 // lods %ds:(%rsi),%rax
8108 .byte 197,252,40,224 // vmovaps %ymm0,%ymm4
8109 .byte 197,252,40,233 // vmovaps %ymm1,%ymm5
8110 .byte 197,252,40,242 // vmovaps %ymm2,%ymm6
8111 .byte 197,252,40,251 // vmovaps %ymm3,%ymm7
8112 .byte 255,224 // jmpq *%rax
8113
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8114HIDDEN _sk_move_dst_src_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8115.globl _sk_move_dst_src_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8116FUNCTION(_sk_move_dst_src_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8117_sk_move_dst_src_hsw:
8118 .byte 72,173 // lods %ds:(%rsi),%rax
8119 .byte 197,252,40,196 // vmovaps %ymm4,%ymm0
8120 .byte 197,252,40,205 // vmovaps %ymm5,%ymm1
8121 .byte 197,252,40,214 // vmovaps %ymm6,%ymm2
8122 .byte 197,252,40,223 // vmovaps %ymm7,%ymm3
8123 .byte 255,224 // jmpq *%rax
8124
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8125HIDDEN _sk_premul_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8126.globl _sk_premul_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8127FUNCTION(_sk_premul_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8128_sk_premul_hsw:
8129 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
8130 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
8131 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
8132 .byte 72,173 // lods %ds:(%rsi),%rax
8133 .byte 255,224 // jmpq *%rax
8134
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8135HIDDEN _sk_unpremul_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8136.globl _sk_unpremul_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8137FUNCTION(_sk_unpremul_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8138_sk_unpremul_hsw:
8139 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
8140 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9
8141 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8142 .byte 197,121,110,208 // vmovd %eax,%xmm10
8143 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
8144 .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10
8145 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8
8146 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
8147 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
8148 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
8149 .byte 72,173 // lods %ds:(%rsi),%rax
8150 .byte 255,224 // jmpq *%rax
8151
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8152HIDDEN _sk_from_srgb_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8153.globl _sk_from_srgb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8154FUNCTION(_sk_from_srgb_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8155_sk_from_srgb_hsw:
8156 .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax
8157 .byte 197,121,110,192 // vmovd %eax,%xmm8
8158 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8159 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
8160 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
8161 .byte 184,154,153,153,62 // mov $0x3e99999a,%eax
8162 .byte 197,121,110,216 // vmovd %eax,%xmm11
8163 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
8164 .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax
8165 .byte 197,121,110,224 // vmovd %eax,%xmm12
8166 .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12
8167 .byte 196,65,125,111,235 // vmovdqa %ymm11,%ymm13
8168 .byte 196,66,125,168,236 // vfmadd213ps %ymm12,%ymm0,%ymm13
8169 .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax
8170 .byte 197,121,110,240 // vmovd %eax,%xmm14
8171 .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14
8172 .byte 196,66,45,168,238 // vfmadd213ps %ymm14,%ymm10,%ymm13
8173 .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax
8174 .byte 197,121,110,208 // vmovd %eax,%xmm10
8175 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
8176 .byte 196,193,124,194,194,1 // vcmpltps %ymm10,%ymm0,%ymm0
8177 .byte 196,195,21,74,193,0 // vblendvps %ymm0,%ymm9,%ymm13,%ymm0
8178 .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9
8179 .byte 197,116,89,233 // vmulps %ymm1,%ymm1,%ymm13
8180 .byte 196,65,125,111,251 // vmovdqa %ymm11,%ymm15
8181 .byte 196,66,117,168,252 // vfmadd213ps %ymm12,%ymm1,%ymm15
8182 .byte 196,66,21,168,254 // vfmadd213ps %ymm14,%ymm13,%ymm15
8183 .byte 196,193,116,194,202,1 // vcmpltps %ymm10,%ymm1,%ymm1
8184 .byte 196,195,5,74,201,16 // vblendvps %ymm1,%ymm9,%ymm15,%ymm1
8185 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
8186 .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9
8187 .byte 196,66,109,168,220 // vfmadd213ps %ymm12,%ymm2,%ymm11
8188 .byte 196,66,53,168,222 // vfmadd213ps %ymm14,%ymm9,%ymm11
8189 .byte 196,193,108,194,210,1 // vcmpltps %ymm10,%ymm2,%ymm2
8190 .byte 196,195,37,74,208,32 // vblendvps %ymm2,%ymm8,%ymm11,%ymm2
8191 .byte 72,173 // lods %ds:(%rsi),%rax
8192 .byte 255,224 // jmpq *%rax
8193
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8194HIDDEN _sk_to_srgb_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8195.globl _sk_to_srgb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8196FUNCTION(_sk_to_srgb_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8197_sk_to_srgb_hsw:
8198 .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8
8199 .byte 196,65,124,83,216 // vrcpps %ymm8,%ymm11
8200 .byte 196,65,124,82,224 // vrsqrtps %ymm8,%ymm12
8201 .byte 184,41,92,71,65 // mov $0x41475c29,%eax
8202 .byte 197,121,110,192 // vmovd %eax,%xmm8
8203 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8204 .byte 197,60,89,232 // vmulps %ymm0,%ymm8,%ymm13
8205 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8206 .byte 197,121,110,200 // vmovd %eax,%xmm9
8207 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
8208 .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax
8209 .byte 197,121,110,208 // vmovd %eax,%xmm10
8210 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
8211 .byte 184,206,111,48,63 // mov $0x3f306fce,%eax
8212 .byte 197,121,110,240 // vmovd %eax,%xmm14
8213 .byte 196,66,125,88,246 // vpbroadcastd %xmm14,%ymm14
8214 .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax
8215 .byte 53,0,0,0,128 // xor $0x80000000,%eax
8216 .byte 197,121,110,248 // vmovd %eax,%xmm15
8217 .byte 196,66,125,88,255 // vpbroadcastd %xmm15,%ymm15
8218 .byte 196,66,13,168,223 // vfmadd213ps %ymm15,%ymm14,%ymm11
8219 .byte 196,66,45,184,220 // vfmadd231ps %ymm12,%ymm10,%ymm11
8220 .byte 196,65,52,93,219 // vminps %ymm11,%ymm9,%ymm11
8221 .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax
8222 .byte 197,121,110,224 // vmovd %eax,%xmm12
8223 .byte 196,66,125,88,228 // vpbroadcastd %xmm12,%ymm12
8224 .byte 196,193,124,194,196,1 // vcmpltps %ymm12,%ymm0,%ymm0
8225 .byte 196,195,37,74,197,0 // vblendvps %ymm0,%ymm13,%ymm11,%ymm0
8226 .byte 197,124,82,217 // vrsqrtps %ymm1,%ymm11
8227 .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13
8228 .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11
8229 .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13
8230 .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13
8231 .byte 197,60,89,217 // vmulps %ymm1,%ymm8,%ymm11
8232 .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13
8233 .byte 196,193,116,194,204,1 // vcmpltps %ymm12,%ymm1,%ymm1
8234 .byte 196,195,21,74,203,16 // vblendvps %ymm1,%ymm11,%ymm13,%ymm1
8235 .byte 197,124,82,218 // vrsqrtps %ymm2,%ymm11
8236 .byte 196,65,124,83,235 // vrcpps %ymm11,%ymm13
8237 .byte 196,66,13,168,239 // vfmadd213ps %ymm15,%ymm14,%ymm13
8238 .byte 196,65,124,82,219 // vrsqrtps %ymm11,%ymm11
8239 .byte 196,66,45,184,235 // vfmadd231ps %ymm11,%ymm10,%ymm13
8240 .byte 196,65,52,93,205 // vminps %ymm13,%ymm9,%ymm9
8241 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
8242 .byte 196,193,108,194,212,1 // vcmpltps %ymm12,%ymm2,%ymm2
8243 .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2
8244 .byte 72,173 // lods %ds:(%rsi),%rax
8245 .byte 255,224 // jmpq *%rax
8246
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]8247HIDDEN _sk_rgb_to_hsl_hsw
8248.globl _sk_rgb_to_hsl_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8249FUNCTION(_sk_rgb_to_hsl_hsw)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]8250_sk_rgb_to_hsl_hsw:
8251 .byte 197,252,17,124,36,200 // vmovups %ymm7,-0x38(%rsp)
8252 .byte 197,252,40,254 // vmovaps %ymm6,%ymm7
8253 .byte 197,252,40,245 // vmovaps %ymm5,%ymm6
8254 .byte 197,252,40,236 // vmovaps %ymm4,%ymm5
8255 .byte 197,252,40,227 // vmovaps %ymm3,%ymm4
8256 .byte 197,252,40,216 // vmovaps %ymm0,%ymm3
8257 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8258 .byte 197,121,110,192 // vmovd %eax,%xmm8
8259 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
8260 .byte 184,0,0,192,64 // mov $0x40c00000,%eax
8261 .byte 197,121,110,200 // vmovd %eax,%xmm9
8262 .byte 65,185,0,0,0,64 // mov $0x40000000,%r9d
8263 .byte 184,0,0,128,64 // mov $0x40800000,%eax
8264 .byte 197,121,110,208 // vmovd %eax,%xmm10
8265 .byte 197,100,95,217 // vmaxps %ymm1,%ymm3,%ymm11
8266 .byte 197,36,95,218 // vmaxps %ymm2,%ymm11,%ymm11
8267 .byte 197,100,93,225 // vminps %ymm1,%ymm3,%ymm12
8268 .byte 197,28,93,226 // vminps %ymm2,%ymm12,%ymm12
8269 .byte 196,65,36,92,236 // vsubps %ymm12,%ymm11,%ymm13
8270 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8271 .byte 196,65,60,94,197 // vdivps %ymm13,%ymm8,%ymm8
8272 .byte 197,116,194,242,1 // vcmpltps %ymm2,%ymm1,%ymm14
8273 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
8274 .byte 196,65,4,87,255 // vxorps %ymm15,%ymm15,%ymm15
8275 .byte 196,67,5,74,201,224 // vblendvps %ymm14,%ymm9,%ymm15,%ymm9
8276 .byte 197,116,92,242 // vsubps %ymm2,%ymm1,%ymm14
8277 .byte 196,66,61,168,241 // vfmadd213ps %ymm9,%ymm8,%ymm14
8278 .byte 197,236,92,195 // vsubps %ymm3,%ymm2,%ymm0
8279 .byte 197,100,92,201 // vsubps %ymm1,%ymm3,%ymm9
8280 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
8281 .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9
8282 .byte 196,193,121,110,209 // vmovd %r9d,%xmm2
8283 .byte 196,98,125,88,210 // vpbroadcastd %xmm2,%ymm10
8284 .byte 196,194,61,168,194 // vfmadd213ps %ymm10,%ymm8,%ymm0
8285 .byte 197,164,194,201,0 // vcmpeqps %ymm1,%ymm11,%ymm1
8286 .byte 196,227,53,74,192,16 // vblendvps %ymm1,%ymm0,%ymm9,%ymm0
8287 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
8288 .byte 197,249,110,200 // vmovd %eax,%xmm1
8289 .byte 197,164,194,211,0 // vcmpeqps %ymm3,%ymm11,%ymm2
8290 .byte 196,195,125,74,198,32 // vblendvps %ymm2,%ymm14,%ymm0,%ymm0
8291 .byte 196,193,36,88,220 // vaddps %ymm12,%ymm11,%ymm3
8292 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
8293 .byte 197,228,89,209 // vmulps %ymm1,%ymm3,%ymm2
8294 .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1
8295 .byte 196,65,44,92,195 // vsubps %ymm11,%ymm10,%ymm8
8296 .byte 196,65,60,92,196 // vsubps %ymm12,%ymm8,%ymm8
8297 .byte 196,195,101,74,200,16 // vblendvps %ymm1,%ymm8,%ymm3,%ymm1
8298 .byte 196,193,36,194,220,0 // vcmpeqps %ymm12,%ymm11,%ymm3
8299 .byte 197,148,94,201 // vdivps %ymm1,%ymm13,%ymm1
8300 .byte 196,195,125,74,199,48 // vblendvps %ymm3,%ymm15,%ymm0,%ymm0
8301 .byte 196,195,117,74,207,48 // vblendvps %ymm3,%ymm15,%ymm1,%ymm1
8302 .byte 196,193,121,110,216 // vmovd %r8d,%xmm3
8303 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8304 .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0
8305 .byte 72,173 // lods %ds:(%rsi),%rax
8306 .byte 197,252,40,220 // vmovaps %ymm4,%ymm3
8307 .byte 197,252,40,229 // vmovaps %ymm5,%ymm4
8308 .byte 197,252,40,238 // vmovaps %ymm6,%ymm5
8309 .byte 197,252,40,247 // vmovaps %ymm7,%ymm6
8310 .byte 197,252,16,124,36,200 // vmovups -0x38(%rsp),%ymm7
8311 .byte 255,224 // jmpq *%rax
8312
8313HIDDEN _sk_hsl_to_rgb_hsw
8314.globl _sk_hsl_to_rgb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8315FUNCTION(_sk_hsl_to_rgb_hsw)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]8316_sk_hsl_to_rgb_hsw:
8317 .byte 72,131,236,56 // sub $0x38,%rsp
8318 .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp)
8319 .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp)
8320 .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp)
8321 .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp)
8322 .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp)
8323 .byte 197,252,40,233 // vmovaps %ymm1,%ymm5
8324 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
8325 .byte 197,249,110,200 // vmovd %eax,%xmm1
8326 .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8
8327 .byte 196,193,108,194,200,1 // vcmpltps %ymm8,%ymm2,%ymm1
8328 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8329 .byte 197,249,110,224 // vmovd %eax,%xmm4
8330 .byte 196,98,125,88,212 // vpbroadcastd %xmm4,%ymm10
8331 .byte 197,172,88,229 // vaddps %ymm5,%ymm10,%ymm4
8332 .byte 197,220,89,226 // vmulps %ymm2,%ymm4,%ymm4
8333 .byte 197,84,88,202 // vaddps %ymm2,%ymm5,%ymm9
8334 .byte 196,98,85,188,202 // vfnmadd231ps %ymm2,%ymm5,%ymm9
8335 .byte 196,99,53,74,204,16 // vblendvps %ymm1,%ymm4,%ymm9,%ymm9
8336 .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d
8337 .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax
8338 .byte 197,249,110,200 // vmovd %eax,%xmm1
8339 .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13
8340 .byte 197,148,88,224 // vaddps %ymm0,%ymm13,%ymm4
8341 .byte 184,0,0,0,0 // mov $0x0,%eax
8342 .byte 197,249,110,200 // vmovd %eax,%xmm1
8343 .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12
8344 .byte 197,172,194,204,1 // vcmpltps %ymm4,%ymm10,%ymm1
8345 .byte 196,65,92,92,218 // vsubps %ymm10,%ymm4,%ymm11
8346 .byte 196,195,93,74,203,16 // vblendvps %ymm1,%ymm11,%ymm4,%ymm1
8347 .byte 196,65,92,194,220,1 // vcmpltps %ymm12,%ymm4,%ymm11
8348 .byte 197,44,88,244 // vaddps %ymm4,%ymm10,%ymm14
8349 .byte 196,195,117,74,206,176 // vblendvps %ymm11,%ymm14,%ymm1,%ymm1
8350 .byte 196,193,121,110,216 // vmovd %r8d,%xmm3
8351 .byte 196,98,125,88,219 // vpbroadcastd %xmm3,%ymm11
8352 .byte 196,66,109,170,217 // vfmsub213ps %ymm9,%ymm2,%ymm11
8353 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
8354 .byte 184,0,0,192,64 // mov $0x40c00000,%eax
8355 .byte 197,249,110,216 // vmovd %eax,%xmm3
8356 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8357 .byte 196,65,52,92,243 // vsubps %ymm11,%ymm9,%ymm14
8358 .byte 197,12,89,243 // vmulps %ymm3,%ymm14,%ymm14
8359 .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax
8360 .byte 197,249,110,216 // vmovd %eax,%xmm3
8361 .byte 196,98,125,88,251 // vpbroadcastd %xmm3,%ymm15
8362 .byte 197,132,92,217 // vsubps %ymm1,%ymm15,%ymm3
8363 .byte 196,194,13,168,219 // vfmadd213ps %ymm11,%ymm14,%ymm3
8364 .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7
8365 .byte 196,227,37,74,219,112 // vblendvps %ymm7,%ymm3,%ymm11,%ymm3
8366 .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7
8367 .byte 196,195,101,74,249,112 // vblendvps %ymm7,%ymm9,%ymm3,%ymm7
8368 .byte 196,193,121,110,216 // vmovd %r8d,%xmm3
8369 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8370 .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1
8371 .byte 196,194,13,168,227 // vfmadd213ps %ymm11,%ymm14,%ymm4
8372 .byte 196,227,69,74,228,16 // vblendvps %ymm1,%ymm4,%ymm7,%ymm4
8373 .byte 197,172,194,200,1 // vcmpltps %ymm0,%ymm10,%ymm1
8374 .byte 196,193,124,92,250 // vsubps %ymm10,%ymm0,%ymm7
8375 .byte 196,227,125,74,207,16 // vblendvps %ymm1,%ymm7,%ymm0,%ymm1
8376 .byte 196,193,124,194,252,1 // vcmpltps %ymm12,%ymm0,%ymm7
8377 .byte 197,172,88,240 // vaddps %ymm0,%ymm10,%ymm6
8378 .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1
8379 .byte 197,132,92,241 // vsubps %ymm1,%ymm15,%ymm6
8380 .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6
8381 .byte 196,193,116,194,255,1 // vcmpltps %ymm15,%ymm1,%ymm7
8382 .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6
8383 .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7
8384 .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6
8385 .byte 197,244,194,203,1 // vcmpltps %ymm3,%ymm1,%ymm1
8386 .byte 196,193,124,92,253 // vsubps %ymm13,%ymm0,%ymm7
8387 .byte 196,194,13,168,195 // vfmadd213ps %ymm11,%ymm14,%ymm0
8388 .byte 196,227,77,74,200,16 // vblendvps %ymm1,%ymm0,%ymm6,%ymm1
8389 .byte 197,172,194,199,1 // vcmpltps %ymm7,%ymm10,%ymm0
8390 .byte 196,193,68,92,242 // vsubps %ymm10,%ymm7,%ymm6
8391 .byte 196,227,69,74,198,0 // vblendvps %ymm0,%ymm6,%ymm7,%ymm0
8392 .byte 196,193,68,194,244,1 // vcmpltps %ymm12,%ymm7,%ymm6
8393 .byte 197,44,88,215 // vaddps %ymm7,%ymm10,%ymm10
8394 .byte 196,195,125,74,194,96 // vblendvps %ymm6,%ymm10,%ymm0,%ymm0
8395 .byte 196,194,13,168,251 // vfmadd213ps %ymm11,%ymm14,%ymm7
8396 .byte 197,132,92,240 // vsubps %ymm0,%ymm15,%ymm6
8397 .byte 196,194,13,168,243 // vfmadd213ps %ymm11,%ymm14,%ymm6
8398 .byte 196,65,124,194,215,1 // vcmpltps %ymm15,%ymm0,%ymm10
8399 .byte 196,227,37,74,246,160 // vblendvps %ymm10,%ymm6,%ymm11,%ymm6
8400 .byte 196,65,124,194,192,1 // vcmpltps %ymm8,%ymm0,%ymm8
8401 .byte 196,195,77,74,241,128 // vblendvps %ymm8,%ymm9,%ymm6,%ymm6
8402 .byte 197,252,194,195,1 // vcmpltps %ymm3,%ymm0,%ymm0
8403 .byte 196,227,77,74,223,0 // vblendvps %ymm0,%ymm7,%ymm6,%ymm3
8404 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
8405 .byte 197,212,194,232,0 // vcmpeqps %ymm0,%ymm5,%ymm5
8406 .byte 196,227,93,74,194,80 // vblendvps %ymm5,%ymm2,%ymm4,%ymm0
8407 .byte 196,227,117,74,202,80 // vblendvps %ymm5,%ymm2,%ymm1,%ymm1
8408 .byte 196,227,101,74,210,80 // vblendvps %ymm5,%ymm2,%ymm3,%ymm2
8409 .byte 72,173 // lods %ds:(%rsi),%rax
8410 .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3
8411 .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4
8412 .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5
8413 .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6
8414 .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7
8415 .byte 72,131,196,56 // add $0x38,%rsp
8416 .byte 255,224 // jmpq *%rax
8417
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8418HIDDEN _sk_scale_1_float_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8419.globl _sk_scale_1_float_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8420FUNCTION(_sk_scale_1_float_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8421_sk_scale_1_float_hsw:
8422 .byte 72,173 // lods %ds:(%rsi),%rax
8423 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
8424 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
8425 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
8426 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
8427 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
8428 .byte 72,173 // lods %ds:(%rsi),%rax
8429 .byte 255,224 // jmpq *%rax
8430
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8431HIDDEN _sk_scale_u8_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8432.globl _sk_scale_u8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8433FUNCTION(_sk_scale_u8_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8434_sk_scale_u8_hsw:
8435 .byte 73,137,200 // mov %rcx,%r8
8436 .byte 72,173 // lods %ds:(%rsi),%rax
8437 .byte 72,139,0 // mov (%rax),%rax
8438 .byte 72,1,248 // add %rdi,%rax
8439 .byte 77,133,192 // test %r8,%r8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8440 .byte 117,56 // jne 105a <_sk_scale_u8_hsw+0x48>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8441 .byte 197,122,126,0 // vmovq (%rax),%xmm8
8442 .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
8443 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
8444 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
8445 .byte 197,121,110,200 // vmovd %eax,%xmm9
8446 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
8447 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
8448 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
8449 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
8450 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
8451 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
8452 .byte 72,173 // lods %ds:(%rsi),%rax
8453 .byte 76,137,193 // mov %r8,%rcx
8454 .byte 255,224 // jmpq *%rax
8455 .byte 49,201 // xor %ecx,%ecx
8456 .byte 77,137,194 // mov %r8,%r10
8457 .byte 69,49,201 // xor %r9d,%r9d
8458 .byte 68,15,182,24 // movzbl (%rax),%r11d
8459 .byte 72,255,192 // inc %rax
8460 .byte 73,211,227 // shl %cl,%r11
8461 .byte 77,9,217 // or %r11,%r9
8462 .byte 72,131,193,8 // add $0x8,%rcx
8463 .byte 73,255,202 // dec %r10
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8464 .byte 117,234 // jne 1062 <_sk_scale_u8_hsw+0x50>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8465 .byte 196,65,249,110,193 // vmovq %r9,%xmm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8466 .byte 235,167 // jmp 1026 <_sk_scale_u8_hsw+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8467
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8468HIDDEN _sk_lerp_1_float_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8469.globl _sk_lerp_1_float_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8470FUNCTION(_sk_lerp_1_float_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8471_sk_lerp_1_float_hsw:
8472 .byte 72,173 // lods %ds:(%rsi),%rax
8473 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
8474 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
8475 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
8476 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
8477 .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1
8478 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
8479 .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2
8480 .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3
8481 .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3
8482 .byte 72,173 // lods %ds:(%rsi),%rax
8483 .byte 255,224 // jmpq *%rax
8484
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8485HIDDEN _sk_lerp_u8_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8486.globl _sk_lerp_u8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8487FUNCTION(_sk_lerp_u8_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8488_sk_lerp_u8_hsw:
8489 .byte 73,137,200 // mov %rcx,%r8
8490 .byte 72,173 // lods %ds:(%rsi),%rax
8491 .byte 72,139,0 // mov (%rax),%rax
8492 .byte 72,1,248 // add %rdi,%rax
8493 .byte 77,133,192 // test %r8,%r8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8494 .byte 117,76 // jne 110a <_sk_lerp_u8_hsw+0x5c>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8495 .byte 197,122,126,0 // vmovq (%rax),%xmm8
8496 .byte 196,66,125,49,192 // vpmovzxbd %xmm8,%ymm8
8497 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
8498 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
8499 .byte 197,121,110,200 // vmovd %eax,%xmm9
8500 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
8501 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
8502 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
8503 .byte 196,226,61,168,196 // vfmadd213ps %ymm4,%ymm8,%ymm0
8504 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
8505 .byte 196,226,61,168,205 // vfmadd213ps %ymm5,%ymm8,%ymm1
8506 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
8507 .byte 196,226,61,168,214 // vfmadd213ps %ymm6,%ymm8,%ymm2
8508 .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3
8509 .byte 196,226,61,168,223 // vfmadd213ps %ymm7,%ymm8,%ymm3
8510 .byte 72,173 // lods %ds:(%rsi),%rax
8511 .byte 76,137,193 // mov %r8,%rcx
8512 .byte 255,224 // jmpq *%rax
8513 .byte 49,201 // xor %ecx,%ecx
8514 .byte 77,137,194 // mov %r8,%r10
8515 .byte 69,49,201 // xor %r9d,%r9d
8516 .byte 68,15,182,24 // movzbl (%rax),%r11d
8517 .byte 72,255,192 // inc %rax
8518 .byte 73,211,227 // shl %cl,%r11
8519 .byte 77,9,217 // or %r11,%r9
8520 .byte 72,131,193,8 // add $0x8,%rcx
8521 .byte 73,255,202 // dec %r10
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8522 .byte 117,234 // jne 1112 <_sk_lerp_u8_hsw+0x64>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8523 .byte 196,65,249,110,193 // vmovq %r9,%xmm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8524 .byte 235,147 // jmp 10c2 <_sk_lerp_u8_hsw+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8525
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8526HIDDEN _sk_lerp_565_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8527.globl _sk_lerp_565_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8528FUNCTION(_sk_lerp_565_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8529_sk_lerp_565_hsw:
8530 .byte 72,173 // lods %ds:(%rsi),%rax
8531 .byte 76,139,16 // mov (%rax),%r10
8532 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8533 .byte 15,133,179,0,0,0 // jne 11f0 <_sk_lerp_565_hsw+0xc1>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8534 .byte 196,193,122,111,28,122 // vmovdqu (%r10,%rdi,2),%xmm3
8535 .byte 196,98,125,51,195 // vpmovzxwd %xmm3,%ymm8
8536 .byte 184,0,248,0,0 // mov $0xf800,%eax
8537 .byte 197,249,110,216 // vmovd %eax,%xmm3
8538 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8539 .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
8540 .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9
8541 .byte 184,8,33,132,55 // mov $0x37842108,%eax
8542 .byte 197,249,110,216 // vmovd %eax,%xmm3
8543 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8544 .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9
8545 .byte 184,224,7,0,0 // mov $0x7e0,%eax
8546 .byte 197,249,110,216 // vmovd %eax,%xmm3
8547 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8548 .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
8549 .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10
8550 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
8551 .byte 197,249,110,216 // vmovd %eax,%xmm3
8552 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8553 .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10
8554 .byte 184,31,0,0,0 // mov $0x1f,%eax
8555 .byte 197,249,110,216 // vmovd %eax,%xmm3
8556 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8557 .byte 196,193,101,219,216 // vpand %ymm8,%ymm3,%ymm3
8558 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
8559 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
8560 .byte 197,249,110,216 // vmovd %eax,%xmm3
8561 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8562 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
8563 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
8564 .byte 196,226,53,168,196 // vfmadd213ps %ymm4,%ymm9,%ymm0
8565 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
8566 .byte 196,226,45,168,205 // vfmadd213ps %ymm5,%ymm10,%ymm1
8567 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
8568 .byte 196,226,101,168,214 // vfmadd213ps %ymm6,%ymm3,%ymm2
8569 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8570 .byte 197,249,110,216 // vmovd %eax,%xmm3
8571 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8572 .byte 72,173 // lods %ds:(%rsi),%rax
8573 .byte 255,224 // jmpq *%rax
8574 .byte 65,137,200 // mov %ecx,%r8d
8575 .byte 65,128,224,7 // and $0x7,%r8b
8576 .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
8577 .byte 65,254,200 // dec %r8b
8578 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8579 .byte 15,135,59,255,255,255 // ja 1143 <_sk_lerp_565_hsw+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8580 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8581 .byte 76,141,13,73,0,0,0 // lea 0x49(%rip),%r9 # 125c <_sk_lerp_565_hsw+0x12d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8582 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
8583 .byte 76,1,200 // add %r9,%rax
8584 .byte 255,224 // jmpq *%rax
8585 .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
8586 .byte 196,193,97,196,92,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm3
8587 .byte 196,193,97,196,92,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm3,%xmm3
8588 .byte 196,193,97,196,92,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm3,%xmm3
8589 .byte 196,193,97,196,92,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm3,%xmm3
8590 .byte 196,193,97,196,92,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm3,%xmm3
8591 .byte 196,193,97,196,92,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm3,%xmm3
8592 .byte 196,193,97,196,28,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm3,%xmm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8593 .byte 233,231,254,255,255 // jmpq 1143 <_sk_lerp_565_hsw+0x14>
8594 .byte 244 // hlt
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8595 .byte 255 // (bad)
8596 .byte 255 // (bad)
8597 .byte 255 // (bad)
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8598 .byte 236 // in (%dx),%al
Mike Klein7125ac62017-04-05 15:27:22 -0400[diff] [blame]8599 .byte 255 // (bad)
8600 .byte 255 // (bad)
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8601 .byte 255,228 // jmpq *%rsp
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8602 .byte 255 // (bad)
8603 .byte 255 // (bad)
8604 .byte 255 // (bad)
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8605 .byte 220,255 // fdivr %st,%st(7)
8606 .byte 255 // (bad)
8607 .byte 255,212 // callq *%rsp
8608 .byte 255 // (bad)
8609 .byte 255 // (bad)
8610 .byte 255,204 // dec %esp
8611 .byte 255 // (bad)
8612 .byte 255 // (bad)
8613 .byte 255,192 // inc %eax
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8614 .byte 255 // (bad)
8615 .byte 255 // (bad)
8616 .byte 255 // .byte 0xff
8617
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]8618HIDDEN _sk_load_tables_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8619.globl _sk_load_tables_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8620FUNCTION(_sk_load_tables_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8621_sk_load_tables_hsw:
8622 .byte 73,137,200 // mov %rcx,%r8
8623 .byte 72,173 // lods %ds:(%rsi),%rax
8624 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
8625 .byte 76,3,8 // add (%rax),%r9
8626 .byte 77,133,192 // test %r8,%r8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8627 .byte 117,121 // jne 1306 <_sk_load_tables_hsw+0x8e>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8628 .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
8629 .byte 185,255,0,0,0 // mov $0xff,%ecx
8630 .byte 197,249,110,193 // vmovd %ecx,%xmm0
8631 .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2
8632 .byte 197,237,219,203 // vpand %ymm3,%ymm2,%ymm1
8633 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
8634 .byte 72,139,72,8 // mov 0x8(%rax),%rcx
8635 .byte 76,139,72,16 // mov 0x10(%rax),%r9
8636 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
8637 .byte 196,226,53,146,4,137 // vgatherdps %ymm9,(%rcx,%ymm1,4),%ymm0
8638 .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1
8639 .byte 197,109,219,201 // vpand %ymm1,%ymm2,%ymm9
8640 .byte 196,65,45,118,210 // vpcmpeqd %ymm10,%ymm10,%ymm10
8641 .byte 196,130,45,146,12,137 // vgatherdps %ymm10,(%r9,%ymm9,4),%ymm1
8642 .byte 72,139,64,24 // mov 0x18(%rax),%rax
8643 .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9
8644 .byte 196,65,109,219,201 // vpand %ymm9,%ymm2,%ymm9
8645 .byte 196,162,61,146,20,136 // vgatherdps %ymm8,(%rax,%ymm9,4),%ymm2
8646 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
8647 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
8648 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
8649 .byte 197,249,110,216 // vmovd %eax,%xmm3
8650 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8651 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
8652 .byte 72,173 // lods %ds:(%rsi),%rax
8653 .byte 76,137,193 // mov %r8,%rcx
8654 .byte 255,224 // jmpq *%rax
8655 .byte 185,8,0,0,0 // mov $0x8,%ecx
8656 .byte 68,41,193 // sub %r8d,%ecx
8657 .byte 192,225,3 // shl $0x3,%cl
8658 .byte 73,199,194,255,255,255,255 // mov $0xffffffffffffffff,%r10
8659 .byte 73,211,234 // shr %cl,%r10
8660 .byte 196,193,249,110,194 // vmovq %r10,%xmm0
8661 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
8662 .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8663 .byte 233,99,255,255,255 // jmpq 1292 <_sk_load_tables_hsw+0x1a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]8664
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8665HIDDEN _sk_load_tables_u16_be_hsw
8666.globl _sk_load_tables_u16_be_hsw
8667FUNCTION(_sk_load_tables_u16_be_hsw)
8668_sk_load_tables_u16_be_hsw:
8669 .byte 72,173 // lods %ds:(%rsi),%rax
8670 .byte 76,139,0 // mov (%rax),%r8
8671 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
8672 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8673 .byte 15,133,216,0,0,0 // jne 141d <_sk_load_tables_u16_be_hsw+0xee>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8674 .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8
8675 .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2
8676 .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3
8677 .byte 196,1,122,111,76,72,48 // vmovdqu 0x30(%r8,%r9,2),%xmm9
8678 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
8679 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
8680 .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1
8681 .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3
8682 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
8683 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
8684 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
8685 .byte 197,113,105,235 // vpunpckhwd %xmm3,%xmm1,%xmm13
8686 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
8687 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
8688 .byte 65,184,255,0,0,0 // mov $0xff,%r8d
8689 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
8690 .byte 196,98,125,88,209 // vpbroadcastd %xmm1,%ymm10
8691 .byte 196,193,125,219,202 // vpand %ymm10,%ymm0,%ymm1
8692 .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11
8693 .byte 76,139,64,8 // mov 0x8(%rax),%r8
8694 .byte 76,139,72,16 // mov 0x10(%rax),%r9
8695 .byte 196,65,29,118,228 // vpcmpeqd %ymm12,%ymm12,%ymm12
8696 .byte 196,194,29,146,4,136 // vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0
8697 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
8698 .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1
8699 .byte 196,193,117,219,210 // vpand %ymm10,%ymm1,%ymm2
8700 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
8701 .byte 196,194,61,146,12,145 // vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1
8702 .byte 72,139,64,24 // mov 0x18(%rax),%rax
8703 .byte 196,193,49,108,213 // vpunpcklqdq %xmm13,%xmm9,%xmm2
8704 .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2
8705 .byte 196,65,109,219,194 // vpand %ymm10,%ymm2,%ymm8
8706 .byte 196,162,37,146,20,128 // vgatherdps %ymm11,(%rax,%ymm8,4),%ymm2
8707 .byte 184,128,0,128,55 // mov $0x37800080,%eax
8708 .byte 197,249,110,216 // vmovd %eax,%xmm3
8709 .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8
8710 .byte 196,193,49,109,221 // vpunpckhqdq %xmm13,%xmm9,%xmm3
8711 .byte 197,177,113,243,8 // vpsllw $0x8,%xmm3,%xmm9
8712 .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3
8713 .byte 197,177,235,219 // vpor %xmm3,%xmm9,%xmm3
8714 .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3
8715 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
8716 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
8717 .byte 72,173 // lods %ds:(%rsi),%rax
8718 .byte 255,224 // jmpq *%rax
8719 .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8
8720 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
8721 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8722 .byte 116,85 // je 1483 <_sk_load_tables_u16_be_hsw+0x154>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8723 .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
8724 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8725 .byte 114,72 // jb 1483 <_sk_load_tables_u16_be_hsw+0x154>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8726 .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2
8727 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8728 .byte 116,72 // je 1490 <_sk_load_tables_u16_be_hsw+0x161>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8729 .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
8730 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8731 .byte 114,59 // jb 1490 <_sk_load_tables_u16_be_hsw+0x161>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8732 .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3
8733 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8734 .byte 15,132,250,254,255,255 // je 1360 <_sk_load_tables_u16_be_hsw+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8735 .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
8736 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8737 .byte 15,130,233,254,255,255 // jb 1360 <_sk_load_tables_u16_be_hsw+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8738 .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8739 .byte 233,221,254,255,255 // jmpq 1360 <_sk_load_tables_u16_be_hsw+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8740 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
8741 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8742 .byte 233,208,254,255,255 // jmpq 1360 <_sk_load_tables_u16_be_hsw+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8743 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8744 .byte 233,199,254,255,255 // jmpq 1360 <_sk_load_tables_u16_be_hsw+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8745
8746HIDDEN _sk_load_tables_rgb_u16_be_hsw
8747.globl _sk_load_tables_rgb_u16_be_hsw
8748FUNCTION(_sk_load_tables_rgb_u16_be_hsw)
8749_sk_load_tables_rgb_u16_be_hsw:
8750 .byte 72,173 // lods %ds:(%rsi),%rax
8751 .byte 76,139,0 // mov (%rax),%r8
8752 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
8753 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8754 .byte 15,133,207,0,0,0 // jne 157a <_sk_load_tables_rgb_u16_be_hsw+0xe1>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8755 .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0
8756 .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2
8757 .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1
8758 .byte 196,129,122,111,92,72,32 // vmovdqu 0x20(%r8,%r9,2),%xmm3
8759 .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3
8760 .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8
8761 .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9
8762 .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11
8763 .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10
8764 .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0
8765 .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2
8766 .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1
8767 .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3
8768 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
8769 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
8770 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
8771 .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3
8772 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
8773 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
8774 .byte 65,184,255,0,0,0 // mov $0xff,%r8d
8775 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
8776 .byte 196,98,125,88,209 // vpbroadcastd %xmm1,%ymm10
8777 .byte 196,193,125,219,202 // vpand %ymm10,%ymm0,%ymm1
8778 .byte 196,65,37,118,219 // vpcmpeqd %ymm11,%ymm11,%ymm11
8779 .byte 76,139,64,8 // mov 0x8(%rax),%r8
8780 .byte 76,139,72,16 // mov 0x10(%rax),%r9
8781 .byte 196,65,29,118,228 // vpcmpeqd %ymm12,%ymm12,%ymm12
8782 .byte 196,194,29,146,4,136 // vgatherdps %ymm12,(%r8,%ymm1,4),%ymm0
8783 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
8784 .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1
8785 .byte 196,193,117,219,210 // vpand %ymm10,%ymm1,%ymm2
8786 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
8787 .byte 196,194,61,146,12,145 // vgatherdps %ymm8,(%r9,%ymm2,4),%ymm1
8788 .byte 72,139,64,24 // mov 0x18(%rax),%rax
8789 .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2
8790 .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2
8791 .byte 196,193,109,219,218 // vpand %ymm10,%ymm2,%ymm3
8792 .byte 196,226,37,146,20,152 // vgatherdps %ymm11,(%rax,%ymm3,4),%ymm2
8793 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
8794 .byte 197,249,110,216 // vmovd %eax,%xmm3
8795 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
8796 .byte 72,173 // lods %ds:(%rsi),%rax
8797 .byte 255,224 // jmpq *%rax
8798 .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0
8799 .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
8800 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8801 .byte 117,5 // jne 1593 <_sk_load_tables_rgb_u16_be_hsw+0xfa>
8802 .byte 233,76,255,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8803 .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1
8804 .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
8805 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8806 .byte 114,26 // jb 15c2 <_sk_load_tables_rgb_u16_be_hsw+0x129>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8807 .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1
8808 .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
8809 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8810 .byte 117,10 // jne 15c7 <_sk_load_tables_rgb_u16_be_hsw+0x12e>
8811 .byte 233,29,255,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
8812 .byte 233,24,255,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8813 .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1
8814 .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
8815 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8816 .byte 114,26 // jb 15f6 <_sk_load_tables_rgb_u16_be_hsw+0x15d>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8817 .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1
8818 .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
8819 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8820 .byte 117,10 // jne 15fb <_sk_load_tables_rgb_u16_be_hsw+0x162>
8821 .byte 233,233,254,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
8822 .byte 233,228,254,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8823 .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3
8824 .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
8825 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8826 .byte 114,20 // jb 1624 <_sk_load_tables_rgb_u16_be_hsw+0x18b>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8827 .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3
8828 .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]8829 .byte 233,187,254,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
8830 .byte 233,182,254,255,255 // jmpq 14df <_sk_load_tables_rgb_u16_be_hsw+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]8831
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]8832HIDDEN _sk_byte_tables_hsw
8833.globl _sk_byte_tables_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8834FUNCTION(_sk_byte_tables_hsw)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]8835_sk_byte_tables_hsw:
8836 .byte 85 // push %rbp
8837 .byte 65,87 // push %r15
8838 .byte 65,86 // push %r14
8839 .byte 65,85 // push %r13
8840 .byte 65,84 // push %r12
8841 .byte 83 // push %rbx
8842 .byte 72,173 // lods %ds:(%rsi),%rax
8843 .byte 65,184,0,0,127,67 // mov $0x437f0000,%r8d
8844 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
8845 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
8846 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
8847 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
8848 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
8849 .byte 68,137,197 // mov %r8d,%ebp
8850 .byte 77,137,194 // mov %r8,%r10
8851 .byte 73,193,234,32 // shr $0x20,%r10
8852 .byte 196,193,249,126,192 // vmovq %xmm0,%r8
8853 .byte 69,137,195 // mov %r8d,%r11d
8854 .byte 77,137,199 // mov %r8,%r15
8855 .byte 73,193,239,32 // shr $0x20,%r15
8856 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
8857 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
8858 .byte 69,137,198 // mov %r8d,%r14d
8859 .byte 77,137,196 // mov %r8,%r12
8860 .byte 73,193,236,32 // shr $0x20,%r12
8861 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
8862 .byte 65,137,221 // mov %ebx,%r13d
8863 .byte 72,193,235,32 // shr $0x20,%rbx
8864 .byte 76,139,8 // mov (%rax),%r9
8865 .byte 76,139,64,8 // mov 0x8(%rax),%r8
8866 .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
8867 .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
8868 .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp
8869 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
8870 .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp
8871 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
8872 .byte 67,15,182,44,41 // movzbl (%r9,%r13,1),%ebp
8873 .byte 196,227,121,32,197,4 // vpinsrb $0x4,%ebp,%xmm0,%xmm0
8874 .byte 65,15,182,44,25 // movzbl (%r9,%rbx,1),%ebp
8875 .byte 196,227,121,32,197,5 // vpinsrb $0x5,%ebp,%xmm0,%xmm0
8876 .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp
8877 .byte 196,227,121,32,197,6 // vpinsrb $0x6,%ebp,%xmm0,%xmm0
8878 .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp
8879 .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0
8880 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
8881 .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10
8882 .byte 189,129,128,128,59 // mov $0x3b808081,%ebp
8883 .byte 197,249,110,197 // vmovd %ebp,%xmm0
8884 .byte 196,98,125,88,200 // vpbroadcastd %xmm0,%ymm9
8885 .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0
8886 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
8887 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
8888 .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp
8889 .byte 65,137,233 // mov %ebp,%r9d
8890 .byte 72,193,237,32 // shr $0x20,%rbp
8891 .byte 196,225,249,126,203 // vmovq %xmm1,%rbx
8892 .byte 65,137,218 // mov %ebx,%r10d
8893 .byte 72,193,235,32 // shr $0x20,%rbx
8894 .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1
8895 .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11
8896 .byte 69,137,222 // mov %r11d,%r14d
8897 .byte 73,193,235,32 // shr $0x20,%r11
8898 .byte 196,193,249,126,207 // vmovq %xmm1,%r15
8899 .byte 69,137,252 // mov %r15d,%r12d
8900 .byte 73,193,239,32 // shr $0x20,%r15
8901 .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
8902 .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
8903 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
8904 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
8905 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
8906 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
8907 .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp
8908 .byte 196,227,113,32,205,4 // vpinsrb $0x4,%ebp,%xmm1,%xmm1
8909 .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp
8910 .byte 196,227,113,32,205,5 // vpinsrb $0x5,%ebp,%xmm1,%xmm1
8911 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
8912 .byte 196,227,113,32,205,6 // vpinsrb $0x6,%ebp,%xmm1,%xmm1
8913 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
8914 .byte 196,227,113,32,205,7 // vpinsrb $0x7,%ebp,%xmm1,%xmm1
8915 .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1
8916 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
8917 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
8918 .byte 76,139,64,16 // mov 0x10(%rax),%r8
8919 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
8920 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
8921 .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp
8922 .byte 65,137,233 // mov %ebp,%r9d
8923 .byte 72,193,237,32 // shr $0x20,%rbp
8924 .byte 196,225,249,126,211 // vmovq %xmm2,%rbx
8925 .byte 65,137,218 // mov %ebx,%r10d
8926 .byte 72,193,235,32 // shr $0x20,%rbx
8927 .byte 196,227,125,57,210,1 // vextracti128 $0x1,%ymm2,%xmm2
8928 .byte 196,195,249,22,211,1 // vpextrq $0x1,%xmm2,%r11
8929 .byte 69,137,222 // mov %r11d,%r14d
8930 .byte 73,193,235,32 // shr $0x20,%r11
8931 .byte 196,193,249,126,215 // vmovq %xmm2,%r15
8932 .byte 69,137,252 // mov %r15d,%r12d
8933 .byte 73,193,239,32 // shr $0x20,%r15
8934 .byte 196,131,121,32,20,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2
8935 .byte 196,195,105,32,20,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2
8936 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
8937 .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2
8938 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
8939 .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2
8940 .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp
8941 .byte 196,227,105,32,213,4 // vpinsrb $0x4,%ebp,%xmm2,%xmm2
8942 .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp
8943 .byte 196,227,105,32,213,5 // vpinsrb $0x5,%ebp,%xmm2,%xmm2
8944 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
8945 .byte 196,227,105,32,213,6 // vpinsrb $0x6,%ebp,%xmm2,%xmm2
8946 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
8947 .byte 196,227,105,32,213,7 // vpinsrb $0x7,%ebp,%xmm2,%xmm2
8948 .byte 196,226,125,49,210 // vpmovzxbd %xmm2,%ymm2
8949 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
8950 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
8951 .byte 72,139,64,24 // mov 0x18(%rax),%rax
8952 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
8953 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3
8954 .byte 196,227,249,22,221,1 // vpextrq $0x1,%xmm3,%rbp
8955 .byte 65,137,232 // mov %ebp,%r8d
8956 .byte 72,193,237,32 // shr $0x20,%rbp
8957 .byte 196,225,249,126,219 // vmovq %xmm3,%rbx
8958 .byte 65,137,217 // mov %ebx,%r9d
8959 .byte 72,193,235,32 // shr $0x20,%rbx
8960 .byte 196,227,125,57,219,1 // vextracti128 $0x1,%ymm3,%xmm3
8961 .byte 196,195,249,22,218,1 // vpextrq $0x1,%xmm3,%r10
8962 .byte 69,137,211 // mov %r10d,%r11d
8963 .byte 73,193,234,32 // shr $0x20,%r10
8964 .byte 196,193,249,126,222 // vmovq %xmm3,%r14
8965 .byte 69,137,247 // mov %r14d,%r15d
8966 .byte 73,193,238,32 // shr $0x20,%r14
8967 .byte 196,163,121,32,28,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3
8968 .byte 196,227,97,32,28,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3
8969 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
8970 .byte 196,227,97,32,219,2 // vpinsrb $0x2,%ebx,%xmm3,%xmm3
8971 .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp
8972 .byte 196,227,97,32,221,3 // vpinsrb $0x3,%ebp,%xmm3,%xmm3
8973 .byte 66,15,182,44,56 // movzbl (%rax,%r15,1),%ebp
8974 .byte 196,227,97,32,221,4 // vpinsrb $0x4,%ebp,%xmm3,%xmm3
8975 .byte 66,15,182,44,48 // movzbl (%rax,%r14,1),%ebp
8976 .byte 196,227,97,32,221,5 // vpinsrb $0x5,%ebp,%xmm3,%xmm3
8977 .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp
8978 .byte 196,227,97,32,221,6 // vpinsrb $0x6,%ebp,%xmm3,%xmm3
8979 .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax
8980 .byte 196,227,97,32,216,7 // vpinsrb $0x7,%eax,%xmm3,%xmm3
8981 .byte 196,226,125,49,219 // vpmovzxbd %xmm3,%ymm3
8982 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
8983 .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3
8984 .byte 72,173 // lods %ds:(%rsi),%rax
8985 .byte 91 // pop %rbx
8986 .byte 65,92 // pop %r12
8987 .byte 65,93 // pop %r13
8988 .byte 65,94 // pop %r14
8989 .byte 65,95 // pop %r15
8990 .byte 93 // pop %rbp
8991 .byte 255,224 // jmpq *%rax
8992
8993HIDDEN _sk_byte_tables_rgb_hsw
8994.globl _sk_byte_tables_rgb_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]8995FUNCTION(_sk_byte_tables_rgb_hsw)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]8996_sk_byte_tables_rgb_hsw:
8997 .byte 85 // push %rbp
8998 .byte 65,87 // push %r15
8999 .byte 65,86 // push %r14
9000 .byte 65,85 // push %r13
9001 .byte 65,84 // push %r12
9002 .byte 83 // push %rbx
9003 .byte 72,173 // lods %ds:(%rsi),%rax
9004 .byte 68,139,64,24 // mov 0x18(%rax),%r8d
9005 .byte 65,255,200 // dec %r8d
9006 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
9007 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9008 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
9009 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
9010 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
9011 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
9012 .byte 68,137,197 // mov %r8d,%ebp
9013 .byte 77,137,194 // mov %r8,%r10
9014 .byte 73,193,234,32 // shr $0x20,%r10
9015 .byte 196,193,249,126,192 // vmovq %xmm0,%r8
9016 .byte 69,137,195 // mov %r8d,%r11d
9017 .byte 77,137,199 // mov %r8,%r15
9018 .byte 73,193,239,32 // shr $0x20,%r15
9019 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
9020 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
9021 .byte 69,137,198 // mov %r8d,%r14d
9022 .byte 77,137,196 // mov %r8,%r12
9023 .byte 73,193,236,32 // shr $0x20,%r12
9024 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
9025 .byte 65,137,221 // mov %ebx,%r13d
9026 .byte 72,193,235,32 // shr $0x20,%rbx
9027 .byte 76,139,8 // mov (%rax),%r9
9028 .byte 76,139,64,8 // mov 0x8(%rax),%r8
9029 .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
9030 .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
9031 .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp
9032 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
9033 .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp
9034 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
9035 .byte 67,15,182,44,41 // movzbl (%r9,%r13,1),%ebp
9036 .byte 196,227,121,32,197,4 // vpinsrb $0x4,%ebp,%xmm0,%xmm0
9037 .byte 65,15,182,44,25 // movzbl (%r9,%rbx,1),%ebp
9038 .byte 196,227,121,32,197,5 // vpinsrb $0x5,%ebp,%xmm0,%xmm0
9039 .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp
9040 .byte 196,227,121,32,197,6 // vpinsrb $0x6,%ebp,%xmm0,%xmm0
9041 .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp
9042 .byte 196,227,121,32,197,7 // vpinsrb $0x7,%ebp,%xmm0,%xmm0
9043 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9044 .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10
9045 .byte 189,129,128,128,59 // mov $0x3b808081,%ebp
9046 .byte 197,249,110,197 // vmovd %ebp,%xmm0
9047 .byte 196,98,125,88,200 // vpbroadcastd %xmm0,%ymm9
9048 .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0
9049 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
9050 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
9051 .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp
9052 .byte 65,137,233 // mov %ebp,%r9d
9053 .byte 72,193,237,32 // shr $0x20,%rbp
9054 .byte 196,225,249,126,203 // vmovq %xmm1,%rbx
9055 .byte 65,137,218 // mov %ebx,%r10d
9056 .byte 72,193,235,32 // shr $0x20,%rbx
9057 .byte 196,227,125,57,201,1 // vextracti128 $0x1,%ymm1,%xmm1
9058 .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11
9059 .byte 69,137,222 // mov %r11d,%r14d
9060 .byte 73,193,235,32 // shr $0x20,%r11
9061 .byte 196,193,249,126,207 // vmovq %xmm1,%r15
9062 .byte 69,137,252 // mov %r15d,%r12d
9063 .byte 73,193,239,32 // shr $0x20,%r15
9064 .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
9065 .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
9066 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
9067 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
9068 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
9069 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
9070 .byte 67,15,182,44,32 // movzbl (%r8,%r12,1),%ebp
9071 .byte 196,227,113,32,205,4 // vpinsrb $0x4,%ebp,%xmm1,%xmm1
9072 .byte 67,15,182,44,56 // movzbl (%r8,%r15,1),%ebp
9073 .byte 196,227,113,32,205,5 // vpinsrb $0x5,%ebp,%xmm1,%xmm1
9074 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
9075 .byte 196,227,113,32,205,6 // vpinsrb $0x6,%ebp,%xmm1,%xmm1
9076 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
9077 .byte 196,227,113,32,205,7 // vpinsrb $0x7,%ebp,%xmm1,%xmm1
9078 .byte 196,226,125,49,201 // vpmovzxbd %xmm1,%ymm1
9079 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
9080 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
9081 .byte 72,139,64,16 // mov 0x10(%rax),%rax
9082 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
9083 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
9084 .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp
9085 .byte 65,137,232 // mov %ebp,%r8d
9086 .byte 72,193,237,32 // shr $0x20,%rbp
9087 .byte 196,225,249,126,211 // vmovq %xmm2,%rbx
9088 .byte 65,137,217 // mov %ebx,%r9d
9089 .byte 72,193,235,32 // shr $0x20,%rbx
9090 .byte 196,227,125,57,210,1 // vextracti128 $0x1,%ymm2,%xmm2
9091 .byte 196,195,249,22,210,1 // vpextrq $0x1,%xmm2,%r10
9092 .byte 69,137,211 // mov %r10d,%r11d
9093 .byte 73,193,234,32 // shr $0x20,%r10
9094 .byte 196,193,249,126,214 // vmovq %xmm2,%r14
9095 .byte 69,137,247 // mov %r14d,%r15d
9096 .byte 73,193,238,32 // shr $0x20,%r14
9097 .byte 196,163,121,32,20,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2
9098 .byte 196,227,105,32,20,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2
9099 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
9100 .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2
9101 .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp
9102 .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2
9103 .byte 66,15,182,44,56 // movzbl (%rax,%r15,1),%ebp
9104 .byte 196,227,105,32,213,4 // vpinsrb $0x4,%ebp,%xmm2,%xmm2
9105 .byte 66,15,182,44,48 // movzbl (%rax,%r14,1),%ebp
9106 .byte 196,227,105,32,213,5 // vpinsrb $0x5,%ebp,%xmm2,%xmm2
9107 .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp
9108 .byte 196,227,105,32,213,6 // vpinsrb $0x6,%ebp,%xmm2,%xmm2
9109 .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax
9110 .byte 196,227,105,32,208,7 // vpinsrb $0x7,%eax,%xmm2,%xmm2
9111 .byte 196,226,125,49,210 // vpmovzxbd %xmm2,%ymm2
9112 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
9113 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
9114 .byte 72,173 // lods %ds:(%rsi),%rax
9115 .byte 91 // pop %rbx
9116 .byte 65,92 // pop %r12
9117 .byte 65,93 // pop %r13
9118 .byte 65,94 // pop %r14
9119 .byte 65,95 // pop %r15
9120 .byte 93 // pop %rbp
9121 .byte 255,224 // jmpq *%rax
9122
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]9123HIDDEN _sk_table_r_hsw
9124.globl _sk_table_r_hsw
9125FUNCTION(_sk_table_r_hsw)
9126_sk_table_r_hsw:
9127 .byte 72,173 // lods %ds:(%rsi),%rax
9128 .byte 76,139,0 // mov (%rax),%r8
9129 .byte 139,64,8 // mov 0x8(%rax),%eax
9130 .byte 255,200 // dec %eax
9131 .byte 197,121,110,192 // vmovd %eax,%xmm8
9132 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9133 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
9134 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
9135 .byte 197,125,91,192 // vcvtps2dq %ymm0,%ymm8
9136 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
9137 .byte 196,130,53,146,4,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm0
9138 .byte 72,173 // lods %ds:(%rsi),%rax
9139 .byte 255,224 // jmpq *%rax
9140
9141HIDDEN _sk_table_g_hsw
9142.globl _sk_table_g_hsw
9143FUNCTION(_sk_table_g_hsw)
9144_sk_table_g_hsw:
9145 .byte 72,173 // lods %ds:(%rsi),%rax
9146 .byte 76,139,0 // mov (%rax),%r8
9147 .byte 139,64,8 // mov 0x8(%rax),%eax
9148 .byte 255,200 // dec %eax
9149 .byte 197,121,110,192 // vmovd %eax,%xmm8
9150 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9151 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
9152 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
9153 .byte 197,125,91,193 // vcvtps2dq %ymm1,%ymm8
9154 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
9155 .byte 196,130,53,146,12,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm1
9156 .byte 72,173 // lods %ds:(%rsi),%rax
9157 .byte 255,224 // jmpq *%rax
9158
9159HIDDEN _sk_table_b_hsw
9160.globl _sk_table_b_hsw
9161FUNCTION(_sk_table_b_hsw)
9162_sk_table_b_hsw:
9163 .byte 72,173 // lods %ds:(%rsi),%rax
9164 .byte 76,139,0 // mov (%rax),%r8
9165 .byte 139,64,8 // mov 0x8(%rax),%eax
9166 .byte 255,200 // dec %eax
9167 .byte 197,121,110,192 // vmovd %eax,%xmm8
9168 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9169 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
9170 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
9171 .byte 197,125,91,194 // vcvtps2dq %ymm2,%ymm8
9172 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
9173 .byte 196,130,53,146,20,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm2
9174 .byte 72,173 // lods %ds:(%rsi),%rax
9175 .byte 255,224 // jmpq *%rax
9176
9177HIDDEN _sk_table_a_hsw
9178.globl _sk_table_a_hsw
9179FUNCTION(_sk_table_a_hsw)
9180_sk_table_a_hsw:
9181 .byte 72,173 // lods %ds:(%rsi),%rax
9182 .byte 76,139,0 // mov (%rax),%r8
9183 .byte 139,64,8 // mov 0x8(%rax),%eax
9184 .byte 255,200 // dec %eax
9185 .byte 197,121,110,192 // vmovd %eax,%xmm8
9186 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9187 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
9188 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
9189 .byte 197,125,91,195 // vcvtps2dq %ymm3,%ymm8
9190 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
9191 .byte 196,130,53,146,28,128 // vgatherdps %ymm9,(%r8,%ymm8,4),%ymm3
9192 .byte 72,173 // lods %ds:(%rsi),%rax
9193 .byte 255,224 // jmpq *%rax
9194
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9195HIDDEN _sk_parametric_r_hsw
9196.globl _sk_parametric_r_hsw
9197FUNCTION(_sk_parametric_r_hsw)
9198_sk_parametric_r_hsw:
9199 .byte 72,173 // lods %ds:(%rsi),%rax
9200 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
9201 .byte 196,65,124,194,192,2 // vcmpleps %ymm8,%ymm0,%ymm8
9202 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
9203 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
9204 .byte 196,66,125,168,202 // vfmadd213ps %ymm10,%ymm0,%ymm9
9205 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
9206 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
9207 .byte 196,66,125,168,211 // vfmadd213ps %ymm11,%ymm0,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9208 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9209 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9210 .byte 196,98,125,24,37,243,36,0,0 // vbroadcastss 0x24f3(%rip),%ymm12 # 4108 <_sk_callback_hsw+0x12e>
9211 .byte 196,98,125,24,45,238,36,0,0 // vbroadcastss 0x24ee(%rip),%ymm13 # 410c <_sk_callback_hsw+0x132>
9212 .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
9213 .byte 196,98,125,24,45,228,36,0,0 // vbroadcastss 0x24e4(%rip),%ymm13 # 4110 <_sk_callback_hsw+0x136>
9214 .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
9215 .byte 196,98,125,24,45,218,36,0,0 // vbroadcastss 0x24da(%rip),%ymm13 # 4114 <_sk_callback_hsw+0x13a>
9216 .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
9217 .byte 196,98,125,24,29,208,36,0,0 // vbroadcastss 0x24d0(%rip),%ymm11 # 4118 <_sk_callback_hsw+0x13e>
9218 .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
9219 .byte 196,98,125,24,37,198,36,0,0 // vbroadcastss 0x24c6(%rip),%ymm12 # 411c <_sk_callback_hsw+0x142>
9220 .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
9221 .byte 196,98,125,24,37,188,36,0,0 // vbroadcastss 0x24bc(%rip),%ymm12 # 4120 <_sk_callback_hsw+0x146>
9222 .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9223 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9224 .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
9225 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10
9226 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10
9227 .byte 196,98,125,24,29,157,36,0,0 // vbroadcastss 0x249d(%rip),%ymm11 # 4124 <_sk_callback_hsw+0x14a>
9228 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
9229 .byte 196,98,125,24,29,147,36,0,0 // vbroadcastss 0x2493(%rip),%ymm11 # 4128 <_sk_callback_hsw+0x14e>
9230 .byte 196,98,45,172,216 // vfnmadd213ps %ymm0,%ymm10,%ymm11
9231 .byte 196,226,125,24,5,137,36,0,0 // vbroadcastss 0x2489(%rip),%ymm0 # 412c <_sk_callback_hsw+0x152>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9232 .byte 196,193,124,92,194 // vsubps %ymm10,%ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9233 .byte 196,98,125,24,21,127,36,0,0 // vbroadcastss 0x247f(%rip),%ymm10 # 4130 <_sk_callback_hsw+0x156>
9234 .byte 197,172,94,192 // vdivps %ymm0,%ymm10,%ymm0
9235 .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
9236 .byte 196,98,125,24,21,114,36,0,0 // vbroadcastss 0x2472(%rip),%ymm10 # 4134 <_sk_callback_hsw+0x15a>
9237 .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9238 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
9239 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
9240 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
9241 .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0
9242 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
9243 .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8
9244 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9245 .byte 197,249,110,192 // vmovd %eax,%xmm0
9246 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9247 .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0
9248 .byte 72,173 // lods %ds:(%rsi),%rax
9249 .byte 255,224 // jmpq *%rax
9250
9251HIDDEN _sk_parametric_g_hsw
9252.globl _sk_parametric_g_hsw
9253FUNCTION(_sk_parametric_g_hsw)
9254_sk_parametric_g_hsw:
9255 .byte 72,173 // lods %ds:(%rsi),%rax
9256 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
9257 .byte 196,65,116,194,192,2 // vcmpleps %ymm8,%ymm1,%ymm8
9258 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
9259 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
9260 .byte 196,66,117,168,202 // vfmadd213ps %ymm10,%ymm1,%ymm9
9261 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
9262 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
9263 .byte 196,66,117,168,211 // vfmadd213ps %ymm11,%ymm1,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9264 .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9265 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9266 .byte 196,98,125,24,37,249,35,0,0 // vbroadcastss 0x23f9(%rip),%ymm12 # 4138 <_sk_callback_hsw+0x15e>
9267 .byte 196,98,125,24,45,244,35,0,0 // vbroadcastss 0x23f4(%rip),%ymm13 # 413c <_sk_callback_hsw+0x162>
9268 .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
9269 .byte 196,98,125,24,45,234,35,0,0 // vbroadcastss 0x23ea(%rip),%ymm13 # 4140 <_sk_callback_hsw+0x166>
9270 .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
9271 .byte 196,98,125,24,45,224,35,0,0 // vbroadcastss 0x23e0(%rip),%ymm13 # 4144 <_sk_callback_hsw+0x16a>
9272 .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
9273 .byte 196,98,125,24,29,214,35,0,0 // vbroadcastss 0x23d6(%rip),%ymm11 # 4148 <_sk_callback_hsw+0x16e>
9274 .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
9275 .byte 196,98,125,24,37,204,35,0,0 // vbroadcastss 0x23cc(%rip),%ymm12 # 414c <_sk_callback_hsw+0x172>
9276 .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
9277 .byte 196,98,125,24,37,194,35,0,0 // vbroadcastss 0x23c2(%rip),%ymm12 # 4150 <_sk_callback_hsw+0x176>
9278 .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9279 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9280 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1
9281 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10
9282 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
9283 .byte 196,98,125,24,29,163,35,0,0 // vbroadcastss 0x23a3(%rip),%ymm11 # 4154 <_sk_callback_hsw+0x17a>
9284 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
9285 .byte 196,98,125,24,29,153,35,0,0 // vbroadcastss 0x2399(%rip),%ymm11 # 4158 <_sk_callback_hsw+0x17e>
9286 .byte 196,98,45,172,217 // vfnmadd213ps %ymm1,%ymm10,%ymm11
9287 .byte 196,226,125,24,13,143,35,0,0 // vbroadcastss 0x238f(%rip),%ymm1 # 415c <_sk_callback_hsw+0x182>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9288 .byte 196,193,116,92,202 // vsubps %ymm10,%ymm1,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9289 .byte 196,98,125,24,21,133,35,0,0 // vbroadcastss 0x2385(%rip),%ymm10 # 4160 <_sk_callback_hsw+0x186>
9290 .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
9291 .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
9292 .byte 196,98,125,24,21,120,35,0,0 // vbroadcastss 0x2378(%rip),%ymm10 # 4164 <_sk_callback_hsw+0x18a>
9293 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9294 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
9295 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
9296 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
9297 .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1
9298 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
9299 .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8
9300 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9301 .byte 197,249,110,200 // vmovd %eax,%xmm1
9302 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9303 .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1
9304 .byte 72,173 // lods %ds:(%rsi),%rax
9305 .byte 255,224 // jmpq *%rax
9306
9307HIDDEN _sk_parametric_b_hsw
9308.globl _sk_parametric_b_hsw
9309FUNCTION(_sk_parametric_b_hsw)
9310_sk_parametric_b_hsw:
9311 .byte 72,173 // lods %ds:(%rsi),%rax
9312 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
9313 .byte 196,65,108,194,192,2 // vcmpleps %ymm8,%ymm2,%ymm8
9314 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
9315 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
9316 .byte 196,66,109,168,202 // vfmadd213ps %ymm10,%ymm2,%ymm9
9317 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
9318 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
9319 .byte 196,66,109,168,211 // vfmadd213ps %ymm11,%ymm2,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9320 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9321 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9322 .byte 196,98,125,24,37,255,34,0,0 // vbroadcastss 0x22ff(%rip),%ymm12 # 4168 <_sk_callback_hsw+0x18e>
9323 .byte 196,98,125,24,45,250,34,0,0 // vbroadcastss 0x22fa(%rip),%ymm13 # 416c <_sk_callback_hsw+0x192>
9324 .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
9325 .byte 196,98,125,24,45,240,34,0,0 // vbroadcastss 0x22f0(%rip),%ymm13 # 4170 <_sk_callback_hsw+0x196>
9326 .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
9327 .byte 196,98,125,24,45,230,34,0,0 // vbroadcastss 0x22e6(%rip),%ymm13 # 4174 <_sk_callback_hsw+0x19a>
9328 .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
9329 .byte 196,98,125,24,29,220,34,0,0 // vbroadcastss 0x22dc(%rip),%ymm11 # 4178 <_sk_callback_hsw+0x19e>
9330 .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
9331 .byte 196,98,125,24,37,210,34,0,0 // vbroadcastss 0x22d2(%rip),%ymm12 # 417c <_sk_callback_hsw+0x1a2>
9332 .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
9333 .byte 196,98,125,24,37,200,34,0,0 // vbroadcastss 0x22c8(%rip),%ymm12 # 4180 <_sk_callback_hsw+0x1a6>
9334 .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9335 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9336 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
9337 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10
9338 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10
9339 .byte 196,98,125,24,29,169,34,0,0 // vbroadcastss 0x22a9(%rip),%ymm11 # 4184 <_sk_callback_hsw+0x1aa>
9340 .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
9341 .byte 196,98,125,24,29,159,34,0,0 // vbroadcastss 0x229f(%rip),%ymm11 # 4188 <_sk_callback_hsw+0x1ae>
9342 .byte 196,98,45,172,218 // vfnmadd213ps %ymm2,%ymm10,%ymm11
9343 .byte 196,226,125,24,21,149,34,0,0 // vbroadcastss 0x2295(%rip),%ymm2 # 418c <_sk_callback_hsw+0x1b2>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9344 .byte 196,193,108,92,210 // vsubps %ymm10,%ymm2,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9345 .byte 196,98,125,24,21,139,34,0,0 // vbroadcastss 0x228b(%rip),%ymm10 # 4190 <_sk_callback_hsw+0x1b6>
9346 .byte 197,172,94,210 // vdivps %ymm2,%ymm10,%ymm2
9347 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
9348 .byte 196,98,125,24,21,126,34,0,0 // vbroadcastss 0x227e(%rip),%ymm10 # 4194 <_sk_callback_hsw+0x1ba>
9349 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9350 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
9351 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
9352 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
9353 .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2
9354 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
9355 .byte 196,65,108,95,192 // vmaxps %ymm8,%ymm2,%ymm8
9356 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9357 .byte 197,249,110,208 // vmovd %eax,%xmm2
9358 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
9359 .byte 197,188,93,210 // vminps %ymm2,%ymm8,%ymm2
9360 .byte 72,173 // lods %ds:(%rsi),%rax
9361 .byte 255,224 // jmpq *%rax
9362
9363HIDDEN _sk_parametric_a_hsw
9364.globl _sk_parametric_a_hsw
9365FUNCTION(_sk_parametric_a_hsw)
9366_sk_parametric_a_hsw:
9367 .byte 72,173 // lods %ds:(%rsi),%rax
9368 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
9369 .byte 196,65,100,194,192,2 // vcmpleps %ymm8,%ymm3,%ymm8
9370 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
9371 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
9372 .byte 196,66,101,168,202 // vfmadd213ps %ymm10,%ymm3,%ymm9
9373 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
9374 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
9375 .byte 196,66,101,168,211 // vfmadd213ps %ymm11,%ymm3,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9376 .byte 196,226,125,24,24 // vbroadcastss (%rax),%ymm3
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9377 .byte 196,65,124,91,218 // vcvtdq2ps %ymm10,%ymm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9378 .byte 196,98,125,24,37,5,34,0,0 // vbroadcastss 0x2205(%rip),%ymm12 # 4198 <_sk_callback_hsw+0x1be>
9379 .byte 196,98,125,24,45,0,34,0,0 // vbroadcastss 0x2200(%rip),%ymm13 # 419c <_sk_callback_hsw+0x1c2>
9380 .byte 196,65,44,84,213 // vandps %ymm13,%ymm10,%ymm10
9381 .byte 196,98,125,24,45,246,33,0,0 // vbroadcastss 0x21f6(%rip),%ymm13 # 41a0 <_sk_callback_hsw+0x1c6>
9382 .byte 196,65,44,86,213 // vorps %ymm13,%ymm10,%ymm10
9383 .byte 196,98,125,24,45,236,33,0,0 // vbroadcastss 0x21ec(%rip),%ymm13 # 41a4 <_sk_callback_hsw+0x1ca>
9384 .byte 196,66,37,184,236 // vfmadd231ps %ymm12,%ymm11,%ymm13
9385 .byte 196,98,125,24,29,226,33,0,0 // vbroadcastss 0x21e2(%rip),%ymm11 # 41a8 <_sk_callback_hsw+0x1ce>
9386 .byte 196,66,45,172,221 // vfnmadd213ps %ymm13,%ymm10,%ymm11
9387 .byte 196,98,125,24,37,216,33,0,0 // vbroadcastss 0x21d8(%rip),%ymm12 # 41ac <_sk_callback_hsw+0x1d2>
9388 .byte 196,65,44,88,212 // vaddps %ymm12,%ymm10,%ymm10
9389 .byte 196,98,125,24,37,206,33,0,0 // vbroadcastss 0x21ce(%rip),%ymm12 # 41b0 <_sk_callback_hsw+0x1d6>
9390 .byte 196,65,28,94,210 // vdivps %ymm10,%ymm12,%ymm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9391 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9392 .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
9393 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10
9394 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10
9395 .byte 196,98,125,24,29,175,33,0,0 // vbroadcastss 0x21af(%rip),%ymm11 # 41b4 <_sk_callback_hsw+0x1da>
9396 .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
9397 .byte 196,98,125,24,29,165,33,0,0 // vbroadcastss 0x21a5(%rip),%ymm11 # 41b8 <_sk_callback_hsw+0x1de>
9398 .byte 196,98,45,172,219 // vfnmadd213ps %ymm3,%ymm10,%ymm11
9399 .byte 196,226,125,24,29,155,33,0,0 // vbroadcastss 0x219b(%rip),%ymm3 # 41bc <_sk_callback_hsw+0x1e2>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9400 .byte 196,193,100,92,218 // vsubps %ymm10,%ymm3,%ymm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9401 .byte 196,98,125,24,21,145,33,0,0 // vbroadcastss 0x2191(%rip),%ymm10 # 41c0 <_sk_callback_hsw+0x1e6>
9402 .byte 197,172,94,219 // vdivps %ymm3,%ymm10,%ymm3
9403 .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
9404 .byte 196,98,125,24,21,132,33,0,0 // vbroadcastss 0x2184(%rip),%ymm10 # 41c4 <_sk_callback_hsw+0x1ea>
9405 .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]9406 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3
9407 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
9408 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3
9409 .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3
9410 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
9411 .byte 196,65,100,95,192 // vmaxps %ymm8,%ymm3,%ymm8
9412 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9413 .byte 197,249,110,216 // vmovd %eax,%xmm3
9414 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9415 .byte 197,188,93,219 // vminps %ymm3,%ymm8,%ymm3
9416 .byte 72,173 // lods %ds:(%rsi),%rax
9417 .byte 255,224 // jmpq *%rax
9418
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9419HIDDEN _sk_lab_to_xyz_hsw
9420.globl _sk_lab_to_xyz_hsw
9421FUNCTION(_sk_lab_to_xyz_hsw)
9422_sk_lab_to_xyz_hsw:
9423 .byte 184,0,0,200,66 // mov $0x42c80000,%eax
9424 .byte 197,121,110,192 // vmovd %eax,%xmm8
9425 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9426 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
9427 .byte 197,121,110,200 // vmovd %eax,%xmm9
9428 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
9429 .byte 184,0,0,0,67 // mov $0x43000000,%eax
9430 .byte 197,121,110,208 // vmovd %eax,%xmm10
9431 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
9432 .byte 196,194,53,170,202 // vfmsub213ps %ymm10,%ymm9,%ymm1
9433 .byte 196,194,53,170,210 // vfmsub213ps %ymm10,%ymm9,%ymm2
9434 .byte 184,0,0,128,65 // mov $0x41800000,%eax
9435 .byte 197,121,110,200 // vmovd %eax,%xmm9
9436 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
9437 .byte 196,66,125,168,193 // vfmadd213ps %ymm9,%ymm0,%ymm8
9438 .byte 184,203,61,13,60 // mov $0x3c0d3dcb,%eax
9439 .byte 197,249,110,192 // vmovd %eax,%xmm0
9440 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9441 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
9442 .byte 184,111,18,3,59 // mov $0x3b03126f,%eax
9443 .byte 197,249,110,192 // vmovd %eax,%xmm0
9444 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9445 .byte 196,194,117,168,192 // vfmadd213ps %ymm8,%ymm1,%ymm0
9446 .byte 184,10,215,163,59 // mov $0x3ba3d70a,%eax
9447 .byte 197,249,110,200 // vmovd %eax,%xmm1
9448 .byte 196,98,125,88,201 // vpbroadcastd %xmm1,%ymm9
9449 .byte 196,66,109,172,200 // vfnmadd213ps %ymm8,%ymm2,%ymm9
9450 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1
9451 .byte 197,124,89,217 // vmulps %ymm1,%ymm0,%ymm11
9452 .byte 184,194,24,17,60 // mov $0x3c1118c2,%eax
9453 .byte 197,249,110,208 // vmovd %eax,%xmm2
9454 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
9455 .byte 196,65,108,194,211,1 // vcmpltps %ymm11,%ymm2,%ymm10
9456 .byte 184,203,61,13,62 // mov $0x3e0d3dcb,%eax
9457 .byte 197,249,110,200 // vmovd %eax,%xmm1
9458 .byte 196,98,125,88,225 // vpbroadcastd %xmm1,%ymm12
9459 .byte 196,193,124,92,196 // vsubps %ymm12,%ymm0,%ymm0
9460 .byte 184,80,128,3,62 // mov $0x3e038050,%eax
9461 .byte 197,249,110,200 // vmovd %eax,%xmm1
9462 .byte 196,98,125,88,233 // vpbroadcastd %xmm1,%ymm13
9463 .byte 196,193,124,89,197 // vmulps %ymm13,%ymm0,%ymm0
9464 .byte 196,67,125,74,219,160 // vblendvps %ymm10,%ymm11,%ymm0,%ymm11
9465 .byte 196,193,60,89,200 // vmulps %ymm8,%ymm8,%ymm1
9466 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
9467 .byte 197,108,194,209,1 // vcmpltps %ymm1,%ymm2,%ymm10
9468 .byte 196,65,60,92,196 // vsubps %ymm12,%ymm8,%ymm8
9469 .byte 196,65,60,89,197 // vmulps %ymm13,%ymm8,%ymm8
9470 .byte 196,227,61,74,201,160 // vblendvps %ymm10,%ymm1,%ymm8,%ymm1
9471 .byte 196,65,52,89,193 // vmulps %ymm9,%ymm9,%ymm8
9472 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
9473 .byte 196,193,108,194,208,1 // vcmpltps %ymm8,%ymm2,%ymm2
9474 .byte 196,65,52,92,204 // vsubps %ymm12,%ymm9,%ymm9
9475 .byte 196,65,52,89,205 // vmulps %ymm13,%ymm9,%ymm9
9476 .byte 196,67,53,74,192,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm8
9477 .byte 184,31,215,118,63 // mov $0x3f76d71f,%eax
9478 .byte 197,249,110,192 // vmovd %eax,%xmm0
9479 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9480 .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0
9481 .byte 184,246,64,83,63 // mov $0x3f5340f6,%eax
9482 .byte 197,249,110,208 // vmovd %eax,%xmm2
9483 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
9484 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
9485 .byte 72,173 // lods %ds:(%rsi),%rax
9486 .byte 255,224 // jmpq *%rax
9487
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]9488HIDDEN _sk_load_a8_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9489.globl _sk_load_a8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9490FUNCTION(_sk_load_a8_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9491_sk_load_a8_hsw:
9492 .byte 73,137,200 // mov %rcx,%r8
9493 .byte 72,173 // lods %ds:(%rsi),%rax
9494 .byte 72,139,0 // mov (%rax),%rax
9495 .byte 72,1,248 // add %rdi,%rax
9496 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9497 .byte 117,50 // jne 21e9 <_sk_load_a8_hsw+0x42>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9498 .byte 197,250,126,0 // vmovq (%rax),%xmm0
9499 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9500 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9501 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
9502 .byte 197,249,110,200 // vmovd %eax,%xmm1
9503 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9504 .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
9505 .byte 72,173 // lods %ds:(%rsi),%rax
9506 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
9507 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
9508 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
9509 .byte 76,137,193 // mov %r8,%rcx
9510 .byte 255,224 // jmpq *%rax
9511 .byte 49,201 // xor %ecx,%ecx
9512 .byte 77,137,194 // mov %r8,%r10
9513 .byte 69,49,201 // xor %r9d,%r9d
9514 .byte 68,15,182,24 // movzbl (%rax),%r11d
9515 .byte 72,255,192 // inc %rax
9516 .byte 73,211,227 // shl %cl,%r11
9517 .byte 77,9,217 // or %r11,%r9
9518 .byte 72,131,193,8 // add $0x8,%rcx
9519 .byte 73,255,202 // dec %r10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9520 .byte 117,234 // jne 21f1 <_sk_load_a8_hsw+0x4a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9521 .byte 196,193,249,110,193 // vmovq %r9,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9522 .byte 235,173 // jmp 21bb <_sk_load_a8_hsw+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9523
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9524HIDDEN _sk_gather_a8_hsw
9525.globl _sk_gather_a8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9526FUNCTION(_sk_gather_a8_hsw)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9527_sk_gather_a8_hsw:
9528 .byte 65,87 // push %r15
9529 .byte 65,86 // push %r14
9530 .byte 65,84 // push %r12
9531 .byte 83 // push %rbx
9532 .byte 72,173 // lods %ds:(%rsi),%rax
9533 .byte 76,139,0 // mov (%rax),%r8
9534 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
9535 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
9536 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
9537 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
9538 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
9539 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
9540 .byte 65,137,193 // mov %eax,%r9d
9541 .byte 72,193,232,32 // shr $0x20,%rax
9542 .byte 196,193,249,126,194 // vmovq %xmm0,%r10
9543 .byte 69,137,211 // mov %r10d,%r11d
9544 .byte 73,193,234,32 // shr $0x20,%r10
9545 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
9546 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
9547 .byte 65,137,222 // mov %ebx,%r14d
9548 .byte 72,193,235,32 // shr $0x20,%rbx
9549 .byte 196,193,249,126,199 // vmovq %xmm0,%r15
9550 .byte 69,137,252 // mov %r15d,%r12d
9551 .byte 73,193,239,32 // shr $0x20,%r15
9552 .byte 196,131,121,32,4,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0
9553 .byte 196,131,121,32,4,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0
9554 .byte 71,15,182,12,8 // movzbl (%r8,%r9,1),%r9d
9555 .byte 196,195,121,32,193,2 // vpinsrb $0x2,%r9d,%xmm0,%xmm0
9556 .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax
9557 .byte 196,227,121,32,192,3 // vpinsrb $0x3,%eax,%xmm0,%xmm0
9558 .byte 67,15,182,4,32 // movzbl (%r8,%r12,1),%eax
9559 .byte 196,227,121,32,192,4 // vpinsrb $0x4,%eax,%xmm0,%xmm0
9560 .byte 67,15,182,4,56 // movzbl (%r8,%r15,1),%eax
9561 .byte 196,227,121,32,192,5 // vpinsrb $0x5,%eax,%xmm0,%xmm0
9562 .byte 67,15,182,4,48 // movzbl (%r8,%r14,1),%eax
9563 .byte 196,227,121,32,192,6 // vpinsrb $0x6,%eax,%xmm0,%xmm0
9564 .byte 65,15,182,4,24 // movzbl (%r8,%rbx,1),%eax
9565 .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0
9566 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9567 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9568 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
9569 .byte 197,249,110,200 // vmovd %eax,%xmm1
9570 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9571 .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
9572 .byte 72,173 // lods %ds:(%rsi),%rax
9573 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
9574 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
9575 .byte 197,237,239,210 // vpxor %ymm2,%ymm2,%ymm2
9576 .byte 91 // pop %rbx
9577 .byte 65,92 // pop %r12
9578 .byte 65,94 // pop %r14
9579 .byte 65,95 // pop %r15
9580 .byte 255,224 // jmpq *%rax
9581
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]9582HIDDEN _sk_store_a8_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9583.globl _sk_store_a8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9584FUNCTION(_sk_store_a8_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9585_sk_store_a8_hsw:
9586 .byte 72,173 // lods %ds:(%rsi),%rax
9587 .byte 76,139,8 // mov (%rax),%r9
9588 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
9589 .byte 197,121,110,192 // vmovd %eax,%xmm8
9590 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9591 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
9592 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
9593 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
9594 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
9595 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
9596 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9597 .byte 117,10 // jne 2326 <_sk_store_a8_hsw+0x3b>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9598 .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1)
9599 .byte 72,173 // lods %ds:(%rsi),%rax
9600 .byte 255,224 // jmpq *%rax
9601 .byte 65,137,200 // mov %ecx,%r8d
9602 .byte 65,128,224,7 // and $0x7,%r8b
9603 .byte 65,254,200 // dec %r8b
9604 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9605 .byte 119,236 // ja 2322 <_sk_store_a8_hsw+0x37>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9606 .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
9607 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9608 .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2388 <_sk_store_a8_hsw+0x9d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9609 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
9610 .byte 76,1,192 // add %r8,%rax
9611 .byte 255,224 // jmpq *%rax
9612 .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1)
9613 .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1)
9614 .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1)
9615 .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1)
9616 .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
9617 .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
9618 .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9619 .byte 235,154 // jmp 2322 <_sk_store_a8_hsw+0x37>
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9620 .byte 247,255 // idiv %edi
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]9621 .byte 255 // (bad)
9622 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9623 .byte 239 // out %eax,(%dx)
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]9624 .byte 255 // (bad)
9625 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9626 .byte 255,231 // jmpq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9627 .byte 255 // (bad)
9628 .byte 255 // (bad)
9629 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9630 .byte 223,255 // (bad)
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]9631 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9632 .byte 255,215 // callq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9633 .byte 255 // (bad)
9634 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9635 .byte 255,207 // dec %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9636 .byte 255 // (bad)
9637 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]9638 .byte 255,199 // inc %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9639 .byte 255 // (bad)
9640 .byte 255 // (bad)
9641 .byte 255 // .byte 0xff
9642
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]9643HIDDEN _sk_load_g8_hsw
9644.globl _sk_load_g8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9645FUNCTION(_sk_load_g8_hsw)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]9646_sk_load_g8_hsw:
9647 .byte 73,137,200 // mov %rcx,%r8
9648 .byte 72,173 // lods %ds:(%rsi),%rax
9649 .byte 72,139,0 // mov (%rax),%rax
9650 .byte 72,1,248 // add %rdi,%rax
9651 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9652 .byte 117,60 // jne 23f0 <_sk_load_g8_hsw+0x4c>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]9653 .byte 197,250,126,0 // vmovq (%rax),%xmm0
9654 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9655 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9656 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
9657 .byte 197,249,110,200 // vmovd %eax,%xmm1
9658 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9659 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
9660 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9661 .byte 197,249,110,200 // vmovd %eax,%xmm1
9662 .byte 196,226,125,88,217 // vpbroadcastd %xmm1,%ymm3
9663 .byte 72,173 // lods %ds:(%rsi),%rax
9664 .byte 76,137,193 // mov %r8,%rcx
9665 .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
9666 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
9667 .byte 255,224 // jmpq *%rax
9668 .byte 49,201 // xor %ecx,%ecx
9669 .byte 77,137,194 // mov %r8,%r10
9670 .byte 69,49,201 // xor %r9d,%r9d
9671 .byte 68,15,182,24 // movzbl (%rax),%r11d
9672 .byte 72,255,192 // inc %rax
9673 .byte 73,211,227 // shl %cl,%r11
9674 .byte 77,9,217 // or %r11,%r9
9675 .byte 72,131,193,8 // add $0x8,%rcx
9676 .byte 73,255,202 // dec %r10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9677 .byte 117,234 // jne 23f8 <_sk_load_g8_hsw+0x54>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]9678 .byte 196,193,249,110,193 // vmovq %r9,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9679 .byte 235,163 // jmp 23b8 <_sk_load_g8_hsw+0x14>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9680
9681HIDDEN _sk_gather_g8_hsw
9682.globl _sk_gather_g8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9683FUNCTION(_sk_gather_g8_hsw)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9684_sk_gather_g8_hsw:
9685 .byte 65,87 // push %r15
9686 .byte 65,86 // push %r14
9687 .byte 65,84 // push %r12
9688 .byte 83 // push %rbx
9689 .byte 72,173 // lods %ds:(%rsi),%rax
9690 .byte 76,139,0 // mov (%rax),%r8
9691 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
9692 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
9693 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
9694 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
9695 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
9696 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
9697 .byte 65,137,193 // mov %eax,%r9d
9698 .byte 72,193,232,32 // shr $0x20,%rax
9699 .byte 196,193,249,126,194 // vmovq %xmm0,%r10
9700 .byte 69,137,211 // mov %r10d,%r11d
9701 .byte 73,193,234,32 // shr $0x20,%r10
9702 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
9703 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
9704 .byte 65,137,222 // mov %ebx,%r14d
9705 .byte 72,193,235,32 // shr $0x20,%rbx
9706 .byte 196,193,249,126,199 // vmovq %xmm0,%r15
9707 .byte 69,137,252 // mov %r15d,%r12d
9708 .byte 73,193,239,32 // shr $0x20,%r15
9709 .byte 196,131,121,32,4,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm0
9710 .byte 196,131,121,32,4,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm0,%xmm0
9711 .byte 71,15,182,12,8 // movzbl (%r8,%r9,1),%r9d
9712 .byte 196,195,121,32,193,2 // vpinsrb $0x2,%r9d,%xmm0,%xmm0
9713 .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax
9714 .byte 196,227,121,32,192,3 // vpinsrb $0x3,%eax,%xmm0,%xmm0
9715 .byte 67,15,182,4,32 // movzbl (%r8,%r12,1),%eax
9716 .byte 196,227,121,32,192,4 // vpinsrb $0x4,%eax,%xmm0,%xmm0
9717 .byte 67,15,182,4,56 // movzbl (%r8,%r15,1),%eax
9718 .byte 196,227,121,32,192,5 // vpinsrb $0x5,%eax,%xmm0,%xmm0
9719 .byte 67,15,182,4,48 // movzbl (%r8,%r14,1),%eax
9720 .byte 196,227,121,32,192,6 // vpinsrb $0x6,%eax,%xmm0,%xmm0
9721 .byte 65,15,182,4,24 // movzbl (%r8,%rbx,1),%eax
9722 .byte 196,227,121,32,192,7 // vpinsrb $0x7,%eax,%xmm0,%xmm0
9723 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9724 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9725 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
9726 .byte 197,249,110,200 // vmovd %eax,%xmm1
9727 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9728 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
9729 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9730 .byte 197,249,110,200 // vmovd %eax,%xmm1
9731 .byte 196,226,125,88,217 // vpbroadcastd %xmm1,%ymm3
9732 .byte 72,173 // lods %ds:(%rsi),%rax
9733 .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
9734 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
9735 .byte 91 // pop %rbx
9736 .byte 65,92 // pop %r12
9737 .byte 65,94 // pop %r14
9738 .byte 65,95 // pop %r15
9739 .byte 255,224 // jmpq *%rax
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]9740
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]9741HIDDEN _sk_gather_i8_hsw
9742.globl _sk_gather_i8_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9743FUNCTION(_sk_gather_i8_hsw)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]9744_sk_gather_i8_hsw:
9745 .byte 72,173 // lods %ds:(%rsi),%rax
9746 .byte 73,137,192 // mov %rax,%r8
9747 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9748 .byte 116,5 // je 250b <_sk_gather_i8_hsw+0xf>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]9749 .byte 76,137,192 // mov %r8,%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9750 .byte 235,2 // jmp 250d <_sk_gather_i8_hsw+0x11>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]9751 .byte 72,173 // lods %ds:(%rsi),%rax
9752 .byte 65,87 // push %r15
9753 .byte 65,86 // push %r14
9754 .byte 65,85 // push %r13
9755 .byte 65,84 // push %r12
9756 .byte 83 // push %rbx
9757 .byte 76,139,8 // mov (%rax),%r9
9758 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
9759 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
9760 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
9761 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
9762 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
9763 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
9764 .byte 65,137,194 // mov %eax,%r10d
9765 .byte 72,193,232,32 // shr $0x20,%rax
9766 .byte 196,193,249,126,195 // vmovq %xmm0,%r11
9767 .byte 69,137,222 // mov %r11d,%r14d
9768 .byte 73,193,235,32 // shr $0x20,%r11
9769 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
9770 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
9771 .byte 65,137,223 // mov %ebx,%r15d
9772 .byte 72,193,235,32 // shr $0x20,%rbx
9773 .byte 196,193,249,126,196 // vmovq %xmm0,%r12
9774 .byte 69,137,229 // mov %r12d,%r13d
9775 .byte 73,193,236,32 // shr $0x20,%r12
9776 .byte 196,131,121,32,4,49,0 // vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0
9777 .byte 196,131,121,32,4,25,1 // vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0
9778 .byte 196,131,121,32,4,17,2 // vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0
9779 .byte 196,195,121,32,4,1,3 // vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0
9780 .byte 196,131,121,32,4,41,4 // vpinsrb $0x4,(%r9,%r13,1),%xmm0,%xmm0
9781 .byte 196,131,121,32,4,33,5 // vpinsrb $0x5,(%r9,%r12,1),%xmm0,%xmm0
9782 .byte 196,131,121,32,4,57,6 // vpinsrb $0x6,(%r9,%r15,1),%xmm0,%xmm0
9783 .byte 196,195,121,32,4,25,7 // vpinsrb $0x7,(%r9,%rbx,1),%xmm0,%xmm0
9784 .byte 196,226,125,49,192 // vpmovzxbd %xmm0,%ymm0
9785 .byte 73,139,64,8 // mov 0x8(%r8),%rax
9786 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
9787 .byte 196,226,117,144,28,128 // vpgatherdd %ymm1,(%rax,%ymm0,4),%ymm3
9788 .byte 184,255,0,0,0 // mov $0xff,%eax
9789 .byte 197,249,110,192 // vmovd %eax,%xmm0
9790 .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2
9791 .byte 197,237,219,195 // vpand %ymm3,%ymm2,%ymm0
9792 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9793 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
9794 .byte 197,249,110,200 // vmovd %eax,%xmm1
9795 .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8
9796 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
9797 .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1
9798 .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
9799 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
9800 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
9801 .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9
9802 .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
9803 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
9804 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
9805 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
9806 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
9807 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
9808 .byte 72,173 // lods %ds:(%rsi),%rax
9809 .byte 91 // pop %rbx
9810 .byte 65,92 // pop %r12
9811 .byte 65,93 // pop %r13
9812 .byte 65,94 // pop %r14
9813 .byte 65,95 // pop %r15
9814 .byte 255,224 // jmpq *%rax
9815
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]9816HIDDEN _sk_load_565_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9817.globl _sk_load_565_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9818FUNCTION(_sk_load_565_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9819_sk_load_565_hsw:
9820 .byte 72,173 // lods %ds:(%rsi),%rax
9821 .byte 76,139,16 // mov (%rax),%r10
9822 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9823 .byte 15,133,149,0,0,0 // jne 26bf <_sk_load_565_hsw+0xa3>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9824 .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
9825 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
9826 .byte 184,0,248,0,0 // mov $0xf800,%eax
9827 .byte 197,249,110,192 // vmovd %eax,%xmm0
9828 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9829 .byte 197,253,219,194 // vpand %ymm2,%ymm0,%ymm0
9830 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9831 .byte 184,8,33,132,55 // mov $0x37842108,%eax
9832 .byte 197,249,110,200 // vmovd %eax,%xmm1
9833 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9834 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
9835 .byte 184,224,7,0,0 // mov $0x7e0,%eax
9836 .byte 197,249,110,200 // vmovd %eax,%xmm1
9837 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9838 .byte 197,245,219,202 // vpand %ymm2,%ymm1,%ymm1
9839 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
9840 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
9841 .byte 197,249,110,216 // vmovd %eax,%xmm3
9842 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9843 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
9844 .byte 184,31,0,0,0 // mov $0x1f,%eax
9845 .byte 197,249,110,216 // vmovd %eax,%xmm3
9846 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9847 .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
9848 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
9849 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
9850 .byte 197,249,110,216 // vmovd %eax,%xmm3
9851 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9852 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
9853 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9854 .byte 197,249,110,216 // vmovd %eax,%xmm3
9855 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9856 .byte 72,173 // lods %ds:(%rsi),%rax
9857 .byte 255,224 // jmpq *%rax
9858 .byte 65,137,200 // mov %ecx,%r8d
9859 .byte 65,128,224,7 // and $0x7,%r8b
9860 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
9861 .byte 65,254,200 // dec %r8b
9862 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9863 .byte 15,135,89,255,255,255 // ja 2630 <_sk_load_565_hsw+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9864 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9865 .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 272c <_sk_load_565_hsw+0x110>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9866 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
9867 .byte 76,1,200 // add %r9,%rax
9868 .byte 255,224 // jmpq *%rax
9869 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
9870 .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
9871 .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
9872 .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
9873 .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
9874 .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
9875 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
9876 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9877 .byte 233,5,255,255,255 // jmpq 2630 <_sk_load_565_hsw+0x14>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9878 .byte 144 // nop
9879 .byte 243,255 // repz (bad)
9880 .byte 255 // (bad)
9881 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]9882 .byte 235,255 // jmp 2731 <_sk_load_565_hsw+0x115>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9883 .byte 255 // (bad)
9884 .byte 255,227 // jmpq *%rbx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9885 .byte 255 // (bad)
9886 .byte 255 // (bad)
9887 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9888 .byte 219,255 // (bad)
9889 .byte 255 // (bad)
9890 .byte 255,211 // callq *%rbx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9891 .byte 255 // (bad)
9892 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9893 .byte 255,203 // dec %ebx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9894 .byte 255 // (bad)
9895 .byte 255 // (bad)
9896 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9897 .byte 191 // .byte 0xbf
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9898 .byte 255 // (bad)
9899 .byte 255 // (bad)
9900 .byte 255 // .byte 0xff
9901
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9902HIDDEN _sk_gather_565_hsw
9903.globl _sk_gather_565_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9904FUNCTION(_sk_gather_565_hsw)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]9905_sk_gather_565_hsw:
9906 .byte 65,87 // push %r15
9907 .byte 65,86 // push %r14
9908 .byte 65,84 // push %r12
9909 .byte 83 // push %rbx
9910 .byte 72,173 // lods %ds:(%rsi),%rax
9911 .byte 76,139,0 // mov (%rax),%r8
9912 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
9913 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
9914 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
9915 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
9916 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
9917 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
9918 .byte 65,137,193 // mov %eax,%r9d
9919 .byte 72,193,232,32 // shr $0x20,%rax
9920 .byte 196,193,249,126,194 // vmovq %xmm0,%r10
9921 .byte 69,137,211 // mov %r10d,%r11d
9922 .byte 73,193,234,32 // shr $0x20,%r10
9923 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
9924 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
9925 .byte 65,137,222 // mov %ebx,%r14d
9926 .byte 72,193,235,32 // shr $0x20,%rbx
9927 .byte 196,193,249,126,199 // vmovq %xmm0,%r15
9928 .byte 69,137,252 // mov %r15d,%r12d
9929 .byte 73,193,239,32 // shr $0x20,%r15
9930 .byte 71,15,183,20,80 // movzwl (%r8,%r10,2),%r10d
9931 .byte 71,15,183,28,88 // movzwl (%r8,%r11,2),%r11d
9932 .byte 196,193,121,110,195 // vmovd %r11d,%xmm0
9933 .byte 196,193,121,196,194,1 // vpinsrw $0x1,%r10d,%xmm0,%xmm0
9934 .byte 71,15,183,12,72 // movzwl (%r8,%r9,2),%r9d
9935 .byte 196,193,121,196,193,2 // vpinsrw $0x2,%r9d,%xmm0,%xmm0
9936 .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax
9937 .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
9938 .byte 67,15,183,4,96 // movzwl (%r8,%r12,2),%eax
9939 .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
9940 .byte 67,15,183,4,120 // movzwl (%r8,%r15,2),%eax
9941 .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
9942 .byte 67,15,183,4,112 // movzwl (%r8,%r14,2),%eax
9943 .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
9944 .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax
9945 .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
9946 .byte 196,226,125,51,208 // vpmovzxwd %xmm0,%ymm2
9947 .byte 184,0,248,0,0 // mov $0xf800,%eax
9948 .byte 197,249,110,192 // vmovd %eax,%xmm0
9949 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
9950 .byte 197,253,219,194 // vpand %ymm2,%ymm0,%ymm0
9951 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
9952 .byte 184,8,33,132,55 // mov $0x37842108,%eax
9953 .byte 197,249,110,200 // vmovd %eax,%xmm1
9954 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9955 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
9956 .byte 184,224,7,0,0 // mov $0x7e0,%eax
9957 .byte 197,249,110,200 // vmovd %eax,%xmm1
9958 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
9959 .byte 197,245,219,202 // vpand %ymm2,%ymm1,%ymm1
9960 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
9961 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
9962 .byte 197,249,110,216 // vmovd %eax,%xmm3
9963 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9964 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
9965 .byte 184,31,0,0,0 // mov $0x1f,%eax
9966 .byte 197,249,110,216 // vmovd %eax,%xmm3
9967 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9968 .byte 197,229,219,210 // vpand %ymm2,%ymm3,%ymm2
9969 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
9970 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
9971 .byte 197,249,110,216 // vmovd %eax,%xmm3
9972 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9973 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
9974 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
9975 .byte 197,249,110,216 // vmovd %eax,%xmm3
9976 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
9977 .byte 72,173 // lods %ds:(%rsi),%rax
9978 .byte 91 // pop %rbx
9979 .byte 65,92 // pop %r12
9980 .byte 65,94 // pop %r14
9981 .byte 65,95 // pop %r15
9982 .byte 255,224 // jmpq *%rax
9983
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]9984HIDDEN _sk_store_565_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9985.globl _sk_store_565_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]9986FUNCTION(_sk_store_565_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]9987_sk_store_565_hsw:
9988 .byte 72,173 // lods %ds:(%rsi),%rax
9989 .byte 76,139,8 // mov (%rax),%r9
9990 .byte 184,0,0,248,65 // mov $0x41f80000,%eax
9991 .byte 197,121,110,192 // vmovd %eax,%xmm8
9992 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
9993 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
9994 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
9995 .byte 196,193,53,114,241,11 // vpslld $0xb,%ymm9,%ymm9
9996 .byte 184,0,0,124,66 // mov $0x427c0000,%eax
9997 .byte 197,121,110,208 // vmovd %eax,%xmm10
9998 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
9999 .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10
10000 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10001 .byte 196,193,45,114,242,5 // vpslld $0x5,%ymm10,%ymm10
10002 .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9
10003 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
10004 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
10005 .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
10006 .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
10007 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
10008 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10009 .byte 117,10 // jne 28f7 <_sk_store_565_hsw+0x6c>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10010 .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
10011 .byte 72,173 // lods %ds:(%rsi),%rax
10012 .byte 255,224 // jmpq *%rax
10013 .byte 65,137,200 // mov %ecx,%r8d
10014 .byte 65,128,224,7 // and $0x7,%r8b
10015 .byte 65,254,200 // dec %r8b
10016 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10017 .byte 119,236 // ja 28f3 <_sk_store_565_hsw+0x68>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10018 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10019 .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2954 <_sk_store_565_hsw+0xc9>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10020 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
10021 .byte 76,1,192 // add %r8,%rax
10022 .byte 255,224 // jmpq *%rax
10023 .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
10024 .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
10025 .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
10026 .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
10027 .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
10028 .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
10029 .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10030 .byte 235,159 // jmp 28f3 <_sk_store_565_hsw+0x68>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10031 .byte 247,255 // idiv %edi
10032 .byte 255 // (bad)
10033 .byte 255 // (bad)
10034 .byte 239 // out %eax,(%dx)
10035 .byte 255 // (bad)
10036 .byte 255 // (bad)
10037 .byte 255,231 // jmpq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10038 .byte 255 // (bad)
10039 .byte 255 // (bad)
10040 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10041 .byte 223,255 // (bad)
10042 .byte 255 // (bad)
10043 .byte 255,215 // callq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10044 .byte 255 // (bad)
10045 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10046 .byte 255,207 // dec %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10047 .byte 255 // (bad)
10048 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10049 .byte 255,199 // inc %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10050 .byte 255 // (bad)
10051 .byte 255 // (bad)
10052 .byte 255 // .byte 0xff
10053
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10054HIDDEN _sk_load_4444_hsw
10055.globl _sk_load_4444_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10056FUNCTION(_sk_load_4444_hsw)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10057_sk_load_4444_hsw:
10058 .byte 72,173 // lods %ds:(%rsi),%rax
10059 .byte 76,139,16 // mov (%rax),%r10
10060 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10061 .byte 15,133,179,0,0,0 // jne 2a31 <_sk_load_4444_hsw+0xc1>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10062 .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
10063 .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9
10064 .byte 184,0,240,0,0 // mov $0xf000,%eax
10065 .byte 197,249,110,192 // vmovd %eax,%xmm0
10066 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
10067 .byte 196,193,125,219,193 // vpand %ymm9,%ymm0,%ymm0
10068 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10069 .byte 184,137,136,136,55 // mov $0x37888889,%eax
10070 .byte 197,249,110,200 // vmovd %eax,%xmm1
10071 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
10072 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
10073 .byte 184,0,15,0,0 // mov $0xf00,%eax
10074 .byte 197,249,110,200 // vmovd %eax,%xmm1
10075 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
10076 .byte 196,193,117,219,201 // vpand %ymm9,%ymm1,%ymm1
10077 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10078 .byte 184,137,136,136,57 // mov $0x39888889,%eax
10079 .byte 197,249,110,208 // vmovd %eax,%xmm2
10080 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10081 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
10082 .byte 184,240,0,0,0 // mov $0xf0,%eax
10083 .byte 197,249,110,208 // vmovd %eax,%xmm2
10084 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10085 .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
10086 .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
10087 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
10088 .byte 197,249,110,208 // vmovd %eax,%xmm2
10089 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10090 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
10091 .byte 184,15,0,0,0 // mov $0xf,%eax
10092 .byte 197,249,110,216 // vmovd %eax,%xmm3
10093 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
10094 .byte 196,193,101,219,217 // vpand %ymm9,%ymm3,%ymm3
10095 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
10096 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
10097 .byte 197,249,110,216 // vmovd %eax,%xmm3
10098 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
10099 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
10100 .byte 72,173 // lods %ds:(%rsi),%rax
10101 .byte 255,224 // jmpq *%rax
10102 .byte 65,137,200 // mov %ecx,%r8d
10103 .byte 65,128,224,7 // and $0x7,%r8b
10104 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
10105 .byte 65,254,200 // dec %r8b
10106 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10107 .byte 15,135,59,255,255,255 // ja 2984 <_sk_load_4444_hsw+0x14>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10108 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10109 .byte 76,141,13,76,0,0,0 // lea 0x4c(%rip),%r9 # 2aa0 <_sk_load_4444_hsw+0x130>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10110 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
10111 .byte 76,1,200 // add %r9,%rax
10112 .byte 255,224 // jmpq *%rax
10113 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
10114 .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
10115 .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
10116 .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
10117 .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
10118 .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
10119 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
10120 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10121 .byte 233,231,254,255,255 // jmpq 2984 <_sk_load_4444_hsw+0x14>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10122 .byte 15,31,0 // nopl (%rax)
10123 .byte 241 // icebp
10124 .byte 255 // (bad)
10125 .byte 255 // (bad)
10126 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10127 .byte 233,255,255,255,225 // jmpq ffffffffe2002aa8 <_sk_callback_hsw+0xffffffffe1ffeace>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10128 .byte 255 // (bad)
10129 .byte 255 // (bad)
10130 .byte 255 // (bad)
10131 .byte 217,255 // fcos
10132 .byte 255 // (bad)
10133 .byte 255,209 // callq *%rcx
10134 .byte 255 // (bad)
10135 .byte 255 // (bad)
10136 .byte 255,201 // dec %ecx
10137 .byte 255 // (bad)
10138 .byte 255 // (bad)
10139 .byte 255 // (bad)
10140 .byte 189 // .byte 0xbd
10141 .byte 255 // (bad)
10142 .byte 255 // (bad)
10143 .byte 255 // .byte 0xff
10144
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10145HIDDEN _sk_gather_4444_hsw
10146.globl _sk_gather_4444_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10147FUNCTION(_sk_gather_4444_hsw)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10148_sk_gather_4444_hsw:
10149 .byte 65,87 // push %r15
10150 .byte 65,86 // push %r14
10151 .byte 65,84 // push %r12
10152 .byte 83 // push %rbx
10153 .byte 72,173 // lods %ds:(%rsi),%rax
10154 .byte 76,139,0 // mov (%rax),%r8
10155 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
10156 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
10157 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
10158 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
10159 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
10160 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
10161 .byte 65,137,193 // mov %eax,%r9d
10162 .byte 72,193,232,32 // shr $0x20,%rax
10163 .byte 196,193,249,126,194 // vmovq %xmm0,%r10
10164 .byte 69,137,211 // mov %r10d,%r11d
10165 .byte 73,193,234,32 // shr $0x20,%r10
10166 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
10167 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
10168 .byte 65,137,222 // mov %ebx,%r14d
10169 .byte 72,193,235,32 // shr $0x20,%rbx
10170 .byte 196,193,249,126,199 // vmovq %xmm0,%r15
10171 .byte 69,137,252 // mov %r15d,%r12d
10172 .byte 73,193,239,32 // shr $0x20,%r15
10173 .byte 71,15,183,20,80 // movzwl (%r8,%r10,2),%r10d
10174 .byte 71,15,183,28,88 // movzwl (%r8,%r11,2),%r11d
10175 .byte 196,193,121,110,195 // vmovd %r11d,%xmm0
10176 .byte 196,193,121,196,194,1 // vpinsrw $0x1,%r10d,%xmm0,%xmm0
10177 .byte 71,15,183,12,72 // movzwl (%r8,%r9,2),%r9d
10178 .byte 196,193,121,196,193,2 // vpinsrw $0x2,%r9d,%xmm0,%xmm0
10179 .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax
10180 .byte 197,249,196,192,3 // vpinsrw $0x3,%eax,%xmm0,%xmm0
10181 .byte 67,15,183,4,96 // movzwl (%r8,%r12,2),%eax
10182 .byte 197,249,196,192,4 // vpinsrw $0x4,%eax,%xmm0,%xmm0
10183 .byte 67,15,183,4,120 // movzwl (%r8,%r15,2),%eax
10184 .byte 197,249,196,192,5 // vpinsrw $0x5,%eax,%xmm0,%xmm0
10185 .byte 67,15,183,4,112 // movzwl (%r8,%r14,2),%eax
10186 .byte 197,249,196,192,6 // vpinsrw $0x6,%eax,%xmm0,%xmm0
10187 .byte 65,15,183,4,88 // movzwl (%r8,%rbx,2),%eax
10188 .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
10189 .byte 196,98,125,51,200 // vpmovzxwd %xmm0,%ymm9
10190 .byte 184,0,240,0,0 // mov $0xf000,%eax
10191 .byte 197,249,110,192 // vmovd %eax,%xmm0
10192 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
10193 .byte 196,193,125,219,193 // vpand %ymm9,%ymm0,%ymm0
10194 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10195 .byte 184,137,136,136,55 // mov $0x37888889,%eax
10196 .byte 197,249,110,200 // vmovd %eax,%xmm1
10197 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
10198 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
10199 .byte 184,0,15,0,0 // mov $0xf00,%eax
10200 .byte 197,249,110,200 // vmovd %eax,%xmm1
10201 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
10202 .byte 196,193,117,219,201 // vpand %ymm9,%ymm1,%ymm1
10203 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10204 .byte 184,137,136,136,57 // mov $0x39888889,%eax
10205 .byte 197,249,110,208 // vmovd %eax,%xmm2
10206 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10207 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
10208 .byte 184,240,0,0,0 // mov $0xf0,%eax
10209 .byte 197,249,110,208 // vmovd %eax,%xmm2
10210 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10211 .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
10212 .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
10213 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
10214 .byte 197,249,110,208 // vmovd %eax,%xmm2
10215 .byte 196,226,125,88,210 // vpbroadcastd %xmm2,%ymm2
10216 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
10217 .byte 184,15,0,0,0 // mov $0xf,%eax
10218 .byte 197,249,110,216 // vmovd %eax,%xmm3
10219 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
10220 .byte 196,193,101,219,217 // vpand %ymm9,%ymm3,%ymm3
10221 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
10222 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
10223 .byte 197,249,110,216 // vmovd %eax,%xmm3
10224 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
10225 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
10226 .byte 72,173 // lods %ds:(%rsi),%rax
10227 .byte 91 // pop %rbx
10228 .byte 65,92 // pop %r12
10229 .byte 65,94 // pop %r14
10230 .byte 65,95 // pop %r15
10231 .byte 255,224 // jmpq *%rax
10232
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10233HIDDEN _sk_store_4444_hsw
10234.globl _sk_store_4444_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10235FUNCTION(_sk_store_4444_hsw)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10236_sk_store_4444_hsw:
10237 .byte 72,173 // lods %ds:(%rsi),%rax
10238 .byte 76,139,8 // mov (%rax),%r9
10239 .byte 184,0,0,112,65 // mov $0x41700000,%eax
10240 .byte 197,121,110,192 // vmovd %eax,%xmm8
10241 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
10242 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
10243 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
10244 .byte 196,193,53,114,241,12 // vpslld $0xc,%ymm9,%ymm9
10245 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
10246 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10247 .byte 196,193,45,114,242,8 // vpslld $0x8,%ymm10,%ymm10
10248 .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9
10249 .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
10250 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10251 .byte 196,193,45,114,242,4 // vpslld $0x4,%ymm10,%ymm10
10252 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
10253 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
10254 .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
10255 .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
10256 .byte 196,67,125,57,193,1 // vextracti128 $0x1,%ymm8,%xmm9
10257 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
10258 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10259 .byte 117,10 // jne 2c8f <_sk_store_4444_hsw+0x72>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10260 .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
10261 .byte 72,173 // lods %ds:(%rsi),%rax
10262 .byte 255,224 // jmpq *%rax
10263 .byte 65,137,200 // mov %ecx,%r8d
10264 .byte 65,128,224,7 // and $0x7,%r8b
10265 .byte 65,254,200 // dec %r8b
10266 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10267 .byte 119,236 // ja 2c8b <_sk_store_4444_hsw+0x6e>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10268 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10269 .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 2cec <_sk_store_4444_hsw+0xcf>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10270 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
10271 .byte 76,1,192 // add %r8,%rax
10272 .byte 255,224 // jmpq *%rax
10273 .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
10274 .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
10275 .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
10276 .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
10277 .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
10278 .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
10279 .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10280 .byte 235,159 // jmp 2c8b <_sk_store_4444_hsw+0x6e>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10281 .byte 247,255 // idiv %edi
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10282 .byte 255 // (bad)
10283 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10284 .byte 239 // out %eax,(%dx)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10285 .byte 255 // (bad)
10286 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10287 .byte 255,231 // jmpq *%rdi
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10288 .byte 255 // (bad)
10289 .byte 255 // (bad)
10290 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10291 .byte 223,255 // (bad)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10292 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10293 .byte 255,215 // callq *%rdi
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10294 .byte 255 // (bad)
10295 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10296 .byte 255,207 // dec %edi
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10297 .byte 255 // (bad)
10298 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]10299 .byte 255,199 // inc %edi
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]10300 .byte 255 // (bad)
10301 .byte 255 // (bad)
10302 .byte 255 // .byte 0xff
10303
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10304HIDDEN _sk_load_8888_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10305.globl _sk_load_8888_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10306FUNCTION(_sk_load_8888_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10307_sk_load_8888_hsw:
10308 .byte 73,137,200 // mov %rcx,%r8
10309 .byte 72,173 // lods %ds:(%rsi),%rax
10310 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
10311 .byte 76,3,8 // add (%rax),%r9
10312 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10313 .byte 117,104 // jne 2d85 <_sk_load_8888_hsw+0x7d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10314 .byte 196,193,126,111,25 // vmovdqu (%r9),%ymm3
10315 .byte 184,255,0,0,0 // mov $0xff,%eax
10316 .byte 197,249,110,192 // vmovd %eax,%xmm0
10317 .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2
10318 .byte 197,237,219,195 // vpand %ymm3,%ymm2,%ymm0
10319 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10320 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
10321 .byte 197,249,110,200 // vmovd %eax,%xmm1
10322 .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8
10323 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
10324 .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1
10325 .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
10326 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10327 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
10328 .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9
10329 .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
10330 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
10331 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
10332 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
10333 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
10334 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
10335 .byte 72,173 // lods %ds:(%rsi),%rax
10336 .byte 76,137,193 // mov %r8,%rcx
10337 .byte 255,224 // jmpq *%rax
10338 .byte 185,8,0,0,0 // mov $0x8,%ecx
10339 .byte 68,41,193 // sub %r8d,%ecx
10340 .byte 192,225,3 // shl $0x3,%cl
10341 .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax
10342 .byte 72,211,232 // shr %cl,%rax
10343 .byte 196,225,249,110,192 // vmovq %rax,%xmm0
10344 .byte 196,226,125,33,192 // vpmovsxbd %xmm0,%ymm0
10345 .byte 196,194,125,140,25 // vpmaskmovd (%r9),%ymm0,%ymm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10346 .byte 233,116,255,255,255 // jmpq 2d22 <_sk_load_8888_hsw+0x1a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10347
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]10348HIDDEN _sk_gather_8888_hsw
10349.globl _sk_gather_8888_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10350FUNCTION(_sk_gather_8888_hsw)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]10351_sk_gather_8888_hsw:
10352 .byte 72,173 // lods %ds:(%rsi),%rax
10353 .byte 76,139,0 // mov (%rax),%r8
10354 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
10355 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
10356 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
10357 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
10358 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
10359 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
10360 .byte 196,194,117,144,28,128 // vpgatherdd %ymm1,(%r8,%ymm0,4),%ymm3
10361 .byte 184,255,0,0,0 // mov $0xff,%eax
10362 .byte 197,249,110,192 // vmovd %eax,%xmm0
10363 .byte 196,226,125,88,208 // vpbroadcastd %xmm0,%ymm2
10364 .byte 197,237,219,195 // vpand %ymm3,%ymm2,%ymm0
10365 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10366 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
10367 .byte 197,249,110,200 // vmovd %eax,%xmm1
10368 .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8
10369 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
10370 .byte 197,245,114,211,8 // vpsrld $0x8,%ymm3,%ymm1
10371 .byte 197,237,219,201 // vpand %ymm1,%ymm2,%ymm1
10372 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10373 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
10374 .byte 197,181,114,211,16 // vpsrld $0x10,%ymm3,%ymm9
10375 .byte 196,193,109,219,209 // vpand %ymm9,%ymm2,%ymm2
10376 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
10377 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
10378 .byte 197,229,114,211,24 // vpsrld $0x18,%ymm3,%ymm3
10379 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
10380 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
10381 .byte 72,173 // lods %ds:(%rsi),%rax
10382 .byte 255,224 // jmpq *%rax
10383
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10384HIDDEN _sk_store_8888_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10385.globl _sk_store_8888_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10386FUNCTION(_sk_store_8888_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10387_sk_store_8888_hsw:
10388 .byte 73,137,200 // mov %rcx,%r8
10389 .byte 72,173 // lods %ds:(%rsi),%rax
10390 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
10391 .byte 76,3,8 // add (%rax),%r9
10392 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
10393 .byte 197,121,110,192 // vmovd %eax,%xmm8
10394 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
10395 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
10396 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
10397 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
10398 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10399 .byte 196,193,45,114,242,8 // vpslld $0x8,%ymm10,%ymm10
10400 .byte 196,65,45,235,201 // vpor %ymm9,%ymm10,%ymm9
10401 .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
10402 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10403 .byte 196,193,45,114,242,16 // vpslld $0x10,%ymm10,%ymm10
10404 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
10405 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
10406 .byte 196,193,61,114,240,24 // vpslld $0x18,%ymm8,%ymm8
10407 .byte 196,65,45,235,192 // vpor %ymm8,%ymm10,%ymm8
10408 .byte 196,65,53,235,192 // vpor %ymm8,%ymm9,%ymm8
10409 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10410 .byte 117,12 // jne 2ea8 <_sk_store_8888_hsw+0x74>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10411 .byte 196,65,126,127,1 // vmovdqu %ymm8,(%r9)
10412 .byte 72,173 // lods %ds:(%rsi),%rax
10413 .byte 76,137,193 // mov %r8,%rcx
10414 .byte 255,224 // jmpq *%rax
10415 .byte 185,8,0,0,0 // mov $0x8,%ecx
10416 .byte 68,41,193 // sub %r8d,%ecx
10417 .byte 192,225,3 // shl $0x3,%cl
10418 .byte 72,199,192,255,255,255,255 // mov $0xffffffffffffffff,%rax
10419 .byte 72,211,232 // shr %cl,%rax
10420 .byte 196,97,249,110,200 // vmovq %rax,%xmm9
10421 .byte 196,66,125,33,201 // vpmovsxbd %xmm9,%ymm9
10422 .byte 196,66,53,142,1 // vpmaskmovd %ymm8,%ymm9,(%r9)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10423 .byte 235,211 // jmp 2ea1 <_sk_store_8888_hsw+0x6d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10424
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10425HIDDEN _sk_load_f16_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10426.globl _sk_load_f16_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10427FUNCTION(_sk_load_f16_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10428_sk_load_f16_hsw:
10429 .byte 72,173 // lods %ds:(%rsi),%rax
10430 .byte 72,139,0 // mov (%rax),%rax
10431 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10432 .byte 117,97 // jne 2f39 <_sk_load_f16_hsw+0x6b>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10433 .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
10434 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
10435 .byte 197,249,16,92,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm3
10436 .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9
10437 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
10438 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
10439 .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1
10440 .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3
10441 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
10442 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
10443 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
10444 .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3
10445 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
10446 .byte 196,226,125,19,192 // vcvtph2ps %xmm0,%ymm0
10447 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
10448 .byte 196,226,125,19,201 // vcvtph2ps %xmm1,%ymm1
10449 .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2
10450 .byte 196,226,125,19,210 // vcvtph2ps %xmm2,%ymm2
10451 .byte 197,177,109,219 // vpunpckhqdq %xmm3,%xmm9,%xmm3
10452 .byte 196,226,125,19,219 // vcvtph2ps %xmm3,%ymm3
10453 .byte 72,173 // lods %ds:(%rsi),%rax
10454 .byte 255,224 // jmpq *%rax
10455 .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
10456 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
10457 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10458 .byte 116,79 // je 2f98 <_sk_load_f16_hsw+0xca>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10459 .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
10460 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10461 .byte 114,67 // jb 2f98 <_sk_load_f16_hsw+0xca>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10462 .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
10463 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10464 .byte 116,68 // je 2fa5 <_sk_load_f16_hsw+0xd7>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10465 .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
10466 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10467 .byte 114,56 // jb 2fa5 <_sk_load_f16_hsw+0xd7>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10468 .byte 197,251,16,92,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm3
10469 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10470 .byte 15,132,114,255,255,255 // je 2eef <_sk_load_f16_hsw+0x21>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10471 .byte 197,225,22,92,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm3,%xmm3
10472 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10473 .byte 15,130,98,255,255,255 // jb 2eef <_sk_load_f16_hsw+0x21>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10474 .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10475 .byte 233,87,255,255,255 // jmpq 2eef <_sk_load_f16_hsw+0x21>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10476 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
10477 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10478 .byte 233,74,255,255,255 // jmpq 2eef <_sk_load_f16_hsw+0x21>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10479 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10480 .byte 233,65,255,255,255 // jmpq 2eef <_sk_load_f16_hsw+0x21>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10481
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]10482HIDDEN _sk_gather_f16_hsw
10483.globl _sk_gather_f16_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10484FUNCTION(_sk_gather_f16_hsw)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]10485_sk_gather_f16_hsw:
10486 .byte 72,173 // lods %ds:(%rsi),%rax
10487 .byte 76,139,0 // mov (%rax),%r8
10488 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
10489 .byte 196,226,125,88,80,16 // vpbroadcastd 0x10(%rax),%ymm2
10490 .byte 196,226,109,64,201 // vpmulld %ymm1,%ymm2,%ymm1
10491 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
10492 .byte 197,245,254,192 // vpaddd %ymm0,%ymm1,%ymm0
10493 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
10494 .byte 197,237,118,210 // vpcmpeqd %ymm2,%ymm2,%ymm2
10495 .byte 196,194,237,144,28,192 // vpgatherdq %ymm2,(%r8,%xmm0,8),%ymm3
10496 .byte 196,227,125,57,192,1 // vextracti128 $0x1,%ymm0,%xmm0
10497 .byte 196,194,245,144,20,192 // vpgatherdq %ymm1,(%r8,%xmm0,8),%ymm2
10498 .byte 196,227,125,57,216,1 // vextracti128 $0x1,%ymm3,%xmm0
10499 .byte 196,227,125,57,209,1 // vextracti128 $0x1,%ymm2,%xmm1
10500 .byte 197,97,97,192 // vpunpcklwd %xmm0,%xmm3,%xmm8
10501 .byte 197,225,105,192 // vpunpckhwd %xmm0,%xmm3,%xmm0
10502 .byte 197,233,97,217 // vpunpcklwd %xmm1,%xmm2,%xmm3
10503 .byte 197,233,105,201 // vpunpckhwd %xmm1,%xmm2,%xmm1
10504 .byte 197,57,97,200 // vpunpcklwd %xmm0,%xmm8,%xmm9
10505 .byte 197,57,105,192 // vpunpckhwd %xmm0,%xmm8,%xmm8
10506 .byte 197,225,97,209 // vpunpcklwd %xmm1,%xmm3,%xmm2
10507 .byte 197,225,105,217 // vpunpckhwd %xmm1,%xmm3,%xmm3
10508 .byte 197,177,108,194 // vpunpcklqdq %xmm2,%xmm9,%xmm0
10509 .byte 196,226,125,19,192 // vcvtph2ps %xmm0,%ymm0
10510 .byte 197,177,109,202 // vpunpckhqdq %xmm2,%xmm9,%xmm1
10511 .byte 196,226,125,19,201 // vcvtph2ps %xmm1,%ymm1
10512 .byte 197,185,108,211 // vpunpcklqdq %xmm3,%xmm8,%xmm2
10513 .byte 196,226,125,19,210 // vcvtph2ps %xmm2,%ymm2
10514 .byte 197,185,109,219 // vpunpckhqdq %xmm3,%xmm8,%xmm3
10515 .byte 196,226,125,19,219 // vcvtph2ps %xmm3,%ymm3
10516 .byte 72,173 // lods %ds:(%rsi),%rax
10517 .byte 255,224 // jmpq *%rax
10518
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10519HIDDEN _sk_store_f16_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10520.globl _sk_store_f16_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10521FUNCTION(_sk_store_f16_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10522_sk_store_f16_hsw:
10523 .byte 72,173 // lods %ds:(%rsi),%rax
10524 .byte 72,139,0 // mov (%rax),%rax
10525 .byte 196,195,125,29,192,4 // vcvtps2ph $0x4,%ymm0,%xmm8
10526 .byte 196,195,125,29,201,4 // vcvtps2ph $0x4,%ymm1,%xmm9
10527 .byte 196,195,125,29,210,4 // vcvtps2ph $0x4,%ymm2,%xmm10
10528 .byte 196,195,125,29,219,4 // vcvtps2ph $0x4,%ymm3,%xmm11
10529 .byte 196,65,57,97,225 // vpunpcklwd %xmm9,%xmm8,%xmm12
10530 .byte 196,65,57,105,193 // vpunpckhwd %xmm9,%xmm8,%xmm8
10531 .byte 196,65,41,97,203 // vpunpcklwd %xmm11,%xmm10,%xmm9
10532 .byte 196,65,41,105,235 // vpunpckhwd %xmm11,%xmm10,%xmm13
10533 .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11
10534 .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10
10535 .byte 196,65,57,98,205 // vpunpckldq %xmm13,%xmm8,%xmm9
10536 .byte 196,65,57,106,197 // vpunpckhdq %xmm13,%xmm8,%xmm8
10537 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10538 .byte 117,27 // jne 309d <_sk_store_f16_hsw+0x65>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10539 .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
10540 .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
10541 .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
10542 .byte 197,122,127,68,248,48 // vmovdqu %xmm8,0x30(%rax,%rdi,8)
10543 .byte 72,173 // lods %ds:(%rsi),%rax
10544 .byte 255,224 // jmpq *%rax
10545 .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
10546 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10547 .byte 116,241 // je 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10548 .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
10549 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10550 .byte 114,229 // jb 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10551 .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10552 .byte 116,221 // je 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10553 .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
10554 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10555 .byte 114,209 // jb 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10556 .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10557 .byte 116,201 // je 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10558 .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
10559 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10560 .byte 114,189 // jb 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10561 .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10562 .byte 235,181 // jmp 3099 <_sk_store_f16_hsw+0x61>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10563
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10564HIDDEN _sk_load_u16_be_hsw
10565.globl _sk_load_u16_be_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10566FUNCTION(_sk_load_u16_be_hsw)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10567_sk_load_u16_be_hsw:
10568 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10569 .byte 76,139,0 // mov (%rax),%r8
10570 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10571 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10572 .byte 15,133,205,0,0,0 // jne 31c7 <_sk_load_u16_be_hsw+0xe3>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10573 .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8
10574 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2
10575 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3
10576 .byte 196,65,122,111,76,64,48 // vmovdqu 0x30(%r8,%rax,2),%xmm9
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10577 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
10578 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
10579 .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1
10580 .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3
10581 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
10582 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
10583 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
10584 .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11
10585 .byte 184,128,0,128,55 // mov $0x37800080,%eax
10586 .byte 197,249,110,192 // vmovd %eax,%xmm0
10587 .byte 196,98,125,88,208 // vpbroadcastd %xmm0,%ymm10
10588 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
10589 .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1
10590 .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0
10591 .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
10592 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
10593 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10594 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
10595 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
10596 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
10597 .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1
10598 .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1
10599 .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1
10600 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10601 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
10602 .byte 196,193,49,108,211 // vpunpcklqdq %xmm11,%xmm9,%xmm2
10603 .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3
10604 .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2
10605 .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2
10606 .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2
10607 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
10608 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
10609 .byte 196,193,49,109,219 // vpunpckhqdq %xmm11,%xmm9,%xmm3
10610 .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8
10611 .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3
10612 .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3
10613 .byte 196,226,125,51,219 // vpmovzxwd %xmm3,%ymm3
10614 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
10615 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3
10616 .byte 72,173 // lods %ds:(%rsi),%rax
10617 .byte 255,224 // jmpq *%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10618 .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10619 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
10620 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10621 .byte 116,85 // je 322d <_sk_load_u16_be_hsw+0x149>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10622 .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10623 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10624 .byte 114,72 // jb 322d <_sk_load_u16_be_hsw+0x149>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10625 .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10626 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10627 .byte 116,72 // je 323a <_sk_load_u16_be_hsw+0x156>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10628 .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10629 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10630 .byte 114,59 // jb 323a <_sk_load_u16_be_hsw+0x156>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10631 .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10632 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10633 .byte 15,132,5,255,255,255 // je 3115 <_sk_load_u16_be_hsw+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10634 .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10635 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10636 .byte 15,130,244,254,255,255 // jb 3115 <_sk_load_u16_be_hsw+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10637 .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10638 .byte 233,232,254,255,255 // jmpq 3115 <_sk_load_u16_be_hsw+0x31>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10639 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
10640 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10641 .byte 233,219,254,255,255 // jmpq 3115 <_sk_load_u16_be_hsw+0x31>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10642 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10643 .byte 233,210,254,255,255 // jmpq 3115 <_sk_load_u16_be_hsw+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10644
10645HIDDEN _sk_load_rgb_u16_be_hsw
10646.globl _sk_load_rgb_u16_be_hsw
10647FUNCTION(_sk_load_rgb_u16_be_hsw)
10648_sk_load_rgb_u16_be_hsw:
10649 .byte 72,173 // lods %ds:(%rsi),%rax
10650 .byte 76,139,0 // mov (%rax),%r8
10651 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax
10652 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10653 .byte 15,133,211,0,0,0 // jne 3328 <_sk_load_rgb_u16_be_hsw+0xe5>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10654 .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0
10655 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2
10656 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1
10657 .byte 196,193,122,111,92,64,32 // vmovdqu 0x20(%r8,%rax,2),%xmm3
10658 .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3
10659 .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8
10660 .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9
10661 .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11
10662 .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10
10663 .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0
10664 .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2
10665 .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1
10666 .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3
10667 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
10668 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
10669 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
10670 .byte 197,241,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm3
10671 .byte 184,128,0,128,55 // mov $0x37800080,%eax
10672 .byte 197,249,110,192 // vmovd %eax,%xmm0
10673 .byte 196,98,125,88,208 // vpbroadcastd %xmm0,%ymm10
10674 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
10675 .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1
10676 .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0
10677 .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
10678 .byte 196,226,125,51,192 // vpmovzxwd %xmm0,%ymm0
10679 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
10680 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
10681 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
10682 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
10683 .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1
10684 .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1
10685 .byte 196,226,125,51,201 // vpmovzxwd %xmm1,%ymm1
10686 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
10687 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
10688 .byte 197,177,108,211 // vpunpcklqdq %xmm3,%xmm9,%xmm2
10689 .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3
10690 .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2
10691 .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2
10692 .byte 196,226,125,51,210 // vpmovzxwd %xmm2,%ymm2
10693 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
10694 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
10695 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
10696 .byte 197,249,110,216 // vmovd %eax,%xmm3
10697 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
10698 .byte 72,173 // lods %ds:(%rsi),%rax
10699 .byte 255,224 // jmpq *%rax
10700 .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0
10701 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
10702 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10703 .byte 117,5 // jne 3341 <_sk_load_rgb_u16_be_hsw+0xfe>
10704 .byte 233,72,255,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10705 .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1
10706 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
10707 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10708 .byte 114,26 // jb 3370 <_sk_load_rgb_u16_be_hsw+0x12d>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10709 .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1
10710 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
10711 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10712 .byte 117,10 // jne 3375 <_sk_load_rgb_u16_be_hsw+0x132>
10713 .byte 233,25,255,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
10714 .byte 233,20,255,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10715 .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1
10716 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
10717 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10718 .byte 114,26 // jb 33a4 <_sk_load_rgb_u16_be_hsw+0x161>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10719 .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1
10720 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
10721 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10722 .byte 117,10 // jne 33a9 <_sk_load_rgb_u16_be_hsw+0x166>
10723 .byte 233,229,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
10724 .byte 233,224,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10725 .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3
10726 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
10727 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10728 .byte 114,20 // jb 33d2 <_sk_load_rgb_u16_be_hsw+0x18f>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10729 .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3
10730 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10731 .byte 233,183,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
10732 .byte 233,178,254,255,255 // jmpq 3289 <_sk_load_rgb_u16_be_hsw+0x46>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10733
10734HIDDEN _sk_store_u16_be_hsw
10735.globl _sk_store_u16_be_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10736FUNCTION(_sk_store_u16_be_hsw)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10737_sk_store_u16_be_hsw:
10738 .byte 72,173 // lods %ds:(%rsi),%rax
10739 .byte 76,139,0 // mov (%rax),%r8
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10740 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10741 .byte 184,0,255,127,71 // mov $0x477fff00,%eax
10742 .byte 197,121,110,192 // vmovd %eax,%xmm8
10743 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
10744 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
10745 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
10746 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10
10747 .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9
10748 .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10
10749 .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9
10750 .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9
10751 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
10752 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
10753 .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11
10754 .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10
10755 .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11
10756 .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10
10757 .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10
10758 .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11
10759 .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11
10760 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12
10761 .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11
10762 .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12
10763 .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11
10764 .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11
10765 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
10766 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
10767 .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12
10768 .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8
10769 .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12
10770 .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8
10771 .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8
10772 .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12
10773 .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13
10774 .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9
10775 .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8
10776 .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11
10777 .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10
10778 .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9
10779 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8
10780 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10781 .byte 117,31 // jne 34d2 <_sk_store_u16_be_hsw+0xfb>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10782 .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2)
10783 .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2)
10784 .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2)
10785 .byte 196,1,122,127,68,72,48 // vmovdqu %xmm8,0x30(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10786 .byte 72,173 // lods %ds:(%rsi),%rax
10787 .byte 255,224 // jmpq *%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10788 .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10789 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10790 .byte 116,240 // je 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10791 .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10792 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10793 .byte 114,227 // jb 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10794 .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10795 .byte 116,218 // je 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10796 .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10797 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10798 .byte 114,205 // jb 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10799 .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10800 .byte 116,196 // je 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10801 .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10802 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10803 .byte 114,183 // jb 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10804 .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10805 .byte 235,174 // jmp 34ce <_sk_store_u16_be_hsw+0xf7>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]10806
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10807HIDDEN _sk_load_f32_hsw
10808.globl _sk_load_f32_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10809FUNCTION(_sk_load_f32_hsw)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10810_sk_load_f32_hsw:
10811 .byte 72,173 // lods %ds:(%rsi),%rax
10812 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10813 .byte 119,110 // ja 3596 <_sk_load_f32_hsw+0x76>
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10814 .byte 76,139,0 // mov (%rax),%r8
10815 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10816 .byte 76,141,21,134,0,0,0 // lea 0x86(%rip),%r10 # 35c0 <_sk_load_f32_hsw+0xa0>
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10817 .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax
10818 .byte 76,1,208 // add %r10,%rax
10819 .byte 255,224 // jmpq *%rax
10820 .byte 196,3,125,24,68,136,112,1 // vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8
10821 .byte 196,131,125,24,92,136,96,1 // vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3
10822 .byte 196,131,125,24,76,136,80,1 // vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1
10823 .byte 196,131,125,24,84,136,64,1 // vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2
10824 .byte 196,129,121,16,68,136,48 // vmovupd 0x30(%r8,%r9,4),%xmm0
10825 .byte 196,195,125,13,192,12 // vblendpd $0xc,%ymm8,%ymm0,%ymm0
10826 .byte 196,1,121,16,68,136,32 // vmovupd 0x20(%r8,%r9,4),%xmm8
10827 .byte 196,99,61,13,203,12 // vblendpd $0xc,%ymm3,%ymm8,%ymm9
10828 .byte 196,129,121,16,92,136,16 // vmovupd 0x10(%r8,%r9,4),%xmm3
10829 .byte 196,99,101,13,209,12 // vblendpd $0xc,%ymm1,%ymm3,%ymm10
10830 .byte 196,129,121,16,12,136 // vmovupd (%r8,%r9,4),%xmm1
10831 .byte 196,227,117,13,202,12 // vblendpd $0xc,%ymm2,%ymm1,%ymm1
10832 .byte 196,193,116,20,210 // vunpcklps %ymm10,%ymm1,%ymm2
10833 .byte 196,193,116,21,218 // vunpckhps %ymm10,%ymm1,%ymm3
10834 .byte 197,180,20,200 // vunpcklps %ymm0,%ymm9,%ymm1
10835 .byte 197,52,21,192 // vunpckhps %ymm0,%ymm9,%ymm8
10836 .byte 197,237,20,193 // vunpcklpd %ymm1,%ymm2,%ymm0
10837 .byte 197,237,21,201 // vunpckhpd %ymm1,%ymm2,%ymm1
10838 .byte 196,193,101,20,208 // vunpcklpd %ymm8,%ymm3,%ymm2
10839 .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3
10840 .byte 72,173 // lods %ds:(%rsi),%rax
10841 .byte 255,224 // jmpq *%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10842 .byte 102,144 // xchg %ax,%ax
10843 .byte 131,255,255 // cmp $0xffffffff,%edi
10844 .byte 255,202 // dec %edx
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10845 .byte 255 // (bad)
10846 .byte 255 // (bad)
10847 .byte 255 // (bad)
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10848 .byte 189,255,255,255,176 // mov $0xb0ffffff,%ebp
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10849 .byte 255 // (bad)
10850 .byte 255 // (bad)
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10851 .byte 255,163,255,255,255,155 // jmpq *-0x64000001(%rbx)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10852 .byte 255 // (bad)
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]10853 .byte 255 // (bad)
10854 .byte 255,147,255,255,255,139 // callq *-0x74000001(%rbx)
10855 .byte 255 // (bad)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]10856 .byte 255 // (bad)
10857 .byte 255 // .byte 0xff
10858
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10859HIDDEN _sk_store_f32_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10860.globl _sk_store_f32_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10861FUNCTION(_sk_store_f32_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10862_sk_store_f32_hsw:
10863 .byte 72,173 // lods %ds:(%rsi),%rax
10864 .byte 76,139,0 // mov (%rax),%r8
10865 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
10866 .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8
10867 .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11
10868 .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9
10869 .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12
10870 .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10
10871 .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9
10872 .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
10873 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
10874 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10875 .byte 117,55 // jne 364d <_sk_store_f32_hsw+0x6d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10876 .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
10877 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
10878 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
10879 .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8
10880 .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4)
10881 .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4)
10882 .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4)
10883 .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4)
10884 .byte 72,173 // lods %ds:(%rsi),%rax
10885 .byte 255,224 // jmpq *%rax
10886 .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
10887 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10888 .byte 116,240 // je 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10889 .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
10890 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10891 .byte 114,227 // jb 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10892 .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10893 .byte 116,218 // je 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10894 .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
10895 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10896 .byte 114,205 // jb 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10897 .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10898 .byte 116,195 // je 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10899 .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
10900 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10901 .byte 114,181 // jb 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10902 .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]10903 .byte 235,171 // jmp 3649 <_sk_store_f32_hsw+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10904
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10905HIDDEN _sk_clamp_x_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10906.globl _sk_clamp_x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10907FUNCTION(_sk_clamp_x_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10908_sk_clamp_x_hsw:
10909 .byte 72,173 // lods %ds:(%rsi),%rax
10910 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
10911 .byte 197,188,95,192 // vmaxps %ymm0,%ymm8,%ymm0
10912 .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8
10913 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
10914 .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8
10915 .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
10916 .byte 72,173 // lods %ds:(%rsi),%rax
10917 .byte 255,224 // jmpq *%rax
10918
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10919HIDDEN _sk_clamp_y_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10920.globl _sk_clamp_y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10921FUNCTION(_sk_clamp_y_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10922_sk_clamp_y_hsw:
10923 .byte 72,173 // lods %ds:(%rsi),%rax
10924 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
10925 .byte 197,188,95,201 // vmaxps %ymm1,%ymm8,%ymm1
10926 .byte 196,98,125,88,0 // vpbroadcastd (%rax),%ymm8
10927 .byte 196,65,53,118,201 // vpcmpeqd %ymm9,%ymm9,%ymm9
10928 .byte 196,65,61,254,193 // vpaddd %ymm9,%ymm8,%ymm8
10929 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
10930 .byte 72,173 // lods %ds:(%rsi),%rax
10931 .byte 255,224 // jmpq *%rax
10932
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10933HIDDEN _sk_repeat_x_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10934.globl _sk_repeat_x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10935FUNCTION(_sk_repeat_x_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10936_sk_repeat_x_hsw:
10937 .byte 72,173 // lods %ds:(%rsi),%rax
10938 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
10939 .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9
10940 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
10941 .byte 196,98,61,172,200 // vfnmadd213ps %ymm0,%ymm8,%ymm9
10942 .byte 197,253,118,192 // vpcmpeqd %ymm0,%ymm0,%ymm0
10943 .byte 197,189,254,192 // vpaddd %ymm0,%ymm8,%ymm0
10944 .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0
10945 .byte 72,173 // lods %ds:(%rsi),%rax
10946 .byte 255,224 // jmpq *%rax
10947
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10948HIDDEN _sk_repeat_y_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10949.globl _sk_repeat_y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10950FUNCTION(_sk_repeat_y_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10951_sk_repeat_y_hsw:
10952 .byte 72,173 // lods %ds:(%rsi),%rax
10953 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
10954 .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9
10955 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
10956 .byte 196,98,61,172,201 // vfnmadd213ps %ymm1,%ymm8,%ymm9
10957 .byte 197,245,118,201 // vpcmpeqd %ymm1,%ymm1,%ymm1
10958 .byte 197,189,254,201 // vpaddd %ymm1,%ymm8,%ymm1
10959 .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1
10960 .byte 72,173 // lods %ds:(%rsi),%rax
10961 .byte 255,224 // jmpq *%rax
10962
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10963HIDDEN _sk_mirror_x_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10964.globl _sk_mirror_x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10965FUNCTION(_sk_mirror_x_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10966_sk_mirror_x_hsw:
10967 .byte 72,173 // lods %ds:(%rsi),%rax
10968 .byte 197,122,16,0 // vmovss (%rax),%xmm8
10969 .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9
10970 .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10
10971 .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0
10972 .byte 196,226,125,24,192 // vbroadcastss %xmm0,%ymm0
10973 .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8
10974 .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8
10975 .byte 196,66,125,172,194 // vfnmadd213ps %ymm10,%ymm0,%ymm8
10976 .byte 196,193,60,92,193 // vsubps %ymm9,%ymm8,%ymm0
10977 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
10978 .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8
10979 .byte 197,188,84,192 // vandps %ymm0,%ymm8,%ymm0
10980 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
10981 .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8
10982 .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
10983 .byte 72,173 // lods %ds:(%rsi),%rax
10984 .byte 255,224 // jmpq *%rax
10985
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]10986HIDDEN _sk_mirror_y_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10987.globl _sk_mirror_y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]10988FUNCTION(_sk_mirror_y_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]10989_sk_mirror_y_hsw:
10990 .byte 72,173 // lods %ds:(%rsi),%rax
10991 .byte 197,122,16,0 // vmovss (%rax),%xmm8
10992 .byte 196,66,125,24,200 // vbroadcastss %xmm8,%ymm9
10993 .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10
10994 .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1
10995 .byte 196,226,125,24,201 // vbroadcastss %xmm1,%ymm1
10996 .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8
10997 .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8
10998 .byte 196,66,117,172,194 // vfnmadd213ps %ymm10,%ymm1,%ymm8
10999 .byte 196,193,60,92,201 // vsubps %ymm9,%ymm8,%ymm1
11000 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
11001 .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8
11002 .byte 197,188,84,201 // vandps %ymm1,%ymm8,%ymm1
11003 .byte 196,65,61,118,192 // vpcmpeqd %ymm8,%ymm8,%ymm8
11004 .byte 196,65,53,254,192 // vpaddd %ymm8,%ymm9,%ymm8
11005 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
11006 .byte 72,173 // lods %ds:(%rsi),%rax
11007 .byte 255,224 // jmpq *%rax
11008
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11009HIDDEN _sk_luminance_to_alpha_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11010.globl _sk_luminance_to_alpha_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11011FUNCTION(_sk_luminance_to_alpha_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11012_sk_luminance_to_alpha_hsw:
11013 .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax
11014 .byte 197,249,110,216 // vmovd %eax,%xmm3
11015 .byte 196,98,125,88,195 // vpbroadcastd %xmm3,%ymm8
11016 .byte 184,89,23,55,63 // mov $0x3f371759,%eax
11017 .byte 197,249,110,216 // vmovd %eax,%xmm3
11018 .byte 196,226,125,88,219 // vpbroadcastd %xmm3,%ymm3
11019 .byte 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1
11020 .byte 196,98,125,168,193 // vfmadd213ps %ymm1,%ymm0,%ymm8
11021 .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax
11022 .byte 197,249,110,192 // vmovd %eax,%xmm0
11023 .byte 196,226,125,88,216 // vpbroadcastd %xmm0,%ymm3
11024 .byte 196,194,109,168,216 // vfmadd213ps %ymm8,%ymm2,%ymm3
11025 .byte 72,173 // lods %ds:(%rsi),%rax
11026 .byte 197,253,239,192 // vpxor %ymm0,%ymm0,%ymm0
11027 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11028 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11029 .byte 255,224 // jmpq *%rax
11030
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11031HIDDEN _sk_matrix_2x3_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11032.globl _sk_matrix_2x3_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11033FUNCTION(_sk_matrix_2x3_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11034_sk_matrix_2x3_hsw:
11035 .byte 72,173 // lods %ds:(%rsi),%rax
11036 .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9
11037 .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10
11038 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
11039 .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8
11040 .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8
11041 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
11042 .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11
11043 .byte 196,98,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm9
11044 .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9
11045 .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9
11046 .byte 72,173 // lods %ds:(%rsi),%rax
11047 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
11048 .byte 197,124,41,201 // vmovaps %ymm9,%ymm1
11049 .byte 255,224 // jmpq *%rax
11050
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11051HIDDEN _sk_matrix_3x4_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11052.globl _sk_matrix_3x4_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11053FUNCTION(_sk_matrix_3x4_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11054_sk_matrix_3x4_hsw:
11055 .byte 72,173 // lods %ds:(%rsi),%rax
11056 .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9
11057 .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10
11058 .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11
11059 .byte 196,98,125,24,64,36 // vbroadcastss 0x24(%rax),%ymm8
11060 .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8
11061 .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8
11062 .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8
11063 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
11064 .byte 196,98,125,24,88,16 // vbroadcastss 0x10(%rax),%ymm11
11065 .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12
11066 .byte 196,98,125,24,72,40 // vbroadcastss 0x28(%rax),%ymm9
11067 .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9
11068 .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9
11069 .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9
11070 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
11071 .byte 196,98,125,24,96,20 // vbroadcastss 0x14(%rax),%ymm12
11072 .byte 196,98,125,24,104,32 // vbroadcastss 0x20(%rax),%ymm13
11073 .byte 196,98,125,24,80,44 // vbroadcastss 0x2c(%rax),%ymm10
11074 .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10
11075 .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10
11076 .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10
11077 .byte 72,173 // lods %ds:(%rsi),%rax
11078 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
11079 .byte 197,124,41,201 // vmovaps %ymm9,%ymm1
11080 .byte 197,124,41,210 // vmovaps %ymm10,%ymm2
11081 .byte 255,224 // jmpq *%rax
11082
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11083HIDDEN _sk_matrix_4x5_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11084.globl _sk_matrix_4x5_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11085FUNCTION(_sk_matrix_4x5_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11086_sk_matrix_4x5_hsw:
11087 .byte 72,173 // lods %ds:(%rsi),%rax
11088 .byte 196,98,125,24,8 // vbroadcastss (%rax),%ymm9
11089 .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10
11090 .byte 196,98,125,24,88,32 // vbroadcastss 0x20(%rax),%ymm11
11091 .byte 196,98,125,24,96,48 // vbroadcastss 0x30(%rax),%ymm12
11092 .byte 196,98,125,24,64,64 // vbroadcastss 0x40(%rax),%ymm8
11093 .byte 196,66,101,184,196 // vfmadd231ps %ymm12,%ymm3,%ymm8
11094 .byte 196,66,109,184,195 // vfmadd231ps %ymm11,%ymm2,%ymm8
11095 .byte 196,66,117,184,194 // vfmadd231ps %ymm10,%ymm1,%ymm8
11096 .byte 196,66,125,184,193 // vfmadd231ps %ymm9,%ymm0,%ymm8
11097 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
11098 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
11099 .byte 196,98,125,24,96,36 // vbroadcastss 0x24(%rax),%ymm12
11100 .byte 196,98,125,24,104,52 // vbroadcastss 0x34(%rax),%ymm13
11101 .byte 196,98,125,24,72,68 // vbroadcastss 0x44(%rax),%ymm9
11102 .byte 196,66,101,184,205 // vfmadd231ps %ymm13,%ymm3,%ymm9
11103 .byte 196,66,109,184,204 // vfmadd231ps %ymm12,%ymm2,%ymm9
11104 .byte 196,66,117,184,203 // vfmadd231ps %ymm11,%ymm1,%ymm9
11105 .byte 196,66,125,184,202 // vfmadd231ps %ymm10,%ymm0,%ymm9
11106 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
11107 .byte 196,98,125,24,96,24 // vbroadcastss 0x18(%rax),%ymm12
11108 .byte 196,98,125,24,104,40 // vbroadcastss 0x28(%rax),%ymm13
11109 .byte 196,98,125,24,112,56 // vbroadcastss 0x38(%rax),%ymm14
11110 .byte 196,98,125,24,80,72 // vbroadcastss 0x48(%rax),%ymm10
11111 .byte 196,66,101,184,214 // vfmadd231ps %ymm14,%ymm3,%ymm10
11112 .byte 196,66,109,184,213 // vfmadd231ps %ymm13,%ymm2,%ymm10
11113 .byte 196,66,117,184,212 // vfmadd231ps %ymm12,%ymm1,%ymm10
11114 .byte 196,66,125,184,211 // vfmadd231ps %ymm11,%ymm0,%ymm10
11115 .byte 196,98,125,24,96,12 // vbroadcastss 0xc(%rax),%ymm12
11116 .byte 196,98,125,24,104,28 // vbroadcastss 0x1c(%rax),%ymm13
11117 .byte 196,98,125,24,112,44 // vbroadcastss 0x2c(%rax),%ymm14
11118 .byte 196,98,125,24,120,60 // vbroadcastss 0x3c(%rax),%ymm15
11119 .byte 196,98,125,24,88,76 // vbroadcastss 0x4c(%rax),%ymm11
11120 .byte 196,66,101,184,223 // vfmadd231ps %ymm15,%ymm3,%ymm11
11121 .byte 196,66,109,184,222 // vfmadd231ps %ymm14,%ymm2,%ymm11
11122 .byte 196,66,117,184,221 // vfmadd231ps %ymm13,%ymm1,%ymm11
11123 .byte 196,66,125,184,220 // vfmadd231ps %ymm12,%ymm0,%ymm11
11124 .byte 72,173 // lods %ds:(%rsi),%rax
11125 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
11126 .byte 197,124,41,201 // vmovaps %ymm9,%ymm1
11127 .byte 197,124,41,210 // vmovaps %ymm10,%ymm2
11128 .byte 197,124,41,219 // vmovaps %ymm11,%ymm3
11129 .byte 255,224 // jmpq *%rax
11130
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11131HIDDEN _sk_matrix_perspective_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11132.globl _sk_matrix_perspective_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11133FUNCTION(_sk_matrix_perspective_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11134_sk_matrix_perspective_hsw:
11135 .byte 72,173 // lods %ds:(%rsi),%rax
11136 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
11137 .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9
11138 .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10
11139 .byte 196,66,117,184,209 // vfmadd231ps %ymm9,%ymm1,%ymm10
11140 .byte 196,66,125,184,208 // vfmadd231ps %ymm8,%ymm0,%ymm10
11141 .byte 196,98,125,24,64,12 // vbroadcastss 0xc(%rax),%ymm8
11142 .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9
11143 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
11144 .byte 196,66,117,184,217 // vfmadd231ps %ymm9,%ymm1,%ymm11
11145 .byte 196,66,125,184,216 // vfmadd231ps %ymm8,%ymm0,%ymm11
11146 .byte 196,98,125,24,64,24 // vbroadcastss 0x18(%rax),%ymm8
11147 .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9
11148 .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12
11149 .byte 196,66,117,184,225 // vfmadd231ps %ymm9,%ymm1,%ymm12
11150 .byte 196,66,125,184,224 // vfmadd231ps %ymm8,%ymm0,%ymm12
11151 .byte 196,193,124,83,204 // vrcpps %ymm12,%ymm1
11152 .byte 197,172,89,193 // vmulps %ymm1,%ymm10,%ymm0
11153 .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1
11154 .byte 72,173 // lods %ds:(%rsi),%rax
11155 .byte 255,224 // jmpq *%rax
11156
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11157HIDDEN _sk_linear_gradient_hsw
11158.globl _sk_linear_gradient_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11159FUNCTION(_sk_linear_gradient_hsw)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11160_sk_linear_gradient_hsw:
11161 .byte 72,173 // lods %ds:(%rsi),%rax
11162 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
11163 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
11164 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
11165 .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9
11166 .byte 76,139,0 // mov (%rax),%r8
11167 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]11168 .byte 15,132,143,0,0,0 // je 3ad9 <_sk_linear_gradient_hsw+0xb5>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11169 .byte 72,139,64,8 // mov 0x8(%rax),%rax
11170 .byte 72,131,192,32 // add $0x20,%rax
11171 .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
11172 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11173 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11174 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11175 .byte 196,98,125,24,104,224 // vbroadcastss -0x20(%rax),%ymm13
11176 .byte 196,65,124,194,237,1 // vcmpltps %ymm13,%ymm0,%ymm13
11177 .byte 196,98,125,24,112,228 // vbroadcastss -0x1c(%rax),%ymm14
11178 .byte 196,67,13,74,228,208 // vblendvps %ymm13,%ymm12,%ymm14,%ymm12
11179 .byte 196,98,125,24,112,232 // vbroadcastss -0x18(%rax),%ymm14
11180 .byte 196,227,13,74,201,208 // vblendvps %ymm13,%ymm1,%ymm14,%ymm1
11181 .byte 196,98,125,24,112,236 // vbroadcastss -0x14(%rax),%ymm14
11182 .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2
11183 .byte 196,98,125,24,112,240 // vbroadcastss -0x10(%rax),%ymm14
11184 .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3
11185 .byte 196,98,125,24,112,244 // vbroadcastss -0xc(%rax),%ymm14
11186 .byte 196,67,13,74,192,208 // vblendvps %ymm13,%ymm8,%ymm14,%ymm8
11187 .byte 196,98,125,24,112,248 // vbroadcastss -0x8(%rax),%ymm14
11188 .byte 196,67,13,74,219,208 // vblendvps %ymm13,%ymm11,%ymm14,%ymm11
11189 .byte 196,98,125,24,112,252 // vbroadcastss -0x4(%rax),%ymm14
11190 .byte 196,67,13,74,210,208 // vblendvps %ymm13,%ymm10,%ymm14,%ymm10
11191 .byte 196,98,125,24,48 // vbroadcastss (%rax),%ymm14
11192 .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9
11193 .byte 72,131,192,36 // add $0x24,%rax
11194 .byte 73,255,200 // dec %r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]11195 .byte 117,140 // jne 3a63 <_sk_linear_gradient_hsw+0x3f>
11196 .byte 235,17 // jmp 3aea <_sk_linear_gradient_hsw+0xc6>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11197 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11198 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11199 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11200 .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
11201 .byte 196,66,125,184,196 // vfmadd231ps %ymm12,%ymm0,%ymm8
11202 .byte 196,194,125,168,203 // vfmadd213ps %ymm11,%ymm0,%ymm1
11203 .byte 196,194,125,168,210 // vfmadd213ps %ymm10,%ymm0,%ymm2
11204 .byte 196,194,125,168,217 // vfmadd213ps %ymm9,%ymm0,%ymm3
11205 .byte 72,173 // lods %ds:(%rsi),%rax
11206 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
11207 .byte 255,224 // jmpq *%rax
11208
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11209HIDDEN _sk_linear_gradient_2stops_hsw
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11210.globl _sk_linear_gradient_2stops_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11211FUNCTION(_sk_linear_gradient_2stops_hsw)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11212_sk_linear_gradient_2stops_hsw:
11213 .byte 72,173 // lods %ds:(%rsi),%rax
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11214 .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1
11215 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11216 .byte 196,98,125,184,193 // vfmadd231ps %ymm1,%ymm0,%ymm8
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11217 .byte 196,226,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm2
11218 .byte 196,226,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11219 .byte 196,226,125,184,202 // vfmadd231ps %ymm2,%ymm0,%ymm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11220 .byte 196,226,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm3
11221 .byte 196,226,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11222 .byte 196,226,125,184,211 // vfmadd231ps %ymm3,%ymm0,%ymm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]11223 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
11224 .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11225 .byte 196,194,125,184,217 // vfmadd231ps %ymm9,%ymm0,%ymm3
11226 .byte 72,173 // lods %ds:(%rsi),%rax
11227 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
11228 .byte 255,224 // jmpq *%rax
11229
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11230HIDDEN _sk_save_xy_hsw
11231.globl _sk_save_xy_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11232FUNCTION(_sk_save_xy_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11233_sk_save_xy_hsw:
11234 .byte 72,173 // lods %ds:(%rsi),%rax
11235 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11236 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11237 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11238 .byte 197,60,88,200 // vaddps %ymm0,%ymm8,%ymm9
11239 .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10
11240 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9
11241 .byte 197,60,88,193 // vaddps %ymm1,%ymm8,%ymm8
11242 .byte 196,67,125,8,208,1 // vroundps $0x1,%ymm8,%ymm10
11243 .byte 196,65,60,92,194 // vsubps %ymm10,%ymm8,%ymm8
11244 .byte 197,252,17,0 // vmovups %ymm0,(%rax)
11245 .byte 197,252,17,72,32 // vmovups %ymm1,0x20(%rax)
11246 .byte 197,124,17,72,64 // vmovups %ymm9,0x40(%rax)
11247 .byte 197,124,17,64,96 // vmovups %ymm8,0x60(%rax)
11248 .byte 72,173 // lods %ds:(%rsi),%rax
11249 .byte 255,224 // jmpq *%rax
11250
11251HIDDEN _sk_accumulate_hsw
11252.globl _sk_accumulate_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11253FUNCTION(_sk_accumulate_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11254_sk_accumulate_hsw:
11255 .byte 72,173 // lods %ds:(%rsi),%rax
11256 .byte 197,124,16,128,128,0,0,0 // vmovups 0x80(%rax),%ymm8
11257 .byte 197,60,89,128,160,0,0,0 // vmulps 0xa0(%rax),%ymm8,%ymm8
11258 .byte 196,226,61,184,224 // vfmadd231ps %ymm0,%ymm8,%ymm4
11259 .byte 196,226,61,184,233 // vfmadd231ps %ymm1,%ymm8,%ymm5
11260 .byte 196,226,61,184,242 // vfmadd231ps %ymm2,%ymm8,%ymm6
11261 .byte 196,98,101,168,199 // vfmadd213ps %ymm7,%ymm3,%ymm8
11262 .byte 72,173 // lods %ds:(%rsi),%rax
11263 .byte 197,124,41,199 // vmovaps %ymm8,%ymm7
11264 .byte 255,224 // jmpq *%rax
11265
11266HIDDEN _sk_bilinear_nx_hsw
11267.globl _sk_bilinear_nx_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11268FUNCTION(_sk_bilinear_nx_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11269_sk_bilinear_nx_hsw:
11270 .byte 72,173 // lods %ds:(%rsi),%rax
11271 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
11272 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11273 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
11274 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
11275 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11276 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11277 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11278 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
11279 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
11280 .byte 72,173 // lods %ds:(%rsi),%rax
11281 .byte 255,224 // jmpq *%rax
11282
11283HIDDEN _sk_bilinear_px_hsw
11284.globl _sk_bilinear_px_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11285FUNCTION(_sk_bilinear_px_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11286_sk_bilinear_px_hsw:
11287 .byte 72,173 // lods %ds:(%rsi),%rax
11288 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11289 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11290 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
11291 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
11292 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
11293 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
11294 .byte 72,173 // lods %ds:(%rsi),%rax
11295 .byte 255,224 // jmpq *%rax
11296
11297HIDDEN _sk_bilinear_ny_hsw
11298.globl _sk_bilinear_ny_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11299FUNCTION(_sk_bilinear_ny_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11300_sk_bilinear_ny_hsw:
11301 .byte 72,173 // lods %ds:(%rsi),%rax
11302 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
11303 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11304 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
11305 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
11306 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11307 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11308 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11309 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
11310 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
11311 .byte 72,173 // lods %ds:(%rsi),%rax
11312 .byte 255,224 // jmpq *%rax
11313
11314HIDDEN _sk_bilinear_py_hsw
11315.globl _sk_bilinear_py_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11316FUNCTION(_sk_bilinear_py_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11317_sk_bilinear_py_hsw:
11318 .byte 72,173 // lods %ds:(%rsi),%rax
11319 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11320 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11321 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
11322 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
11323 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
11324 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
11325 .byte 72,173 // lods %ds:(%rsi),%rax
11326 .byte 255,224 // jmpq *%rax
11327
11328HIDDEN _sk_bicubic_n3x_hsw
11329.globl _sk_bicubic_n3x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11330FUNCTION(_sk_bicubic_n3x_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11331_sk_bicubic_n3x_hsw:
11332 .byte 72,173 // lods %ds:(%rsi),%rax
11333 .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d
11334 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11335 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
11336 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
11337 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11338 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11339 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11340 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
11341 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
11342 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
11343 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11344 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11345 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
11346 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11347 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11348 .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10
11349 .byte 196,65,44,89,193 // vmulps %ymm9,%ymm10,%ymm8
11350 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
11351 .byte 72,173 // lods %ds:(%rsi),%rax
11352 .byte 255,224 // jmpq *%rax
11353
11354HIDDEN _sk_bicubic_n1x_hsw
11355.globl _sk_bicubic_n1x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11356FUNCTION(_sk_bicubic_n1x_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11357_sk_bicubic_n1x_hsw:
11358 .byte 72,173 // lods %ds:(%rsi),%rax
11359 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
11360 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11361 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
11362 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
11363 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11364 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11365 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11366 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
11367 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
11368 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
11369 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
11370 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11371 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11372 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11373 .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9
11374 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11375 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11376 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11377 .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10
11378 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
11379 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
11380 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
11381 .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
11382 .byte 197,124,17,136,128,0,0,0 // vmovups %ymm9,0x80(%rax)
11383 .byte 72,173 // lods %ds:(%rsi),%rax
11384 .byte 255,224 // jmpq *%rax
11385
11386HIDDEN _sk_bicubic_p1x_hsw
11387.globl _sk_bicubic_p1x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11388FUNCTION(_sk_bicubic_p1x_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11389_sk_bicubic_p1x_hsw:
11390 .byte 72,173 // lods %ds:(%rsi),%rax
11391 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11392 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11393 .byte 196,98,125,88,192 // vpbroadcastd %xmm0,%ymm8
11394 .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
11395 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
11396 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
11397 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11398 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11399 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11400 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11401 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11402 .byte 196,66,53,168,211 // vfmadd213ps %ymm11,%ymm9,%ymm10
11403 .byte 196,66,53,168,208 // vfmadd213ps %ymm8,%ymm9,%ymm10
11404 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
11405 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11406 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11407 .byte 196,66,53,184,194 // vfmadd231ps %ymm10,%ymm9,%ymm8
11408 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
11409 .byte 72,173 // lods %ds:(%rsi),%rax
11410 .byte 255,224 // jmpq *%rax
11411
11412HIDDEN _sk_bicubic_p3x_hsw
11413.globl _sk_bicubic_p3x_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11414FUNCTION(_sk_bicubic_p3x_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11415_sk_bicubic_p3x_hsw:
11416 .byte 72,173 // lods %ds:(%rsi),%rax
11417 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11418 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
11419 .byte 196,226,125,88,192 // vpbroadcastd %xmm0,%ymm0
11420 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
11421 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
11422 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
11423 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
11424 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11425 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11426 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
11427 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11428 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11429 .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10
11430 .byte 196,65,52,89,194 // vmulps %ymm10,%ymm9,%ymm8
11431 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
11432 .byte 72,173 // lods %ds:(%rsi),%rax
11433 .byte 255,224 // jmpq *%rax
11434
11435HIDDEN _sk_bicubic_n3y_hsw
11436.globl _sk_bicubic_n3y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11437FUNCTION(_sk_bicubic_n3y_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11438_sk_bicubic_n3y_hsw:
11439 .byte 72,173 // lods %ds:(%rsi),%rax
11440 .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d
11441 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11442 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
11443 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
11444 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11445 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11446 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11447 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
11448 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
11449 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
11450 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11451 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11452 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
11453 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11454 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11455 .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10
11456 .byte 196,65,44,89,193 // vmulps %ymm9,%ymm10,%ymm8
11457 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
11458 .byte 72,173 // lods %ds:(%rsi),%rax
11459 .byte 255,224 // jmpq *%rax
11460
11461HIDDEN _sk_bicubic_n1y_hsw
11462.globl _sk_bicubic_n1y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11463FUNCTION(_sk_bicubic_n1y_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11464_sk_bicubic_n1y_hsw:
11465 .byte 72,173 // lods %ds:(%rsi),%rax
11466 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
11467 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11468 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
11469 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
11470 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
11471 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11472 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11473 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
11474 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
11475 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
11476 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
11477 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11478 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11479 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11480 .byte 196,66,61,168,202 // vfmadd213ps %ymm10,%ymm8,%ymm9
11481 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11482 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11483 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11484 .byte 196,66,61,184,209 // vfmadd231ps %ymm9,%ymm8,%ymm10
11485 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
11486 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
11487 .byte 196,66,125,88,201 // vpbroadcastd %xmm9,%ymm9
11488 .byte 196,66,61,184,202 // vfmadd231ps %ymm10,%ymm8,%ymm9
11489 .byte 197,124,17,136,160,0,0,0 // vmovups %ymm9,0xa0(%rax)
11490 .byte 72,173 // lods %ds:(%rsi),%rax
11491 .byte 255,224 // jmpq *%rax
11492
11493HIDDEN _sk_bicubic_p1y_hsw
11494.globl _sk_bicubic_p1y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11495FUNCTION(_sk_bicubic_p1y_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11496_sk_bicubic_p1y_hsw:
11497 .byte 72,173 // lods %ds:(%rsi),%rax
11498 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
11499 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11500 .byte 196,98,125,88,193 // vpbroadcastd %xmm1,%ymm8
11501 .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
11502 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
11503 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
11504 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11505 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11506 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11507 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11508 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11509 .byte 196,66,53,168,211 // vfmadd213ps %ymm11,%ymm9,%ymm10
11510 .byte 196,66,53,168,208 // vfmadd213ps %ymm8,%ymm9,%ymm10
11511 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
11512 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
11513 .byte 196,66,125,88,192 // vpbroadcastd %xmm8,%ymm8
11514 .byte 196,66,53,184,194 // vfmadd231ps %ymm10,%ymm9,%ymm8
11515 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
11516 .byte 72,173 // lods %ds:(%rsi),%rax
11517 .byte 255,224 // jmpq *%rax
11518
11519HIDDEN _sk_bicubic_p3y_hsw
11520.globl _sk_bicubic_p3y_hsw
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11521FUNCTION(_sk_bicubic_p3y_hsw)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]11522_sk_bicubic_p3y_hsw:
11523 .byte 72,173 // lods %ds:(%rsi),%rax
11524 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
11525 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
11526 .byte 196,226,125,88,201 // vpbroadcastd %xmm1,%ymm1
11527 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
11528 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
11529 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
11530 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
11531 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
11532 .byte 196,66,125,88,210 // vpbroadcastd %xmm10,%ymm10
11533 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
11534 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
11535 .byte 196,66,125,88,219 // vpbroadcastd %xmm11,%ymm11
11536 .byte 196,66,61,168,211 // vfmadd213ps %ymm11,%ymm8,%ymm10
11537 .byte 196,65,52,89,194 // vmulps %ymm10,%ymm9,%ymm8
11538 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
11539 .byte 72,173 // lods %ds:(%rsi),%rax
11540 .byte 255,224 // jmpq *%rax
11541
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11542HIDDEN _sk_callback_hsw
11543.globl _sk_callback_hsw
11544FUNCTION(_sk_callback_hsw)
11545_sk_callback_hsw:
11546 .byte 65,87 // push %r15
11547 .byte 65,86 // push %r14
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11548 .byte 65,85 // push %r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11549 .byte 65,84 // push %r12
11550 .byte 83 // push %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11551 .byte 72,129,236,144,0,0,0 // sub $0x90,%rsp
11552 .byte 197,252,17,124,36,96 // vmovups %ymm7,0x60(%rsp)
11553 .byte 197,252,17,116,36,64 // vmovups %ymm6,0x40(%rsp)
11554 .byte 197,252,17,108,36,32 // vmovups %ymm5,0x20(%rsp)
11555 .byte 197,252,17,36,36 // vmovups %ymm4,(%rsp)
11556 .byte 73,137,205 // mov %rcx,%r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11557 .byte 73,137,214 // mov %rdx,%r14
11558 .byte 73,137,255 // mov %rdi,%r15
11559 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11560 .byte 72,137,195 // mov %rax,%rbx
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11561 .byte 73,137,244 // mov %rsi,%r12
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11562 .byte 197,252,20,225 // vunpcklps %ymm1,%ymm0,%ymm4
11563 .byte 197,252,21,193 // vunpckhps %ymm1,%ymm0,%ymm0
11564 .byte 197,236,20,203 // vunpcklps %ymm3,%ymm2,%ymm1
11565 .byte 197,236,21,211 // vunpckhps %ymm3,%ymm2,%ymm2
11566 .byte 197,221,20,217 // vunpcklpd %ymm1,%ymm4,%ymm3
11567 .byte 197,221,21,201 // vunpckhpd %ymm1,%ymm4,%ymm1
11568 .byte 197,253,20,226 // vunpcklpd %ymm2,%ymm0,%ymm4
11569 .byte 197,253,21,194 // vunpckhpd %ymm2,%ymm0,%ymm0
11570 .byte 196,227,101,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm3,%ymm2
11571 .byte 196,227,93,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm4,%ymm5
11572 .byte 196,227,101,6,201,49 // vperm2f128 $0x31,%ymm1,%ymm3,%ymm1
11573 .byte 196,227,93,6,192,49 // vperm2f128 $0x31,%ymm0,%ymm4,%ymm0
11574 .byte 197,253,17,83,8 // vmovupd %ymm2,0x8(%rbx)
11575 .byte 197,253,17,107,40 // vmovupd %ymm5,0x28(%rbx)
11576 .byte 197,253,17,75,72 // vmovupd %ymm1,0x48(%rbx)
11577 .byte 197,253,17,67,104 // vmovupd %ymm0,0x68(%rbx)
11578 .byte 77,133,237 // test %r13,%r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11579 .byte 190,8,0,0,0 // mov $0x8,%esi
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11580 .byte 65,15,69,245 // cmovne %r13d,%esi
11581 .byte 72,137,223 // mov %rbx,%rdi
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11582 .byte 197,248,119 // vzeroupper
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11583 .byte 255,19 // callq *(%rbx)
11584 .byte 72,139,131,136,0,0,0 // mov 0x88(%rbx),%rax
11585 .byte 197,248,16,0 // vmovups (%rax),%xmm0
11586 .byte 197,248,16,72,16 // vmovups 0x10(%rax),%xmm1
11587 .byte 197,248,16,80,32 // vmovups 0x20(%rax),%xmm2
11588 .byte 197,248,16,88,48 // vmovups 0x30(%rax),%xmm3
11589 .byte 196,227,101,24,88,112,1 // vinsertf128 $0x1,0x70(%rax),%ymm3,%ymm3
11590 .byte 196,227,109,24,80,96,1 // vinsertf128 $0x1,0x60(%rax),%ymm2,%ymm2
11591 .byte 196,227,117,24,72,80,1 // vinsertf128 $0x1,0x50(%rax),%ymm1,%ymm1
11592 .byte 196,227,125,24,64,64,1 // vinsertf128 $0x1,0x40(%rax),%ymm0,%ymm0
11593 .byte 197,252,20,225 // vunpcklps %ymm1,%ymm0,%ymm4
11594 .byte 197,252,21,233 // vunpckhps %ymm1,%ymm0,%ymm5
11595 .byte 197,236,20,203 // vunpcklps %ymm3,%ymm2,%ymm1
11596 .byte 197,236,21,219 // vunpckhps %ymm3,%ymm2,%ymm3
11597 .byte 197,221,20,193 // vunpcklpd %ymm1,%ymm4,%ymm0
11598 .byte 197,221,21,201 // vunpckhpd %ymm1,%ymm4,%ymm1
11599 .byte 197,213,20,211 // vunpcklpd %ymm3,%ymm5,%ymm2
11600 .byte 197,213,21,219 // vunpckhpd %ymm3,%ymm5,%ymm3
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11601 .byte 76,137,230 // mov %r12,%rsi
11602 .byte 72,173 // lods %ds:(%rsi),%rax
11603 .byte 76,137,255 // mov %r15,%rdi
11604 .byte 76,137,242 // mov %r14,%rdx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11605 .byte 76,137,233 // mov %r13,%rcx
11606 .byte 197,252,16,36,36 // vmovups (%rsp),%ymm4
11607 .byte 197,252,16,108,36,32 // vmovups 0x20(%rsp),%ymm5
11608 .byte 197,252,16,116,36,64 // vmovups 0x40(%rsp),%ymm6
11609 .byte 197,252,16,124,36,96 // vmovups 0x60(%rsp),%ymm7
11610 .byte 72,129,196,144,0,0,0 // add $0x90,%rsp
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11611 .byte 91 // pop %rbx
11612 .byte 65,92 // pop %r12
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]11613 .byte 65,93 // pop %r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]11614 .byte 65,94 // pop %r14
11615 .byte 65,95 // pop %r15
11616 .byte 255,224 // jmpq *%rax
11617
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]11618BALIGN4
11619 .byte 0,0 // add %al,(%rax)
11620 .byte 0,63 // add %bh,(%rdi)
11621 .byte 0,0 // add %al,(%rax)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]11622 .byte 128,63,0 // cmpb $0x0,(%rdi)
11623 .byte 0,0 // add %al,(%rax)
11624 .byte 52,255 // xor $0xff,%al
11625 .byte 255 // (bad)
11626 .byte 127,0 // jg 4110 <.literal4+0x10>
11627 .byte 0,0 // add %al,(%rax)
11628 .byte 0,63 // add %bh,(%rdi)
11629 .byte 119,115 // ja 4189 <.literal4+0x89>
11630 .byte 248 // clc
11631 .byte 194,117,191 // retq $0xbf75
11632 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
11633 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
11634 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
11635 .byte 154 // (bad)
11636 .byte 64,254 // rex (bad)
11637 .byte 210,221 // rcr %cl,%ch
11638 .byte 65,0,0 // add %al,(%r8)
11639 .byte 0,75,0 // add %cl,0x0(%rbx)
11640 .byte 0,0 // add %al,(%rax)
11641 .byte 52,255 // xor $0xff,%al
11642 .byte 255 // (bad)
11643 .byte 127,0 // jg 4140 <.literal4+0x40>
11644 .byte 0,0 // add %al,(%rax)
11645 .byte 0,63 // add %bh,(%rdi)
11646 .byte 119,115 // ja 41b9 <.literal4+0xb9>
11647 .byte 248 // clc
11648 .byte 194,117,191 // retq $0xbf75
11649 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
11650 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
11651 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
11652 .byte 154 // (bad)
11653 .byte 64,254 // rex (bad)
11654 .byte 210,221 // rcr %cl,%ch
11655 .byte 65,0,0 // add %al,(%r8)
11656 .byte 0,75,0 // add %cl,0x0(%rbx)
11657 .byte 0,0 // add %al,(%rax)
11658 .byte 52,255 // xor $0xff,%al
11659 .byte 255 // (bad)
11660 .byte 127,0 // jg 4170 <.literal4+0x70>
11661 .byte 0,0 // add %al,(%rax)
11662 .byte 0,63 // add %bh,(%rdi)
11663 .byte 119,115 // ja 41e9 <_sk_callback_hsw+0x20f>
11664 .byte 248 // clc
11665 .byte 194,117,191 // retq $0xbf75
11666 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
11667 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
11668 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
11669 .byte 154 // (bad)
11670 .byte 64,254 // rex (bad)
11671 .byte 210,221 // rcr %cl,%ch
11672 .byte 65,0,0 // add %al,(%r8)
11673 .byte 0,75,0 // add %cl,0x0(%rbx)
11674 .byte 0,0 // add %al,(%rax)
11675 .byte 52,255 // xor $0xff,%al
11676 .byte 255 // (bad)
11677 .byte 127,0 // jg 41a0 <.literal4+0xa0>
11678 .byte 0,0 // add %al,(%rax)
11679 .byte 0,63 // add %bh,(%rdi)
11680 .byte 119,115 // ja 4219 <_sk_callback_hsw+0x23f>
11681 .byte 248 // clc
11682 .byte 194,117,191 // retq $0xbf75
11683 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
11684 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
11685 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
11686 .byte 154 // (bad)
11687 .byte 64,254 // rex (bad)
11688 .byte 210,221 // rcr %cl,%ch
11689 .byte 65,0,0 // add %al,(%r8)
11690 .byte 0 // .byte 0x0
11691 .byte 75 // rex.WXB
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]11692BALIGN32
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]11693
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11694HIDDEN _sk_start_pipeline_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11695.globl _sk_start_pipeline_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11696FUNCTION(_sk_start_pipeline_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11697_sk_start_pipeline_avx:
11698 .byte 65,87 // push %r15
11699 .byte 65,86 // push %r14
11700 .byte 65,85 // push %r13
11701 .byte 65,84 // push %r12
11702 .byte 83 // push %rbx
11703 .byte 73,137,205 // mov %rcx,%r13
11704 .byte 73,137,214 // mov %rdx,%r14
11705 .byte 72,137,251 // mov %rdi,%rbx
11706 .byte 72,173 // lods %ds:(%rsi),%rax
11707 .byte 73,137,199 // mov %rax,%r15
11708 .byte 73,137,244 // mov %rsi,%r12
11709 .byte 72,141,67,8 // lea 0x8(%rbx),%rax
11710 .byte 76,57,232 // cmp %r13,%rax
11711 .byte 118,5 // jbe 28 <_sk_start_pipeline_avx+0x28>
11712 .byte 72,137,223 // mov %rbx,%rdi
11713 .byte 235,65 // jmp 69 <_sk_start_pipeline_avx+0x69>
11714 .byte 185,0,0,0,0 // mov $0x0,%ecx
11715 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
11716 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11717 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11718 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11719 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
11720 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
11721 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
11722 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
11723 .byte 72,137,223 // mov %rbx,%rdi
11724 .byte 76,137,230 // mov %r12,%rsi
11725 .byte 76,137,242 // mov %r14,%rdx
11726 .byte 65,255,215 // callq *%r15
11727 .byte 72,141,123,8 // lea 0x8(%rbx),%rdi
11728 .byte 72,131,195,16 // add $0x10,%rbx
11729 .byte 76,57,235 // cmp %r13,%rbx
11730 .byte 72,137,251 // mov %rdi,%rbx
11731 .byte 118,191 // jbe 28 <_sk_start_pipeline_avx+0x28>
11732 .byte 76,137,233 // mov %r13,%rcx
11733 .byte 72,41,249 // sub %rdi,%rcx
11734 .byte 116,41 // je 9a <_sk_start_pipeline_avx+0x9a>
11735 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
11736 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11737 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11738 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11739 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
11740 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
11741 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
11742 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
11743 .byte 76,137,230 // mov %r12,%rsi
11744 .byte 76,137,242 // mov %r14,%rdx
11745 .byte 65,255,215 // callq *%r15
11746 .byte 76,137,232 // mov %r13,%rax
11747 .byte 91 // pop %rbx
11748 .byte 65,92 // pop %r12
11749 .byte 65,93 // pop %r13
11750 .byte 65,94 // pop %r14
11751 .byte 65,95 // pop %r15
11752 .byte 197,248,119 // vzeroupper
11753 .byte 195 // retq
11754
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11755HIDDEN _sk_just_return_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11756.globl _sk_just_return_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11757FUNCTION(_sk_just_return_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11758_sk_just_return_avx:
11759 .byte 195 // retq
11760
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11761HIDDEN _sk_seed_shader_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11762.globl _sk_seed_shader_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11763FUNCTION(_sk_seed_shader_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11764_sk_seed_shader_avx:
11765 .byte 72,173 // lods %ds:(%rsi),%rax
11766 .byte 197,249,110,199 // vmovd %edi,%xmm0
11767 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
11768 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
11769 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]11770 .byte 196,226,125,24,13,135,91,0,0 // vbroadcastss 0x5b87(%rip),%ymm1 # 5c50 <_sk_callback_avx+0x126>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11771 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
11772 .byte 197,252,88,2 // vaddps (%rdx),%ymm0,%ymm0
11773 .byte 196,226,125,24,16 // vbroadcastss (%rax),%ymm2
11774 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
11775 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11776 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]11777 .byte 196,226,125,24,21,107,91,0,0 // vbroadcastss 0x5b6b(%rip),%ymm2 # 5c54 <_sk_callback_avx+0x12a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11778 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11779 .byte 197,220,87,228 // vxorps %ymm4,%ymm4,%ymm4
11780 .byte 197,212,87,237 // vxorps %ymm5,%ymm5,%ymm5
11781 .byte 197,204,87,246 // vxorps %ymm6,%ymm6,%ymm6
11782 .byte 197,196,87,255 // vxorps %ymm7,%ymm7,%ymm7
11783 .byte 255,224 // jmpq *%rax
11784
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11785HIDDEN _sk_constant_color_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11786.globl _sk_constant_color_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11787FUNCTION(_sk_constant_color_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11788_sk_constant_color_avx:
11789 .byte 72,173 // lods %ds:(%rsi),%rax
11790 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
11791 .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1
11792 .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2
11793 .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3
11794 .byte 72,173 // lods %ds:(%rsi),%rax
11795 .byte 255,224 // jmpq *%rax
11796
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11797HIDDEN _sk_clear_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11798.globl _sk_clear_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11799FUNCTION(_sk_clear_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11800_sk_clear_avx:
11801 .byte 72,173 // lods %ds:(%rsi),%rax
11802 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
11803 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
11804 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
11805 .byte 197,228,87,219 // vxorps %ymm3,%ymm3,%ymm3
11806 .byte 255,224 // jmpq *%rax
11807
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11808HIDDEN _sk_srcatop_avx
11809.globl _sk_srcatop_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11810FUNCTION(_sk_srcatop_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11811_sk_srcatop_avx:
11812 .byte 197,124,89,199 // vmulps %ymm7,%ymm0,%ymm8
11813 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11814 .byte 197,249,110,192 // vmovd %eax,%xmm0
11815 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
11816 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
11817 .byte 197,124,92,203 // vsubps %ymm3,%ymm0,%ymm9
11818 .byte 197,180,89,196 // vmulps %ymm4,%ymm9,%ymm0
11819 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0
11820 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
11821 .byte 197,52,89,197 // vmulps %ymm5,%ymm9,%ymm8
11822 .byte 196,193,116,88,200 // vaddps %ymm8,%ymm1,%ymm1
11823 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
11824 .byte 197,52,89,198 // vmulps %ymm6,%ymm9,%ymm8
11825 .byte 196,193,108,88,208 // vaddps %ymm8,%ymm2,%ymm2
11826 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
11827 .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8
11828 .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3
11829 .byte 72,173 // lods %ds:(%rsi),%rax
11830 .byte 255,224 // jmpq *%rax
11831
11832HIDDEN _sk_dstatop_avx
11833.globl _sk_dstatop_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11834FUNCTION(_sk_dstatop_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11835_sk_dstatop_avx:
11836 .byte 197,100,89,196 // vmulps %ymm4,%ymm3,%ymm8
11837 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11838 .byte 197,121,110,200 // vmovd %eax,%xmm9
11839 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
11840 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
11841 .byte 197,52,92,207 // vsubps %ymm7,%ymm9,%ymm9
11842 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
11843 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0
11844 .byte 197,100,89,197 // vmulps %ymm5,%ymm3,%ymm8
11845 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
11846 .byte 197,188,88,201 // vaddps %ymm1,%ymm8,%ymm1
11847 .byte 197,100,89,198 // vmulps %ymm6,%ymm3,%ymm8
11848 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
11849 .byte 197,188,88,210 // vaddps %ymm2,%ymm8,%ymm2
11850 .byte 197,100,89,199 // vmulps %ymm7,%ymm3,%ymm8
11851 .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3
11852 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
11853 .byte 72,173 // lods %ds:(%rsi),%rax
11854 .byte 255,224 // jmpq *%rax
11855
11856HIDDEN _sk_srcin_avx
11857.globl _sk_srcin_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11858FUNCTION(_sk_srcin_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11859_sk_srcin_avx:
11860 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
11861 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
11862 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
11863 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
11864 .byte 72,173 // lods %ds:(%rsi),%rax
11865 .byte 255,224 // jmpq *%rax
11866
11867HIDDEN _sk_dstin_avx
11868.globl _sk_dstin_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11869FUNCTION(_sk_dstin_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11870_sk_dstin_avx:
11871 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
11872 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
11873 .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2
11874 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
11875 .byte 72,173 // lods %ds:(%rsi),%rax
11876 .byte 255,224 // jmpq *%rax
11877
11878HIDDEN _sk_srcout_avx
11879.globl _sk_srcout_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11880FUNCTION(_sk_srcout_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11881_sk_srcout_avx:
11882 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11883 .byte 197,121,110,192 // vmovd %eax,%xmm8
11884 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
11885 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
11886 .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
11887 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
11888 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
11889 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
11890 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
11891 .byte 72,173 // lods %ds:(%rsi),%rax
11892 .byte 255,224 // jmpq *%rax
11893
11894HIDDEN _sk_dstout_avx
11895.globl _sk_dstout_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11896FUNCTION(_sk_dstout_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11897_sk_dstout_avx:
11898 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11899 .byte 197,249,110,192 // vmovd %eax,%xmm0
11900 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
11901 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
11902 .byte 197,252,92,219 // vsubps %ymm3,%ymm0,%ymm3
11903 .byte 197,228,89,196 // vmulps %ymm4,%ymm3,%ymm0
11904 .byte 197,228,89,205 // vmulps %ymm5,%ymm3,%ymm1
11905 .byte 197,228,89,214 // vmulps %ymm6,%ymm3,%ymm2
11906 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11907 .byte 72,173 // lods %ds:(%rsi),%rax
11908 .byte 255,224 // jmpq *%rax
11909
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11910HIDDEN _sk_srcover_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11911.globl _sk_srcover_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11912FUNCTION(_sk_srcover_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11913_sk_srcover_avx:
11914 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11915 .byte 197,121,110,192 // vmovd %eax,%xmm8
11916 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
11917 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
11918 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
11919 .byte 197,60,89,204 // vmulps %ymm4,%ymm8,%ymm9
11920 .byte 197,180,88,192 // vaddps %ymm0,%ymm9,%ymm0
11921 .byte 197,60,89,205 // vmulps %ymm5,%ymm8,%ymm9
11922 .byte 197,180,88,201 // vaddps %ymm1,%ymm9,%ymm1
11923 .byte 197,60,89,206 // vmulps %ymm6,%ymm8,%ymm9
11924 .byte 197,180,88,210 // vaddps %ymm2,%ymm9,%ymm2
11925 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
11926 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
11927 .byte 72,173 // lods %ds:(%rsi),%rax
11928 .byte 255,224 // jmpq *%rax
11929
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]11930HIDDEN _sk_dstover_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11931.globl _sk_dstover_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11932FUNCTION(_sk_dstover_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]11933_sk_dstover_avx:
11934 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11935 .byte 197,121,110,192 // vmovd %eax,%xmm8
11936 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
11937 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
11938 .byte 197,60,92,199 // vsubps %ymm7,%ymm8,%ymm8
11939 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
11940 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
11941 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
11942 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
11943 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
11944 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
11945 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
11946 .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
11947 .byte 72,173 // lods %ds:(%rsi),%rax
11948 .byte 255,224 // jmpq *%rax
11949
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11950HIDDEN _sk_modulate_avx
11951.globl _sk_modulate_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11952FUNCTION(_sk_modulate_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11953_sk_modulate_avx:
11954 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
11955 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
11956 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
11957 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
11958 .byte 72,173 // lods %ds:(%rsi),%rax
11959 .byte 255,224 // jmpq *%rax
11960
11961HIDDEN _sk_multiply_avx
11962.globl _sk_multiply_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11963FUNCTION(_sk_multiply_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11964_sk_multiply_avx:
11965 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
11966 .byte 197,121,110,192 // vmovd %eax,%xmm8
11967 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
11968 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
11969 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
11970 .byte 197,52,89,208 // vmulps %ymm0,%ymm9,%ymm10
11971 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
11972 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
11973 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
11974 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
11975 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
11976 .byte 197,52,89,209 // vmulps %ymm1,%ymm9,%ymm10
11977 .byte 197,60,89,221 // vmulps %ymm5,%ymm8,%ymm11
11978 .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10
11979 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
11980 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
11981 .byte 197,52,89,210 // vmulps %ymm2,%ymm9,%ymm10
11982 .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11
11983 .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10
11984 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
11985 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
11986 .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9
11987 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
11988 .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
11989 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
11990 .byte 196,193,100,88,216 // vaddps %ymm8,%ymm3,%ymm3
11991 .byte 72,173 // lods %ds:(%rsi),%rax
11992 .byte 255,224 // jmpq *%rax
11993
11994HIDDEN _sk_plus__avx
11995.globl _sk_plus__avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]11996FUNCTION(_sk_plus__avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]11997_sk_plus__avx:
11998 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
11999 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
12000 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
12001 .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
12002 .byte 72,173 // lods %ds:(%rsi),%rax
12003 .byte 255,224 // jmpq *%rax
12004
12005HIDDEN _sk_screen_avx
12006.globl _sk_screen_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12007FUNCTION(_sk_screen_avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]12008_sk_screen_avx:
12009 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
12010 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
12011 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
12012 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
12013 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
12014 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
12015 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
12016 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
12017 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
12018 .byte 197,100,88,199 // vaddps %ymm7,%ymm3,%ymm8
12019 .byte 197,228,89,223 // vmulps %ymm7,%ymm3,%ymm3
12020 .byte 197,188,92,219 // vsubps %ymm3,%ymm8,%ymm3
12021 .byte 72,173 // lods %ds:(%rsi),%rax
12022 .byte 255,224 // jmpq *%rax
12023
12024HIDDEN _sk_xor__avx
12025.globl _sk_xor__avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12026FUNCTION(_sk_xor__avx)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]12027_sk_xor__avx:
12028 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12029 .byte 197,121,110,192 // vmovd %eax,%xmm8
12030 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12031 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12032 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
12033 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
12034 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12035 .byte 197,60,89,212 // vmulps %ymm4,%ymm8,%ymm10
12036 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
12037 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
12038 .byte 197,60,89,213 // vmulps %ymm5,%ymm8,%ymm10
12039 .byte 197,172,88,201 // vaddps %ymm1,%ymm10,%ymm1
12040 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
12041 .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10
12042 .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2
12043 .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3
12044 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12045 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12046 .byte 72,173 // lods %ds:(%rsi),%rax
12047 .byte 255,224 // jmpq *%rax
12048
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]12049HIDDEN _sk_darken_avx
12050.globl _sk_darken_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12051FUNCTION(_sk_darken_avx)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]12052_sk_darken_avx:
12053 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
12054 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
12055 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
12056 .byte 196,193,124,95,193 // vmaxps %ymm9,%ymm0,%ymm0
12057 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
12058 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
12059 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
12060 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
12061 .byte 196,193,116,95,201 // vmaxps %ymm9,%ymm1,%ymm1
12062 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
12063 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
12064 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
12065 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
12066 .byte 196,193,108,95,209 // vmaxps %ymm9,%ymm2,%ymm2
12067 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
12068 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12069 .byte 197,121,110,192 // vmovd %eax,%xmm8
12070 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12071 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12072 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12073 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12074 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12075 .byte 72,173 // lods %ds:(%rsi),%rax
12076 .byte 255,224 // jmpq *%rax
12077
12078HIDDEN _sk_lighten_avx
12079.globl _sk_lighten_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12080FUNCTION(_sk_lighten_avx)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]12081_sk_lighten_avx:
12082 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
12083 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
12084 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
12085 .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0
12086 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
12087 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
12088 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
12089 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
12090 .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1
12091 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
12092 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
12093 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
12094 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
12095 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
12096 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
12097 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12098 .byte 197,121,110,192 // vmovd %eax,%xmm8
12099 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12100 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12101 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12102 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12103 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12104 .byte 72,173 // lods %ds:(%rsi),%rax
12105 .byte 255,224 // jmpq *%rax
12106
12107HIDDEN _sk_difference_avx
12108.globl _sk_difference_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12109FUNCTION(_sk_difference_avx)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]12110_sk_difference_avx:
12111 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
12112 .byte 197,252,89,199 // vmulps %ymm7,%ymm0,%ymm0
12113 .byte 197,100,89,204 // vmulps %ymm4,%ymm3,%ymm9
12114 .byte 196,193,124,93,193 // vminps %ymm9,%ymm0,%ymm0
12115 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
12116 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
12117 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
12118 .byte 197,244,89,207 // vmulps %ymm7,%ymm1,%ymm1
12119 .byte 197,100,89,205 // vmulps %ymm5,%ymm3,%ymm9
12120 .byte 196,193,116,93,201 // vminps %ymm9,%ymm1,%ymm1
12121 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
12122 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
12123 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
12124 .byte 197,236,89,215 // vmulps %ymm7,%ymm2,%ymm2
12125 .byte 197,100,89,206 // vmulps %ymm6,%ymm3,%ymm9
12126 .byte 196,193,108,93,209 // vminps %ymm9,%ymm2,%ymm2
12127 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
12128 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
12129 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12130 .byte 197,121,110,192 // vmovd %eax,%xmm8
12131 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12132 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12133 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12134 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12135 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12136 .byte 72,173 // lods %ds:(%rsi),%rax
12137 .byte 255,224 // jmpq *%rax
12138
12139HIDDEN _sk_exclusion_avx
12140.globl _sk_exclusion_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12141FUNCTION(_sk_exclusion_avx)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]12142_sk_exclusion_avx:
12143 .byte 197,124,88,196 // vaddps %ymm4,%ymm0,%ymm8
12144 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
12145 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
12146 .byte 197,188,92,192 // vsubps %ymm0,%ymm8,%ymm0
12147 .byte 197,116,88,197 // vaddps %ymm5,%ymm1,%ymm8
12148 .byte 197,244,89,205 // vmulps %ymm5,%ymm1,%ymm1
12149 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
12150 .byte 197,188,92,201 // vsubps %ymm1,%ymm8,%ymm1
12151 .byte 197,108,88,198 // vaddps %ymm6,%ymm2,%ymm8
12152 .byte 197,236,89,214 // vmulps %ymm6,%ymm2,%ymm2
12153 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
12154 .byte 197,188,92,210 // vsubps %ymm2,%ymm8,%ymm2
12155 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12156 .byte 197,121,110,192 // vmovd %eax,%xmm8
12157 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12158 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12159 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12160 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12161 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12162 .byte 72,173 // lods %ds:(%rsi),%rax
12163 .byte 255,224 // jmpq *%rax
12164
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12165HIDDEN _sk_colorburn_avx
12166.globl _sk_colorburn_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12167FUNCTION(_sk_colorburn_avx)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12168_sk_colorburn_avx:
12169 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12170 .byte 197,121,110,192 // vmovd %eax,%xmm8
12171 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12172 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12173 .byte 197,60,92,207 // vsubps %ymm7,%ymm8,%ymm9
12174 .byte 197,52,89,216 // vmulps %ymm0,%ymm9,%ymm11
12175 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
12176 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12177 .byte 197,60,89,228 // vmulps %ymm4,%ymm8,%ymm12
12178 .byte 197,68,92,236 // vsubps %ymm4,%ymm7,%ymm13
12179 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
12180 .byte 197,20,94,232 // vdivps %ymm0,%ymm13,%ymm13
12181 .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13
12182 .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13
12183 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
12184 .byte 196,65,20,88,235 // vaddps %ymm11,%ymm13,%ymm13
12185 .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13
12186 .byte 197,28,88,224 // vaddps %ymm0,%ymm12,%ymm12
12187 .byte 196,193,124,194,194,0 // vcmpeqps %ymm10,%ymm0,%ymm0
12188 .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0
12189 .byte 197,92,194,231,0 // vcmpeqps %ymm7,%ymm4,%ymm12
12190 .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11
12191 .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0
12192 .byte 197,52,89,217 // vmulps %ymm1,%ymm9,%ymm11
12193 .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12
12194 .byte 197,68,92,237 // vsubps %ymm5,%ymm7,%ymm13
12195 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
12196 .byte 197,20,94,233 // vdivps %ymm1,%ymm13,%ymm13
12197 .byte 196,65,68,93,237 // vminps %ymm13,%ymm7,%ymm13
12198 .byte 196,65,68,92,237 // vsubps %ymm13,%ymm7,%ymm13
12199 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
12200 .byte 196,65,36,88,237 // vaddps %ymm13,%ymm11,%ymm13
12201 .byte 196,65,28,88,237 // vaddps %ymm13,%ymm12,%ymm13
12202 .byte 197,28,88,225 // vaddps %ymm1,%ymm12,%ymm12
12203 .byte 196,193,116,194,202,0 // vcmpeqps %ymm10,%ymm1,%ymm1
12204 .byte 196,195,21,74,204,16 // vblendvps %ymm1,%ymm12,%ymm13,%ymm1
12205 .byte 197,84,194,231,0 // vcmpeqps %ymm7,%ymm5,%ymm12
12206 .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11
12207 .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1
12208 .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9
12209 .byte 196,65,108,194,210,0 // vcmpeqps %ymm10,%ymm2,%ymm10
12210 .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11
12211 .byte 197,68,92,230 // vsubps %ymm6,%ymm7,%ymm12
12212 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
12213 .byte 197,28,94,226 // vdivps %ymm2,%ymm12,%ymm12
12214 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
12215 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
12216 .byte 196,65,68,92,228 // vsubps %ymm12,%ymm7,%ymm12
12217 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
12218 .byte 196,65,52,88,228 // vaddps %ymm12,%ymm9,%ymm12
12219 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
12220 .byte 196,227,37,74,210,160 // vblendvps %ymm10,%ymm2,%ymm11,%ymm2
12221 .byte 197,76,194,215,0 // vcmpeqps %ymm7,%ymm6,%ymm10
12222 .byte 197,52,88,206 // vaddps %ymm6,%ymm9,%ymm9
12223 .byte 196,195,109,74,209,160 // vblendvps %ymm10,%ymm9,%ymm2,%ymm2
12224 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12225 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12226 .byte 72,173 // lods %ds:(%rsi),%rax
12227 .byte 255,224 // jmpq *%rax
12228
12229HIDDEN _sk_colordodge_avx
12230.globl _sk_colordodge_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12231FUNCTION(_sk_colordodge_avx)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12232_sk_colordodge_avx:
12233 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
12234 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12235 .byte 197,121,110,200 // vmovd %eax,%xmm9
12236 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
12237 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
12238 .byte 197,52,92,215 // vsubps %ymm7,%ymm9,%ymm10
12239 .byte 197,44,89,216 // vmulps %ymm0,%ymm10,%ymm11
12240 .byte 197,52,92,203 // vsubps %ymm3,%ymm9,%ymm9
12241 .byte 197,100,89,228 // vmulps %ymm4,%ymm3,%ymm12
12242 .byte 197,100,92,232 // vsubps %ymm0,%ymm3,%ymm13
12243 .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12
12244 .byte 197,52,89,236 // vmulps %ymm4,%ymm9,%ymm13
12245 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
12246 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
12247 .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12
12248 .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12
12249 .byte 197,20,88,232 // vaddps %ymm0,%ymm13,%ymm13
12250 .byte 197,252,194,195,0 // vcmpeqps %ymm3,%ymm0,%ymm0
12251 .byte 196,195,29,74,197,0 // vblendvps %ymm0,%ymm13,%ymm12,%ymm0
12252 .byte 196,65,92,194,224,0 // vcmpeqps %ymm8,%ymm4,%ymm12
12253 .byte 197,36,88,220 // vaddps %ymm4,%ymm11,%ymm11
12254 .byte 196,195,125,74,195,192 // vblendvps %ymm12,%ymm11,%ymm0,%ymm0
12255 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
12256 .byte 197,100,89,229 // vmulps %ymm5,%ymm3,%ymm12
12257 .byte 197,100,92,233 // vsubps %ymm1,%ymm3,%ymm13
12258 .byte 196,65,28,94,229 // vdivps %ymm13,%ymm12,%ymm12
12259 .byte 197,52,89,237 // vmulps %ymm5,%ymm9,%ymm13
12260 .byte 196,65,68,93,228 // vminps %ymm12,%ymm7,%ymm12
12261 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
12262 .byte 196,65,28,88,227 // vaddps %ymm11,%ymm12,%ymm12
12263 .byte 196,65,20,88,228 // vaddps %ymm12,%ymm13,%ymm12
12264 .byte 197,20,88,233 // vaddps %ymm1,%ymm13,%ymm13
12265 .byte 197,244,194,203,0 // vcmpeqps %ymm3,%ymm1,%ymm1
12266 .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1
12267 .byte 196,65,84,194,224,0 // vcmpeqps %ymm8,%ymm5,%ymm12
12268 .byte 197,36,88,221 // vaddps %ymm5,%ymm11,%ymm11
12269 .byte 196,195,117,74,203,192 // vblendvps %ymm12,%ymm11,%ymm1,%ymm1
12270 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
12271 .byte 197,100,89,222 // vmulps %ymm6,%ymm3,%ymm11
12272 .byte 197,100,92,226 // vsubps %ymm2,%ymm3,%ymm12
12273 .byte 196,65,36,94,220 // vdivps %ymm12,%ymm11,%ymm11
12274 .byte 197,52,89,230 // vmulps %ymm6,%ymm9,%ymm12
12275 .byte 196,65,68,93,219 // vminps %ymm11,%ymm7,%ymm11
12276 .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11
12277 .byte 196,65,44,88,219 // vaddps %ymm11,%ymm10,%ymm11
12278 .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11
12279 .byte 197,28,88,226 // vaddps %ymm2,%ymm12,%ymm12
12280 .byte 197,236,194,211,0 // vcmpeqps %ymm3,%ymm2,%ymm2
12281 .byte 196,195,37,74,212,32 // vblendvps %ymm2,%ymm12,%ymm11,%ymm2
12282 .byte 196,65,76,194,192,0 // vcmpeqps %ymm8,%ymm6,%ymm8
12283 .byte 197,44,88,214 // vaddps %ymm6,%ymm10,%ymm10
12284 .byte 196,195,109,74,210,128 // vblendvps %ymm8,%ymm10,%ymm2,%ymm2
12285 .byte 197,52,89,199 // vmulps %ymm7,%ymm9,%ymm8
12286 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12287 .byte 72,173 // lods %ds:(%rsi),%rax
12288 .byte 255,224 // jmpq *%rax
12289
12290HIDDEN _sk_hardlight_avx
12291.globl _sk_hardlight_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12292FUNCTION(_sk_hardlight_avx)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12293_sk_hardlight_avx:
12294 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12295 .byte 197,121,110,192 // vmovd %eax,%xmm8
12296 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12297 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12298 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
12299 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9
12300 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12301 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
12302 .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11
12303 .byte 197,124,88,200 // vaddps %ymm0,%ymm0,%ymm9
12304 .byte 197,52,194,227,2 // vcmpleps %ymm3,%ymm9,%ymm12
12305 .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9
12306 .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13
12307 .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9
12308 .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14
12309 .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0
12310 .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0
12311 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
12312 .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0
12313 .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0
12314 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
12315 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
12316 .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12
12317 .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11
12318 .byte 197,116,88,225 // vaddps %ymm1,%ymm1,%ymm12
12319 .byte 197,28,194,227,2 // vcmpleps %ymm3,%ymm12,%ymm12
12320 .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13
12321 .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13
12322 .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14
12323 .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1
12324 .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
12325 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
12326 .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1
12327 .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1
12328 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
12329 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
12330 .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11
12331 .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10
12332 .byte 197,108,88,218 // vaddps %ymm2,%ymm2,%ymm11
12333 .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11
12334 .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12
12335 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
12336 .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13
12337 .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2
12338 .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2
12339 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
12340 .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2
12341 .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2
12342 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
12343 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12344 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12345 .byte 72,173 // lods %ds:(%rsi),%rax
12346 .byte 255,224 // jmpq *%rax
12347
12348HIDDEN _sk_overlay_avx
12349.globl _sk_overlay_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12350FUNCTION(_sk_overlay_avx)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12351_sk_overlay_avx:
12352 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12353 .byte 197,121,110,192 // vmovd %eax,%xmm8
12354 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12355 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12356 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
12357 .byte 197,44,89,200 // vmulps %ymm0,%ymm10,%ymm9
12358 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12359 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
12360 .byte 196,65,52,88,219 // vaddps %ymm11,%ymm9,%ymm11
12361 .byte 197,92,88,204 // vaddps %ymm4,%ymm4,%ymm9
12362 .byte 197,52,194,231,2 // vcmpleps %ymm7,%ymm9,%ymm12
12363 .byte 197,124,89,204 // vmulps %ymm4,%ymm0,%ymm9
12364 .byte 196,65,52,88,233 // vaddps %ymm9,%ymm9,%ymm13
12365 .byte 197,100,89,207 // vmulps %ymm7,%ymm3,%ymm9
12366 .byte 197,68,92,244 // vsubps %ymm4,%ymm7,%ymm14
12367 .byte 197,228,92,192 // vsubps %ymm0,%ymm3,%ymm0
12368 .byte 196,193,124,89,198 // vmulps %ymm14,%ymm0,%ymm0
12369 .byte 197,252,88,192 // vaddps %ymm0,%ymm0,%ymm0
12370 .byte 197,180,92,192 // vsubps %ymm0,%ymm9,%ymm0
12371 .byte 196,195,125,74,197,192 // vblendvps %ymm12,%ymm13,%ymm0,%ymm0
12372 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
12373 .byte 197,44,89,217 // vmulps %ymm1,%ymm10,%ymm11
12374 .byte 197,60,89,229 // vmulps %ymm5,%ymm8,%ymm12
12375 .byte 196,65,28,88,219 // vaddps %ymm11,%ymm12,%ymm11
12376 .byte 197,84,88,229 // vaddps %ymm5,%ymm5,%ymm12
12377 .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12
12378 .byte 197,116,89,237 // vmulps %ymm5,%ymm1,%ymm13
12379 .byte 196,65,20,88,237 // vaddps %ymm13,%ymm13,%ymm13
12380 .byte 197,68,92,245 // vsubps %ymm5,%ymm7,%ymm14
12381 .byte 197,228,92,201 // vsubps %ymm1,%ymm3,%ymm1
12382 .byte 196,193,116,89,206 // vmulps %ymm14,%ymm1,%ymm1
12383 .byte 197,244,88,201 // vaddps %ymm1,%ymm1,%ymm1
12384 .byte 197,180,92,201 // vsubps %ymm1,%ymm9,%ymm1
12385 .byte 196,195,117,74,205,192 // vblendvps %ymm12,%ymm13,%ymm1,%ymm1
12386 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
12387 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
12388 .byte 197,60,89,222 // vmulps %ymm6,%ymm8,%ymm11
12389 .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10
12390 .byte 197,76,88,222 // vaddps %ymm6,%ymm6,%ymm11
12391 .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11
12392 .byte 197,108,89,230 // vmulps %ymm6,%ymm2,%ymm12
12393 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
12394 .byte 197,68,92,238 // vsubps %ymm6,%ymm7,%ymm13
12395 .byte 197,228,92,210 // vsubps %ymm2,%ymm3,%ymm2
12396 .byte 196,193,108,89,213 // vmulps %ymm13,%ymm2,%ymm2
12397 .byte 197,236,88,210 // vaddps %ymm2,%ymm2,%ymm2
12398 .byte 197,180,92,210 // vsubps %ymm2,%ymm9,%ymm2
12399 .byte 196,195,109,74,212,176 // vblendvps %ymm11,%ymm12,%ymm2,%ymm2
12400 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
12401 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12402 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12403 .byte 72,173 // lods %ds:(%rsi),%rax
12404 .byte 255,224 // jmpq *%rax
12405
12406HIDDEN _sk_softlight_avx
12407.globl _sk_softlight_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12408FUNCTION(_sk_softlight_avx)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12409_sk_softlight_avx:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]12410 .byte 197,252,17,84,36,200 // vmovups %ymm2,-0x38(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12411 .byte 197,252,40,209 // vmovaps %ymm1,%ymm2
12412 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
12413 .byte 197,52,194,215,1 // vcmpltps %ymm7,%ymm9,%ymm10
12414 .byte 197,92,94,199 // vdivps %ymm7,%ymm4,%ymm8
12415 .byte 196,67,53,74,216,160 // vblendvps %ymm10,%ymm8,%ymm9,%ymm11
12416 .byte 196,65,36,88,195 // vaddps %ymm11,%ymm11,%ymm8
12417 .byte 196,65,60,88,224 // vaddps %ymm8,%ymm8,%ymm12
12418 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12419 .byte 197,121,110,192 // vmovd %eax,%xmm8
12420 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12421 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12422 .byte 196,65,28,89,236 // vmulps %ymm12,%ymm12,%ymm13
12423 .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
12424 .byte 196,65,36,92,232 // vsubps %ymm8,%ymm11,%ymm13
12425 .byte 196,65,28,89,237 // vmulps %ymm13,%ymm12,%ymm13
12426 .byte 184,0,0,224,64 // mov $0x40e00000,%eax
12427 .byte 197,121,110,224 // vmovd %eax,%xmm12
12428 .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12
12429 .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12
12430 .byte 196,65,36,89,244 // vmulps %ymm12,%ymm11,%ymm14
12431 .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
12432 .byte 196,65,124,82,243 // vrsqrtps %ymm11,%ymm14
12433 .byte 196,65,124,83,246 // vrcpps %ymm14,%ymm14
12434 .byte 196,65,12,92,243 // vsubps %ymm11,%ymm14,%ymm14
12435 .byte 197,92,88,252 // vaddps %ymm4,%ymm4,%ymm15
12436 .byte 196,65,4,88,255 // vaddps %ymm15,%ymm15,%ymm15
12437 .byte 197,4,194,255,2 // vcmpleps %ymm7,%ymm15,%ymm15
12438 .byte 196,67,13,74,237,240 // vblendvps %ymm15,%ymm13,%ymm14,%ymm13
12439 .byte 197,124,88,240 // vaddps %ymm0,%ymm0,%ymm14
12440 .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15
12441 .byte 196,65,60,92,219 // vsubps %ymm11,%ymm8,%ymm11
12442 .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11
12443 .byte 197,36,88,219 // vaddps %ymm3,%ymm11,%ymm11
12444 .byte 197,36,89,220 // vmulps %ymm4,%ymm11,%ymm11
12445 .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15
12446 .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13
12447 .byte 197,100,89,252 // vmulps %ymm4,%ymm3,%ymm15
12448 .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13
12449 .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14
12450 .byte 196,195,21,74,203,224 // vblendvps %ymm14,%ymm11,%ymm13,%ymm1
12451 .byte 197,84,94,239 // vdivps %ymm7,%ymm5,%ymm13
12452 .byte 196,67,53,74,237,160 // vblendvps %ymm10,%ymm13,%ymm9,%ymm13
12453 .byte 196,65,20,88,245 // vaddps %ymm13,%ymm13,%ymm14
12454 .byte 196,65,12,88,246 // vaddps %ymm14,%ymm14,%ymm14
12455 .byte 196,65,12,89,254 // vmulps %ymm14,%ymm14,%ymm15
12456 .byte 196,65,12,88,247 // vaddps %ymm15,%ymm14,%ymm14
12457 .byte 196,65,20,92,248 // vsubps %ymm8,%ymm13,%ymm15
12458 .byte 196,65,4,89,246 // vmulps %ymm14,%ymm15,%ymm14
12459 .byte 196,65,28,89,253 // vmulps %ymm13,%ymm12,%ymm15
12460 .byte 196,65,4,88,246 // vaddps %ymm14,%ymm15,%ymm14
12461 .byte 196,65,124,82,253 // vrsqrtps %ymm13,%ymm15
12462 .byte 196,65,124,83,255 // vrcpps %ymm15,%ymm15
12463 .byte 196,65,4,92,253 // vsubps %ymm13,%ymm15,%ymm15
12464 .byte 197,84,88,221 // vaddps %ymm5,%ymm5,%ymm11
12465 .byte 196,65,36,88,219 // vaddps %ymm11,%ymm11,%ymm11
12466 .byte 197,36,194,223,2 // vcmpleps %ymm7,%ymm11,%ymm11
12467 .byte 196,67,5,74,222,176 // vblendvps %ymm11,%ymm14,%ymm15,%ymm11
12468 .byte 197,108,88,242 // vaddps %ymm2,%ymm2,%ymm14
12469 .byte 196,65,60,92,237 // vsubps %ymm13,%ymm8,%ymm13
12470 .byte 197,12,92,251 // vsubps %ymm3,%ymm14,%ymm15
12471 .byte 196,65,4,89,237 // vmulps %ymm13,%ymm15,%ymm13
12472 .byte 197,4,89,255 // vmulps %ymm7,%ymm15,%ymm15
12473 .byte 196,65,4,89,219 // vmulps %ymm11,%ymm15,%ymm11
12474 .byte 197,100,89,253 // vmulps %ymm5,%ymm3,%ymm15
12475 .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11
12476 .byte 197,20,88,235 // vaddps %ymm3,%ymm13,%ymm13
12477 .byte 197,20,89,237 // vmulps %ymm5,%ymm13,%ymm13
12478 .byte 197,12,194,243,2 // vcmpleps %ymm3,%ymm14,%ymm14
12479 .byte 196,67,37,74,237,224 // vblendvps %ymm14,%ymm13,%ymm11,%ymm13
12480 .byte 197,76,94,223 // vdivps %ymm7,%ymm6,%ymm11
12481 .byte 196,67,53,74,203,160 // vblendvps %ymm10,%ymm11,%ymm9,%ymm9
12482 .byte 196,65,52,88,209 // vaddps %ymm9,%ymm9,%ymm10
12483 .byte 196,65,44,88,210 // vaddps %ymm10,%ymm10,%ymm10
12484 .byte 196,65,44,89,218 // vmulps %ymm10,%ymm10,%ymm11
12485 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
12486 .byte 196,65,52,92,216 // vsubps %ymm8,%ymm9,%ymm11
12487 .byte 196,65,36,89,210 // vmulps %ymm10,%ymm11,%ymm10
12488 .byte 196,65,28,89,217 // vmulps %ymm9,%ymm12,%ymm11
12489 .byte 196,65,36,88,210 // vaddps %ymm10,%ymm11,%ymm10
12490 .byte 196,65,124,82,217 // vrsqrtps %ymm9,%ymm11
12491 .byte 196,65,124,83,219 // vrcpps %ymm11,%ymm11
12492 .byte 196,65,36,92,217 // vsubps %ymm9,%ymm11,%ymm11
12493 .byte 197,76,88,230 // vaddps %ymm6,%ymm6,%ymm12
12494 .byte 196,65,28,88,228 // vaddps %ymm12,%ymm12,%ymm12
12495 .byte 197,28,194,231,2 // vcmpleps %ymm7,%ymm12,%ymm12
12496 .byte 196,67,37,74,210,192 // vblendvps %ymm12,%ymm10,%ymm11,%ymm10
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]12497 .byte 197,124,16,116,36,200 // vmovups -0x38(%rsp),%ymm14
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12498 .byte 196,65,12,88,222 // vaddps %ymm14,%ymm14,%ymm11
12499 .byte 197,36,92,227 // vsubps %ymm3,%ymm11,%ymm12
12500 .byte 196,65,60,92,201 // vsubps %ymm9,%ymm8,%ymm9
12501 .byte 196,65,28,89,201 // vmulps %ymm9,%ymm12,%ymm9
12502 .byte 197,28,89,231 // vmulps %ymm7,%ymm12,%ymm12
12503 .byte 196,65,28,89,210 // vmulps %ymm10,%ymm12,%ymm10
12504 .byte 197,100,89,230 // vmulps %ymm6,%ymm3,%ymm12
12505 .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10
12506 .byte 197,52,88,203 // vaddps %ymm3,%ymm9,%ymm9
12507 .byte 197,52,89,206 // vmulps %ymm6,%ymm9,%ymm9
12508 .byte 197,36,194,219,2 // vcmpleps %ymm3,%ymm11,%ymm11
12509 .byte 196,67,45,74,201,176 // vblendvps %ymm11,%ymm9,%ymm10,%ymm9
12510 .byte 197,60,92,215 // vsubps %ymm7,%ymm8,%ymm10
12511 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
12512 .byte 197,60,92,195 // vsubps %ymm3,%ymm8,%ymm8
12513 .byte 197,60,89,220 // vmulps %ymm4,%ymm8,%ymm11
12514 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
12515 .byte 197,244,88,192 // vaddps %ymm0,%ymm1,%ymm0
12516 .byte 197,172,89,202 // vmulps %ymm2,%ymm10,%ymm1
12517 .byte 197,188,89,213 // vmulps %ymm5,%ymm8,%ymm2
12518 .byte 197,236,88,201 // vaddps %ymm1,%ymm2,%ymm1
12519 .byte 196,193,116,88,205 // vaddps %ymm13,%ymm1,%ymm1
12520 .byte 196,193,44,89,214 // vmulps %ymm14,%ymm10,%ymm2
12521 .byte 197,60,89,214 // vmulps %ymm6,%ymm8,%ymm10
12522 .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2
12523 .byte 196,193,108,88,209 // vaddps %ymm9,%ymm2,%ymm2
12524 .byte 197,60,89,199 // vmulps %ymm7,%ymm8,%ymm8
12525 .byte 197,188,88,219 // vaddps %ymm3,%ymm8,%ymm3
12526 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]12527 .byte 255,224 // jmpq *%rax
12528
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12529HIDDEN _sk_clamp_0_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12530.globl _sk_clamp_0_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12531FUNCTION(_sk_clamp_0_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12532_sk_clamp_0_avx:
12533 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
12534 .byte 196,193,124,95,192 // vmaxps %ymm8,%ymm0,%ymm0
12535 .byte 196,193,116,95,200 // vmaxps %ymm8,%ymm1,%ymm1
12536 .byte 196,193,108,95,208 // vmaxps %ymm8,%ymm2,%ymm2
12537 .byte 196,193,100,95,216 // vmaxps %ymm8,%ymm3,%ymm3
12538 .byte 72,173 // lods %ds:(%rsi),%rax
12539 .byte 255,224 // jmpq *%rax
12540
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12541HIDDEN _sk_clamp_1_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12542.globl _sk_clamp_1_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12543FUNCTION(_sk_clamp_1_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12544_sk_clamp_1_avx:
12545 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12546 .byte 197,121,110,192 // vmovd %eax,%xmm8
12547 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12548 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12549 .byte 196,193,124,93,192 // vminps %ymm8,%ymm0,%ymm0
12550 .byte 196,193,116,93,200 // vminps %ymm8,%ymm1,%ymm1
12551 .byte 196,193,108,93,208 // vminps %ymm8,%ymm2,%ymm2
12552 .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
12553 .byte 72,173 // lods %ds:(%rsi),%rax
12554 .byte 255,224 // jmpq *%rax
12555
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12556HIDDEN _sk_clamp_a_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12557.globl _sk_clamp_a_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12558FUNCTION(_sk_clamp_a_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12559_sk_clamp_a_avx:
12560 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12561 .byte 197,121,110,192 // vmovd %eax,%xmm8
12562 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12563 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12564 .byte 196,193,100,93,216 // vminps %ymm8,%ymm3,%ymm3
12565 .byte 197,252,93,195 // vminps %ymm3,%ymm0,%ymm0
12566 .byte 197,244,93,203 // vminps %ymm3,%ymm1,%ymm1
12567 .byte 197,236,93,211 // vminps %ymm3,%ymm2,%ymm2
12568 .byte 72,173 // lods %ds:(%rsi),%rax
12569 .byte 255,224 // jmpq *%rax
12570
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12571HIDDEN _sk_set_rgb_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12572.globl _sk_set_rgb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12573FUNCTION(_sk_set_rgb_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12574_sk_set_rgb_avx:
12575 .byte 72,173 // lods %ds:(%rsi),%rax
12576 .byte 196,226,125,24,0 // vbroadcastss (%rax),%ymm0
12577 .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1
12578 .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2
12579 .byte 72,173 // lods %ds:(%rsi),%rax
12580 .byte 255,224 // jmpq *%rax
12581
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12582HIDDEN _sk_swap_rb_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12583.globl _sk_swap_rb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12584FUNCTION(_sk_swap_rb_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12585_sk_swap_rb_avx:
12586 .byte 197,124,40,192 // vmovaps %ymm0,%ymm8
12587 .byte 72,173 // lods %ds:(%rsi),%rax
12588 .byte 197,252,40,194 // vmovaps %ymm2,%ymm0
12589 .byte 197,124,41,194 // vmovaps %ymm8,%ymm2
12590 .byte 255,224 // jmpq *%rax
12591
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12592HIDDEN _sk_swap_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12593.globl _sk_swap_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12594FUNCTION(_sk_swap_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12595_sk_swap_avx:
12596 .byte 197,124,40,195 // vmovaps %ymm3,%ymm8
12597 .byte 197,124,40,202 // vmovaps %ymm2,%ymm9
12598 .byte 197,124,40,209 // vmovaps %ymm1,%ymm10
12599 .byte 197,124,40,216 // vmovaps %ymm0,%ymm11
12600 .byte 72,173 // lods %ds:(%rsi),%rax
12601 .byte 197,252,40,196 // vmovaps %ymm4,%ymm0
12602 .byte 197,252,40,205 // vmovaps %ymm5,%ymm1
12603 .byte 197,252,40,214 // vmovaps %ymm6,%ymm2
12604 .byte 197,252,40,223 // vmovaps %ymm7,%ymm3
12605 .byte 197,124,41,220 // vmovaps %ymm11,%ymm4
12606 .byte 197,124,41,213 // vmovaps %ymm10,%ymm5
12607 .byte 197,124,41,206 // vmovaps %ymm9,%ymm6
12608 .byte 197,124,41,199 // vmovaps %ymm8,%ymm7
12609 .byte 255,224 // jmpq *%rax
12610
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12611HIDDEN _sk_move_src_dst_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12612.globl _sk_move_src_dst_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12613FUNCTION(_sk_move_src_dst_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12614_sk_move_src_dst_avx:
12615 .byte 72,173 // lods %ds:(%rsi),%rax
12616 .byte 197,252,40,224 // vmovaps %ymm0,%ymm4
12617 .byte 197,252,40,233 // vmovaps %ymm1,%ymm5
12618 .byte 197,252,40,242 // vmovaps %ymm2,%ymm6
12619 .byte 197,252,40,251 // vmovaps %ymm3,%ymm7
12620 .byte 255,224 // jmpq *%rax
12621
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12622HIDDEN _sk_move_dst_src_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12623.globl _sk_move_dst_src_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12624FUNCTION(_sk_move_dst_src_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12625_sk_move_dst_src_avx:
12626 .byte 72,173 // lods %ds:(%rsi),%rax
12627 .byte 197,252,40,196 // vmovaps %ymm4,%ymm0
12628 .byte 197,252,40,205 // vmovaps %ymm5,%ymm1
12629 .byte 197,252,40,214 // vmovaps %ymm6,%ymm2
12630 .byte 197,252,40,223 // vmovaps %ymm7,%ymm3
12631 .byte 255,224 // jmpq *%rax
12632
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12633HIDDEN _sk_premul_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12634.globl _sk_premul_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12635FUNCTION(_sk_premul_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12636_sk_premul_avx:
12637 .byte 197,252,89,195 // vmulps %ymm3,%ymm0,%ymm0
12638 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
12639 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
12640 .byte 72,173 // lods %ds:(%rsi),%rax
12641 .byte 255,224 // jmpq *%rax
12642
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12643HIDDEN _sk_unpremul_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12644.globl _sk_unpremul_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12645FUNCTION(_sk_unpremul_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12646_sk_unpremul_avx:
12647 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
12648 .byte 196,65,100,194,200,0 // vcmpeqps %ymm8,%ymm3,%ymm9
12649 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12650 .byte 197,121,110,208 // vmovd %eax,%xmm10
12651 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
12652 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
12653 .byte 197,44,94,211 // vdivps %ymm3,%ymm10,%ymm10
12654 .byte 196,67,45,74,192,144 // vblendvps %ymm9,%ymm8,%ymm10,%ymm8
12655 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
12656 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
12657 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
12658 .byte 72,173 // lods %ds:(%rsi),%rax
12659 .byte 255,224 // jmpq *%rax
12660
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12661HIDDEN _sk_from_srgb_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12662.globl _sk_from_srgb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12663FUNCTION(_sk_from_srgb_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12664_sk_from_srgb_avx:
12665 .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax
12666 .byte 197,121,110,192 // vmovd %eax,%xmm8
12667 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12668 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12669 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
12670 .byte 197,124,89,208 // vmulps %ymm0,%ymm0,%ymm10
12671 .byte 184,154,153,153,62 // mov $0x3e99999a,%eax
12672 .byte 197,121,110,216 // vmovd %eax,%xmm11
12673 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
12674 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
12675 .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax
12676 .byte 197,121,110,224 // vmovd %eax,%xmm12
12677 .byte 196,67,121,4,228,0 // vpermilps $0x0,%xmm12,%xmm12
12678 .byte 196,67,29,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm12,%ymm12
12679 .byte 197,36,89,232 // vmulps %ymm0,%ymm11,%ymm13
12680 .byte 196,65,20,88,236 // vaddps %ymm12,%ymm13,%ymm13
12681 .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax
12682 .byte 197,121,110,240 // vmovd %eax,%xmm14
12683 .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14
12684 .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14
12685 .byte 196,65,44,89,213 // vmulps %ymm13,%ymm10,%ymm10
12686 .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10
12687 .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax
12688 .byte 197,121,110,232 // vmovd %eax,%xmm13
12689 .byte 196,67,121,4,237,0 // vpermilps $0x0,%xmm13,%xmm13
12690 .byte 196,67,21,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm13,%ymm13
12691 .byte 196,193,124,194,197,1 // vcmpltps %ymm13,%ymm0,%ymm0
12692 .byte 196,195,45,74,193,0 // vblendvps %ymm0,%ymm9,%ymm10,%ymm0
12693 .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9
12694 .byte 197,116,89,209 // vmulps %ymm1,%ymm1,%ymm10
12695 .byte 197,36,89,249 // vmulps %ymm1,%ymm11,%ymm15
12696 .byte 196,65,28,88,255 // vaddps %ymm15,%ymm12,%ymm15
12697 .byte 196,65,44,89,215 // vmulps %ymm15,%ymm10,%ymm10
12698 .byte 196,65,12,88,210 // vaddps %ymm10,%ymm14,%ymm10
12699 .byte 196,193,116,194,205,1 // vcmpltps %ymm13,%ymm1,%ymm1
12700 .byte 196,195,45,74,201,16 // vblendvps %ymm1,%ymm9,%ymm10,%ymm1
12701 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
12702 .byte 197,108,89,202 // vmulps %ymm2,%ymm2,%ymm9
12703 .byte 197,36,89,210 // vmulps %ymm2,%ymm11,%ymm10
12704 .byte 196,65,28,88,210 // vaddps %ymm10,%ymm12,%ymm10
12705 .byte 196,65,52,89,202 // vmulps %ymm10,%ymm9,%ymm9
12706 .byte 196,65,12,88,201 // vaddps %ymm9,%ymm14,%ymm9
12707 .byte 196,193,108,194,213,1 // vcmpltps %ymm13,%ymm2,%ymm2
12708 .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2
12709 .byte 72,173 // lods %ds:(%rsi),%rax
12710 .byte 255,224 // jmpq *%rax
12711
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12712HIDDEN _sk_to_srgb_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12713.globl _sk_to_srgb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12714FUNCTION(_sk_to_srgb_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12715_sk_to_srgb_avx:
12716 .byte 197,124,82,192 // vrsqrtps %ymm0,%ymm8
12717 .byte 196,65,124,83,232 // vrcpps %ymm8,%ymm13
12718 .byte 196,65,124,82,240 // vrsqrtps %ymm8,%ymm14
12719 .byte 184,41,92,71,65 // mov $0x41475c29,%eax
12720 .byte 197,121,110,192 // vmovd %eax,%xmm8
12721 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
12722 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
12723 .byte 197,60,89,224 // vmulps %ymm0,%ymm8,%ymm12
12724 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12725 .byte 197,121,110,200 // vmovd %eax,%xmm9
12726 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
12727 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
12728 .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax
12729 .byte 197,121,110,208 // vmovd %eax,%xmm10
12730 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
12731 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
12732 .byte 184,206,111,48,63 // mov $0x3f306fce,%eax
12733 .byte 197,121,110,216 // vmovd %eax,%xmm11
12734 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
12735 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
12736 .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax
12737 .byte 53,0,0,0,128 // xor $0x80000000,%eax
12738 .byte 197,121,110,248 // vmovd %eax,%xmm15
12739 .byte 196,67,121,4,255,0 // vpermilps $0x0,%xmm15,%xmm15
12740 .byte 196,67,5,24,255,1 // vinsertf128 $0x1,%xmm15,%ymm15,%ymm15
12741 .byte 196,65,20,89,235 // vmulps %ymm11,%ymm13,%ymm13
12742 .byte 196,65,20,88,239 // vaddps %ymm15,%ymm13,%ymm13
12743 .byte 196,65,12,89,242 // vmulps %ymm10,%ymm14,%ymm14
12744 .byte 196,65,12,88,237 // vaddps %ymm13,%ymm14,%ymm13
12745 .byte 196,65,52,93,237 // vminps %ymm13,%ymm9,%ymm13
12746 .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax
12747 .byte 197,121,110,240 // vmovd %eax,%xmm14
12748 .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14
12749 .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14
12750 .byte 196,193,124,194,198,1 // vcmpltps %ymm14,%ymm0,%ymm0
12751 .byte 196,195,21,74,196,0 // vblendvps %ymm0,%ymm12,%ymm13,%ymm0
12752 .byte 197,124,82,225 // vrsqrtps %ymm1,%ymm12
12753 .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13
12754 .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12
12755 .byte 196,65,36,89,237 // vmulps %ymm13,%ymm11,%ymm13
12756 .byte 196,65,4,88,237 // vaddps %ymm13,%ymm15,%ymm13
12757 .byte 196,65,44,89,228 // vmulps %ymm12,%ymm10,%ymm12
12758 .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
12759 .byte 197,60,89,233 // vmulps %ymm1,%ymm8,%ymm13
12760 .byte 196,65,52,93,228 // vminps %ymm12,%ymm9,%ymm12
12761 .byte 196,193,116,194,206,1 // vcmpltps %ymm14,%ymm1,%ymm1
12762 .byte 196,195,29,74,205,16 // vblendvps %ymm1,%ymm13,%ymm12,%ymm1
12763 .byte 197,124,82,226 // vrsqrtps %ymm2,%ymm12
12764 .byte 196,65,124,83,236 // vrcpps %ymm12,%ymm13
12765 .byte 196,65,36,89,221 // vmulps %ymm13,%ymm11,%ymm11
12766 .byte 196,65,4,88,219 // vaddps %ymm11,%ymm15,%ymm11
12767 .byte 196,65,124,82,228 // vrsqrtps %ymm12,%ymm12
12768 .byte 196,65,44,89,212 // vmulps %ymm12,%ymm10,%ymm10
12769 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
12770 .byte 196,65,52,93,202 // vminps %ymm10,%ymm9,%ymm9
12771 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
12772 .byte 196,193,108,194,214,1 // vcmpltps %ymm14,%ymm2,%ymm2
12773 .byte 196,195,53,74,208,32 // vblendvps %ymm2,%ymm8,%ymm9,%ymm2
12774 .byte 72,173 // lods %ds:(%rsi),%rax
12775 .byte 255,224 // jmpq *%rax
12776
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]12777HIDDEN _sk_rgb_to_hsl_avx
12778.globl _sk_rgb_to_hsl_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12779FUNCTION(_sk_rgb_to_hsl_avx)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]12780_sk_rgb_to_hsl_avx:
12781 .byte 197,124,95,193 // vmaxps %ymm1,%ymm0,%ymm8
12782 .byte 197,60,95,194 // vmaxps %ymm2,%ymm8,%ymm8
12783 .byte 197,124,93,201 // vminps %ymm1,%ymm0,%ymm9
12784 .byte 197,52,93,202 // vminps %ymm2,%ymm9,%ymm9
12785 .byte 196,65,60,92,209 // vsubps %ymm9,%ymm8,%ymm10
12786 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12787 .byte 197,121,110,216 // vmovd %eax,%xmm11
12788 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
12789 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
12790 .byte 196,65,36,94,218 // vdivps %ymm10,%ymm11,%ymm11
12791 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
12792 .byte 197,116,92,226 // vsubps %ymm2,%ymm1,%ymm12
12793 .byte 196,65,28,89,227 // vmulps %ymm11,%ymm12,%ymm12
12794 .byte 65,185,0,0,192,64 // mov $0x40c00000,%r9d
12795 .byte 197,108,92,232 // vsubps %ymm0,%ymm2,%ymm13
12796 .byte 196,65,20,89,235 // vmulps %ymm11,%ymm13,%ymm13
12797 .byte 65,186,0,0,0,64 // mov $0x40000000,%r10d
12798 .byte 197,124,92,241 // vsubps %ymm1,%ymm0,%ymm14
12799 .byte 196,65,12,89,219 // vmulps %ymm11,%ymm14,%ymm11
12800 .byte 184,0,0,128,64 // mov $0x40800000,%eax
12801 .byte 197,121,110,240 // vmovd %eax,%xmm14
12802 .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14
12803 .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14
12804 .byte 196,65,36,88,222 // vaddps %ymm14,%ymm11,%ymm11
12805 .byte 196,65,121,110,242 // vmovd %r10d,%xmm14
12806 .byte 197,244,194,210,1 // vcmpltps %ymm2,%ymm1,%ymm2
12807 .byte 197,188,194,201,0 // vcmpeqps %ymm1,%ymm8,%ymm1
12808 .byte 196,67,121,4,246,0 // vpermilps $0x0,%xmm14,%xmm14
12809 .byte 196,67,13,24,246,1 // vinsertf128 $0x1,%xmm14,%ymm14,%ymm14
12810 .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
12811 .byte 196,67,37,74,221,16 // vblendvps %ymm1,%ymm13,%ymm11,%ymm11
12812 .byte 196,193,121,110,201 // vmovd %r9d,%xmm1
12813 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
12814 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
12815 .byte 196,65,20,87,237 // vxorps %ymm13,%ymm13,%ymm13
12816 .byte 196,227,21,74,201,32 // vblendvps %ymm2,%ymm1,%ymm13,%ymm1
12817 .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1
12818 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
12819 .byte 197,249,110,208 // vmovd %eax,%xmm2
12820 .byte 197,188,194,192,0 // vcmpeqps %ymm0,%ymm8,%ymm0
12821 .byte 196,227,37,74,193,0 // vblendvps %ymm0,%ymm1,%ymm11,%ymm0
12822 .byte 196,193,60,88,201 // vaddps %ymm9,%ymm8,%ymm1
12823 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
12824 .byte 196,99,109,24,218,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm11
12825 .byte 196,193,116,89,211 // vmulps %ymm11,%ymm1,%ymm2
12826 .byte 197,36,194,218,1 // vcmpltps %ymm2,%ymm11,%ymm11
12827 .byte 196,65,12,92,224 // vsubps %ymm8,%ymm14,%ymm12
12828 .byte 196,65,28,92,225 // vsubps %ymm9,%ymm12,%ymm12
12829 .byte 196,195,117,74,204,176 // vblendvps %ymm11,%ymm12,%ymm1,%ymm1
12830 .byte 196,65,60,194,193,0 // vcmpeqps %ymm9,%ymm8,%ymm8
12831 .byte 197,172,94,201 // vdivps %ymm1,%ymm10,%ymm1
12832 .byte 196,67,125,74,205,128 // vblendvps %ymm8,%ymm13,%ymm0,%ymm9
12833 .byte 196,195,117,74,205,128 // vblendvps %ymm8,%ymm13,%ymm1,%ymm1
12834 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
12835 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12836 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
12837 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
12838 .byte 72,173 // lods %ds:(%rsi),%rax
12839 .byte 255,224 // jmpq *%rax
12840
12841HIDDEN _sk_hsl_to_rgb_avx
12842.globl _sk_hsl_to_rgb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12843FUNCTION(_sk_hsl_to_rgb_avx)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]12844_sk_hsl_to_rgb_avx:
12845 .byte 72,131,236,56 // sub $0x38,%rsp
12846 .byte 197,252,17,60,36 // vmovups %ymm7,(%rsp)
12847 .byte 197,252,17,116,36,224 // vmovups %ymm6,-0x20(%rsp)
12848 .byte 197,252,17,108,36,192 // vmovups %ymm5,-0x40(%rsp)
12849 .byte 197,252,17,100,36,160 // vmovups %ymm4,-0x60(%rsp)
12850 .byte 197,252,17,92,36,128 // vmovups %ymm3,-0x80(%rsp)
12851 .byte 197,252,40,226 // vmovaps %ymm2,%ymm4
12852 .byte 197,252,40,233 // vmovaps %ymm1,%ymm5
12853 .byte 197,252,40,216 // vmovaps %ymm0,%ymm3
12854 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
12855 .byte 197,249,110,192 // vmovd %eax,%xmm0
12856 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12857 .byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8
12858 .byte 196,193,92,194,192,1 // vcmpltps %ymm8,%ymm4,%ymm0
12859 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
12860 .byte 197,249,110,208 // vmovd %eax,%xmm2
12861 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
12862 .byte 196,99,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm10
12863 .byte 197,172,88,213 // vaddps %ymm5,%ymm10,%ymm2
12864 .byte 197,236,89,212 // vmulps %ymm4,%ymm2,%ymm2
12865 .byte 197,84,88,204 // vaddps %ymm4,%ymm5,%ymm9
12866 .byte 197,84,89,220 // vmulps %ymm4,%ymm5,%ymm11
12867 .byte 196,65,52,92,203 // vsubps %ymm11,%ymm9,%ymm9
12868 .byte 196,99,53,74,202,0 // vblendvps %ymm0,%ymm2,%ymm9,%ymm9
12869 .byte 65,184,0,0,0,64 // mov $0x40000000,%r8d
12870 .byte 184,171,170,170,62 // mov $0x3eaaaaab,%eax
12871 .byte 197,249,110,192 // vmovd %eax,%xmm0
12872 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12873 .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12
12874 .byte 197,28,88,251 // vaddps %ymm3,%ymm12,%ymm15
12875 .byte 184,0,0,0,0 // mov $0x0,%eax
12876 .byte 197,249,110,192 // vmovd %eax,%xmm0
12877 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12878 .byte 196,99,125,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm13
12879 .byte 196,193,44,194,199,1 // vcmpltps %ymm15,%ymm10,%ymm0
12880 .byte 196,193,4,92,210 // vsubps %ymm10,%ymm15,%ymm2
12881 .byte 196,227,5,74,194,0 // vblendvps %ymm0,%ymm2,%ymm15,%ymm0
12882 .byte 196,193,4,194,213,1 // vcmpltps %ymm13,%ymm15,%ymm2
12883 .byte 196,65,44,88,223 // vaddps %ymm15,%ymm10,%ymm11
12884 .byte 196,195,125,74,203,32 // vblendvps %ymm2,%ymm11,%ymm0,%ymm1
12885 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
12886 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12887 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
12888 .byte 197,252,89,196 // vmulps %ymm4,%ymm0,%ymm0
12889 .byte 196,65,124,92,217 // vsubps %ymm9,%ymm0,%ymm11
12890 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
12891 .byte 184,0,0,192,64 // mov $0x40c00000,%eax
12892 .byte 197,249,110,192 // vmovd %eax,%xmm0
12893 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12894 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
12895 .byte 196,193,52,92,211 // vsubps %ymm11,%ymm9,%ymm2
12896 .byte 197,108,89,240 // vmulps %ymm0,%ymm2,%ymm14
12897 .byte 184,171,170,42,63 // mov $0x3f2aaaab,%eax
12898 .byte 197,249,110,192 // vmovd %eax,%xmm0
12899 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
12900 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
12901 .byte 197,252,92,209 // vsubps %ymm1,%ymm0,%ymm2
12902 .byte 197,140,89,210 // vmulps %ymm2,%ymm14,%ymm2
12903 .byte 197,164,88,210 // vaddps %ymm2,%ymm11,%ymm2
12904 .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7
12905 .byte 196,227,37,74,210,112 // vblendvps %ymm7,%ymm2,%ymm11,%ymm2
12906 .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7
12907 .byte 196,195,109,74,249,112 // vblendvps %ymm7,%ymm9,%ymm2,%ymm7
12908 .byte 196,193,121,110,208 // vmovd %r8d,%xmm2
12909 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
12910 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
12911 .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1
12912 .byte 196,65,4,89,254 // vmulps %ymm14,%ymm15,%ymm15
12913 .byte 196,65,36,88,255 // vaddps %ymm15,%ymm11,%ymm15
12914 .byte 196,67,69,74,255,16 // vblendvps %ymm1,%ymm15,%ymm7,%ymm15
12915 .byte 197,172,194,203,1 // vcmpltps %ymm3,%ymm10,%ymm1
12916 .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7
12917 .byte 196,227,101,74,207,16 // vblendvps %ymm1,%ymm7,%ymm3,%ymm1
12918 .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7
12919 .byte 197,172,88,243 // vaddps %ymm3,%ymm10,%ymm6
12920 .byte 196,227,117,74,206,112 // vblendvps %ymm7,%ymm6,%ymm1,%ymm1
12921 .byte 197,252,92,241 // vsubps %ymm1,%ymm0,%ymm6
12922 .byte 197,140,89,246 // vmulps %ymm6,%ymm14,%ymm6
12923 .byte 197,164,88,246 // vaddps %ymm6,%ymm11,%ymm6
12924 .byte 197,244,194,248,1 // vcmpltps %ymm0,%ymm1,%ymm7
12925 .byte 196,227,37,74,246,112 // vblendvps %ymm7,%ymm6,%ymm11,%ymm6
12926 .byte 196,193,116,194,248,1 // vcmpltps %ymm8,%ymm1,%ymm7
12927 .byte 196,195,77,74,241,112 // vblendvps %ymm7,%ymm9,%ymm6,%ymm6
12928 .byte 197,244,194,202,1 // vcmpltps %ymm2,%ymm1,%ymm1
12929 .byte 197,140,89,251 // vmulps %ymm3,%ymm14,%ymm7
12930 .byte 197,164,88,255 // vaddps %ymm7,%ymm11,%ymm7
12931 .byte 196,227,77,74,207,16 // vblendvps %ymm1,%ymm7,%ymm6,%ymm1
12932 .byte 196,193,100,92,220 // vsubps %ymm12,%ymm3,%ymm3
12933 .byte 197,172,194,243,1 // vcmpltps %ymm3,%ymm10,%ymm6
12934 .byte 196,193,100,92,250 // vsubps %ymm10,%ymm3,%ymm7
12935 .byte 196,227,101,74,247,96 // vblendvps %ymm6,%ymm7,%ymm3,%ymm6
12936 .byte 196,193,100,194,253,1 // vcmpltps %ymm13,%ymm3,%ymm7
12937 .byte 197,44,88,211 // vaddps %ymm3,%ymm10,%ymm10
12938 .byte 196,195,77,74,242,112 // vblendvps %ymm7,%ymm10,%ymm6,%ymm6
12939 .byte 197,204,194,248,1 // vcmpltps %ymm0,%ymm6,%ymm7
12940 .byte 197,252,92,198 // vsubps %ymm6,%ymm0,%ymm0
12941 .byte 197,140,89,192 // vmulps %ymm0,%ymm14,%ymm0
12942 .byte 197,164,88,192 // vaddps %ymm0,%ymm11,%ymm0
12943 .byte 196,227,37,74,192,112 // vblendvps %ymm7,%ymm0,%ymm11,%ymm0
12944 .byte 196,193,76,194,248,1 // vcmpltps %ymm8,%ymm6,%ymm7
12945 .byte 196,195,125,74,193,112 // vblendvps %ymm7,%ymm9,%ymm0,%ymm0
12946 .byte 197,204,194,210,1 // vcmpltps %ymm2,%ymm6,%ymm2
12947 .byte 196,193,100,89,222 // vmulps %ymm14,%ymm3,%ymm3
12948 .byte 197,164,88,219 // vaddps %ymm3,%ymm11,%ymm3
12949 .byte 196,227,125,74,211,32 // vblendvps %ymm2,%ymm3,%ymm0,%ymm2
12950 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
12951 .byte 197,212,194,216,0 // vcmpeqps %ymm0,%ymm5,%ymm3
12952 .byte 196,227,5,74,196,48 // vblendvps %ymm3,%ymm4,%ymm15,%ymm0
12953 .byte 196,227,117,74,204,48 // vblendvps %ymm3,%ymm4,%ymm1,%ymm1
12954 .byte 196,227,109,74,212,48 // vblendvps %ymm3,%ymm4,%ymm2,%ymm2
12955 .byte 72,173 // lods %ds:(%rsi),%rax
12956 .byte 197,252,16,92,36,128 // vmovups -0x80(%rsp),%ymm3
12957 .byte 197,252,16,100,36,160 // vmovups -0x60(%rsp),%ymm4
12958 .byte 197,252,16,108,36,192 // vmovups -0x40(%rsp),%ymm5
12959 .byte 197,252,16,116,36,224 // vmovups -0x20(%rsp),%ymm6
12960 .byte 197,252,16,60,36 // vmovups (%rsp),%ymm7
12961 .byte 72,131,196,56 // add $0x38,%rsp
12962 .byte 255,224 // jmpq *%rax
12963
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12964HIDDEN _sk_scale_1_float_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12965.globl _sk_scale_1_float_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12966FUNCTION(_sk_scale_1_float_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12967_sk_scale_1_float_avx:
12968 .byte 72,173 // lods %ds:(%rsi),%rax
12969 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
12970 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
12971 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
12972 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
12973 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
12974 .byte 72,173 // lods %ds:(%rsi),%rax
12975 .byte 255,224 // jmpq *%rax
12976
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]12977HIDDEN _sk_scale_u8_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12978.globl _sk_scale_u8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]12979FUNCTION(_sk_scale_u8_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12980_sk_scale_u8_avx:
12981 .byte 73,137,200 // mov %rcx,%r8
12982 .byte 72,173 // lods %ds:(%rsi),%rax
12983 .byte 72,139,0 // mov (%rax),%rax
12984 .byte 72,1,248 // add %rdi,%rax
12985 .byte 77,133,192 // test %r8,%r8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]12986 .byte 117,80 // jne 12e6 <_sk_scale_u8_avx+0x60>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]12987 .byte 197,122,126,0 // vmovq (%rax),%xmm8
12988 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
12989 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
12990 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
12991 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
12992 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
12993 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
12994 .byte 197,121,110,200 // vmovd %eax,%xmm9
12995 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
12996 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
12997 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
12998 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
12999 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
13000 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
13001 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
13002 .byte 72,173 // lods %ds:(%rsi),%rax
13003 .byte 76,137,193 // mov %r8,%rcx
13004 .byte 255,224 // jmpq *%rax
13005 .byte 49,201 // xor %ecx,%ecx
13006 .byte 77,137,194 // mov %r8,%r10
13007 .byte 69,49,201 // xor %r9d,%r9d
13008 .byte 68,15,182,24 // movzbl (%rax),%r11d
13009 .byte 72,255,192 // inc %rax
13010 .byte 73,211,227 // shl %cl,%r11
13011 .byte 77,9,217 // or %r11,%r9
13012 .byte 72,131,193,8 // add $0x8,%rcx
13013 .byte 73,255,202 // dec %r10
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13014 .byte 117,234 // jne 12ee <_sk_scale_u8_avx+0x68>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13015 .byte 196,65,249,110,193 // vmovq %r9,%xmm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13016 .byte 235,143 // jmp 129a <_sk_scale_u8_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13017
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]13018HIDDEN _sk_lerp_1_float_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13019.globl _sk_lerp_1_float_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13020FUNCTION(_sk_lerp_1_float_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13021_sk_lerp_1_float_avx:
13022 .byte 72,173 // lods %ds:(%rsi),%rax
13023 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
13024 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
13025 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
13026 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
13027 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
13028 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
13029 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
13030 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
13031 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
13032 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
13033 .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3
13034 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
13035 .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
13036 .byte 72,173 // lods %ds:(%rsi),%rax
13037 .byte 255,224 // jmpq *%rax
13038
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]13039HIDDEN _sk_lerp_u8_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13040.globl _sk_lerp_u8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13041FUNCTION(_sk_lerp_u8_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13042_sk_lerp_u8_avx:
13043 .byte 73,137,200 // mov %rcx,%r8
13044 .byte 72,173 // lods %ds:(%rsi),%rax
13045 .byte 72,139,0 // mov (%rax),%rax
13046 .byte 72,1,248 // add %rdi,%rax
13047 .byte 77,133,192 // test %r8,%r8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13048 .byte 117,116 // jne 13ce <_sk_lerp_u8_avx+0x84>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13049 .byte 197,122,126,0 // vmovq (%rax),%xmm8
13050 .byte 196,66,121,49,200 // vpmovzxbd %xmm8,%xmm9
13051 .byte 196,67,121,4,192,229 // vpermilps $0xe5,%xmm8,%xmm8
13052 .byte 196,66,121,49,192 // vpmovzxbd %xmm8,%xmm8
13053 .byte 196,67,53,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm9,%ymm8
13054 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
13055 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
13056 .byte 197,121,110,200 // vmovd %eax,%xmm9
13057 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
13058 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
13059 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
13060 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
13061 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
13062 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
13063 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
13064 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
13065 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
13066 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
13067 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
13068 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
13069 .byte 197,228,92,223 // vsubps %ymm7,%ymm3,%ymm3
13070 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
13071 .byte 197,228,88,223 // vaddps %ymm7,%ymm3,%ymm3
13072 .byte 72,173 // lods %ds:(%rsi),%rax
13073 .byte 76,137,193 // mov %r8,%rcx
13074 .byte 255,224 // jmpq *%rax
13075 .byte 49,201 // xor %ecx,%ecx
13076 .byte 77,137,194 // mov %r8,%r10
13077 .byte 69,49,201 // xor %r9d,%r9d
13078 .byte 68,15,182,24 // movzbl (%rax),%r11d
13079 .byte 72,255,192 // inc %rax
13080 .byte 73,211,227 // shl %cl,%r11
13081 .byte 77,9,217 // or %r11,%r9
13082 .byte 72,131,193,8 // add $0x8,%rcx
13083 .byte 73,255,202 // dec %r10
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13084 .byte 117,234 // jne 13d6 <_sk_lerp_u8_avx+0x8c>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13085 .byte 196,65,249,110,193 // vmovq %r9,%xmm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13086 .byte 233,104,255,255,255 // jmpq 135e <_sk_lerp_u8_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13087
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]13088HIDDEN _sk_lerp_565_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13089.globl _sk_lerp_565_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13090FUNCTION(_sk_lerp_565_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13091_sk_lerp_565_avx:
13092 .byte 72,173 // lods %ds:(%rsi),%rax
13093 .byte 76,139,16 // mov (%rax),%r10
13094 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13095 .byte 15,133,250,0,0,0 // jne 14fe <_sk_lerp_565_avx+0x108>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13096 .byte 196,65,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm8
13097 .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
13098 .byte 197,185,105,219 // vpunpckhwd %xmm3,%xmm8,%xmm3
13099 .byte 196,66,121,51,192 // vpmovzxwd %xmm8,%xmm8
13100 .byte 196,99,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm8
13101 .byte 184,0,248,0,0 // mov $0xf800,%eax
13102 .byte 197,249,110,216 // vmovd %eax,%xmm3
13103 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
13104 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13105 .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
13106 .byte 197,124,91,203 // vcvtdq2ps %ymm3,%ymm9
13107 .byte 184,8,33,132,55 // mov $0x37842108,%eax
13108 .byte 197,249,110,216 // vmovd %eax,%xmm3
13109 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13110 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13111 .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9
13112 .byte 184,224,7,0,0 // mov $0x7e0,%eax
13113 .byte 197,249,110,216 // vmovd %eax,%xmm3
13114 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
13115 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13116 .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
13117 .byte 197,124,91,211 // vcvtdq2ps %ymm3,%ymm10
13118 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
13119 .byte 197,249,110,216 // vmovd %eax,%xmm3
13120 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13121 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13122 .byte 197,44,89,211 // vmulps %ymm3,%ymm10,%ymm10
13123 .byte 184,31,0,0,0 // mov $0x1f,%eax
13124 .byte 197,249,110,216 // vmovd %eax,%xmm3
13125 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
13126 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13127 .byte 196,193,100,84,216 // vandps %ymm8,%ymm3,%ymm3
13128 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
13129 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
13130 .byte 197,249,110,216 // vmovd %eax,%xmm3
13131 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13132 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13133 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
13134 .byte 197,252,92,196 // vsubps %ymm4,%ymm0,%ymm0
13135 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
13136 .byte 197,252,88,196 // vaddps %ymm4,%ymm0,%ymm0
13137 .byte 197,244,92,205 // vsubps %ymm5,%ymm1,%ymm1
13138 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1
13139 .byte 197,244,88,205 // vaddps %ymm5,%ymm1,%ymm1
13140 .byte 197,236,92,214 // vsubps %ymm6,%ymm2,%ymm2
13141 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
13142 .byte 197,236,88,214 // vaddps %ymm6,%ymm2,%ymm2
13143 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
13144 .byte 197,249,110,216 // vmovd %eax,%xmm3
13145 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13146 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13147 .byte 72,173 // lods %ds:(%rsi),%rax
13148 .byte 255,224 // jmpq *%rax
13149 .byte 65,137,200 // mov %ecx,%r8d
13150 .byte 65,128,224,7 // and $0x7,%r8b
13151 .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
13152 .byte 65,254,200 // dec %r8b
13153 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13154 .byte 15,135,243,254,255,255 // ja 140a <_sk_lerp_565_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13155 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13156 .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 156c <_sk_lerp_565_avx+0x176>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13157 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
13158 .byte 76,1,200 // add %r9,%rax
13159 .byte 255,224 // jmpq *%rax
13160 .byte 197,225,239,219 // vpxor %xmm3,%xmm3,%xmm3
13161 .byte 196,65,97,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm3,%xmm8
13162 .byte 196,65,57,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm8,%xmm8
13163 .byte 196,65,57,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm8,%xmm8
13164 .byte 196,65,57,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm8,%xmm8
13165 .byte 196,65,57,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm8,%xmm8
13166 .byte 196,65,57,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm8,%xmm8
13167 .byte 196,65,57,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm8,%xmm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13168 .byte 233,159,254,255,255 // jmpq 140a <_sk_lerp_565_avx+0x14>
13169 .byte 144 // nop
13170 .byte 243,255 // repz (bad)
13171 .byte 255 // (bad)
13172 .byte 255 // (bad)
13173 .byte 235,255 // jmp 1571 <_sk_lerp_565_avx+0x17b>
13174 .byte 255 // (bad)
13175 .byte 255,227 // jmpq *%rbx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13176 .byte 255 // (bad)
13177 .byte 255 // (bad)
13178 .byte 255 // (bad)
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13179 .byte 219,255 // (bad)
13180 .byte 255 // (bad)
13181 .byte 255,211 // callq *%rbx
13182 .byte 255 // (bad)
13183 .byte 255 // (bad)
13184 .byte 255,203 // dec %ebx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13185 .byte 255 // (bad)
13186 .byte 255 // (bad)
13187 .byte 255 // (bad)
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13188 .byte 191 // .byte 0xbf
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13189 .byte 255 // (bad)
13190 .byte 255 // (bad)
13191 .byte 255 // .byte 0xff
13192
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]13193HIDDEN _sk_load_tables_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13194.globl _sk_load_tables_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13195FUNCTION(_sk_load_tables_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13196_sk_load_tables_avx:
13197 .byte 85 // push %rbp
13198 .byte 65,87 // push %r15
13199 .byte 65,86 // push %r14
13200 .byte 65,85 // push %r13
13201 .byte 65,84 // push %r12
13202 .byte 83 // push %rbx
13203 .byte 72,173 // lods %ds:(%rsi),%rax
13204 .byte 76,139,0 // mov (%rax),%r8
13205 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13206 .byte 15,133,56,2,0,0 // jne 17d8 <_sk_load_tables_avx+0x250>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13207 .byte 196,65,124,16,4,184 // vmovups (%r8,%rdi,4),%ymm8
13208 .byte 187,255,0,0,0 // mov $0xff,%ebx
13209 .byte 197,249,110,195 // vmovd %ebx,%xmm0
13210 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
13211 .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
13212 .byte 196,193,52,84,192 // vandps %ymm8,%ymm9,%ymm0
13213 .byte 196,193,249,126,193 // vmovq %xmm0,%r9
13214 .byte 69,137,203 // mov %r9d,%r11d
13215 .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10
13216 .byte 69,137,214 // mov %r10d,%r14d
13217 .byte 73,193,234,32 // shr $0x20,%r10
13218 .byte 73,193,233,32 // shr $0x20,%r9
13219 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
13220 .byte 196,193,249,126,196 // vmovq %xmm0,%r12
13221 .byte 69,137,231 // mov %r12d,%r15d
13222 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
13223 .byte 65,137,221 // mov %ebx,%r13d
13224 .byte 72,193,235,32 // shr $0x20,%rbx
13225 .byte 73,193,236,32 // shr $0x20,%r12
13226 .byte 72,139,104,8 // mov 0x8(%rax),%rbp
13227 .byte 76,139,64,16 // mov 0x10(%rax),%r8
13228 .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0
13229 .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
13230 .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1
13231 .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0
13232 .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1
13233 .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0
13234 .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1
13235 .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
13236 .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3
13237 .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1
13238 .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3
13239 .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1
13240 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
13241 .byte 196,193,113,114,208,8 // vpsrld $0x8,%xmm8,%xmm1
13242 .byte 196,67,125,25,194,1 // vextractf128 $0x1,%ymm8,%xmm10
13243 .byte 196,193,105,114,210,8 // vpsrld $0x8,%xmm10,%xmm2
13244 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
13245 .byte 197,180,84,201 // vandps %ymm1,%ymm9,%ymm1
13246 .byte 196,193,249,126,201 // vmovq %xmm1,%r9
13247 .byte 69,137,203 // mov %r9d,%r11d
13248 .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10
13249 .byte 69,137,214 // mov %r10d,%r14d
13250 .byte 73,193,234,32 // shr $0x20,%r10
13251 .byte 73,193,233,32 // shr $0x20,%r9
13252 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
13253 .byte 196,225,249,126,205 // vmovq %xmm1,%rbp
13254 .byte 65,137,239 // mov %ebp,%r15d
13255 .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx
13256 .byte 65,137,220 // mov %ebx,%r12d
13257 .byte 72,193,235,32 // shr $0x20,%rbx
13258 .byte 72,193,237,32 // shr $0x20,%rbp
13259 .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1
13260 .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1
13261 .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2
13262 .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1
13263 .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2
13264 .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1
13265 .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2
13266 .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2
13267 .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3
13268 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13269 .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3
13270 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13271 .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1
13272 .byte 72,139,64,24 // mov 0x18(%rax),%rax
13273 .byte 196,193,105,114,208,16 // vpsrld $0x10,%xmm8,%xmm2
13274 .byte 196,193,97,114,210,16 // vpsrld $0x10,%xmm10,%xmm3
13275 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
13276 .byte 197,180,84,210 // vandps %ymm2,%ymm9,%ymm2
13277 .byte 196,193,249,126,208 // vmovq %xmm2,%r8
13278 .byte 69,137,194 // mov %r8d,%r10d
13279 .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9
13280 .byte 69,137,203 // mov %r9d,%r11d
13281 .byte 73,193,233,32 // shr $0x20,%r9
13282 .byte 73,193,232,32 // shr $0x20,%r8
13283 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
13284 .byte 196,225,249,126,213 // vmovq %xmm2,%rbp
13285 .byte 65,137,238 // mov %ebp,%r14d
13286 .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx
13287 .byte 65,137,223 // mov %ebx,%r15d
13288 .byte 72,193,235,32 // shr $0x20,%rbx
13289 .byte 72,193,237,32 // shr $0x20,%rbp
13290 .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2
13291 .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2
13292 .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3
13293 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13294 .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3
13295 .byte 196,99,105,33,203,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm9
13296 .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3
13297 .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3
13298 .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2
13299 .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2
13300 .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3
13301 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13302 .byte 196,195,109,24,209,1 // vinsertf128 $0x1,%xmm9,%ymm2,%ymm2
13303 .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8
13304 .byte 196,193,97,114,210,24 // vpsrld $0x18,%xmm10,%xmm3
13305 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
13306 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
13307 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
13308 .byte 197,249,110,216 // vmovd %eax,%xmm3
13309 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13310 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13311 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
13312 .byte 72,173 // lods %ds:(%rsi),%rax
13313 .byte 91 // pop %rbx
13314 .byte 65,92 // pop %r12
13315 .byte 65,93 // pop %r13
13316 .byte 65,94 // pop %r14
13317 .byte 65,95 // pop %r15
13318 .byte 93 // pop %rbp
13319 .byte 255,224 // jmpq *%rax
13320 .byte 137,203 // mov %ecx,%ebx
13321 .byte 128,227,7 // and $0x7,%bl
13322 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
13323 .byte 254,203 // dec %bl
13324 .byte 128,251,6 // cmp $0x6,%bl
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13325 .byte 15,135,185,253,255,255 // ja 15a6 <_sk_load_tables_avx+0x1e>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13326 .byte 15,182,219 // movzbl %bl,%ebx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13327 .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 1880 <_sk_load_tables_avx+0x2f8>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13328 .byte 73,99,28,153 // movslq (%r9,%rbx,4),%rbx
13329 .byte 76,1,203 // add %r9,%rbx
13330 .byte 255,227 // jmpq *%rbx
13331 .byte 196,193,121,110,68,184,24 // vmovd 0x18(%r8,%rdi,4),%xmm0
13332 .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0
13333 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
13334 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
13335 .byte 196,99,117,12,192,64 // vblendps $0x40,%ymm0,%ymm1,%ymm8
13336 .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0
13337 .byte 196,195,121,34,68,184,20,1 // vpinsrd $0x1,0x14(%r8,%rdi,4),%xmm0,%xmm0
13338 .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8
13339 .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0
13340 .byte 196,195,121,34,68,184,16,0 // vpinsrd $0x0,0x10(%r8,%rdi,4),%xmm0,%xmm0
13341 .byte 196,99,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm8
13342 .byte 196,195,57,34,68,184,12,3 // vpinsrd $0x3,0xc(%r8,%rdi,4),%xmm8,%xmm0
13343 .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
13344 .byte 196,195,57,34,68,184,8,2 // vpinsrd $0x2,0x8(%r8,%rdi,4),%xmm8,%xmm0
13345 .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
13346 .byte 196,195,57,34,68,184,4,1 // vpinsrd $0x1,0x4(%r8,%rdi,4),%xmm8,%xmm0
13347 .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
13348 .byte 196,195,57,34,4,184,0 // vpinsrd $0x0,(%r8,%rdi,4),%xmm8,%xmm0
13349 .byte 196,99,61,12,192,15 // vblendps $0xf,%ymm0,%ymm8,%ymm8
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13350 .byte 233,38,253,255,255 // jmpq 15a6 <_sk_load_tables_avx+0x1e>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]13351 .byte 238 // out %al,(%dx)
13352 .byte 255 // (bad)
13353 .byte 255 // (bad)
13354 .byte 255,224 // jmpq *%rax
13355 .byte 255 // (bad)
13356 .byte 255 // (bad)
13357 .byte 255,210 // callq *%rdx
13358 .byte 255 // (bad)
13359 .byte 255 // (bad)
13360 .byte 255,196 // inc %esp
13361 .byte 255 // (bad)
13362 .byte 255 // (bad)
13363 .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax)
13364 .byte 255 // (bad)
13365 .byte 255 // (bad)
13366 .byte 255 // .byte 0xff
13367 .byte 128,255,255 // cmp $0xff,%bh
13368 .byte 255 // .byte 0xff
13369
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13370HIDDEN _sk_load_tables_u16_be_avx
13371.globl _sk_load_tables_u16_be_avx
13372FUNCTION(_sk_load_tables_u16_be_avx)
13373_sk_load_tables_u16_be_avx:
13374 .byte 72,173 // lods %ds:(%rsi),%rax
13375 .byte 76,139,0 // mov (%rax),%r8
13376 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
13377 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13378 .byte 15,133,165,2,0,0 // jne 1b57 <_sk_load_tables_u16_be_avx+0x2bb>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13379 .byte 196,1,121,16,4,72 // vmovupd (%r8,%r9,2),%xmm8
13380 .byte 196,129,121,16,84,72,16 // vmovupd 0x10(%r8,%r9,2),%xmm2
13381 .byte 196,129,121,16,92,72,32 // vmovupd 0x20(%r8,%r9,2),%xmm3
13382 .byte 196,1,122,111,76,72,48 // vmovdqu 0x30(%r8,%r9,2),%xmm9
13383 .byte 85 // push %rbp
13384 .byte 65,87 // push %r15
13385 .byte 65,86 // push %r14
13386 .byte 65,85 // push %r13
13387 .byte 65,84 // push %r12
13388 .byte 83 // push %rbx
13389 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
13390 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
13391 .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1
13392 .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3
13393 .byte 197,121,97,218 // vpunpcklwd %xmm2,%xmm0,%xmm11
13394 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
13395 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
13396 .byte 197,113,105,211 // vpunpckhwd %xmm3,%xmm1,%xmm10
13397 .byte 197,161,108,194 // vpunpcklqdq %xmm2,%xmm11,%xmm0
13398 .byte 196,65,57,239,192 // vpxor %xmm8,%xmm8,%xmm8
13399 .byte 196,193,121,105,200 // vpunpckhwd %xmm8,%xmm0,%xmm1
13400 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
13401 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
13402 .byte 187,255,0,0,0 // mov $0xff,%ebx
13403 .byte 197,249,110,203 // vmovd %ebx,%xmm1
13404 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
13405 .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12
13406 .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0
13407 .byte 196,193,249,126,193 // vmovq %xmm0,%r9
13408 .byte 69,137,203 // mov %r9d,%r11d
13409 .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10
13410 .byte 69,137,214 // mov %r10d,%r14d
13411 .byte 73,193,234,32 // shr $0x20,%r10
13412 .byte 73,193,233,32 // shr $0x20,%r9
13413 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
13414 .byte 196,193,249,126,196 // vmovq %xmm0,%r12
13415 .byte 69,137,231 // mov %r12d,%r15d
13416 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
13417 .byte 65,137,221 // mov %ebx,%r13d
13418 .byte 72,193,235,32 // shr $0x20,%rbx
13419 .byte 73,193,236,32 // shr $0x20,%r12
13420 .byte 72,139,104,8 // mov 0x8(%rax),%rbp
13421 .byte 76,139,64,16 // mov 0x10(%rax),%r8
13422 .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0
13423 .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
13424 .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1
13425 .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0
13426 .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1
13427 .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0
13428 .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1
13429 .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
13430 .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3
13431 .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1
13432 .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3
13433 .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1
13434 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
13435 .byte 197,161,109,202 // vpunpckhqdq %xmm2,%xmm11,%xmm1
13436 .byte 196,193,113,105,208 // vpunpckhwd %xmm8,%xmm1,%xmm2
13437 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
13438 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
13439 .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1
13440 .byte 196,193,249,126,201 // vmovq %xmm1,%r9
13441 .byte 69,137,203 // mov %r9d,%r11d
13442 .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10
13443 .byte 69,137,214 // mov %r10d,%r14d
13444 .byte 73,193,234,32 // shr $0x20,%r10
13445 .byte 73,193,233,32 // shr $0x20,%r9
13446 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
13447 .byte 196,225,249,126,205 // vmovq %xmm1,%rbp
13448 .byte 65,137,239 // mov %ebp,%r15d
13449 .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx
13450 .byte 65,137,220 // mov %ebx,%r12d
13451 .byte 72,193,235,32 // shr $0x20,%rbx
13452 .byte 72,193,237,32 // shr $0x20,%rbp
13453 .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1
13454 .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1
13455 .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2
13456 .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1
13457 .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2
13458 .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1
13459 .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2
13460 .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2
13461 .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3
13462 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13463 .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3
13464 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13465 .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1
13466 .byte 72,139,64,24 // mov 0x18(%rax),%rax
13467 .byte 196,193,49,108,210 // vpunpcklqdq %xmm10,%xmm9,%xmm2
13468 .byte 196,193,105,105,216 // vpunpckhwd %xmm8,%xmm2,%xmm3
13469 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
13470 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
13471 .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2
13472 .byte 196,193,249,126,208 // vmovq %xmm2,%r8
13473 .byte 69,137,194 // mov %r8d,%r10d
13474 .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9
13475 .byte 69,137,203 // mov %r9d,%r11d
13476 .byte 73,193,233,32 // shr $0x20,%r9
13477 .byte 73,193,232,32 // shr $0x20,%r8
13478 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
13479 .byte 196,225,249,126,213 // vmovq %xmm2,%rbp
13480 .byte 65,137,238 // mov %ebp,%r14d
13481 .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx
13482 .byte 65,137,223 // mov %ebx,%r15d
13483 .byte 72,193,235,32 // shr $0x20,%rbx
13484 .byte 72,193,237,32 // shr $0x20,%rbp
13485 .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2
13486 .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2
13487 .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3
13488 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13489 .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3
13490 .byte 196,99,105,33,219,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm11
13491 .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3
13492 .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3
13493 .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2
13494 .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2
13495 .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3
13496 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13497 .byte 196,195,109,24,211,1 // vinsertf128 $0x1,%xmm11,%ymm2,%ymm2
13498 .byte 184,128,0,128,55 // mov $0x37800080,%eax
13499 .byte 197,249,110,216 // vmovd %eax,%xmm3
13500 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13501 .byte 196,99,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm11
13502 .byte 196,193,49,109,218 // vpunpckhqdq %xmm10,%xmm9,%xmm3
13503 .byte 197,177,113,243,8 // vpsllw $0x8,%xmm3,%xmm9
13504 .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3
13505 .byte 197,177,235,219 // vpor %xmm3,%xmm9,%xmm3
13506 .byte 196,65,97,105,192 // vpunpckhwd %xmm8,%xmm3,%xmm8
13507 .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3
13508 .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3
13509 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
13510 .byte 196,193,100,89,219 // vmulps %ymm11,%ymm3,%ymm3
13511 .byte 72,173 // lods %ds:(%rsi),%rax
13512 .byte 91 // pop %rbx
13513 .byte 65,92 // pop %r12
13514 .byte 65,93 // pop %r13
13515 .byte 65,94 // pop %r14
13516 .byte 65,95 // pop %r15
13517 .byte 93 // pop %rbp
13518 .byte 255,224 // jmpq *%rax
13519 .byte 196,1,123,16,4,72 // vmovsd (%r8,%r9,2),%xmm8
13520 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
13521 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13522 .byte 116,85 // je 1bbd <_sk_load_tables_u16_be_avx+0x321>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13523 .byte 196,1,57,22,68,72,8 // vmovhpd 0x8(%r8,%r9,2),%xmm8,%xmm8
13524 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13525 .byte 114,72 // jb 1bbd <_sk_load_tables_u16_be_avx+0x321>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13526 .byte 196,129,123,16,84,72,16 // vmovsd 0x10(%r8,%r9,2),%xmm2
13527 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13528 .byte 116,72 // je 1bca <_sk_load_tables_u16_be_avx+0x32e>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13529 .byte 196,129,105,22,84,72,24 // vmovhpd 0x18(%r8,%r9,2),%xmm2,%xmm2
13530 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13531 .byte 114,59 // jb 1bca <_sk_load_tables_u16_be_avx+0x32e>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13532 .byte 196,129,123,16,92,72,32 // vmovsd 0x20(%r8,%r9,2),%xmm3
13533 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13534 .byte 15,132,45,253,255,255 // je 18cd <_sk_load_tables_u16_be_avx+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13535 .byte 196,129,97,22,92,72,40 // vmovhpd 0x28(%r8,%r9,2),%xmm3,%xmm3
13536 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13537 .byte 15,130,28,253,255,255 // jb 18cd <_sk_load_tables_u16_be_avx+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13538 .byte 196,1,122,126,76,72,48 // vmovq 0x30(%r8,%r9,2),%xmm9
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13539 .byte 233,16,253,255,255 // jmpq 18cd <_sk_load_tables_u16_be_avx+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13540 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
13541 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13542 .byte 233,3,253,255,255 // jmpq 18cd <_sk_load_tables_u16_be_avx+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13543 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13544 .byte 233,250,252,255,255 // jmpq 18cd <_sk_load_tables_u16_be_avx+0x31>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13545
13546HIDDEN _sk_load_tables_rgb_u16_be_avx
13547.globl _sk_load_tables_rgb_u16_be_avx
13548FUNCTION(_sk_load_tables_rgb_u16_be_avx)
13549_sk_load_tables_rgb_u16_be_avx:
13550 .byte 72,173 // lods %ds:(%rsi),%rax
13551 .byte 76,139,0 // mov (%rax),%r8
13552 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
13553 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13554 .byte 15,133,146,2,0,0 // jne 1e77 <_sk_load_tables_rgb_u16_be_avx+0x2a4>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13555 .byte 196,129,122,111,4,72 // vmovdqu (%r8,%r9,2),%xmm0
13556 .byte 196,129,122,111,84,72,12 // vmovdqu 0xc(%r8,%r9,2),%xmm2
13557 .byte 196,129,122,111,76,72,24 // vmovdqu 0x18(%r8,%r9,2),%xmm1
13558 .byte 196,129,122,111,92,72,32 // vmovdqu 0x20(%r8,%r9,2),%xmm3
13559 .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3
13560 .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8
13561 .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9
13562 .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11
13563 .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10
13564 .byte 85 // push %rbp
13565 .byte 65,87 // push %r15
13566 .byte 65,86 // push %r14
13567 .byte 65,85 // push %r13
13568 .byte 65,84 // push %r12
13569 .byte 83 // push %rbx
13570 .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0
13571 .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2
13572 .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1
13573 .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3
13574 .byte 197,121,97,218 // vpunpcklwd %xmm2,%xmm0,%xmm11
13575 .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8
13576 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
13577 .byte 197,113,105,211 // vpunpckhwd %xmm3,%xmm1,%xmm10
13578 .byte 197,161,108,194 // vpunpcklqdq %xmm2,%xmm11,%xmm0
13579 .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12
13580 .byte 196,193,121,105,204 // vpunpckhwd %xmm12,%xmm0,%xmm1
13581 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
13582 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
13583 .byte 187,255,0,0,0 // mov $0xff,%ebx
13584 .byte 197,249,110,203 // vmovd %ebx,%xmm1
13585 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
13586 .byte 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9
13587 .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0
13588 .byte 196,193,249,126,193 // vmovq %xmm0,%r9
13589 .byte 69,137,203 // mov %r9d,%r11d
13590 .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10
13591 .byte 69,137,214 // mov %r10d,%r14d
13592 .byte 73,193,234,32 // shr $0x20,%r10
13593 .byte 73,193,233,32 // shr $0x20,%r9
13594 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
13595 .byte 196,193,249,126,196 // vmovq %xmm0,%r12
13596 .byte 69,137,231 // mov %r12d,%r15d
13597 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
13598 .byte 65,137,221 // mov %ebx,%r13d
13599 .byte 72,193,235,32 // shr $0x20,%rbx
13600 .byte 73,193,236,32 // shr $0x20,%r12
13601 .byte 72,139,104,8 // mov 0x8(%rax),%rbp
13602 .byte 76,139,64,16 // mov 0x10(%rax),%r8
13603 .byte 196,161,122,16,68,189,0 // vmovss 0x0(%rbp,%r15,4),%xmm0
13604 .byte 196,163,121,33,68,165,0,16 // vinsertps $0x10,0x0(%rbp,%r12,4),%xmm0,%xmm0
13605 .byte 196,161,122,16,76,173,0 // vmovss 0x0(%rbp,%r13,4),%xmm1
13606 .byte 196,227,121,33,193,32 // vinsertps $0x20,%xmm1,%xmm0,%xmm0
13607 .byte 197,250,16,76,157,0 // vmovss 0x0(%rbp,%rbx,4),%xmm1
13608 .byte 196,227,121,33,193,48 // vinsertps $0x30,%xmm1,%xmm0,%xmm0
13609 .byte 196,161,122,16,76,157,0 // vmovss 0x0(%rbp,%r11,4),%xmm1
13610 .byte 196,163,113,33,76,141,0,16 // vinsertps $0x10,0x0(%rbp,%r9,4),%xmm1,%xmm1
13611 .byte 196,161,122,16,92,181,0 // vmovss 0x0(%rbp,%r14,4),%xmm3
13612 .byte 196,227,113,33,203,32 // vinsertps $0x20,%xmm3,%xmm1,%xmm1
13613 .byte 196,161,122,16,92,149,0 // vmovss 0x0(%rbp,%r10,4),%xmm3
13614 .byte 196,227,113,33,203,48 // vinsertps $0x30,%xmm3,%xmm1,%xmm1
13615 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
13616 .byte 197,161,109,202 // vpunpckhqdq %xmm2,%xmm11,%xmm1
13617 .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2
13618 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
13619 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
13620 .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1
13621 .byte 196,193,249,126,201 // vmovq %xmm1,%r9
13622 .byte 69,137,203 // mov %r9d,%r11d
13623 .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10
13624 .byte 69,137,214 // mov %r10d,%r14d
13625 .byte 73,193,234,32 // shr $0x20,%r10
13626 .byte 73,193,233,32 // shr $0x20,%r9
13627 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
13628 .byte 196,225,249,126,205 // vmovq %xmm1,%rbp
13629 .byte 65,137,239 // mov %ebp,%r15d
13630 .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx
13631 .byte 65,137,220 // mov %ebx,%r12d
13632 .byte 72,193,235,32 // shr $0x20,%rbx
13633 .byte 72,193,237,32 // shr $0x20,%rbp
13634 .byte 196,129,122,16,12,184 // vmovss (%r8,%r15,4),%xmm1
13635 .byte 196,195,113,33,12,168,16 // vinsertps $0x10,(%r8,%rbp,4),%xmm1,%xmm1
13636 .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2
13637 .byte 196,227,113,33,202,32 // vinsertps $0x20,%xmm2,%xmm1,%xmm1
13638 .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2
13639 .byte 196,227,113,33,202,48 // vinsertps $0x30,%xmm2,%xmm1,%xmm1
13640 .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2
13641 .byte 196,131,105,33,20,136,16 // vinsertps $0x10,(%r8,%r9,4),%xmm2,%xmm2
13642 .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3
13643 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13644 .byte 196,129,122,16,28,144 // vmovss (%r8,%r10,4),%xmm3
13645 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13646 .byte 196,227,109,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm2,%ymm1
13647 .byte 72,139,64,24 // mov 0x18(%rax),%rax
13648 .byte 196,193,57,108,210 // vpunpcklqdq %xmm10,%xmm8,%xmm2
13649 .byte 196,193,105,105,220 // vpunpckhwd %xmm12,%xmm2,%xmm3
13650 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
13651 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
13652 .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2
13653 .byte 196,193,249,126,208 // vmovq %xmm2,%r8
13654 .byte 69,137,194 // mov %r8d,%r10d
13655 .byte 196,195,249,22,209,1 // vpextrq $0x1,%xmm2,%r9
13656 .byte 69,137,203 // mov %r9d,%r11d
13657 .byte 73,193,233,32 // shr $0x20,%r9
13658 .byte 73,193,232,32 // shr $0x20,%r8
13659 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
13660 .byte 196,225,249,126,213 // vmovq %xmm2,%rbp
13661 .byte 65,137,238 // mov %ebp,%r14d
13662 .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx
13663 .byte 65,137,223 // mov %ebx,%r15d
13664 .byte 72,193,235,32 // shr $0x20,%rbx
13665 .byte 72,193,237,32 // shr $0x20,%rbp
13666 .byte 196,161,122,16,20,176 // vmovss (%rax,%r14,4),%xmm2
13667 .byte 196,227,105,33,20,168,16 // vinsertps $0x10,(%rax,%rbp,4),%xmm2,%xmm2
13668 .byte 196,161,122,16,28,184 // vmovss (%rax,%r15,4),%xmm3
13669 .byte 196,227,105,33,211,32 // vinsertps $0x20,%xmm3,%xmm2,%xmm2
13670 .byte 197,250,16,28,152 // vmovss (%rax,%rbx,4),%xmm3
13671 .byte 196,99,105,33,195,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm8
13672 .byte 196,161,122,16,28,144 // vmovss (%rax,%r10,4),%xmm3
13673 .byte 196,163,97,33,28,128,16 // vinsertps $0x10,(%rax,%r8,4),%xmm3,%xmm3
13674 .byte 196,161,122,16,20,152 // vmovss (%rax,%r11,4),%xmm2
13675 .byte 196,227,97,33,210,32 // vinsertps $0x20,%xmm2,%xmm3,%xmm2
13676 .byte 196,161,122,16,28,136 // vmovss (%rax,%r9,4),%xmm3
13677 .byte 196,227,105,33,211,48 // vinsertps $0x30,%xmm3,%xmm2,%xmm2
13678 .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2
13679 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
13680 .byte 197,249,110,216 // vmovd %eax,%xmm3
13681 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
13682 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
13683 .byte 72,173 // lods %ds:(%rsi),%rax
13684 .byte 91 // pop %rbx
13685 .byte 65,92 // pop %r12
13686 .byte 65,93 // pop %r13
13687 .byte 65,94 // pop %r14
13688 .byte 65,95 // pop %r15
13689 .byte 93 // pop %rbp
13690 .byte 255,224 // jmpq *%rax
13691 .byte 196,129,121,110,4,72 // vmovd (%r8,%r9,2),%xmm0
13692 .byte 196,129,121,196,68,72,4,2 // vpinsrw $0x2,0x4(%r8,%r9,2),%xmm0,%xmm0
13693 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13694 .byte 117,5 // jne 1e90 <_sk_load_tables_rgb_u16_be_avx+0x2bd>
13695 .byte 233,137,253,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13696 .byte 196,129,121,110,76,72,6 // vmovd 0x6(%r8,%r9,2),%xmm1
13697 .byte 196,1,113,196,68,72,10,2 // vpinsrw $0x2,0xa(%r8,%r9,2),%xmm1,%xmm8
13698 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13699 .byte 114,26 // jb 1ebf <_sk_load_tables_rgb_u16_be_avx+0x2ec>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13700 .byte 196,129,121,110,76,72,12 // vmovd 0xc(%r8,%r9,2),%xmm1
13701 .byte 196,129,113,196,84,72,16,2 // vpinsrw $0x2,0x10(%r8,%r9,2),%xmm1,%xmm2
13702 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13703 .byte 117,10 // jne 1ec4 <_sk_load_tables_rgb_u16_be_avx+0x2f1>
13704 .byte 233,90,253,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
13705 .byte 233,85,253,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13706 .byte 196,129,121,110,76,72,18 // vmovd 0x12(%r8,%r9,2),%xmm1
13707 .byte 196,1,113,196,76,72,22,2 // vpinsrw $0x2,0x16(%r8,%r9,2),%xmm1,%xmm9
13708 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13709 .byte 114,26 // jb 1ef3 <_sk_load_tables_rgb_u16_be_avx+0x320>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13710 .byte 196,129,121,110,76,72,24 // vmovd 0x18(%r8,%r9,2),%xmm1
13711 .byte 196,129,113,196,76,72,28,2 // vpinsrw $0x2,0x1c(%r8,%r9,2),%xmm1,%xmm1
13712 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13713 .byte 117,10 // jne 1ef8 <_sk_load_tables_rgb_u16_be_avx+0x325>
13714 .byte 233,38,253,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
13715 .byte 233,33,253,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13716 .byte 196,129,121,110,92,72,30 // vmovd 0x1e(%r8,%r9,2),%xmm3
13717 .byte 196,1,97,196,92,72,34,2 // vpinsrw $0x2,0x22(%r8,%r9,2),%xmm3,%xmm11
13718 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13719 .byte 114,20 // jb 1f21 <_sk_load_tables_rgb_u16_be_avx+0x34e>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13720 .byte 196,129,121,110,92,72,36 // vmovd 0x24(%r8,%r9,2),%xmm3
13721 .byte 196,129,97,196,92,72,40,2 // vpinsrw $0x2,0x28(%r8,%r9,2),%xmm3,%xmm3
Mike Kleinc7be0032017-04-25 15:51:23 -0400[diff] [blame]13722 .byte 233,248,252,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
13723 .byte 233,243,252,255,255 // jmpq 1c19 <_sk_load_tables_rgb_u16_be_avx+0x46>
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]13724
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]13725HIDDEN _sk_byte_tables_avx
13726.globl _sk_byte_tables_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13727FUNCTION(_sk_byte_tables_avx)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]13728_sk_byte_tables_avx:
13729 .byte 85 // push %rbp
13730 .byte 65,87 // push %r15
13731 .byte 65,86 // push %r14
13732 .byte 65,85 // push %r13
13733 .byte 65,84 // push %r12
13734 .byte 83 // push %rbx
13735 .byte 72,173 // lods %ds:(%rsi),%rax
13736 .byte 65,184,0,0,127,67 // mov $0x437f0000,%r8d
13737 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
13738 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
13739 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
13740 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
13741 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
13742 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
13743 .byte 68,137,197 // mov %r8d,%ebp
13744 .byte 77,137,194 // mov %r8,%r10
13745 .byte 73,193,234,32 // shr $0x20,%r10
13746 .byte 196,193,249,126,192 // vmovq %xmm0,%r8
13747 .byte 69,137,195 // mov %r8d,%r11d
13748 .byte 77,137,199 // mov %r8,%r15
13749 .byte 73,193,239,32 // shr $0x20,%r15
13750 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
13751 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
13752 .byte 69,137,198 // mov %r8d,%r14d
13753 .byte 77,137,196 // mov %r8,%r12
13754 .byte 73,193,236,32 // shr $0x20,%r12
13755 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
13756 .byte 65,137,221 // mov %ebx,%r13d
13757 .byte 72,193,235,32 // shr $0x20,%rbx
13758 .byte 76,139,8 // mov (%rax),%r9
13759 .byte 76,139,64,8 // mov 0x8(%rax),%r8
13760 .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
13761 .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
13762 .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp
13763 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
13764 .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp
13765 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
13766 .byte 196,98,121,49,200 // vpmovzxbd %xmm0,%xmm9
13767 .byte 196,131,121,32,4,41,0 // vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0
13768 .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0
13769 .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp
13770 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
13771 .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp
13772 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
13773 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
13774 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0
13775 .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10
13776 .byte 189,129,128,128,59 // mov $0x3b808081,%ebp
13777 .byte 197,249,110,197 // vmovd %ebp,%xmm0
13778 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
13779 .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
13780 .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0
13781 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
13782 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
13783 .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp
13784 .byte 65,137,233 // mov %ebp,%r9d
13785 .byte 72,193,237,32 // shr $0x20,%rbp
13786 .byte 196,225,249,126,203 // vmovq %xmm1,%rbx
13787 .byte 65,137,218 // mov %ebx,%r10d
13788 .byte 72,193,235,32 // shr $0x20,%rbx
13789 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
13790 .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11
13791 .byte 69,137,222 // mov %r11d,%r14d
13792 .byte 73,193,235,32 // shr $0x20,%r11
13793 .byte 196,193,249,126,207 // vmovq %xmm1,%r15
13794 .byte 69,137,252 // mov %r15d,%r12d
13795 .byte 73,193,239,32 // shr $0x20,%r15
13796 .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
13797 .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
13798 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
13799 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
13800 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
13801 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
13802 .byte 196,98,121,49,209 // vpmovzxbd %xmm1,%xmm10
13803 .byte 196,131,121,32,12,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1
13804 .byte 196,131,113,32,12,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1
13805 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
13806 .byte 196,227,113,32,205,2 // vpinsrb $0x2,%ebp,%xmm1,%xmm1
13807 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
13808 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
13809 .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1
13810 .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
13811 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
13812 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
13813 .byte 76,139,64,16 // mov 0x10(%rax),%r8
13814 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
13815 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
13816 .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp
13817 .byte 65,137,233 // mov %ebp,%r9d
13818 .byte 72,193,237,32 // shr $0x20,%rbp
13819 .byte 196,225,249,126,211 // vmovq %xmm2,%rbx
13820 .byte 65,137,218 // mov %ebx,%r10d
13821 .byte 72,193,235,32 // shr $0x20,%rbx
13822 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
13823 .byte 196,195,249,22,211,1 // vpextrq $0x1,%xmm2,%r11
13824 .byte 69,137,222 // mov %r11d,%r14d
13825 .byte 73,193,235,32 // shr $0x20,%r11
13826 .byte 196,193,249,126,215 // vmovq %xmm2,%r15
13827 .byte 69,137,252 // mov %r15d,%r12d
13828 .byte 73,193,239,32 // shr $0x20,%r15
13829 .byte 196,131,121,32,20,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm2
13830 .byte 196,195,105,32,20,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm2,%xmm2
13831 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
13832 .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2
13833 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
13834 .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2
13835 .byte 196,98,121,49,210 // vpmovzxbd %xmm2,%xmm10
13836 .byte 196,131,121,32,20,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm2
13837 .byte 196,131,105,32,20,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm2,%xmm2
13838 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
13839 .byte 196,227,105,32,213,2 // vpinsrb $0x2,%ebp,%xmm2,%xmm2
13840 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
13841 .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2
13842 .byte 196,226,121,49,210 // vpmovzxbd %xmm2,%xmm2
13843 .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2
13844 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
13845 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
13846 .byte 72,139,64,24 // mov 0x18(%rax),%rax
13847 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
13848 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3
13849 .byte 196,227,249,22,221,1 // vpextrq $0x1,%xmm3,%rbp
13850 .byte 65,137,232 // mov %ebp,%r8d
13851 .byte 72,193,237,32 // shr $0x20,%rbp
13852 .byte 196,225,249,126,219 // vmovq %xmm3,%rbx
13853 .byte 65,137,217 // mov %ebx,%r9d
13854 .byte 72,193,235,32 // shr $0x20,%rbx
13855 .byte 196,227,125,25,219,1 // vextractf128 $0x1,%ymm3,%xmm3
13856 .byte 196,195,249,22,218,1 // vpextrq $0x1,%xmm3,%r10
13857 .byte 69,137,211 // mov %r10d,%r11d
13858 .byte 73,193,234,32 // shr $0x20,%r10
13859 .byte 196,193,249,126,222 // vmovq %xmm3,%r14
13860 .byte 69,137,247 // mov %r14d,%r15d
13861 .byte 73,193,238,32 // shr $0x20,%r14
13862 .byte 196,163,121,32,28,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm3
13863 .byte 196,227,97,32,28,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm3,%xmm3
13864 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
13865 .byte 196,227,97,32,219,2 // vpinsrb $0x2,%ebx,%xmm3,%xmm3
13866 .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp
13867 .byte 196,227,97,32,221,3 // vpinsrb $0x3,%ebp,%xmm3,%xmm3
13868 .byte 196,98,121,49,195 // vpmovzxbd %xmm3,%xmm8
13869 .byte 196,163,121,32,28,56,0 // vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm3
13870 .byte 196,163,97,32,28,48,1 // vpinsrb $0x1,(%rax,%r14,1),%xmm3,%xmm3
13871 .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp
13872 .byte 196,227,97,32,221,2 // vpinsrb $0x2,%ebp,%xmm3,%xmm3
13873 .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax
13874 .byte 196,227,97,32,216,3 // vpinsrb $0x3,%eax,%xmm3,%xmm3
13875 .byte 196,226,121,49,219 // vpmovzxbd %xmm3,%xmm3
13876 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
13877 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
13878 .byte 197,180,89,219 // vmulps %ymm3,%ymm9,%ymm3
13879 .byte 72,173 // lods %ds:(%rsi),%rax
13880 .byte 91 // pop %rbx
13881 .byte 65,92 // pop %r12
13882 .byte 65,93 // pop %r13
13883 .byte 65,94 // pop %r14
13884 .byte 65,95 // pop %r15
13885 .byte 93 // pop %rbp
13886 .byte 255,224 // jmpq *%rax
13887
13888HIDDEN _sk_byte_tables_rgb_avx
13889.globl _sk_byte_tables_rgb_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]13890FUNCTION(_sk_byte_tables_rgb_avx)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]13891_sk_byte_tables_rgb_avx:
13892 .byte 85 // push %rbp
13893 .byte 65,87 // push %r15
13894 .byte 65,86 // push %r14
13895 .byte 65,85 // push %r13
13896 .byte 65,84 // push %r12
13897 .byte 83 // push %rbx
13898 .byte 72,173 // lods %ds:(%rsi),%rax
13899 .byte 68,139,64,24 // mov 0x18(%rax),%r8d
13900 .byte 65,255,200 // dec %r8d
13901 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
13902 .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8
13903 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
13904 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
13905 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
13906 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
13907 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
13908 .byte 68,137,197 // mov %r8d,%ebp
13909 .byte 77,137,194 // mov %r8,%r10
13910 .byte 73,193,234,32 // shr $0x20,%r10
13911 .byte 196,193,249,126,192 // vmovq %xmm0,%r8
13912 .byte 69,137,195 // mov %r8d,%r11d
13913 .byte 77,137,199 // mov %r8,%r15
13914 .byte 73,193,239,32 // shr $0x20,%r15
13915 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
13916 .byte 196,195,249,22,192,1 // vpextrq $0x1,%xmm0,%r8
13917 .byte 69,137,198 // mov %r8d,%r14d
13918 .byte 77,137,196 // mov %r8,%r12
13919 .byte 73,193,236,32 // shr $0x20,%r12
13920 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
13921 .byte 65,137,221 // mov %ebx,%r13d
13922 .byte 72,193,235,32 // shr $0x20,%rbx
13923 .byte 76,139,8 // mov (%rax),%r9
13924 .byte 76,139,64,8 // mov 0x8(%rax),%r8
13925 .byte 196,131,121,32,4,25,0 // vpinsrb $0x0,(%r9,%r11,1),%xmm0,%xmm0
13926 .byte 196,131,121,32,4,57,1 // vpinsrb $0x1,(%r9,%r15,1),%xmm0,%xmm0
13927 .byte 65,15,182,44,41 // movzbl (%r9,%rbp,1),%ebp
13928 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
13929 .byte 67,15,182,44,17 // movzbl (%r9,%r10,1),%ebp
13930 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
13931 .byte 196,98,121,49,200 // vpmovzxbd %xmm0,%xmm9
13932 .byte 196,131,121,32,4,41,0 // vpinsrb $0x0,(%r9,%r13,1),%xmm0,%xmm0
13933 .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0
13934 .byte 67,15,182,44,49 // movzbl (%r9,%r14,1),%ebp
13935 .byte 196,227,121,32,197,2 // vpinsrb $0x2,%ebp,%xmm0,%xmm0
13936 .byte 67,15,182,44,33 // movzbl (%r9,%r12,1),%ebp
13937 .byte 196,227,121,32,197,3 // vpinsrb $0x3,%ebp,%xmm0,%xmm0
13938 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
13939 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0
13940 .byte 197,124,91,208 // vcvtdq2ps %ymm0,%ymm10
13941 .byte 189,129,128,128,59 // mov $0x3b808081,%ebp
13942 .byte 197,249,110,197 // vmovd %ebp,%xmm0
13943 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
13944 .byte 196,99,125,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm9
13945 .byte 196,193,44,89,193 // vmulps %ymm9,%ymm10,%ymm0
13946 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
13947 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
13948 .byte 196,227,249,22,205,1 // vpextrq $0x1,%xmm1,%rbp
13949 .byte 65,137,233 // mov %ebp,%r9d
13950 .byte 72,193,237,32 // shr $0x20,%rbp
13951 .byte 196,225,249,126,203 // vmovq %xmm1,%rbx
13952 .byte 65,137,218 // mov %ebx,%r10d
13953 .byte 72,193,235,32 // shr $0x20,%rbx
13954 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
13955 .byte 196,195,249,22,203,1 // vpextrq $0x1,%xmm1,%r11
13956 .byte 69,137,222 // mov %r11d,%r14d
13957 .byte 73,193,235,32 // shr $0x20,%r11
13958 .byte 196,193,249,126,207 // vmovq %xmm1,%r15
13959 .byte 69,137,252 // mov %r15d,%r12d
13960 .byte 73,193,239,32 // shr $0x20,%r15
13961 .byte 196,131,121,32,12,16,0 // vpinsrb $0x0,(%r8,%r10,1),%xmm0,%xmm1
13962 .byte 196,195,113,32,12,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm1,%xmm1
13963 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
13964 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
13965 .byte 65,15,182,44,40 // movzbl (%r8,%rbp,1),%ebp
13966 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
13967 .byte 196,98,121,49,209 // vpmovzxbd %xmm1,%xmm10
13968 .byte 196,131,121,32,12,32,0 // vpinsrb $0x0,(%r8,%r12,1),%xmm0,%xmm1
13969 .byte 196,131,113,32,12,56,1 // vpinsrb $0x1,(%r8,%r15,1),%xmm1,%xmm1
13970 .byte 67,15,182,44,48 // movzbl (%r8,%r14,1),%ebp
13971 .byte 196,227,113,32,205,2 // vpinsrb $0x2,%ebp,%xmm1,%xmm1
13972 .byte 67,15,182,44,24 // movzbl (%r8,%r11,1),%ebp
13973 .byte 196,227,113,32,205,3 // vpinsrb $0x3,%ebp,%xmm1,%xmm1
13974 .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1
13975 .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
13976 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
13977 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
13978 .byte 72,139,64,16 // mov 0x10(%rax),%rax
13979 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
13980 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
13981 .byte 196,227,249,22,213,1 // vpextrq $0x1,%xmm2,%rbp
13982 .byte 65,137,232 // mov %ebp,%r8d
13983 .byte 72,193,237,32 // shr $0x20,%rbp
13984 .byte 196,225,249,126,211 // vmovq %xmm2,%rbx
13985 .byte 65,137,217 // mov %ebx,%r9d
13986 .byte 72,193,235,32 // shr $0x20,%rbx
13987 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
13988 .byte 196,195,249,22,210,1 // vpextrq $0x1,%xmm2,%r10
13989 .byte 69,137,211 // mov %r10d,%r11d
13990 .byte 73,193,234,32 // shr $0x20,%r10
13991 .byte 196,193,249,126,214 // vmovq %xmm2,%r14
13992 .byte 69,137,247 // mov %r14d,%r15d
13993 .byte 73,193,238,32 // shr $0x20,%r14
13994 .byte 196,163,121,32,20,8,0 // vpinsrb $0x0,(%rax,%r9,1),%xmm0,%xmm2
13995 .byte 196,227,105,32,20,24,1 // vpinsrb $0x1,(%rax,%rbx,1),%xmm2,%xmm2
13996 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
13997 .byte 196,227,105,32,211,2 // vpinsrb $0x2,%ebx,%xmm2,%xmm2
13998 .byte 15,182,44,40 // movzbl (%rax,%rbp,1),%ebp
13999 .byte 196,227,105,32,213,3 // vpinsrb $0x3,%ebp,%xmm2,%xmm2
14000 .byte 196,98,121,49,194 // vpmovzxbd %xmm2,%xmm8
14001 .byte 196,163,121,32,20,56,0 // vpinsrb $0x0,(%rax,%r15,1),%xmm0,%xmm2
14002 .byte 196,163,105,32,20,48,1 // vpinsrb $0x1,(%rax,%r14,1),%xmm2,%xmm2
14003 .byte 66,15,182,44,24 // movzbl (%rax,%r11,1),%ebp
14004 .byte 196,227,105,32,213,2 // vpinsrb $0x2,%ebp,%xmm2,%xmm2
14005 .byte 66,15,182,4,16 // movzbl (%rax,%r10,1),%eax
14006 .byte 196,227,105,32,208,3 // vpinsrb $0x3,%eax,%xmm2,%xmm2
14007 .byte 196,226,121,49,210 // vpmovzxbd %xmm2,%xmm2
14008 .byte 196,227,61,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm8,%ymm2
14009 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
14010 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
14011 .byte 72,173 // lods %ds:(%rsi),%rax
14012 .byte 91 // pop %rbx
14013 .byte 65,92 // pop %r12
14014 .byte 65,93 // pop %r13
14015 .byte 65,94 // pop %r14
14016 .byte 65,95 // pop %r15
14017 .byte 93 // pop %rbp
14018 .byte 255,224 // jmpq *%rax
14019
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]14020HIDDEN _sk_table_r_avx
14021.globl _sk_table_r_avx
14022FUNCTION(_sk_table_r_avx)
14023_sk_table_r_avx:
14024 .byte 65,87 // push %r15
14025 .byte 65,86 // push %r14
14026 .byte 65,84 // push %r12
14027 .byte 83 // push %rbx
14028 .byte 72,173 // lods %ds:(%rsi),%rax
14029 .byte 76,139,0 // mov (%rax),%r8
14030 .byte 139,64,8 // mov 0x8(%rax),%eax
14031 .byte 255,200 // dec %eax
14032 .byte 197,121,110,192 // vmovd %eax,%xmm8
14033 .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8
14034 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14035 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
14036 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
14037 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
14038 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
14039 .byte 65,137,193 // mov %eax,%r9d
14040 .byte 72,193,232,32 // shr $0x20,%rax
14041 .byte 196,193,249,126,194 // vmovq %xmm0,%r10
14042 .byte 69,137,211 // mov %r10d,%r11d
14043 .byte 73,193,234,32 // shr $0x20,%r10
14044 .byte 196,227,125,25,192,1 // vextractf128 $0x1,%ymm0,%xmm0
14045 .byte 196,227,249,22,195,1 // vpextrq $0x1,%xmm0,%rbx
14046 .byte 65,137,222 // mov %ebx,%r14d
14047 .byte 72,193,235,32 // shr $0x20,%rbx
14048 .byte 196,193,249,126,199 // vmovq %xmm0,%r15
14049 .byte 69,137,252 // mov %r15d,%r12d
14050 .byte 73,193,239,32 // shr $0x20,%r15
14051 .byte 196,129,122,16,4,160 // vmovss (%r8,%r12,4),%xmm0
14052 .byte 196,3,121,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm0,%xmm8
14053 .byte 196,129,122,16,4,176 // vmovss (%r8,%r14,4),%xmm0
14054 .byte 196,99,57,33,192,32 // vinsertps $0x20,%xmm0,%xmm8,%xmm8
14055 .byte 196,193,122,16,4,152 // vmovss (%r8,%rbx,4),%xmm0
14056 .byte 196,99,57,33,192,48 // vinsertps $0x30,%xmm0,%xmm8,%xmm8
14057 .byte 196,129,122,16,4,152 // vmovss (%r8,%r11,4),%xmm0
14058 .byte 196,3,121,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm0,%xmm9
14059 .byte 196,129,122,16,4,136 // vmovss (%r8,%r9,4),%xmm0
14060 .byte 196,99,49,33,200,32 // vinsertps $0x20,%xmm0,%xmm9,%xmm9
14061 .byte 196,193,122,16,4,128 // vmovss (%r8,%rax,4),%xmm0
14062 .byte 196,227,49,33,192,48 // vinsertps $0x30,%xmm0,%xmm9,%xmm0
14063 .byte 196,195,125,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm0,%ymm0
14064 .byte 72,173 // lods %ds:(%rsi),%rax
14065 .byte 91 // pop %rbx
14066 .byte 65,92 // pop %r12
14067 .byte 65,94 // pop %r14
14068 .byte 65,95 // pop %r15
14069 .byte 255,224 // jmpq *%rax
14070
14071HIDDEN _sk_table_g_avx
14072.globl _sk_table_g_avx
14073FUNCTION(_sk_table_g_avx)
14074_sk_table_g_avx:
14075 .byte 65,87 // push %r15
14076 .byte 65,86 // push %r14
14077 .byte 65,84 // push %r12
14078 .byte 83 // push %rbx
14079 .byte 72,173 // lods %ds:(%rsi),%rax
14080 .byte 76,139,0 // mov (%rax),%r8
14081 .byte 139,64,8 // mov 0x8(%rax),%eax
14082 .byte 255,200 // dec %eax
14083 .byte 197,121,110,192 // vmovd %eax,%xmm8
14084 .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8
14085 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14086 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
14087 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
14088 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
14089 .byte 196,227,249,22,200,1 // vpextrq $0x1,%xmm1,%rax
14090 .byte 65,137,193 // mov %eax,%r9d
14091 .byte 72,193,232,32 // shr $0x20,%rax
14092 .byte 196,193,249,126,202 // vmovq %xmm1,%r10
14093 .byte 69,137,211 // mov %r10d,%r11d
14094 .byte 73,193,234,32 // shr $0x20,%r10
14095 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
14096 .byte 196,227,249,22,203,1 // vpextrq $0x1,%xmm1,%rbx
14097 .byte 65,137,222 // mov %ebx,%r14d
14098 .byte 72,193,235,32 // shr $0x20,%rbx
14099 .byte 196,193,249,126,207 // vmovq %xmm1,%r15
14100 .byte 69,137,252 // mov %r15d,%r12d
14101 .byte 73,193,239,32 // shr $0x20,%r15
14102 .byte 196,129,122,16,12,160 // vmovss (%r8,%r12,4),%xmm1
14103 .byte 196,3,113,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm1,%xmm8
14104 .byte 196,129,122,16,12,176 // vmovss (%r8,%r14,4),%xmm1
14105 .byte 196,99,57,33,193,32 // vinsertps $0x20,%xmm1,%xmm8,%xmm8
14106 .byte 196,193,122,16,12,152 // vmovss (%r8,%rbx,4),%xmm1
14107 .byte 196,99,57,33,193,48 // vinsertps $0x30,%xmm1,%xmm8,%xmm8
14108 .byte 196,129,122,16,12,152 // vmovss (%r8,%r11,4),%xmm1
14109 .byte 196,3,113,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm1,%xmm9
14110 .byte 196,129,122,16,12,136 // vmovss (%r8,%r9,4),%xmm1
14111 .byte 196,99,49,33,201,32 // vinsertps $0x20,%xmm1,%xmm9,%xmm9
14112 .byte 196,193,122,16,12,128 // vmovss (%r8,%rax,4),%xmm1
14113 .byte 196,227,49,33,201,48 // vinsertps $0x30,%xmm1,%xmm9,%xmm1
14114 .byte 196,195,117,24,200,1 // vinsertf128 $0x1,%xmm8,%ymm1,%ymm1
14115 .byte 72,173 // lods %ds:(%rsi),%rax
14116 .byte 91 // pop %rbx
14117 .byte 65,92 // pop %r12
14118 .byte 65,94 // pop %r14
14119 .byte 65,95 // pop %r15
14120 .byte 255,224 // jmpq *%rax
14121
14122HIDDEN _sk_table_b_avx
14123.globl _sk_table_b_avx
14124FUNCTION(_sk_table_b_avx)
14125_sk_table_b_avx:
14126 .byte 65,87 // push %r15
14127 .byte 65,86 // push %r14
14128 .byte 65,84 // push %r12
14129 .byte 83 // push %rbx
14130 .byte 72,173 // lods %ds:(%rsi),%rax
14131 .byte 76,139,0 // mov (%rax),%r8
14132 .byte 139,64,8 // mov 0x8(%rax),%eax
14133 .byte 255,200 // dec %eax
14134 .byte 197,121,110,192 // vmovd %eax,%xmm8
14135 .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8
14136 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14137 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
14138 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
14139 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
14140 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
14141 .byte 65,137,193 // mov %eax,%r9d
14142 .byte 72,193,232,32 // shr $0x20,%rax
14143 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
14144 .byte 69,137,211 // mov %r10d,%r11d
14145 .byte 73,193,234,32 // shr $0x20,%r10
14146 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
14147 .byte 196,227,249,22,211,1 // vpextrq $0x1,%xmm2,%rbx
14148 .byte 65,137,222 // mov %ebx,%r14d
14149 .byte 72,193,235,32 // shr $0x20,%rbx
14150 .byte 196,193,249,126,215 // vmovq %xmm2,%r15
14151 .byte 69,137,252 // mov %r15d,%r12d
14152 .byte 73,193,239,32 // shr $0x20,%r15
14153 .byte 196,129,122,16,20,160 // vmovss (%r8,%r12,4),%xmm2
14154 .byte 196,3,105,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm2,%xmm8
14155 .byte 196,129,122,16,20,176 // vmovss (%r8,%r14,4),%xmm2
14156 .byte 196,99,57,33,194,32 // vinsertps $0x20,%xmm2,%xmm8,%xmm8
14157 .byte 196,193,122,16,20,152 // vmovss (%r8,%rbx,4),%xmm2
14158 .byte 196,99,57,33,194,48 // vinsertps $0x30,%xmm2,%xmm8,%xmm8
14159 .byte 196,129,122,16,20,152 // vmovss (%r8,%r11,4),%xmm2
14160 .byte 196,3,105,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm2,%xmm9
14161 .byte 196,129,122,16,20,136 // vmovss (%r8,%r9,4),%xmm2
14162 .byte 196,99,49,33,202,32 // vinsertps $0x20,%xmm2,%xmm9,%xmm9
14163 .byte 196,193,122,16,20,128 // vmovss (%r8,%rax,4),%xmm2
14164 .byte 196,227,49,33,210,48 // vinsertps $0x30,%xmm2,%xmm9,%xmm2
14165 .byte 196,195,109,24,208,1 // vinsertf128 $0x1,%xmm8,%ymm2,%ymm2
14166 .byte 72,173 // lods %ds:(%rsi),%rax
14167 .byte 91 // pop %rbx
14168 .byte 65,92 // pop %r12
14169 .byte 65,94 // pop %r14
14170 .byte 65,95 // pop %r15
14171 .byte 255,224 // jmpq *%rax
14172
14173HIDDEN _sk_table_a_avx
14174.globl _sk_table_a_avx
14175FUNCTION(_sk_table_a_avx)
14176_sk_table_a_avx:
14177 .byte 65,87 // push %r15
14178 .byte 65,86 // push %r14
14179 .byte 65,84 // push %r12
14180 .byte 83 // push %rbx
14181 .byte 72,173 // lods %ds:(%rsi),%rax
14182 .byte 76,139,0 // mov (%rax),%r8
14183 .byte 139,64,8 // mov 0x8(%rax),%eax
14184 .byte 255,200 // dec %eax
14185 .byte 197,121,110,192 // vmovd %eax,%xmm8
14186 .byte 196,65,121,112,192,0 // vpshufd $0x0,%xmm8,%xmm8
14187 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14188 .byte 196,65,124,91,192 // vcvtdq2ps %ymm8,%ymm8
14189 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
14190 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3
14191 .byte 196,227,249,22,216,1 // vpextrq $0x1,%xmm3,%rax
14192 .byte 65,137,193 // mov %eax,%r9d
14193 .byte 72,193,232,32 // shr $0x20,%rax
14194 .byte 196,193,249,126,218 // vmovq %xmm3,%r10
14195 .byte 69,137,211 // mov %r10d,%r11d
14196 .byte 73,193,234,32 // shr $0x20,%r10
14197 .byte 196,227,125,25,219,1 // vextractf128 $0x1,%ymm3,%xmm3
14198 .byte 196,227,249,22,219,1 // vpextrq $0x1,%xmm3,%rbx
14199 .byte 65,137,222 // mov %ebx,%r14d
14200 .byte 72,193,235,32 // shr $0x20,%rbx
14201 .byte 196,193,249,126,223 // vmovq %xmm3,%r15
14202 .byte 69,137,252 // mov %r15d,%r12d
14203 .byte 73,193,239,32 // shr $0x20,%r15
14204 .byte 196,129,122,16,28,160 // vmovss (%r8,%r12,4),%xmm3
14205 .byte 196,3,97,33,4,184,16 // vinsertps $0x10,(%r8,%r15,4),%xmm3,%xmm8
14206 .byte 196,129,122,16,28,176 // vmovss (%r8,%r14,4),%xmm3
14207 .byte 196,99,57,33,195,32 // vinsertps $0x20,%xmm3,%xmm8,%xmm8
14208 .byte 196,193,122,16,28,152 // vmovss (%r8,%rbx,4),%xmm3
14209 .byte 196,99,57,33,195,48 // vinsertps $0x30,%xmm3,%xmm8,%xmm8
14210 .byte 196,129,122,16,28,152 // vmovss (%r8,%r11,4),%xmm3
14211 .byte 196,3,97,33,12,144,16 // vinsertps $0x10,(%r8,%r10,4),%xmm3,%xmm9
14212 .byte 196,129,122,16,28,136 // vmovss (%r8,%r9,4),%xmm3
14213 .byte 196,99,49,33,203,32 // vinsertps $0x20,%xmm3,%xmm9,%xmm9
14214 .byte 196,193,122,16,28,128 // vmovss (%r8,%rax,4),%xmm3
14215 .byte 196,227,49,33,219,48 // vinsertps $0x30,%xmm3,%xmm9,%xmm3
14216 .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3
14217 .byte 72,173 // lods %ds:(%rsi),%rax
14218 .byte 91 // pop %rbx
14219 .byte 65,92 // pop %r12
14220 .byte 65,94 // pop %r14
14221 .byte 65,95 // pop %r15
14222 .byte 255,224 // jmpq *%rax
14223
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14224HIDDEN _sk_parametric_r_avx
14225.globl _sk_parametric_r_avx
14226FUNCTION(_sk_parametric_r_avx)
14227_sk_parametric_r_avx:
14228 .byte 72,173 // lods %ds:(%rsi),%rax
14229 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
14230 .byte 196,65,124,194,192,2 // vcmpleps %ymm8,%ymm0,%ymm8
14231 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
14232 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
14233 .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9
14234 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
14235 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
14236 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
14237 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14238 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
14239 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
14240 .byte 197,124,91,216 // vcvtdq2ps %ymm0,%ymm11
14241 .byte 196,98,125,24,37,198,52,0,0 // vbroadcastss 0x34c6(%rip),%ymm12 # 5c58 <_sk_callback_avx+0x12e>
14242 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
14243 .byte 196,98,125,24,37,188,52,0,0 // vbroadcastss 0x34bc(%rip),%ymm12 # 5c5c <_sk_callback_avx+0x132>
14244 .byte 196,193,124,84,196 // vandps %ymm12,%ymm0,%ymm0
14245 .byte 196,98,125,24,37,178,52,0,0 // vbroadcastss 0x34b2(%rip),%ymm12 # 5c60 <_sk_callback_avx+0x136>
14246 .byte 196,193,124,86,196 // vorps %ymm12,%ymm0,%ymm0
14247 .byte 196,98,125,24,37,168,52,0,0 // vbroadcastss 0x34a8(%rip),%ymm12 # 5c64 <_sk_callback_avx+0x13a>
14248 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
14249 .byte 196,98,125,24,37,158,52,0,0 // vbroadcastss 0x349e(%rip),%ymm12 # 5c68 <_sk_callback_avx+0x13e>
14250 .byte 196,65,124,89,228 // vmulps %ymm12,%ymm0,%ymm12
14251 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
14252 .byte 196,98,125,24,37,143,52,0,0 // vbroadcastss 0x348f(%rip),%ymm12 # 5c6c <_sk_callback_avx+0x142>
14253 .byte 196,193,124,88,196 // vaddps %ymm12,%ymm0,%ymm0
14254 .byte 196,98,125,24,37,133,52,0,0 // vbroadcastss 0x3485(%rip),%ymm12 # 5c70 <_sk_callback_avx+0x146>
14255 .byte 197,156,94,192 // vdivps %ymm0,%ymm12,%ymm0
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]14256 .byte 197,164,92,192 // vsubps %ymm0,%ymm11,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14257 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
14258 .byte 196,99,125,8,208,1 // vroundps $0x1,%ymm0,%ymm10
14259 .byte 196,65,124,92,210 // vsubps %ymm10,%ymm0,%ymm10
14260 .byte 196,98,125,24,29,105,52,0,0 // vbroadcastss 0x3469(%rip),%ymm11 # 5c74 <_sk_callback_avx+0x14a>
14261 .byte 196,193,124,88,195 // vaddps %ymm11,%ymm0,%ymm0
14262 .byte 196,98,125,24,29,95,52,0,0 // vbroadcastss 0x345f(%rip),%ymm11 # 5c78 <_sk_callback_avx+0x14e>
14263 .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
14264 .byte 196,193,124,92,195 // vsubps %ymm11,%ymm0,%ymm0
14265 .byte 196,98,125,24,29,80,52,0,0 // vbroadcastss 0x3450(%rip),%ymm11 # 5c7c <_sk_callback_avx+0x152>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14266 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14267 .byte 196,98,125,24,29,70,52,0,0 // vbroadcastss 0x3446(%rip),%ymm11 # 5c80 <_sk_callback_avx+0x156>
14268 .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
14269 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
14270 .byte 196,98,125,24,21,55,52,0,0 // vbroadcastss 0x3437(%rip),%ymm10 # 5c84 <_sk_callback_avx+0x15a>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14271 .byte 196,193,124,89,194 // vmulps %ymm10,%ymm0,%ymm0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14272 .byte 197,253,91,192 // vcvtps2dq %ymm0,%ymm0
14273 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
14274 .byte 196,193,124,88,194 // vaddps %ymm10,%ymm0,%ymm0
14275 .byte 196,195,125,74,193,128 // vblendvps %ymm8,%ymm9,%ymm0,%ymm0
14276 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
14277 .byte 196,65,124,95,192 // vmaxps %ymm8,%ymm0,%ymm8
14278 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14279 .byte 197,249,110,192 // vmovd %eax,%xmm0
14280 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
14281 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
14282 .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0
14283 .byte 72,173 // lods %ds:(%rsi),%rax
14284 .byte 255,224 // jmpq *%rax
14285
14286HIDDEN _sk_parametric_g_avx
14287.globl _sk_parametric_g_avx
14288FUNCTION(_sk_parametric_g_avx)
14289_sk_parametric_g_avx:
14290 .byte 72,173 // lods %ds:(%rsi),%rax
14291 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
14292 .byte 196,65,116,194,192,2 // vcmpleps %ymm8,%ymm1,%ymm8
14293 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
14294 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
14295 .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9
14296 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
14297 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
14298 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
14299 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14300 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
14301 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
14302 .byte 197,124,91,217 // vcvtdq2ps %ymm1,%ymm11
14303 .byte 196,98,125,24,37,176,51,0,0 // vbroadcastss 0x33b0(%rip),%ymm12 # 5c88 <_sk_callback_avx+0x15e>
14304 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
14305 .byte 196,98,125,24,37,166,51,0,0 // vbroadcastss 0x33a6(%rip),%ymm12 # 5c8c <_sk_callback_avx+0x162>
14306 .byte 196,193,116,84,204 // vandps %ymm12,%ymm1,%ymm1
14307 .byte 196,98,125,24,37,156,51,0,0 // vbroadcastss 0x339c(%rip),%ymm12 # 5c90 <_sk_callback_avx+0x166>
14308 .byte 196,193,116,86,204 // vorps %ymm12,%ymm1,%ymm1
14309 .byte 196,98,125,24,37,146,51,0,0 // vbroadcastss 0x3392(%rip),%ymm12 # 5c94 <_sk_callback_avx+0x16a>
14310 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
14311 .byte 196,98,125,24,37,136,51,0,0 // vbroadcastss 0x3388(%rip),%ymm12 # 5c98 <_sk_callback_avx+0x16e>
14312 .byte 196,65,116,89,228 // vmulps %ymm12,%ymm1,%ymm12
14313 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
14314 .byte 196,98,125,24,37,121,51,0,0 // vbroadcastss 0x3379(%rip),%ymm12 # 5c9c <_sk_callback_avx+0x172>
14315 .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1
14316 .byte 196,98,125,24,37,111,51,0,0 // vbroadcastss 0x336f(%rip),%ymm12 # 5ca0 <_sk_callback_avx+0x176>
14317 .byte 197,156,94,201 // vdivps %ymm1,%ymm12,%ymm1
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]14318 .byte 197,164,92,201 // vsubps %ymm1,%ymm11,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14319 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
14320 .byte 196,99,125,8,209,1 // vroundps $0x1,%ymm1,%ymm10
14321 .byte 196,65,116,92,210 // vsubps %ymm10,%ymm1,%ymm10
14322 .byte 196,98,125,24,29,83,51,0,0 // vbroadcastss 0x3353(%rip),%ymm11 # 5ca4 <_sk_callback_avx+0x17a>
14323 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
14324 .byte 196,98,125,24,29,73,51,0,0 // vbroadcastss 0x3349(%rip),%ymm11 # 5ca8 <_sk_callback_avx+0x17e>
14325 .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
14326 .byte 196,193,116,92,203 // vsubps %ymm11,%ymm1,%ymm1
14327 .byte 196,98,125,24,29,58,51,0,0 // vbroadcastss 0x333a(%rip),%ymm11 # 5cac <_sk_callback_avx+0x182>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14328 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14329 .byte 196,98,125,24,29,48,51,0,0 // vbroadcastss 0x3330(%rip),%ymm11 # 5cb0 <_sk_callback_avx+0x186>
14330 .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
14331 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
14332 .byte 196,98,125,24,21,33,51,0,0 // vbroadcastss 0x3321(%rip),%ymm10 # 5cb4 <_sk_callback_avx+0x18a>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14333 .byte 196,193,116,89,202 // vmulps %ymm10,%ymm1,%ymm1
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14334 .byte 197,253,91,201 // vcvtps2dq %ymm1,%ymm1
14335 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
14336 .byte 196,193,116,88,202 // vaddps %ymm10,%ymm1,%ymm1
14337 .byte 196,195,117,74,201,128 // vblendvps %ymm8,%ymm9,%ymm1,%ymm1
14338 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
14339 .byte 196,65,116,95,192 // vmaxps %ymm8,%ymm1,%ymm8
14340 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14341 .byte 197,249,110,200 // vmovd %eax,%xmm1
14342 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14343 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14344 .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1
14345 .byte 72,173 // lods %ds:(%rsi),%rax
14346 .byte 255,224 // jmpq *%rax
14347
14348HIDDEN _sk_parametric_b_avx
14349.globl _sk_parametric_b_avx
14350FUNCTION(_sk_parametric_b_avx)
14351_sk_parametric_b_avx:
14352 .byte 72,173 // lods %ds:(%rsi),%rax
14353 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
14354 .byte 196,65,108,194,192,2 // vcmpleps %ymm8,%ymm2,%ymm8
14355 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
14356 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
14357 .byte 197,52,89,202 // vmulps %ymm2,%ymm9,%ymm9
14358 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
14359 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
14360 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
14361 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14362 .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
14363 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
14364 .byte 197,124,91,218 // vcvtdq2ps %ymm2,%ymm11
14365 .byte 196,98,125,24,37,154,50,0,0 // vbroadcastss 0x329a(%rip),%ymm12 # 5cb8 <_sk_callback_avx+0x18e>
14366 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
14367 .byte 196,98,125,24,37,144,50,0,0 // vbroadcastss 0x3290(%rip),%ymm12 # 5cbc <_sk_callback_avx+0x192>
14368 .byte 196,193,108,84,212 // vandps %ymm12,%ymm2,%ymm2
14369 .byte 196,98,125,24,37,134,50,0,0 // vbroadcastss 0x3286(%rip),%ymm12 # 5cc0 <_sk_callback_avx+0x196>
14370 .byte 196,193,108,86,212 // vorps %ymm12,%ymm2,%ymm2
14371 .byte 196,98,125,24,37,124,50,0,0 // vbroadcastss 0x327c(%rip),%ymm12 # 5cc4 <_sk_callback_avx+0x19a>
14372 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
14373 .byte 196,98,125,24,37,114,50,0,0 // vbroadcastss 0x3272(%rip),%ymm12 # 5cc8 <_sk_callback_avx+0x19e>
14374 .byte 196,65,108,89,228 // vmulps %ymm12,%ymm2,%ymm12
14375 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
14376 .byte 196,98,125,24,37,99,50,0,0 // vbroadcastss 0x3263(%rip),%ymm12 # 5ccc <_sk_callback_avx+0x1a2>
14377 .byte 196,193,108,88,212 // vaddps %ymm12,%ymm2,%ymm2
14378 .byte 196,98,125,24,37,89,50,0,0 // vbroadcastss 0x3259(%rip),%ymm12 # 5cd0 <_sk_callback_avx+0x1a6>
14379 .byte 197,156,94,210 // vdivps %ymm2,%ymm12,%ymm2
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]14380 .byte 197,164,92,210 // vsubps %ymm2,%ymm11,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14381 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
14382 .byte 196,99,125,8,210,1 // vroundps $0x1,%ymm2,%ymm10
14383 .byte 196,65,108,92,210 // vsubps %ymm10,%ymm2,%ymm10
14384 .byte 196,98,125,24,29,61,50,0,0 // vbroadcastss 0x323d(%rip),%ymm11 # 5cd4 <_sk_callback_avx+0x1aa>
14385 .byte 196,193,108,88,211 // vaddps %ymm11,%ymm2,%ymm2
14386 .byte 196,98,125,24,29,51,50,0,0 // vbroadcastss 0x3233(%rip),%ymm11 # 5cd8 <_sk_callback_avx+0x1ae>
14387 .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
14388 .byte 196,193,108,92,211 // vsubps %ymm11,%ymm2,%ymm2
14389 .byte 196,98,125,24,29,36,50,0,0 // vbroadcastss 0x3224(%rip),%ymm11 # 5cdc <_sk_callback_avx+0x1b2>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14390 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14391 .byte 196,98,125,24,29,26,50,0,0 // vbroadcastss 0x321a(%rip),%ymm11 # 5ce0 <_sk_callback_avx+0x1b6>
14392 .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
14393 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
14394 .byte 196,98,125,24,21,11,50,0,0 // vbroadcastss 0x320b(%rip),%ymm10 # 5ce4 <_sk_callback_avx+0x1ba>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14395 .byte 196,193,108,89,210 // vmulps %ymm10,%ymm2,%ymm2
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14396 .byte 197,253,91,210 // vcvtps2dq %ymm2,%ymm2
14397 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
14398 .byte 196,193,108,88,210 // vaddps %ymm10,%ymm2,%ymm2
14399 .byte 196,195,109,74,209,128 // vblendvps %ymm8,%ymm9,%ymm2,%ymm2
14400 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
14401 .byte 196,65,108,95,192 // vmaxps %ymm8,%ymm2,%ymm8
14402 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14403 .byte 197,249,110,208 // vmovd %eax,%xmm2
14404 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
14405 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
14406 .byte 197,188,93,210 // vminps %ymm2,%ymm8,%ymm2
14407 .byte 72,173 // lods %ds:(%rsi),%rax
14408 .byte 255,224 // jmpq *%rax
14409
14410HIDDEN _sk_parametric_a_avx
14411.globl _sk_parametric_a_avx
14412FUNCTION(_sk_parametric_a_avx)
14413_sk_parametric_a_avx:
14414 .byte 72,173 // lods %ds:(%rsi),%rax
14415 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
14416 .byte 196,65,100,194,192,2 // vcmpleps %ymm8,%ymm3,%ymm8
14417 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
14418 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
14419 .byte 197,52,89,203 // vmulps %ymm3,%ymm9,%ymm9
14420 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
14421 .byte 196,98,125,24,80,4 // vbroadcastss 0x4(%rax),%ymm10
14422 .byte 196,98,125,24,88,8 // vbroadcastss 0x8(%rax),%ymm11
14423 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14424 .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
14425 .byte 196,98,125,24,16 // vbroadcastss (%rax),%ymm10
14426 .byte 197,124,91,219 // vcvtdq2ps %ymm3,%ymm11
14427 .byte 196,98,125,24,37,132,49,0,0 // vbroadcastss 0x3184(%rip),%ymm12 # 5ce8 <_sk_callback_avx+0x1be>
14428 .byte 196,65,36,89,220 // vmulps %ymm12,%ymm11,%ymm11
14429 .byte 196,98,125,24,37,122,49,0,0 // vbroadcastss 0x317a(%rip),%ymm12 # 5cec <_sk_callback_avx+0x1c2>
14430 .byte 196,193,100,84,220 // vandps %ymm12,%ymm3,%ymm3
14431 .byte 196,98,125,24,37,112,49,0,0 // vbroadcastss 0x3170(%rip),%ymm12 # 5cf0 <_sk_callback_avx+0x1c6>
14432 .byte 196,193,100,86,220 // vorps %ymm12,%ymm3,%ymm3
14433 .byte 196,98,125,24,37,102,49,0,0 // vbroadcastss 0x3166(%rip),%ymm12 # 5cf4 <_sk_callback_avx+0x1ca>
14434 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
14435 .byte 196,98,125,24,37,92,49,0,0 // vbroadcastss 0x315c(%rip),%ymm12 # 5cf8 <_sk_callback_avx+0x1ce>
14436 .byte 196,65,100,89,228 // vmulps %ymm12,%ymm3,%ymm12
14437 .byte 196,65,36,92,220 // vsubps %ymm12,%ymm11,%ymm11
14438 .byte 196,98,125,24,37,77,49,0,0 // vbroadcastss 0x314d(%rip),%ymm12 # 5cfc <_sk_callback_avx+0x1d2>
14439 .byte 196,193,100,88,220 // vaddps %ymm12,%ymm3,%ymm3
14440 .byte 196,98,125,24,37,67,49,0,0 // vbroadcastss 0x3143(%rip),%ymm12 # 5d00 <_sk_callback_avx+0x1d6>
14441 .byte 197,156,94,219 // vdivps %ymm3,%ymm12,%ymm3
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]14442 .byte 197,164,92,219 // vsubps %ymm3,%ymm11,%ymm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14443 .byte 197,172,89,219 // vmulps %ymm3,%ymm10,%ymm3
14444 .byte 196,99,125,8,211,1 // vroundps $0x1,%ymm3,%ymm10
14445 .byte 196,65,100,92,210 // vsubps %ymm10,%ymm3,%ymm10
14446 .byte 196,98,125,24,29,39,49,0,0 // vbroadcastss 0x3127(%rip),%ymm11 # 5d04 <_sk_callback_avx+0x1da>
14447 .byte 196,193,100,88,219 // vaddps %ymm11,%ymm3,%ymm3
14448 .byte 196,98,125,24,29,29,49,0,0 // vbroadcastss 0x311d(%rip),%ymm11 # 5d08 <_sk_callback_avx+0x1de>
14449 .byte 196,65,44,89,219 // vmulps %ymm11,%ymm10,%ymm11
14450 .byte 196,193,100,92,219 // vsubps %ymm11,%ymm3,%ymm3
14451 .byte 196,98,125,24,29,14,49,0,0 // vbroadcastss 0x310e(%rip),%ymm11 # 5d0c <_sk_callback_avx+0x1e2>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14452 .byte 196,65,36,92,210 // vsubps %ymm10,%ymm11,%ymm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14453 .byte 196,98,125,24,29,4,49,0,0 // vbroadcastss 0x3104(%rip),%ymm11 # 5d10 <_sk_callback_avx+0x1e6>
14454 .byte 196,65,36,94,210 // vdivps %ymm10,%ymm11,%ymm10
14455 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3
14456 .byte 196,98,125,24,21,245,48,0,0 // vbroadcastss 0x30f5(%rip),%ymm10 # 5d14 <_sk_callback_avx+0x1ea>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14457 .byte 196,193,100,89,218 // vmulps %ymm10,%ymm3,%ymm3
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]14458 .byte 197,253,91,219 // vcvtps2dq %ymm3,%ymm3
14459 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
14460 .byte 196,193,100,88,218 // vaddps %ymm10,%ymm3,%ymm3
14461 .byte 196,195,101,74,217,128 // vblendvps %ymm8,%ymm9,%ymm3,%ymm3
14462 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
14463 .byte 196,65,100,95,192 // vmaxps %ymm8,%ymm3,%ymm8
14464 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14465 .byte 197,249,110,216 // vmovd %eax,%xmm3
14466 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
14467 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
14468 .byte 197,188,93,219 // vminps %ymm3,%ymm8,%ymm3
14469 .byte 72,173 // lods %ds:(%rsi),%rax
14470 .byte 255,224 // jmpq *%rax
14471
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]14472HIDDEN _sk_lab_to_xyz_avx
14473.globl _sk_lab_to_xyz_avx
14474FUNCTION(_sk_lab_to_xyz_avx)
14475_sk_lab_to_xyz_avx:
14476 .byte 184,0,0,200,66 // mov $0x42c80000,%eax
14477 .byte 197,121,110,192 // vmovd %eax,%xmm8
14478 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
14479 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14480 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
14481 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
14482 .byte 197,249,110,192 // vmovd %eax,%xmm0
14483 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
14484 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
14485 .byte 197,124,89,201 // vmulps %ymm1,%ymm0,%ymm9
14486 .byte 184,0,0,0,67 // mov $0x43000000,%eax
14487 .byte 197,249,110,200 // vmovd %eax,%xmm1
14488 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14489 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14490 .byte 197,52,92,201 // vsubps %ymm1,%ymm9,%ymm9
14491 .byte 197,252,89,194 // vmulps %ymm2,%ymm0,%ymm0
14492 .byte 197,124,92,209 // vsubps %ymm1,%ymm0,%ymm10
14493 .byte 184,0,0,128,65 // mov $0x41800000,%eax
14494 .byte 197,249,110,192 // vmovd %eax,%xmm0
14495 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
14496 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
14497 .byte 197,188,88,192 // vaddps %ymm0,%ymm8,%ymm0
14498 .byte 184,203,61,13,60 // mov $0x3c0d3dcb,%eax
14499 .byte 197,249,110,208 // vmovd %eax,%xmm2
14500 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
14501 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
14502 .byte 197,252,89,194 // vmulps %ymm2,%ymm0,%ymm0
14503 .byte 184,111,18,3,59 // mov $0x3b03126f,%eax
14504 .byte 197,249,110,208 // vmovd %eax,%xmm2
14505 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
14506 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
14507 .byte 197,180,89,210 // vmulps %ymm2,%ymm9,%ymm2
14508 .byte 197,252,88,210 // vaddps %ymm2,%ymm0,%ymm2
14509 .byte 184,10,215,163,59 // mov $0x3ba3d70a,%eax
14510 .byte 197,249,110,200 // vmovd %eax,%xmm1
14511 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14512 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14513 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
14514 .byte 197,124,92,193 // vsubps %ymm1,%ymm0,%ymm8
14515 .byte 197,236,89,202 // vmulps %ymm2,%ymm2,%ymm1
14516 .byte 197,108,89,201 // vmulps %ymm1,%ymm2,%ymm9
14517 .byte 184,194,24,17,60 // mov $0x3c1118c2,%eax
14518 .byte 197,249,110,200 // vmovd %eax,%xmm1
14519 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14520 .byte 196,99,117,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm10
14521 .byte 196,65,44,194,217,1 // vcmpltps %ymm9,%ymm10,%ymm11
14522 .byte 184,203,61,13,62 // mov $0x3e0d3dcb,%eax
14523 .byte 197,249,110,200 // vmovd %eax,%xmm1
14524 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14525 .byte 196,99,117,24,225,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm12
14526 .byte 196,193,108,92,204 // vsubps %ymm12,%ymm2,%ymm1
14527 .byte 184,80,128,3,62 // mov $0x3e038050,%eax
14528 .byte 197,249,110,208 // vmovd %eax,%xmm2
14529 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
14530 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
14531 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
14532 .byte 196,67,117,74,201,176 // vblendvps %ymm11,%ymm9,%ymm1,%ymm9
14533 .byte 197,252,89,200 // vmulps %ymm0,%ymm0,%ymm1
14534 .byte 197,252,89,201 // vmulps %ymm1,%ymm0,%ymm1
14535 .byte 197,44,194,217,1 // vcmpltps %ymm1,%ymm10,%ymm11
14536 .byte 196,193,124,92,196 // vsubps %ymm12,%ymm0,%ymm0
14537 .byte 197,252,89,194 // vmulps %ymm2,%ymm0,%ymm0
14538 .byte 196,227,125,74,201,176 // vblendvps %ymm11,%ymm1,%ymm0,%ymm1
14539 .byte 196,193,60,89,192 // vmulps %ymm8,%ymm8,%ymm0
14540 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
14541 .byte 197,44,194,208,1 // vcmpltps %ymm0,%ymm10,%ymm10
14542 .byte 196,65,60,92,196 // vsubps %ymm12,%ymm8,%ymm8
14543 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
14544 .byte 196,99,109,74,192,160 // vblendvps %ymm10,%ymm0,%ymm2,%ymm8
14545 .byte 184,31,215,118,63 // mov $0x3f76d71f,%eax
14546 .byte 197,249,110,192 // vmovd %eax,%xmm0
14547 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
14548 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
14549 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
14550 .byte 184,246,64,83,63 // mov $0x3f5340f6,%eax
14551 .byte 197,249,110,208 // vmovd %eax,%xmm2
14552 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
14553 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
14554 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
14555 .byte 72,173 // lods %ds:(%rsi),%rax
14556 .byte 255,224 // jmpq *%rax
14557
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]14558HIDDEN _sk_load_a8_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14559.globl _sk_load_a8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14560FUNCTION(_sk_load_a8_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14561_sk_load_a8_avx:
14562 .byte 73,137,200 // mov %rcx,%r8
14563 .byte 72,173 // lods %ds:(%rsi),%rax
14564 .byte 72,139,0 // mov (%rax),%rax
14565 .byte 72,1,248 // add %rdi,%rax
14566 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14567 .byte 117,74 // jne 2e44 <_sk_load_a8_avx+0x5a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14568 .byte 197,250,126,0 // vmovq (%rax),%xmm0
14569 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
14570 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
14571 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14572 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
14573 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14574 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
14575 .byte 197,249,110,200 // vmovd %eax,%xmm1
14576 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14577 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14578 .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
14579 .byte 72,173 // lods %ds:(%rsi),%rax
14580 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
14581 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
14582 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
14583 .byte 76,137,193 // mov %r8,%rcx
14584 .byte 255,224 // jmpq *%rax
14585 .byte 49,201 // xor %ecx,%ecx
14586 .byte 77,137,194 // mov %r8,%r10
14587 .byte 69,49,201 // xor %r9d,%r9d
14588 .byte 68,15,182,24 // movzbl (%rax),%r11d
14589 .byte 72,255,192 // inc %rax
14590 .byte 73,211,227 // shl %cl,%r11
14591 .byte 77,9,217 // or %r11,%r9
14592 .byte 72,131,193,8 // add $0x8,%rcx
14593 .byte 73,255,202 // dec %r10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14594 .byte 117,234 // jne 2e4c <_sk_load_a8_avx+0x62>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14595 .byte 196,193,249,110,193 // vmovq %r9,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14596 .byte 235,149 // jmp 2dfe <_sk_load_a8_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14597
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]14598HIDDEN _sk_gather_a8_avx
14599.globl _sk_gather_a8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14600FUNCTION(_sk_gather_a8_avx)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]14601_sk_gather_a8_avx:
14602 .byte 65,87 // push %r15
14603 .byte 65,86 // push %r14
14604 .byte 65,84 // push %r12
14605 .byte 83 // push %rbx
14606 .byte 72,173 // lods %ds:(%rsi),%rax
14607 .byte 76,139,0 // mov (%rax),%r8
14608 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
14609 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
14610 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
14611 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
14612 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
14613 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
14614 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
14615 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
14616 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
14617 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
14618 .byte 65,137,193 // mov %eax,%r9d
14619 .byte 72,193,232,32 // shr $0x20,%rax
14620 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
14621 .byte 69,137,211 // mov %r10d,%r11d
14622 .byte 73,193,234,32 // shr $0x20,%r10
14623 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
14624 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
14625 .byte 65,137,222 // mov %ebx,%r14d
14626 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
14627 .byte 69,137,252 // mov %r15d,%r12d
14628 .byte 73,193,239,32 // shr $0x20,%r15
14629 .byte 72,193,235,32 // shr $0x20,%rbx
14630 .byte 196,131,121,32,4,48,0 // vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0
14631 .byte 196,195,121,32,4,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0
14632 .byte 67,15,182,28,32 // movzbl (%r8,%r12,1),%ebx
14633 .byte 196,227,121,32,195,2 // vpinsrb $0x2,%ebx,%xmm0,%xmm0
14634 .byte 67,15,182,28,56 // movzbl (%r8,%r15,1),%ebx
14635 .byte 196,227,121,32,195,3 // vpinsrb $0x3,%ebx,%xmm0,%xmm0
14636 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14637 .byte 196,131,121,32,12,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1
14638 .byte 196,131,113,32,12,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1
14639 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
14640 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
14641 .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax
14642 .byte 196,227,113,32,200,3 // vpinsrb $0x3,%eax,%xmm1,%xmm1
14643 .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1
14644 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
14645 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14646 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
14647 .byte 197,249,110,200 // vmovd %eax,%xmm1
14648 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14649 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14650 .byte 197,252,89,217 // vmulps %ymm1,%ymm0,%ymm3
14651 .byte 72,173 // lods %ds:(%rsi),%rax
14652 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
14653 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
14654 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
14655 .byte 91 // pop %rbx
14656 .byte 65,92 // pop %r12
14657 .byte 65,94 // pop %r14
14658 .byte 65,95 // pop %r15
14659 .byte 255,224 // jmpq *%rax
14660
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]14661HIDDEN _sk_store_a8_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14662.globl _sk_store_a8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14663FUNCTION(_sk_store_a8_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14664_sk_store_a8_avx:
14665 .byte 72,173 // lods %ds:(%rsi),%rax
14666 .byte 76,139,8 // mov (%rax),%r9
14667 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
14668 .byte 197,121,110,192 // vmovd %eax,%xmm8
14669 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
14670 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
14671 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
14672 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
14673 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
14674 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
14675 .byte 196,65,57,103,192 // vpackuswb %xmm8,%xmm8,%xmm8
14676 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14677 .byte 117,10 // jne 2fa5 <_sk_store_a8_avx+0x42>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14678 .byte 196,65,123,17,4,57 // vmovsd %xmm8,(%r9,%rdi,1)
14679 .byte 72,173 // lods %ds:(%rsi),%rax
14680 .byte 255,224 // jmpq *%rax
14681 .byte 65,137,200 // mov %ecx,%r8d
14682 .byte 65,128,224,7 // and $0x7,%r8b
14683 .byte 65,254,200 // dec %r8b
14684 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14685 .byte 119,236 // ja 2fa1 <_sk_store_a8_avx+0x3e>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14686 .byte 196,66,121,48,192 // vpmovzxbw %xmm8,%xmm8
14687 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14688 .byte 76,141,5,67,0,0,0 // lea 0x43(%rip),%r8 # 3008 <_sk_store_a8_avx+0xa5>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14689 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
14690 .byte 76,1,192 // add %r8,%rax
14691 .byte 255,224 // jmpq *%rax
14692 .byte 196,67,121,20,68,57,6,12 // vpextrb $0xc,%xmm8,0x6(%r9,%rdi,1)
14693 .byte 196,67,121,20,68,57,5,10 // vpextrb $0xa,%xmm8,0x5(%r9,%rdi,1)
14694 .byte 196,67,121,20,68,57,4,8 // vpextrb $0x8,%xmm8,0x4(%r9,%rdi,1)
14695 .byte 196,67,121,20,68,57,3,6 // vpextrb $0x6,%xmm8,0x3(%r9,%rdi,1)
14696 .byte 196,67,121,20,68,57,2,4 // vpextrb $0x4,%xmm8,0x2(%r9,%rdi,1)
14697 .byte 196,67,121,20,68,57,1,2 // vpextrb $0x2,%xmm8,0x1(%r9,%rdi,1)
14698 .byte 196,67,121,20,4,57,0 // vpextrb $0x0,%xmm8,(%r9,%rdi,1)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14699 .byte 235,154 // jmp 2fa1 <_sk_store_a8_avx+0x3e>
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]14700 .byte 144 // nop
14701 .byte 246,255 // idiv %bh
14702 .byte 255 // (bad)
14703 .byte 255 // (bad)
14704 .byte 238 // out %al,(%dx)
14705 .byte 255 // (bad)
14706 .byte 255 // (bad)
14707 .byte 255,230 // jmpq *%rsi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14708 .byte 255 // (bad)
14709 .byte 255 // (bad)
14710 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]14711 .byte 222,255 // fdivrp %st,%st(7)
14712 .byte 255 // (bad)
14713 .byte 255,214 // callq *%rsi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14714 .byte 255 // (bad)
14715 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]14716 .byte 255,206 // dec %esi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14717 .byte 255 // (bad)
14718 .byte 255 // (bad)
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]14719 .byte 255,198 // inc %esi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14720 .byte 255 // (bad)
14721 .byte 255 // (bad)
14722 .byte 255 // .byte 0xff
14723
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]14724HIDDEN _sk_load_g8_avx
14725.globl _sk_load_g8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14726FUNCTION(_sk_load_g8_avx)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]14727_sk_load_g8_avx:
14728 .byte 73,137,200 // mov %rcx,%r8
14729 .byte 72,173 // lods %ds:(%rsi),%rax
14730 .byte 72,139,0 // mov (%rax),%rax
14731 .byte 72,1,248 // add %rdi,%rax
14732 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14733 .byte 117,91 // jne 308f <_sk_load_g8_avx+0x6b>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]14734 .byte 197,250,126,0 // vmovq (%rax),%xmm0
14735 .byte 196,226,121,49,200 // vpmovzxbd %xmm0,%xmm1
14736 .byte 196,227,121,4,192,229 // vpermilps $0xe5,%xmm0,%xmm0
14737 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14738 .byte 196,227,117,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm1,%ymm0
14739 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14740 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
14741 .byte 197,249,110,200 // vmovd %eax,%xmm1
14742 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14743 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14744 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
14745 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14746 .byte 197,249,110,200 // vmovd %eax,%xmm1
14747 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14748 .byte 196,227,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm3
14749 .byte 72,173 // lods %ds:(%rsi),%rax
14750 .byte 76,137,193 // mov %r8,%rcx
14751 .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
14752 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
14753 .byte 255,224 // jmpq *%rax
14754 .byte 49,201 // xor %ecx,%ecx
14755 .byte 77,137,194 // mov %r8,%r10
14756 .byte 69,49,201 // xor %r9d,%r9d
14757 .byte 68,15,182,24 // movzbl (%rax),%r11d
14758 .byte 72,255,192 // inc %rax
14759 .byte 73,211,227 // shl %cl,%r11
14760 .byte 77,9,217 // or %r11,%r9
14761 .byte 72,131,193,8 // add $0x8,%rcx
14762 .byte 73,255,202 // dec %r10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14763 .byte 117,234 // jne 3097 <_sk_load_g8_avx+0x73>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]14764 .byte 196,193,249,110,193 // vmovq %r9,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14765 .byte 235,132 // jmp 3038 <_sk_load_g8_avx+0x14>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]14766
14767HIDDEN _sk_gather_g8_avx
14768.globl _sk_gather_g8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14769FUNCTION(_sk_gather_g8_avx)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]14770_sk_gather_g8_avx:
14771 .byte 65,87 // push %r15
14772 .byte 65,86 // push %r14
14773 .byte 65,84 // push %r12
14774 .byte 83 // push %rbx
14775 .byte 72,173 // lods %ds:(%rsi),%rax
14776 .byte 76,139,0 // mov (%rax),%r8
14777 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
14778 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
14779 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
14780 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
14781 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
14782 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
14783 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
14784 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
14785 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
14786 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
14787 .byte 65,137,193 // mov %eax,%r9d
14788 .byte 72,193,232,32 // shr $0x20,%rax
14789 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
14790 .byte 69,137,211 // mov %r10d,%r11d
14791 .byte 73,193,234,32 // shr $0x20,%r10
14792 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
14793 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
14794 .byte 65,137,222 // mov %ebx,%r14d
14795 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
14796 .byte 69,137,252 // mov %r15d,%r12d
14797 .byte 73,193,239,32 // shr $0x20,%r15
14798 .byte 72,193,235,32 // shr $0x20,%rbx
14799 .byte 196,131,121,32,4,48,0 // vpinsrb $0x0,(%r8,%r14,1),%xmm0,%xmm0
14800 .byte 196,195,121,32,4,24,1 // vpinsrb $0x1,(%r8,%rbx,1),%xmm0,%xmm0
14801 .byte 67,15,182,28,32 // movzbl (%r8,%r12,1),%ebx
14802 .byte 196,227,121,32,195,2 // vpinsrb $0x2,%ebx,%xmm0,%xmm0
14803 .byte 67,15,182,28,56 // movzbl (%r8,%r15,1),%ebx
14804 .byte 196,227,121,32,195,3 // vpinsrb $0x3,%ebx,%xmm0,%xmm0
14805 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14806 .byte 196,131,121,32,12,24,0 // vpinsrb $0x0,(%r8,%r11,1),%xmm0,%xmm1
14807 .byte 196,131,113,32,12,16,1 // vpinsrb $0x1,(%r8,%r10,1),%xmm1,%xmm1
14808 .byte 67,15,182,28,8 // movzbl (%r8,%r9,1),%ebx
14809 .byte 196,227,113,32,203,2 // vpinsrb $0x2,%ebx,%xmm1,%xmm1
14810 .byte 65,15,182,4,0 // movzbl (%r8,%rax,1),%eax
14811 .byte 196,227,113,32,200,3 // vpinsrb $0x3,%eax,%xmm1,%xmm1
14812 .byte 196,226,121,49,201 // vpmovzxbd %xmm1,%xmm1
14813 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
14814 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14815 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
14816 .byte 197,249,110,200 // vmovd %eax,%xmm1
14817 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14818 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14819 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
14820 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14821 .byte 197,249,110,200 // vmovd %eax,%xmm1
14822 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14823 .byte 196,227,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm3
14824 .byte 72,173 // lods %ds:(%rsi),%rax
14825 .byte 197,252,40,200 // vmovaps %ymm0,%ymm1
14826 .byte 197,252,40,208 // vmovaps %ymm0,%ymm2
14827 .byte 91 // pop %rbx
14828 .byte 65,92 // pop %r12
14829 .byte 65,94 // pop %r14
14830 .byte 65,95 // pop %r15
14831 .byte 255,224 // jmpq *%rax
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]14832
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]14833HIDDEN _sk_gather_i8_avx
14834.globl _sk_gather_i8_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14835FUNCTION(_sk_gather_i8_avx)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]14836_sk_gather_i8_avx:
14837 .byte 72,173 // lods %ds:(%rsi),%rax
14838 .byte 73,137,192 // mov %rax,%r8
14839 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14840 .byte 116,5 // je 31ce <_sk_gather_i8_avx+0xf>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]14841 .byte 76,137,192 // mov %r8,%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14842 .byte 235,2 // jmp 31d0 <_sk_gather_i8_avx+0x11>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]14843 .byte 72,173 // lods %ds:(%rsi),%rax
14844 .byte 65,87 // push %r15
14845 .byte 65,86 // push %r14
14846 .byte 65,85 // push %r13
14847 .byte 65,84 // push %r12
14848 .byte 83 // push %rbx
14849 .byte 76,139,8 // mov (%rax),%r9
14850 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
14851 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
14852 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
14853 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
14854 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
14855 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
14856 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
14857 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
14858 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
14859 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
14860 .byte 65,137,194 // mov %eax,%r10d
14861 .byte 72,193,232,32 // shr $0x20,%rax
14862 .byte 196,193,249,126,211 // vmovq %xmm2,%r11
14863 .byte 69,137,222 // mov %r11d,%r14d
14864 .byte 73,193,235,32 // shr $0x20,%r11
14865 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
14866 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
14867 .byte 65,137,223 // mov %ebx,%r15d
14868 .byte 196,195,249,22,196,1 // vpextrq $0x1,%xmm0,%r12
14869 .byte 69,137,229 // mov %r12d,%r13d
14870 .byte 73,193,236,32 // shr $0x20,%r12
14871 .byte 72,193,235,32 // shr $0x20,%rbx
14872 .byte 196,131,121,32,4,49,0 // vpinsrb $0x0,(%r9,%r14,1),%xmm0,%xmm0
14873 .byte 196,131,121,32,4,25,1 // vpinsrb $0x1,(%r9,%r11,1),%xmm0,%xmm0
14874 .byte 196,131,121,32,4,17,2 // vpinsrb $0x2,(%r9,%r10,1),%xmm0,%xmm0
14875 .byte 196,195,121,32,4,1,3 // vpinsrb $0x3,(%r9,%rax,1),%xmm0,%xmm0
14876 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14877 .byte 196,195,249,22,194,1 // vpextrq $0x1,%xmm0,%r10
14878 .byte 196,193,249,126,195 // vmovq %xmm0,%r11
14879 .byte 196,131,121,32,4,57,0 // vpinsrb $0x0,(%r9,%r15,1),%xmm0,%xmm0
14880 .byte 196,195,121,32,4,25,1 // vpinsrb $0x1,(%r9,%rbx,1),%xmm0,%xmm0
14881 .byte 196,131,121,32,4,41,2 // vpinsrb $0x2,(%r9,%r13,1),%xmm0,%xmm0
14882 .byte 196,131,121,32,4,33,3 // vpinsrb $0x3,(%r9,%r12,1),%xmm0,%xmm0
14883 .byte 196,226,121,49,192 // vpmovzxbd %xmm0,%xmm0
14884 .byte 73,139,88,8 // mov 0x8(%r8),%rbx
14885 .byte 196,193,249,126,193 // vmovq %xmm0,%r9
14886 .byte 69,137,200 // mov %r9d,%r8d
14887 .byte 73,193,233,30 // shr $0x1e,%r9
14888 .byte 196,227,249,22,192,1 // vpextrq $0x1,%xmm0,%rax
14889 .byte 65,137,198 // mov %eax,%r14d
14890 .byte 72,193,232,30 // shr $0x1e,%rax
14891 .byte 69,137,223 // mov %r11d,%r15d
14892 .byte 73,193,235,30 // shr $0x1e,%r11
14893 .byte 69,137,212 // mov %r10d,%r12d
14894 .byte 73,193,234,30 // shr $0x1e,%r10
14895 .byte 196,161,121,110,4,131 // vmovd (%rbx,%r8,4),%xmm0
14896 .byte 196,163,121,34,4,11,1 // vpinsrd $0x1,(%rbx,%r9,1),%xmm0,%xmm0
14897 .byte 196,163,121,34,4,179,2 // vpinsrd $0x2,(%rbx,%r14,4),%xmm0,%xmm0
14898 .byte 196,99,121,34,4,3,3 // vpinsrd $0x3,(%rbx,%rax,1),%xmm0,%xmm8
14899 .byte 196,161,121,110,4,187 // vmovd (%rbx,%r15,4),%xmm0
14900 .byte 196,163,121,34,4,27,1 // vpinsrd $0x1,(%rbx,%r11,1),%xmm0,%xmm0
14901 .byte 196,163,121,34,4,163,2 // vpinsrd $0x2,(%rbx,%r12,4),%xmm0,%xmm0
14902 .byte 196,163,121,34,28,19,3 // vpinsrd $0x3,(%rbx,%r10,1),%xmm0,%xmm3
14903 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
14904 .byte 184,255,0,0,0 // mov $0xff,%eax
14905 .byte 197,249,110,200 // vmovd %eax,%xmm1
14906 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
14907 .byte 196,99,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm11
14908 .byte 197,164,84,192 // vandps %ymm0,%ymm11,%ymm0
14909 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14910 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
14911 .byte 197,249,110,200 // vmovd %eax,%xmm1
14912 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14913 .byte 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9
14914 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
14915 .byte 196,193,41,114,208,8 // vpsrld $0x8,%xmm8,%xmm10
14916 .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1
14917 .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
14918 .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1
14919 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
14920 .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1
14921 .byte 196,193,41,114,208,16 // vpsrld $0x10,%xmm8,%xmm10
14922 .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2
14923 .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2
14924 .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2
14925 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
14926 .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2
14927 .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8
14928 .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3
14929 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
14930 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
14931 .byte 196,193,100,89,217 // vmulps %ymm9,%ymm3,%ymm3
14932 .byte 72,173 // lods %ds:(%rsi),%rax
14933 .byte 91 // pop %rbx
14934 .byte 65,92 // pop %r12
14935 .byte 65,93 // pop %r13
14936 .byte 65,94 // pop %r14
14937 .byte 65,95 // pop %r15
14938 .byte 255,224 // jmpq *%rax
14939
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]14940HIDDEN _sk_load_565_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14941.globl _sk_load_565_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]14942FUNCTION(_sk_load_565_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14943_sk_load_565_avx:
14944 .byte 72,173 // lods %ds:(%rsi),%rax
14945 .byte 76,139,16 // mov (%rax),%r10
14946 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14947 .byte 15,133,209,0,0,0 // jne 346a <_sk_load_565_avx+0xdf>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14948 .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
14949 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
14950 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
14951 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
14952 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
14953 .byte 184,0,248,0,0 // mov $0xf800,%eax
14954 .byte 197,249,110,192 // vmovd %eax,%xmm0
14955 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
14956 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
14957 .byte 197,252,84,194 // vandps %ymm2,%ymm0,%ymm0
14958 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
14959 .byte 184,8,33,132,55 // mov $0x37842108,%eax
14960 .byte 197,249,110,200 // vmovd %eax,%xmm1
14961 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
14962 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14963 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
14964 .byte 184,224,7,0,0 // mov $0x7e0,%eax
14965 .byte 197,249,110,200 // vmovd %eax,%xmm1
14966 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
14967 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
14968 .byte 197,244,84,202 // vandps %ymm2,%ymm1,%ymm1
14969 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
14970 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
14971 .byte 197,249,110,216 // vmovd %eax,%xmm3
14972 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
14973 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
14974 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
14975 .byte 184,31,0,0,0 // mov $0x1f,%eax
14976 .byte 197,249,110,216 // vmovd %eax,%xmm3
14977 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
14978 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
14979 .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
14980 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
14981 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
14982 .byte 197,249,110,216 // vmovd %eax,%xmm3
14983 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
14984 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
14985 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
14986 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
14987 .byte 197,249,110,216 // vmovd %eax,%xmm3
14988 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
14989 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
14990 .byte 72,173 // lods %ds:(%rsi),%rax
14991 .byte 255,224 // jmpq *%rax
14992 .byte 65,137,200 // mov %ecx,%r8d
14993 .byte 65,128,224,7 // and $0x7,%r8b
14994 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
14995 .byte 65,254,200 // dec %r8b
14996 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14997 .byte 15,135,29,255,255,255 // ja 339f <_sk_load_565_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]14998 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]14999 .byte 76,141,13,75,0,0,0 // lea 0x4b(%rip),%r9 # 34d8 <_sk_load_565_avx+0x14d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15000 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
15001 .byte 76,1,200 // add %r9,%rax
15002 .byte 255,224 // jmpq *%rax
15003 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
15004 .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
15005 .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
15006 .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
15007 .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
15008 .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
15009 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
15010 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15011 .byte 233,201,254,255,255 // jmpq 339f <_sk_load_565_avx+0x14>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15012 .byte 102,144 // xchg %ax,%ax
15013 .byte 242,255 // repnz (bad)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15014 .byte 255 // (bad)
15015 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15016 .byte 234 // (bad)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15017 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15018 .byte 255 // (bad)
15019 .byte 255,226 // jmpq *%rdx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15020 .byte 255 // (bad)
15021 .byte 255 // (bad)
15022 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15023 .byte 218,255 // (bad)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15024 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15025 .byte 255,210 // callq *%rdx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15026 .byte 255 // (bad)
15027 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15028 .byte 255,202 // dec %edx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15029 .byte 255 // (bad)
15030 .byte 255 // (bad)
15031 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15032 .byte 190 // .byte 0xbe
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15033 .byte 255 // (bad)
15034 .byte 255 // (bad)
15035 .byte 255 // .byte 0xff
15036
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15037HIDDEN _sk_gather_565_avx
15038.globl _sk_gather_565_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15039FUNCTION(_sk_gather_565_avx)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15040_sk_gather_565_avx:
15041 .byte 85 // push %rbp
15042 .byte 65,87 // push %r15
15043 .byte 65,86 // push %r14
15044 .byte 65,84 // push %r12
15045 .byte 83 // push %rbx
15046 .byte 72,173 // lods %ds:(%rsi),%rax
15047 .byte 76,139,0 // mov (%rax),%r8
15048 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
15049 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
15050 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
15051 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
15052 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
15053 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
15054 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
15055 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
15056 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
15057 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
15058 .byte 65,137,193 // mov %eax,%r9d
15059 .byte 72,193,232,32 // shr $0x20,%rax
15060 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
15061 .byte 69,137,211 // mov %r10d,%r11d
15062 .byte 73,193,234,32 // shr $0x20,%r10
15063 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
15064 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
15065 .byte 65,137,222 // mov %ebx,%r14d
15066 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
15067 .byte 69,137,252 // mov %r15d,%r12d
15068 .byte 73,193,239,32 // shr $0x20,%r15
15069 .byte 72,193,235,32 // shr $0x20,%rbx
15070 .byte 65,15,183,28,88 // movzwl (%r8,%rbx,2),%ebx
15071 .byte 67,15,183,44,112 // movzwl (%r8,%r14,2),%ebp
15072 .byte 197,249,110,197 // vmovd %ebp,%xmm0
15073 .byte 197,249,196,195,1 // vpinsrw $0x1,%ebx,%xmm0,%xmm0
15074 .byte 67,15,183,28,96 // movzwl (%r8,%r12,2),%ebx
15075 .byte 197,249,196,195,2 // vpinsrw $0x2,%ebx,%xmm0,%xmm0
15076 .byte 67,15,183,28,120 // movzwl (%r8,%r15,2),%ebx
15077 .byte 197,249,196,195,3 // vpinsrw $0x3,%ebx,%xmm0,%xmm0
15078 .byte 67,15,183,44,88 // movzwl (%r8,%r11,2),%ebp
15079 .byte 197,249,196,197,4 // vpinsrw $0x4,%ebp,%xmm0,%xmm0
15080 .byte 67,15,183,44,80 // movzwl (%r8,%r10,2),%ebp
15081 .byte 197,249,196,197,5 // vpinsrw $0x5,%ebp,%xmm0,%xmm0
15082 .byte 67,15,183,44,72 // movzwl (%r8,%r9,2),%ebp
15083 .byte 197,249,196,197,6 // vpinsrw $0x6,%ebp,%xmm0,%xmm0
15084 .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax
15085 .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
15086 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
15087 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
15088 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
15089 .byte 196,227,125,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm2
15090 .byte 184,0,248,0,0 // mov $0xf800,%eax
15091 .byte 197,249,110,192 // vmovd %eax,%xmm0
15092 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
15093 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
15094 .byte 197,252,84,194 // vandps %ymm2,%ymm0,%ymm0
15095 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
15096 .byte 184,8,33,132,55 // mov $0x37842108,%eax
15097 .byte 197,249,110,200 // vmovd %eax,%xmm1
15098 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
15099 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15100 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
15101 .byte 184,224,7,0,0 // mov $0x7e0,%eax
15102 .byte 197,249,110,200 // vmovd %eax,%xmm1
15103 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
15104 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15105 .byte 197,244,84,202 // vandps %ymm2,%ymm1,%ymm1
15106 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
15107 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
15108 .byte 197,249,110,216 // vmovd %eax,%xmm3
15109 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
15110 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15111 .byte 197,244,89,203 // vmulps %ymm3,%ymm1,%ymm1
15112 .byte 184,31,0,0,0 // mov $0x1f,%eax
15113 .byte 197,249,110,216 // vmovd %eax,%xmm3
15114 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
15115 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15116 .byte 197,228,84,210 // vandps %ymm2,%ymm3,%ymm2
15117 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
15118 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
15119 .byte 197,249,110,216 // vmovd %eax,%xmm3
15120 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
15121 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15122 .byte 197,236,89,211 // vmulps %ymm3,%ymm2,%ymm2
15123 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
15124 .byte 197,249,110,216 // vmovd %eax,%xmm3
15125 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
15126 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15127 .byte 72,173 // lods %ds:(%rsi),%rax
15128 .byte 91 // pop %rbx
15129 .byte 65,92 // pop %r12
15130 .byte 65,94 // pop %r14
15131 .byte 65,95 // pop %r15
15132 .byte 93 // pop %rbp
15133 .byte 255,224 // jmpq *%rax
15134
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]15135HIDDEN _sk_store_565_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15136.globl _sk_store_565_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15137FUNCTION(_sk_store_565_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15138_sk_store_565_avx:
15139 .byte 72,173 // lods %ds:(%rsi),%rax
15140 .byte 76,139,8 // mov (%rax),%r9
15141 .byte 184,0,0,248,65 // mov $0x41f80000,%eax
15142 .byte 197,121,110,192 // vmovd %eax,%xmm8
15143 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
15144 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
15145 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
15146 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
15147 .byte 196,193,41,114,241,11 // vpslld $0xb,%xmm9,%xmm10
15148 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9
15149 .byte 196,193,49,114,241,11 // vpslld $0xb,%xmm9,%xmm9
15150 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
15151 .byte 184,0,0,124,66 // mov $0x427c0000,%eax
15152 .byte 197,121,110,208 // vmovd %eax,%xmm10
15153 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
15154 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
15155 .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10
15156 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
15157 .byte 196,193,33,114,242,5 // vpslld $0x5,%xmm10,%xmm11
15158 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
15159 .byte 196,193,41,114,242,5 // vpslld $0x5,%xmm10,%xmm10
15160 .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
15161 .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9
15162 .byte 197,60,89,194 // vmulps %ymm2,%ymm8,%ymm8
15163 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
15164 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
15165 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
15166 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
15167 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15168 .byte 117,10 // jne 3723 <_sk_store_565_avx+0x9e>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15169 .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
15170 .byte 72,173 // lods %ds:(%rsi),%rax
15171 .byte 255,224 // jmpq *%rax
15172 .byte 65,137,200 // mov %ecx,%r8d
15173 .byte 65,128,224,7 // and $0x7,%r8b
15174 .byte 65,254,200 // dec %r8b
15175 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15176 .byte 119,236 // ja 371f <_sk_store_565_avx+0x9a>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15177 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15178 .byte 76,141,5,66,0,0,0 // lea 0x42(%rip),%r8 # 3780 <_sk_store_565_avx+0xfb>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15179 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
15180 .byte 76,1,192 // add %r8,%rax
15181 .byte 255,224 // jmpq *%rax
15182 .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
15183 .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
15184 .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
15185 .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
15186 .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
15187 .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
15188 .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15189 .byte 235,159 // jmp 371f <_sk_store_565_avx+0x9a>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15190 .byte 247,255 // idiv %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15191 .byte 255 // (bad)
15192 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15193 .byte 239 // out %eax,(%dx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15194 .byte 255 // (bad)
15195 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15196 .byte 255,231 // jmpq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15197 .byte 255 // (bad)
15198 .byte 255 // (bad)
15199 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15200 .byte 223,255 // (bad)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15201 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15202 .byte 255,215 // callq *%rdi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15203 .byte 255 // (bad)
15204 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15205 .byte 255,207 // dec %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15206 .byte 255 // (bad)
15207 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15208 .byte 255,199 // inc %edi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15209 .byte 255 // (bad)
15210 .byte 255 // (bad)
15211 .byte 255 // .byte 0xff
15212
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15213HIDDEN _sk_load_4444_avx
15214.globl _sk_load_4444_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15215FUNCTION(_sk_load_4444_avx)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15216_sk_load_4444_avx:
15217 .byte 72,173 // lods %ds:(%rsi),%rax
15218 .byte 76,139,16 // mov (%rax),%r10
15219 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15220 .byte 15,133,245,0,0,0 // jne 389f <_sk_load_4444_avx+0x103>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15221 .byte 196,193,122,111,4,122 // vmovdqu (%r10,%rdi,2),%xmm0
15222 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
15223 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
15224 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
15225 .byte 196,99,125,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm9
15226 .byte 184,0,240,0,0 // mov $0xf000,%eax
15227 .byte 197,249,110,192 // vmovd %eax,%xmm0
15228 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
15229 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
15230 .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0
15231 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
15232 .byte 184,137,136,136,55 // mov $0x37888889,%eax
15233 .byte 197,249,110,200 // vmovd %eax,%xmm1
15234 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
15235 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15236 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
15237 .byte 184,0,15,0,0 // mov $0xf00,%eax
15238 .byte 197,249,110,200 // vmovd %eax,%xmm1
15239 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
15240 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15241 .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1
15242 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
15243 .byte 184,137,136,136,57 // mov $0x39888889,%eax
15244 .byte 197,249,110,208 // vmovd %eax,%xmm2
15245 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
15246 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15247 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
15248 .byte 184,240,0,0,0 // mov $0xf0,%eax
15249 .byte 197,249,110,208 // vmovd %eax,%xmm2
15250 .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2
15251 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15252 .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2
15253 .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
15254 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
15255 .byte 197,249,110,208 // vmovd %eax,%xmm2
15256 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
15257 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15258 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
15259 .byte 184,15,0,0,0 // mov $0xf,%eax
15260 .byte 197,249,110,216 // vmovd %eax,%xmm3
15261 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
15262 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15263 .byte 196,193,100,84,217 // vandps %ymm9,%ymm3,%ymm3
15264 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
15265 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
15266 .byte 197,249,110,216 // vmovd %eax,%xmm3
15267 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
15268 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15269 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
15270 .byte 72,173 // lods %ds:(%rsi),%rax
15271 .byte 255,224 // jmpq *%rax
15272 .byte 65,137,200 // mov %ecx,%r8d
15273 .byte 65,128,224,7 // and $0x7,%r8b
15274 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
15275 .byte 65,254,200 // dec %r8b
15276 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15277 .byte 15,135,249,254,255,255 // ja 37b0 <_sk_load_4444_avx+0x14>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15278 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15279 .byte 76,141,13,74,0,0,0 // lea 0x4a(%rip),%r9 # 390c <_sk_load_4444_avx+0x170>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15280 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
15281 .byte 76,1,200 // add %r9,%rax
15282 .byte 255,224 // jmpq *%rax
15283 .byte 197,249,239,192 // vpxor %xmm0,%xmm0,%xmm0
15284 .byte 196,193,121,196,68,122,12,6 // vpinsrw $0x6,0xc(%r10,%rdi,2),%xmm0,%xmm0
15285 .byte 196,193,121,196,68,122,10,5 // vpinsrw $0x5,0xa(%r10,%rdi,2),%xmm0,%xmm0
15286 .byte 196,193,121,196,68,122,8,4 // vpinsrw $0x4,0x8(%r10,%rdi,2),%xmm0,%xmm0
15287 .byte 196,193,121,196,68,122,6,3 // vpinsrw $0x3,0x6(%r10,%rdi,2),%xmm0,%xmm0
15288 .byte 196,193,121,196,68,122,4,2 // vpinsrw $0x2,0x4(%r10,%rdi,2),%xmm0,%xmm0
15289 .byte 196,193,121,196,68,122,2,1 // vpinsrw $0x1,0x2(%r10,%rdi,2),%xmm0,%xmm0
15290 .byte 196,193,121,196,4,122,0 // vpinsrw $0x0,(%r10,%rdi,2),%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15291 .byte 233,165,254,255,255 // jmpq 37b0 <_sk_load_4444_avx+0x14>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15292 .byte 144 // nop
15293 .byte 243,255 // repz (bad)
15294 .byte 255 // (bad)
15295 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15296 .byte 235,255 // jmp 3911 <_sk_load_4444_avx+0x175>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15297 .byte 255 // (bad)
15298 .byte 255,227 // jmpq *%rbx
15299 .byte 255 // (bad)
15300 .byte 255 // (bad)
15301 .byte 255 // (bad)
15302 .byte 219,255 // (bad)
15303 .byte 255 // (bad)
15304 .byte 255,211 // callq *%rbx
15305 .byte 255 // (bad)
15306 .byte 255 // (bad)
15307 .byte 255,203 // dec %ebx
15308 .byte 255 // (bad)
15309 .byte 255 // (bad)
15310 .byte 255 // (bad)
15311 .byte 191 // .byte 0xbf
15312 .byte 255 // (bad)
15313 .byte 255 // (bad)
15314 .byte 255 // .byte 0xff
15315
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15316HIDDEN _sk_gather_4444_avx
15317.globl _sk_gather_4444_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15318FUNCTION(_sk_gather_4444_avx)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15319_sk_gather_4444_avx:
15320 .byte 85 // push %rbp
15321 .byte 65,87 // push %r15
15322 .byte 65,86 // push %r14
15323 .byte 65,84 // push %r12
15324 .byte 83 // push %rbx
15325 .byte 72,173 // lods %ds:(%rsi),%rax
15326 .byte 76,139,0 // mov (%rax),%r8
15327 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
15328 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
15329 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
15330 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
15331 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
15332 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
15333 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
15334 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
15335 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
15336 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
15337 .byte 65,137,193 // mov %eax,%r9d
15338 .byte 72,193,232,32 // shr $0x20,%rax
15339 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
15340 .byte 69,137,211 // mov %r10d,%r11d
15341 .byte 73,193,234,32 // shr $0x20,%r10
15342 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
15343 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
15344 .byte 65,137,222 // mov %ebx,%r14d
15345 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
15346 .byte 69,137,252 // mov %r15d,%r12d
15347 .byte 73,193,239,32 // shr $0x20,%r15
15348 .byte 72,193,235,32 // shr $0x20,%rbx
15349 .byte 65,15,183,28,88 // movzwl (%r8,%rbx,2),%ebx
15350 .byte 67,15,183,44,112 // movzwl (%r8,%r14,2),%ebp
15351 .byte 197,249,110,197 // vmovd %ebp,%xmm0
15352 .byte 197,249,196,195,1 // vpinsrw $0x1,%ebx,%xmm0,%xmm0
15353 .byte 67,15,183,28,96 // movzwl (%r8,%r12,2),%ebx
15354 .byte 197,249,196,195,2 // vpinsrw $0x2,%ebx,%xmm0,%xmm0
15355 .byte 67,15,183,28,120 // movzwl (%r8,%r15,2),%ebx
15356 .byte 197,249,196,195,3 // vpinsrw $0x3,%ebx,%xmm0,%xmm0
15357 .byte 67,15,183,44,88 // movzwl (%r8,%r11,2),%ebp
15358 .byte 197,249,196,197,4 // vpinsrw $0x4,%ebp,%xmm0,%xmm0
15359 .byte 67,15,183,44,80 // movzwl (%r8,%r10,2),%ebp
15360 .byte 197,249,196,197,5 // vpinsrw $0x5,%ebp,%xmm0,%xmm0
15361 .byte 67,15,183,44,72 // movzwl (%r8,%r9,2),%ebp
15362 .byte 197,249,196,197,6 // vpinsrw $0x6,%ebp,%xmm0,%xmm0
15363 .byte 65,15,183,4,64 // movzwl (%r8,%rax,2),%eax
15364 .byte 197,249,196,192,7 // vpinsrw $0x7,%eax,%xmm0,%xmm0
15365 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
15366 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
15367 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
15368 .byte 196,99,125,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm9
15369 .byte 184,0,240,0,0 // mov $0xf000,%eax
15370 .byte 197,249,110,192 // vmovd %eax,%xmm0
15371 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
15372 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
15373 .byte 196,193,124,84,193 // vandps %ymm9,%ymm0,%ymm0
15374 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
15375 .byte 184,137,136,136,55 // mov $0x37888889,%eax
15376 .byte 197,249,110,200 // vmovd %eax,%xmm1
15377 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
15378 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15379 .byte 197,252,89,193 // vmulps %ymm1,%ymm0,%ymm0
15380 .byte 184,0,15,0,0 // mov $0xf00,%eax
15381 .byte 197,249,110,200 // vmovd %eax,%xmm1
15382 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
15383 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
15384 .byte 196,193,116,84,201 // vandps %ymm9,%ymm1,%ymm1
15385 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
15386 .byte 184,137,136,136,57 // mov $0x39888889,%eax
15387 .byte 197,249,110,208 // vmovd %eax,%xmm2
15388 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
15389 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15390 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
15391 .byte 184,240,0,0,0 // mov $0xf0,%eax
15392 .byte 197,249,110,208 // vmovd %eax,%xmm2
15393 .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2
15394 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15395 .byte 196,193,108,84,209 // vandps %ymm9,%ymm2,%ymm2
15396 .byte 197,124,91,194 // vcvtdq2ps %ymm2,%ymm8
15397 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
15398 .byte 197,249,110,208 // vmovd %eax,%xmm2
15399 .byte 196,227,121,4,210,0 // vpermilps $0x0,%xmm2,%xmm2
15400 .byte 196,227,109,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm2,%ymm2
15401 .byte 197,188,89,210 // vmulps %ymm2,%ymm8,%ymm2
15402 .byte 184,15,0,0,0 // mov $0xf,%eax
15403 .byte 197,249,110,216 // vmovd %eax,%xmm3
15404 .byte 197,249,112,219,0 // vpshufd $0x0,%xmm3,%xmm3
15405 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15406 .byte 196,193,100,84,217 // vandps %ymm9,%ymm3,%ymm3
15407 .byte 197,124,91,195 // vcvtdq2ps %ymm3,%ymm8
15408 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
15409 .byte 197,249,110,216 // vmovd %eax,%xmm3
15410 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
15411 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
15412 .byte 197,188,89,219 // vmulps %ymm3,%ymm8,%ymm3
15413 .byte 72,173 // lods %ds:(%rsi),%rax
15414 .byte 91 // pop %rbx
15415 .byte 65,92 // pop %r12
15416 .byte 65,94 // pop %r14
15417 .byte 65,95 // pop %r15
15418 .byte 93 // pop %rbp
15419 .byte 255,224 // jmpq *%rax
15420
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15421HIDDEN _sk_store_4444_avx
15422.globl _sk_store_4444_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15423FUNCTION(_sk_store_4444_avx)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15424_sk_store_4444_avx:
15425 .byte 72,173 // lods %ds:(%rsi),%rax
15426 .byte 76,139,8 // mov (%rax),%r9
15427 .byte 184,0,0,112,65 // mov $0x41700000,%eax
15428 .byte 197,121,110,192 // vmovd %eax,%xmm8
15429 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
15430 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
15431 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
15432 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
15433 .byte 196,193,41,114,241,12 // vpslld $0xc,%xmm9,%xmm10
15434 .byte 196,67,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm9
15435 .byte 196,193,49,114,241,12 // vpslld $0xc,%xmm9,%xmm9
15436 .byte 196,67,45,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm9
15437 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
15438 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
15439 .byte 196,193,33,114,242,8 // vpslld $0x8,%xmm10,%xmm11
15440 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
15441 .byte 196,193,41,114,242,8 // vpslld $0x8,%xmm10,%xmm10
15442 .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
15443 .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9
15444 .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
15445 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
15446 .byte 196,193,33,114,242,4 // vpslld $0x4,%xmm10,%xmm11
15447 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
15448 .byte 196,193,41,114,242,4 // vpslld $0x4,%xmm10,%xmm10
15449 .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
15450 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
15451 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
15452 .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8
15453 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
15454 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
15455 .byte 196,66,57,43,193 // vpackusdw %xmm9,%xmm8,%xmm8
15456 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15457 .byte 117,10 // jne 3b8c <_sk_store_4444_avx+0xaf>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15458 .byte 196,65,122,127,4,121 // vmovdqu %xmm8,(%r9,%rdi,2)
15459 .byte 72,173 // lods %ds:(%rsi),%rax
15460 .byte 255,224 // jmpq *%rax
15461 .byte 65,137,200 // mov %ecx,%r8d
15462 .byte 65,128,224,7 // and $0x7,%r8b
15463 .byte 65,254,200 // dec %r8b
15464 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15465 .byte 119,236 // ja 3b88 <_sk_store_4444_avx+0xab>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15466 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15467 .byte 76,141,5,69,0,0,0 // lea 0x45(%rip),%r8 # 3bec <_sk_store_4444_avx+0x10f>
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15468 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
15469 .byte 76,1,192 // add %r8,%rax
15470 .byte 255,224 // jmpq *%rax
15471 .byte 196,67,121,21,68,121,12,6 // vpextrw $0x6,%xmm8,0xc(%r9,%rdi,2)
15472 .byte 196,67,121,21,68,121,10,5 // vpextrw $0x5,%xmm8,0xa(%r9,%rdi,2)
15473 .byte 196,67,121,21,68,121,8,4 // vpextrw $0x4,%xmm8,0x8(%r9,%rdi,2)
15474 .byte 196,67,121,21,68,121,6,3 // vpextrw $0x3,%xmm8,0x6(%r9,%rdi,2)
15475 .byte 196,67,121,21,68,121,4,2 // vpextrw $0x2,%xmm8,0x4(%r9,%rdi,2)
15476 .byte 196,67,121,21,68,121,2,1 // vpextrw $0x1,%xmm8,0x2(%r9,%rdi,2)
15477 .byte 196,67,121,21,4,121,0 // vpextrw $0x0,%xmm8,(%r9,%rdi,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15478 .byte 235,159 // jmp 3b88 <_sk_store_4444_avx+0xab>
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15479 .byte 15,31,0 // nopl (%rax)
15480 .byte 244 // hlt
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15481 .byte 255 // (bad)
15482 .byte 255 // (bad)
15483 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15484 .byte 236 // in (%dx),%al
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15485 .byte 255 // (bad)
15486 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15487 .byte 255,228 // jmpq *%rsp
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15488 .byte 255 // (bad)
15489 .byte 255 // (bad)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]15490 .byte 255 // (bad)
15491 .byte 220,255 // fdivr %st,%st(7)
15492 .byte 255 // (bad)
15493 .byte 255,212 // callq *%rsp
15494 .byte 255 // (bad)
15495 .byte 255 // (bad)
15496 .byte 255,204 // dec %esp
15497 .byte 255 // (bad)
15498 .byte 255 // (bad)
15499 .byte 255,196 // inc %esp
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]15500 .byte 255 // (bad)
15501 .byte 255 // (bad)
15502 .byte 255 // .byte 0xff
15503
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]15504HIDDEN _sk_load_8888_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15505.globl _sk_load_8888_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15506FUNCTION(_sk_load_8888_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15507_sk_load_8888_avx:
15508 .byte 72,173 // lods %ds:(%rsi),%rax
15509 .byte 76,139,16 // mov (%rax),%r10
15510 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15511 .byte 15,133,157,0,0,0 // jne 3cb3 <_sk_load_8888_avx+0xab>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15512 .byte 196,65,124,16,12,186 // vmovups (%r10,%rdi,4),%ymm9
15513 .byte 184,255,0,0,0 // mov $0xff,%eax
15514 .byte 197,249,110,192 // vmovd %eax,%xmm0
15515 .byte 197,249,112,192,0 // vpshufd $0x0,%xmm0,%xmm0
15516 .byte 196,99,125,24,216,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm11
15517 .byte 196,193,36,84,193 // vandps %ymm9,%ymm11,%ymm0
15518 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
15519 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
15520 .byte 197,249,110,200 // vmovd %eax,%xmm1
15521 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
15522 .byte 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8
15523 .byte 196,193,124,89,192 // vmulps %ymm8,%ymm0,%ymm0
15524 .byte 196,193,41,114,209,8 // vpsrld $0x8,%xmm9,%xmm10
15525 .byte 196,99,125,25,203,1 // vextractf128 $0x1,%ymm9,%xmm3
15526 .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1
15527 .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
15528 .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1
15529 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
15530 .byte 196,193,116,89,200 // vmulps %ymm8,%ymm1,%ymm1
15531 .byte 196,193,41,114,209,16 // vpsrld $0x10,%xmm9,%xmm10
15532 .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2
15533 .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2
15534 .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2
15535 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
15536 .byte 196,193,108,89,208 // vmulps %ymm8,%ymm2,%ymm2
15537 .byte 196,193,49,114,209,24 // vpsrld $0x18,%xmm9,%xmm9
15538 .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3
15539 .byte 196,227,53,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm9,%ymm3
15540 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
15541 .byte 196,193,100,89,216 // vmulps %ymm8,%ymm3,%ymm3
15542 .byte 72,173 // lods %ds:(%rsi),%rax
15543 .byte 255,224 // jmpq *%rax
15544 .byte 65,137,200 // mov %ecx,%r8d
15545 .byte 65,128,224,7 // and $0x7,%r8b
15546 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
15547 .byte 65,254,200 // dec %r8b
15548 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15549 .byte 15,135,80,255,255,255 // ja 3c1c <_sk_load_8888_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15550 .byte 69,15,182,192 // movzbl %r8b,%r8d
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15551 .byte 76,141,13,137,0,0,0 // lea 0x89(%rip),%r9 # 3d60 <_sk_load_8888_avx+0x158>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15552 .byte 75,99,4,129 // movslq (%r9,%r8,4),%rax
15553 .byte 76,1,200 // add %r9,%rax
15554 .byte 255,224 // jmpq *%rax
15555 .byte 196,193,121,110,68,186,24 // vmovd 0x18(%r10,%rdi,4),%xmm0
15556 .byte 197,249,112,192,68 // vpshufd $0x44,%xmm0,%xmm0
15557 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
15558 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
15559 .byte 196,99,117,12,200,64 // vblendps $0x40,%ymm0,%ymm1,%ymm9
15560 .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0
15561 .byte 196,195,121,34,68,186,20,1 // vpinsrd $0x1,0x14(%r10,%rdi,4),%xmm0,%xmm0
15562 .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9
15563 .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0
15564 .byte 196,195,121,34,68,186,16,0 // vpinsrd $0x0,0x10(%r10,%rdi,4),%xmm0,%xmm0
15565 .byte 196,99,53,24,200,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm9
15566 .byte 196,195,49,34,68,186,12,3 // vpinsrd $0x3,0xc(%r10,%rdi,4),%xmm9,%xmm0
15567 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
15568 .byte 196,195,49,34,68,186,8,2 // vpinsrd $0x2,0x8(%r10,%rdi,4),%xmm9,%xmm0
15569 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
15570 .byte 196,195,49,34,68,186,4,1 // vpinsrd $0x1,0x4(%r10,%rdi,4),%xmm9,%xmm0
15571 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
15572 .byte 196,195,49,34,4,186,0 // vpinsrd $0x0,(%r10,%rdi,4),%xmm9,%xmm0
15573 .byte 196,99,53,12,200,15 // vblendps $0xf,%ymm0,%ymm9,%ymm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15574 .byte 233,188,254,255,255 // jmpq 3c1c <_sk_load_8888_avx+0x14>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15575 .byte 238 // out %al,(%dx)
15576 .byte 255 // (bad)
15577 .byte 255 // (bad)
15578 .byte 255,224 // jmpq *%rax
15579 .byte 255 // (bad)
15580 .byte 255 // (bad)
15581 .byte 255,210 // callq *%rdx
15582 .byte 255 // (bad)
15583 .byte 255 // (bad)
15584 .byte 255,196 // inc %esp
15585 .byte 255 // (bad)
15586 .byte 255 // (bad)
15587 .byte 255,176,255,255,255,156 // pushq -0x63000001(%rax)
15588 .byte 255 // (bad)
15589 .byte 255 // (bad)
15590 .byte 255 // .byte 0xff
15591 .byte 128,255,255 // cmp $0xff,%bh
15592 .byte 255 // .byte 0xff
15593
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15594HIDDEN _sk_gather_8888_avx
15595.globl _sk_gather_8888_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15596FUNCTION(_sk_gather_8888_avx)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15597_sk_gather_8888_avx:
15598 .byte 65,87 // push %r15
15599 .byte 65,86 // push %r14
15600 .byte 65,84 // push %r12
15601 .byte 83 // push %rbx
15602 .byte 72,173 // lods %ds:(%rsi),%rax
15603 .byte 76,139,0 // mov (%rax),%r8
15604 .byte 197,254,91,201 // vcvttps2dq %ymm1,%ymm1
15605 .byte 197,249,110,80,16 // vmovd 0x10(%rax),%xmm2
15606 .byte 197,249,112,210,0 // vpshufd $0x0,%xmm2,%xmm2
15607 .byte 196,226,105,64,217 // vpmulld %xmm1,%xmm2,%xmm3
15608 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
15609 .byte 196,226,105,64,201 // vpmulld %xmm1,%xmm2,%xmm1
15610 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
15611 .byte 196,227,125,25,194,1 // vextractf128 $0x1,%ymm0,%xmm2
15612 .byte 197,241,254,202 // vpaddd %xmm2,%xmm1,%xmm1
15613 .byte 196,225,249,126,200 // vmovq %xmm1,%rax
15614 .byte 65,137,193 // mov %eax,%r9d
15615 .byte 72,193,232,32 // shr $0x20,%rax
15616 .byte 196,195,249,22,202,1 // vpextrq $0x1,%xmm1,%r10
15617 .byte 69,137,211 // mov %r10d,%r11d
15618 .byte 73,193,234,32 // shr $0x20,%r10
15619 .byte 197,225,254,192 // vpaddd %xmm0,%xmm3,%xmm0
15620 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
15621 .byte 65,137,222 // mov %ebx,%r14d
15622 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
15623 .byte 69,137,252 // mov %r15d,%r12d
15624 .byte 72,193,235,32 // shr $0x20,%rbx
15625 .byte 73,193,239,32 // shr $0x20,%r15
15626 .byte 196,129,121,110,4,176 // vmovd (%r8,%r14,4),%xmm0
15627 .byte 196,195,121,34,4,152,1 // vpinsrd $0x1,(%r8,%rbx,4),%xmm0,%xmm0
15628 .byte 196,131,121,34,4,160,2 // vpinsrd $0x2,(%r8,%r12,4),%xmm0,%xmm0
15629 .byte 196,3,121,34,4,184,3 // vpinsrd $0x3,(%r8,%r15,4),%xmm0,%xmm8
15630 .byte 196,129,121,110,4,136 // vmovd (%r8,%r9,4),%xmm0
15631 .byte 196,195,121,34,4,128,1 // vpinsrd $0x1,(%r8,%rax,4),%xmm0,%xmm0
15632 .byte 196,131,121,34,4,152,2 // vpinsrd $0x2,(%r8,%r11,4),%xmm0,%xmm0
15633 .byte 196,131,121,34,28,144,3 // vpinsrd $0x3,(%r8,%r10,4),%xmm0,%xmm3
15634 .byte 196,227,61,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm0
15635 .byte 184,255,0,0,0 // mov $0xff,%eax
15636 .byte 197,249,110,200 // vmovd %eax,%xmm1
15637 .byte 197,249,112,201,0 // vpshufd $0x0,%xmm1,%xmm1
15638 .byte 196,99,117,24,217,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm11
15639 .byte 197,164,84,192 // vandps %ymm0,%ymm11,%ymm0
15640 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
15641 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
15642 .byte 197,249,110,200 // vmovd %eax,%xmm1
15643 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
15644 .byte 196,99,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm9
15645 .byte 196,193,124,89,193 // vmulps %ymm9,%ymm0,%ymm0
15646 .byte 196,193,41,114,208,8 // vpsrld $0x8,%xmm8,%xmm10
15647 .byte 197,241,114,211,8 // vpsrld $0x8,%xmm3,%xmm1
15648 .byte 196,227,45,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm10,%ymm1
15649 .byte 197,164,84,201 // vandps %ymm1,%ymm11,%ymm1
15650 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
15651 .byte 196,193,116,89,201 // vmulps %ymm9,%ymm1,%ymm1
15652 .byte 196,193,41,114,208,16 // vpsrld $0x10,%xmm8,%xmm10
15653 .byte 197,233,114,211,16 // vpsrld $0x10,%xmm3,%xmm2
15654 .byte 196,227,45,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm10,%ymm2
15655 .byte 197,164,84,210 // vandps %ymm2,%ymm11,%ymm2
15656 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
15657 .byte 196,193,108,89,209 // vmulps %ymm9,%ymm2,%ymm2
15658 .byte 196,193,57,114,208,24 // vpsrld $0x18,%xmm8,%xmm8
15659 .byte 197,225,114,211,24 // vpsrld $0x18,%xmm3,%xmm3
15660 .byte 196,227,61,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm8,%ymm3
15661 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
15662 .byte 196,193,100,89,217 // vmulps %ymm9,%ymm3,%ymm3
15663 .byte 72,173 // lods %ds:(%rsi),%rax
15664 .byte 91 // pop %rbx
15665 .byte 65,92 // pop %r12
15666 .byte 65,94 // pop %r14
15667 .byte 65,95 // pop %r15
15668 .byte 255,224 // jmpq *%rax
15669
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]15670HIDDEN _sk_store_8888_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15671.globl _sk_store_8888_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15672FUNCTION(_sk_store_8888_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15673_sk_store_8888_avx:
15674 .byte 72,173 // lods %ds:(%rsi),%rax
15675 .byte 76,139,8 // mov (%rax),%r9
15676 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
15677 .byte 197,121,110,192 // vmovd %eax,%xmm8
15678 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
15679 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
15680 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
15681 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
15682 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
15683 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
15684 .byte 196,193,33,114,242,8 // vpslld $0x8,%xmm10,%xmm11
15685 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
15686 .byte 196,193,41,114,242,8 // vpslld $0x8,%xmm10,%xmm10
15687 .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
15688 .byte 196,65,45,86,201 // vorpd %ymm9,%ymm10,%ymm9
15689 .byte 197,60,89,210 // vmulps %ymm2,%ymm8,%ymm10
15690 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
15691 .byte 196,193,33,114,242,16 // vpslld $0x10,%xmm10,%xmm11
15692 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
15693 .byte 196,193,41,114,242,16 // vpslld $0x10,%xmm10,%xmm10
15694 .byte 196,67,37,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm11,%ymm10
15695 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
15696 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
15697 .byte 196,193,33,114,240,24 // vpslld $0x18,%xmm8,%xmm11
15698 .byte 196,67,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm8
15699 .byte 196,193,57,114,240,24 // vpslld $0x18,%xmm8,%xmm8
15700 .byte 196,67,37,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm11,%ymm8
15701 .byte 196,65,45,86,192 // vorpd %ymm8,%ymm10,%ymm8
15702 .byte 196,65,53,86,192 // vorpd %ymm8,%ymm9,%ymm8
15703 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15704 .byte 117,10 // jne 3f61 <_sk_store_8888_avx+0xa4>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15705 .byte 196,65,124,17,4,185 // vmovups %ymm8,(%r9,%rdi,4)
15706 .byte 72,173 // lods %ds:(%rsi),%rax
15707 .byte 255,224 // jmpq *%rax
15708 .byte 65,137,200 // mov %ecx,%r8d
15709 .byte 65,128,224,7 // and $0x7,%r8b
15710 .byte 65,254,200 // dec %r8b
15711 .byte 65,128,248,6 // cmp $0x6,%r8b
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15712 .byte 119,236 // ja 3f5d <_sk_store_8888_avx+0xa0>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15713 .byte 65,15,182,192 // movzbl %r8b,%eax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15714 .byte 76,141,5,84,0,0,0 // lea 0x54(%rip),%r8 # 3fd0 <_sk_store_8888_avx+0x113>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15715 .byte 73,99,4,128 // movslq (%r8,%rax,4),%rax
15716 .byte 76,1,192 // add %r8,%rax
15717 .byte 255,224 // jmpq *%rax
15718 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
15719 .byte 196,67,121,22,76,185,24,2 // vpextrd $0x2,%xmm9,0x18(%r9,%rdi,4)
15720 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
15721 .byte 196,67,121,22,76,185,20,1 // vpextrd $0x1,%xmm9,0x14(%r9,%rdi,4)
15722 .byte 196,67,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm9
15723 .byte 196,65,122,17,76,185,16 // vmovss %xmm9,0x10(%r9,%rdi,4)
15724 .byte 196,67,121,22,68,185,12,3 // vpextrd $0x3,%xmm8,0xc(%r9,%rdi,4)
15725 .byte 196,67,121,22,68,185,8,2 // vpextrd $0x2,%xmm8,0x8(%r9,%rdi,4)
15726 .byte 196,67,121,22,68,185,4,1 // vpextrd $0x1,%xmm8,0x4(%r9,%rdi,4)
15727 .byte 196,65,121,126,4,185 // vmovd %xmm8,(%r9,%rdi,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15728 .byte 235,143 // jmp 3f5d <_sk_store_8888_avx+0xa0>
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15729 .byte 102,144 // xchg %ax,%ax
15730 .byte 246,255 // idiv %bh
15731 .byte 255 // (bad)
15732 .byte 255 // (bad)
15733 .byte 238 // out %al,(%dx)
15734 .byte 255 // (bad)
15735 .byte 255 // (bad)
15736 .byte 255,230 // jmpq *%rsi
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15737 .byte 255 // (bad)
15738 .byte 255 // (bad)
15739 .byte 255 // (bad)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15740 .byte 222,255 // fdivrp %st,%st(7)
15741 .byte 255 // (bad)
15742 .byte 255,209 // callq *%rcx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15743 .byte 255 // (bad)
15744 .byte 255 // (bad)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15745 .byte 255,195 // inc %ebx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15746 .byte 255 // (bad)
15747 .byte 255 // (bad)
15748 .byte 255 // .byte 0xff
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]15749 .byte 181,255 // mov $0xff,%ch
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15750 .byte 255 // (bad)
15751 .byte 255 // .byte 0xff
15752
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]15753HIDDEN _sk_load_f16_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15754.globl _sk_load_f16_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15755FUNCTION(_sk_load_f16_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15756_sk_load_f16_avx:
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15757 .byte 72,131,236,24 // sub $0x18,%rsp
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15758 .byte 72,173 // lods %ds:(%rsi),%rax
15759 .byte 72,139,0 // mov (%rax),%rax
15760 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15761 .byte 197,252,17,124,36,224 // vmovups %ymm7,-0x20(%rsp)
15762 .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp)
15763 .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp)
15764 .byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp)
15765 .byte 15,133,141,2,0,0 // jne 42a3 <_sk_load_f16_avx+0x2b7>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15766 .byte 197,121,16,4,248 // vmovupd (%rax,%rdi,8),%xmm8
15767 .byte 197,249,16,84,248,16 // vmovupd 0x10(%rax,%rdi,8),%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15768 .byte 197,249,16,76,248,32 // vmovupd 0x20(%rax,%rdi,8),%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15769 .byte 197,122,111,76,248,48 // vmovdqu 0x30(%rax,%rdi,8),%xmm9
15770 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
15771 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15772 .byte 196,193,113,97,217 // vpunpcklwd %xmm9,%xmm1,%xmm3
15773 .byte 196,193,113,105,201 // vpunpckhwd %xmm9,%xmm1,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15774 .byte 197,121,97,250 // vpunpcklwd %xmm2,%xmm0,%xmm15
Mike Klein114e6b32017-04-03 22:21:15 -0400[diff] [blame]15775 .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15776 .byte 197,225,97,209 // vpunpcklwd %xmm1,%xmm3,%xmm2
15777 .byte 197,97,105,201 // vpunpckhwd %xmm1,%xmm3,%xmm9
15778 .byte 197,129,108,194 // vpunpcklqdq %xmm2,%xmm15,%xmm0
15779 .byte 197,241,239,201 // vpxor %xmm1,%xmm1,%xmm1
15780 .byte 197,249,105,201 // vpunpckhwd %xmm1,%xmm0,%xmm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15781 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
15782 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15783 .byte 196,98,125,24,37,169,28,0,0 // vbroadcastss 0x1ca9(%rip),%ymm12 # 5d18 <_sk_callback_avx+0x1ee>
15784 .byte 196,193,124,84,204 // vandps %ymm12,%ymm0,%ymm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15785 .byte 197,252,87,193 // vxorps %ymm1,%ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15786 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14
15787 .byte 196,98,121,24,29,149,28,0,0 // vbroadcastss 0x1c95(%rip),%xmm11 # 5d1c <_sk_callback_avx+0x1f2>
15788 .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
15789 .byte 196,98,121,24,45,139,28,0,0 // vbroadcastss 0x1c8b(%rip),%xmm13 # 5d20 <_sk_callback_avx+0x1f6>
15790 .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3
15791 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10
15792 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10
15793 .byte 196,99,45,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm10,%ymm10
15794 .byte 197,225,114,241,16 // vpslld $0x10,%xmm1,%xmm3
15795 .byte 196,227,125,25,201,1 // vextractf128 $0x1,%ymm1,%xmm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15796 .byte 197,241,114,241,16 // vpslld $0x10,%xmm1,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15797 .byte 196,227,101,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm3,%ymm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15798 .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15799 .byte 196,193,97,114,246,13 // vpslld $0xd,%xmm14,%xmm3
15800 .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
15801 .byte 197,252,86,193 // vorps %ymm1,%ymm0,%ymm0
15802 .byte 196,227,125,25,193,1 // vextractf128 $0x1,%ymm0,%xmm1
15803 .byte 196,226,121,24,29,65,28,0,0 // vbroadcastss 0x1c41(%rip),%xmm3 # 5d24 <_sk_callback_avx+0x1fa>
15804 .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1
15805 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0
15806 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
15807 .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14
15808 .byte 196,195,125,74,198,160 // vblendvps %ymm10,%ymm14,%ymm0,%ymm0
15809 .byte 197,129,109,202 // vpunpckhqdq %xmm2,%xmm15,%xmm1
15810 .byte 197,217,239,228 // vpxor %xmm4,%xmm4,%xmm4
15811 .byte 197,241,105,212 // vpunpckhwd %xmm4,%xmm1,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15812 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
15813 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15814 .byte 196,193,116,84,212 // vandps %ymm12,%ymm1,%ymm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15815 .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15816 .byte 196,195,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm10
15817 .byte 196,193,40,87,251 // vxorps %xmm11,%xmm10,%xmm7
15818 .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7
15819 .byte 196,193,112,87,243 // vxorps %xmm11,%xmm1,%xmm6
15820 .byte 197,145,102,246 // vpcmpgtd %xmm6,%xmm13,%xmm6
15821 .byte 196,227,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm6
15822 .byte 197,193,114,242,16 // vpslld $0x10,%xmm2,%xmm7
15823 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15824 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15825 .byte 196,227,69,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm7,%ymm2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15826 .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15827 .byte 196,193,65,114,242,13 // vpslld $0xd,%xmm10,%xmm7
15828 .byte 196,227,117,24,207,1 // vinsertf128 $0x1,%xmm7,%ymm1,%ymm1
15829 .byte 197,244,86,202 // vorps %ymm2,%ymm1,%ymm1
15830 .byte 196,227,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm2
15831 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
15832 .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1
15833 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
15834 .byte 196,195,117,74,206,96 // vblendvps %ymm6,%ymm14,%ymm1,%ymm1
15835 .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2
15836 .byte 197,233,105,244 // vpunpckhwd %xmm4,%xmm2,%xmm6
15837 .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15838 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
15839 .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15840 .byte 196,193,108,84,244 // vandps %ymm12,%ymm2,%ymm6
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]15841 .byte 197,236,87,214 // vxorps %ymm6,%ymm2,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15842 .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7
15843 .byte 196,193,64,87,235 // vxorps %xmm11,%xmm7,%xmm5
15844 .byte 197,145,102,237 // vpcmpgtd %xmm5,%xmm13,%xmm5
15845 .byte 196,193,104,87,227 // vxorps %xmm11,%xmm2,%xmm4
15846 .byte 197,145,102,228 // vpcmpgtd %xmm4,%xmm13,%xmm4
15847 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4
15848 .byte 197,209,114,246,16 // vpslld $0x10,%xmm6,%xmm5
15849 .byte 196,227,125,25,246,1 // vextractf128 $0x1,%ymm6,%xmm6
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]15850 .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15851 .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]15852 .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15853 .byte 197,201,114,247,13 // vpslld $0xd,%xmm7,%xmm6
15854 .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2
15855 .byte 197,236,86,213 // vorps %ymm5,%ymm2,%ymm2
15856 .byte 196,227,125,25,213,1 // vextractf128 $0x1,%ymm2,%xmm5
15857 .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5
15858 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
15859 .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2
15860 .byte 196,195,109,74,214,64 // vblendvps %ymm4,%ymm14,%ymm2,%ymm2
15861 .byte 196,193,57,109,225 // vpunpckhqdq %xmm9,%xmm8,%xmm4
15862 .byte 196,193,89,105,234 // vpunpckhwd %xmm10,%xmm4,%xmm5
15863 .byte 196,226,121,51,228 // vpmovzxwd %xmm4,%xmm4
15864 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4
15865 .byte 196,193,92,84,236 // vandps %ymm12,%ymm4,%ymm5
15866 .byte 197,220,87,229 // vxorps %ymm5,%ymm4,%ymm4
15867 .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6
15868 .byte 196,193,72,87,251 // vxorps %xmm11,%xmm6,%xmm7
15869 .byte 197,17,102,199 // vpcmpgtd %xmm7,%xmm13,%xmm8
15870 .byte 196,193,88,87,251 // vxorps %xmm11,%xmm4,%xmm7
15871 .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7
15872 .byte 196,195,69,24,248,1 // vinsertf128 $0x1,%xmm8,%ymm7,%ymm7
15873 .byte 197,185,114,245,16 // vpslld $0x10,%xmm5,%xmm8
15874 .byte 196,227,125,25,237,1 // vextractf128 $0x1,%ymm5,%xmm5
15875 .byte 197,209,114,245,16 // vpslld $0x10,%xmm5,%xmm5
15876 .byte 196,227,61,24,237,1 // vinsertf128 $0x1,%xmm5,%ymm8,%ymm5
15877 .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4
15878 .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6
15879 .byte 196,227,93,24,230,1 // vinsertf128 $0x1,%xmm6,%ymm4,%ymm4
15880 .byte 197,220,86,229 // vorps %ymm5,%ymm4,%ymm4
15881 .byte 196,227,125,25,229,1 // vextractf128 $0x1,%ymm4,%xmm5
15882 .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5
15883 .byte 197,217,254,219 // vpaddd %xmm3,%xmm4,%xmm3
15884 .byte 196,227,101,24,221,1 // vinsertf128 $0x1,%xmm5,%ymm3,%ymm3
15885 .byte 196,195,101,74,222,112 // vblendvps %ymm7,%ymm14,%ymm3,%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15886 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15887 .byte 197,252,16,100,36,128 // vmovups -0x80(%rsp),%ymm4
15888 .byte 197,252,16,108,36,160 // vmovups -0x60(%rsp),%ymm5
15889 .byte 197,252,16,116,36,192 // vmovups -0x40(%rsp),%ymm6
15890 .byte 197,252,16,124,36,224 // vmovups -0x20(%rsp),%ymm7
15891 .byte 72,131,196,24 // add $0x18,%rsp
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15892 .byte 255,224 // jmpq *%rax
15893 .byte 197,123,16,4,248 // vmovsd (%rax,%rdi,8),%xmm8
15894 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
15895 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15896 .byte 116,79 // je 4302 <_sk_load_f16_avx+0x316>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15897 .byte 197,57,22,68,248,8 // vmovhpd 0x8(%rax,%rdi,8),%xmm8,%xmm8
15898 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15899 .byte 114,67 // jb 4302 <_sk_load_f16_avx+0x316>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15900 .byte 197,251,16,84,248,16 // vmovsd 0x10(%rax,%rdi,8),%xmm2
15901 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15902 .byte 116,68 // je 430f <_sk_load_f16_avx+0x323>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15903 .byte 197,233,22,84,248,24 // vmovhpd 0x18(%rax,%rdi,8),%xmm2,%xmm2
15904 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15905 .byte 114,56 // jb 430f <_sk_load_f16_avx+0x323>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15906 .byte 197,251,16,76,248,32 // vmovsd 0x20(%rax,%rdi,8),%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15907 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15908 .byte 15,132,70,253,255,255 // je 402d <_sk_load_f16_avx+0x41>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15909 .byte 197,241,22,76,248,40 // vmovhpd 0x28(%rax,%rdi,8),%xmm1,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15910 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15911 .byte 15,130,54,253,255,255 // jb 402d <_sk_load_f16_avx+0x41>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15912 .byte 197,122,126,76,248,48 // vmovq 0x30(%rax,%rdi,8),%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15913 .byte 233,43,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15914 .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15915 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15916 .byte 233,30,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15917 .byte 197,241,87,201 // vxorpd %xmm1,%xmm1,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15918 .byte 233,21,253,255,255 // jmpq 402d <_sk_load_f16_avx+0x41>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]15919
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]15920HIDDEN _sk_gather_f16_avx
15921.globl _sk_gather_f16_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]15922FUNCTION(_sk_gather_f16_avx)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]15923_sk_gather_f16_avx:
15924 .byte 65,87 // push %r15
15925 .byte 65,86 // push %r14
15926 .byte 65,84 // push %r12
15927 .byte 83 // push %rbx
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15928 .byte 72,131,236,24 // sub $0x18,%rsp
15929 .byte 197,252,17,124,36,224 // vmovups %ymm7,-0x20(%rsp)
15930 .byte 197,252,17,116,36,192 // vmovups %ymm6,-0x40(%rsp)
15931 .byte 197,252,17,108,36,160 // vmovups %ymm5,-0x60(%rsp)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15932 .byte 197,254,127,100,36,128 // vmovdqu %ymm4,-0x80(%rsp)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]15933 .byte 72,173 // lods %ds:(%rsi),%rax
15934 .byte 76,139,0 // mov (%rax),%r8
15935 .byte 197,254,91,209 // vcvttps2dq %ymm1,%ymm2
15936 .byte 197,249,110,72,16 // vmovd 0x10(%rax),%xmm1
15937 .byte 197,249,112,217,0 // vpshufd $0x0,%xmm1,%xmm3
15938 .byte 196,226,97,64,202 // vpmulld %xmm2,%xmm3,%xmm1
15939 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
15940 .byte 196,226,97,64,210 // vpmulld %xmm2,%xmm3,%xmm2
15941 .byte 197,254,91,192 // vcvttps2dq %ymm0,%ymm0
15942 .byte 196,227,125,25,195,1 // vextractf128 $0x1,%ymm0,%xmm3
15943 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
15944 .byte 196,227,249,22,208,1 // vpextrq $0x1,%xmm2,%rax
15945 .byte 65,137,193 // mov %eax,%r9d
15946 .byte 72,193,232,32 // shr $0x20,%rax
15947 .byte 196,193,249,126,210 // vmovq %xmm2,%r10
15948 .byte 69,137,211 // mov %r10d,%r11d
15949 .byte 73,193,234,32 // shr $0x20,%r10
15950 .byte 197,241,254,192 // vpaddd %xmm0,%xmm1,%xmm0
15951 .byte 196,225,249,126,195 // vmovq %xmm0,%rbx
15952 .byte 65,137,222 // mov %ebx,%r14d
15953 .byte 196,195,249,22,199,1 // vpextrq $0x1,%xmm0,%r15
15954 .byte 69,137,252 // mov %r15d,%r12d
15955 .byte 73,193,239,32 // shr $0x20,%r15
15956 .byte 72,193,235,32 // shr $0x20,%rbx
15957 .byte 196,193,122,126,4,216 // vmovq (%r8,%rbx,8),%xmm0
15958 .byte 196,129,122,126,12,240 // vmovq (%r8,%r14,8),%xmm1
15959 .byte 197,113,108,200 // vpunpcklqdq %xmm0,%xmm1,%xmm9
15960 .byte 196,129,122,126,12,248 // vmovq (%r8,%r15,8),%xmm1
15961 .byte 196,129,122,126,20,224 // vmovq (%r8,%r12,8),%xmm2
15962 .byte 197,233,108,201 // vpunpcklqdq %xmm1,%xmm2,%xmm1
15963 .byte 196,129,122,126,20,208 // vmovq (%r8,%r10,8),%xmm2
15964 .byte 196,129,122,126,28,216 // vmovq (%r8,%r11,8),%xmm3
15965 .byte 197,97,108,210 // vpunpcklqdq %xmm2,%xmm3,%xmm10
15966 .byte 196,65,122,126,4,192 // vmovq (%r8,%rax,8),%xmm8
15967 .byte 196,129,122,126,28,200 // vmovq (%r8,%r9,8),%xmm3
15968 .byte 196,193,97,108,216 // vpunpcklqdq %xmm8,%xmm3,%xmm3
15969 .byte 197,177,97,193 // vpunpcklwd %xmm1,%xmm9,%xmm0
15970 .byte 197,177,105,201 // vpunpckhwd %xmm1,%xmm9,%xmm1
15971 .byte 197,169,97,211 // vpunpcklwd %xmm3,%xmm10,%xmm2
15972 .byte 197,169,105,219 // vpunpckhwd %xmm3,%xmm10,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15973 .byte 197,121,97,249 // vpunpcklwd %xmm1,%xmm0,%xmm15
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]15974 .byte 197,121,105,193 // vpunpckhwd %xmm1,%xmm0,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15975 .byte 197,233,97,203 // vpunpcklwd %xmm3,%xmm2,%xmm1
15976 .byte 197,105,105,203 // vpunpckhwd %xmm3,%xmm2,%xmm9
15977 .byte 197,129,108,193 // vpunpcklqdq %xmm1,%xmm15,%xmm0
15978 .byte 197,233,239,210 // vpxor %xmm2,%xmm2,%xmm2
15979 .byte 197,249,105,210 // vpunpckhwd %xmm2,%xmm0,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15980 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
15981 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15982 .byte 196,98,125,24,37,5,25,0,0 // vbroadcastss 0x1905(%rip),%ymm12 # 5d28 <_sk_callback_avx+0x1fe>
15983 .byte 196,193,124,84,212 // vandps %ymm12,%ymm0,%ymm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15984 .byte 197,252,87,194 // vxorps %ymm2,%ymm0,%ymm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15985 .byte 196,195,125,25,198,1 // vextractf128 $0x1,%ymm0,%xmm14
15986 .byte 196,98,121,24,29,241,24,0,0 // vbroadcastss 0x18f1(%rip),%xmm11 # 5d2c <_sk_callback_avx+0x202>
15987 .byte 196,193,8,87,219 // vxorps %xmm11,%xmm14,%xmm3
15988 .byte 196,98,121,24,45,231,24,0,0 // vbroadcastss 0x18e7(%rip),%xmm13 # 5d30 <_sk_callback_avx+0x206>
15989 .byte 197,145,102,219 // vpcmpgtd %xmm3,%xmm13,%xmm3
15990 .byte 196,65,120,87,211 // vxorps %xmm11,%xmm0,%xmm10
15991 .byte 196,65,17,102,210 // vpcmpgtd %xmm10,%xmm13,%xmm10
15992 .byte 196,99,45,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm10,%ymm10
15993 .byte 197,225,114,242,16 // vpslld $0x10,%xmm2,%xmm3
15994 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15995 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15996 .byte 196,227,101,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm3,%ymm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]15997 .byte 197,249,114,240,13 // vpslld $0xd,%xmm0,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]15998 .byte 196,193,97,114,246,13 // vpslld $0xd,%xmm14,%xmm3
15999 .byte 196,227,125,24,195,1 // vinsertf128 $0x1,%xmm3,%ymm0,%ymm0
16000 .byte 197,252,86,194 // vorps %ymm2,%ymm0,%ymm0
16001 .byte 196,227,125,25,194,1 // vextractf128 $0x1,%ymm0,%xmm2
16002 .byte 196,226,121,24,29,157,24,0,0 // vbroadcastss 0x189d(%rip),%xmm3 # 5d34 <_sk_callback_avx+0x20a>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16003 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16004 .byte 197,249,254,195 // vpaddd %xmm3,%xmm0,%xmm0
16005 .byte 196,227,125,24,194,1 // vinsertf128 $0x1,%xmm2,%ymm0,%ymm0
16006 .byte 196,65,12,87,246 // vxorps %ymm14,%ymm14,%ymm14
16007 .byte 196,195,125,74,198,160 // vblendvps %ymm10,%ymm14,%ymm0,%ymm0
16008 .byte 197,129,109,201 // vpunpckhqdq %xmm1,%xmm15,%xmm1
16009 .byte 197,217,239,228 // vpxor %xmm4,%xmm4,%xmm4
16010 .byte 197,241,105,212 // vpunpckhwd %xmm4,%xmm1,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16011 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
16012 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16013 .byte 196,193,116,84,212 // vandps %ymm12,%ymm1,%ymm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16014 .byte 197,244,87,202 // vxorps %ymm2,%ymm1,%ymm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16015 .byte 196,195,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm10
16016 .byte 196,193,40,87,251 // vxorps %xmm11,%xmm10,%xmm7
16017 .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7
16018 .byte 196,193,112,87,243 // vxorps %xmm11,%xmm1,%xmm6
16019 .byte 197,145,102,246 // vpcmpgtd %xmm6,%xmm13,%xmm6
16020 .byte 196,227,77,24,247,1 // vinsertf128 $0x1,%xmm7,%ymm6,%ymm6
16021 .byte 197,193,114,242,16 // vpslld $0x10,%xmm2,%xmm7
16022 .byte 196,227,125,25,210,1 // vextractf128 $0x1,%ymm2,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16023 .byte 197,233,114,242,16 // vpslld $0x10,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16024 .byte 196,227,69,24,210,1 // vinsertf128 $0x1,%xmm2,%ymm7,%ymm2
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]16025 .byte 197,241,114,241,13 // vpslld $0xd,%xmm1,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16026 .byte 196,193,65,114,242,13 // vpslld $0xd,%xmm10,%xmm7
16027 .byte 196,227,117,24,207,1 // vinsertf128 $0x1,%xmm7,%ymm1,%ymm1
16028 .byte 197,244,86,202 // vorps %ymm2,%ymm1,%ymm1
16029 .byte 196,227,125,25,202,1 // vextractf128 $0x1,%ymm1,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16030 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16031 .byte 197,241,254,203 // vpaddd %xmm3,%xmm1,%xmm1
16032 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
16033 .byte 196,195,117,74,206,96 // vblendvps %ymm6,%ymm14,%ymm1,%ymm1
16034 .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2
16035 .byte 197,233,105,244 // vpunpckhwd %xmm4,%xmm2,%xmm6
16036 .byte 196,65,41,239,210 // vpxor %xmm10,%xmm10,%xmm10
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16037 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16038 .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2
16039 .byte 196,193,108,84,244 // vandps %ymm12,%ymm2,%ymm6
16040 .byte 197,236,87,214 // vxorps %ymm6,%ymm2,%ymm2
16041 .byte 196,227,125,25,215,1 // vextractf128 $0x1,%ymm2,%xmm7
16042 .byte 196,193,64,87,235 // vxorps %xmm11,%xmm7,%xmm5
16043 .byte 197,145,102,237 // vpcmpgtd %xmm5,%xmm13,%xmm5
16044 .byte 196,193,104,87,227 // vxorps %xmm11,%xmm2,%xmm4
16045 .byte 197,145,102,228 // vpcmpgtd %xmm4,%xmm13,%xmm4
16046 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4
16047 .byte 197,209,114,246,16 // vpslld $0x10,%xmm6,%xmm5
16048 .byte 196,227,125,25,246,1 // vextractf128 $0x1,%ymm6,%xmm6
16049 .byte 197,201,114,246,16 // vpslld $0x10,%xmm6,%xmm6
16050 .byte 196,227,85,24,238,1 // vinsertf128 $0x1,%xmm6,%ymm5,%ymm5
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]16051 .byte 197,233,114,242,13 // vpslld $0xd,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16052 .byte 197,201,114,247,13 // vpslld $0xd,%xmm7,%xmm6
16053 .byte 196,227,109,24,214,1 // vinsertf128 $0x1,%xmm6,%ymm2,%ymm2
16054 .byte 197,236,86,213 // vorps %ymm5,%ymm2,%ymm2
16055 .byte 196,227,125,25,213,1 // vextractf128 $0x1,%ymm2,%xmm5
16056 .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5
16057 .byte 197,233,254,211 // vpaddd %xmm3,%xmm2,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16058 .byte 196,227,109,24,213,1 // vinsertf128 $0x1,%xmm5,%ymm2,%ymm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16059 .byte 196,195,109,74,214,64 // vblendvps %ymm4,%ymm14,%ymm2,%ymm2
16060 .byte 196,193,57,109,225 // vpunpckhqdq %xmm9,%xmm8,%xmm4
16061 .byte 196,193,89,105,234 // vpunpckhwd %xmm10,%xmm4,%xmm5
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16062 .byte 196,226,121,51,228 // vpmovzxwd %xmm4,%xmm4
16063 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16064 .byte 196,193,92,84,236 // vandps %ymm12,%ymm4,%ymm5
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16065 .byte 197,220,87,229 // vxorps %ymm5,%ymm4,%ymm4
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16066 .byte 196,227,125,25,230,1 // vextractf128 $0x1,%ymm4,%xmm6
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16067 .byte 196,193,72,87,251 // vxorps %xmm11,%xmm6,%xmm7
16068 .byte 197,17,102,199 // vpcmpgtd %xmm7,%xmm13,%xmm8
16069 .byte 196,193,88,87,251 // vxorps %xmm11,%xmm4,%xmm7
16070 .byte 197,145,102,255 // vpcmpgtd %xmm7,%xmm13,%xmm7
16071 .byte 196,195,69,24,248,1 // vinsertf128 $0x1,%xmm8,%ymm7,%ymm7
16072 .byte 197,185,114,245,16 // vpslld $0x10,%xmm5,%xmm8
16073 .byte 196,227,125,25,237,1 // vextractf128 $0x1,%ymm5,%xmm5
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16074 .byte 197,209,114,245,16 // vpslld $0x10,%xmm5,%xmm5
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16075 .byte 196,227,61,24,237,1 // vinsertf128 $0x1,%xmm5,%ymm8,%ymm5
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16076 .byte 197,217,114,244,13 // vpslld $0xd,%xmm4,%xmm4
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16077 .byte 197,201,114,246,13 // vpslld $0xd,%xmm6,%xmm6
16078 .byte 196,227,93,24,230,1 // vinsertf128 $0x1,%xmm6,%ymm4,%ymm4
16079 .byte 197,220,86,229 // vorps %ymm5,%ymm4,%ymm4
16080 .byte 196,227,125,25,229,1 // vextractf128 $0x1,%ymm4,%xmm5
16081 .byte 197,209,254,235 // vpaddd %xmm3,%xmm5,%xmm5
16082 .byte 197,217,254,219 // vpaddd %xmm3,%xmm4,%xmm3
16083 .byte 196,227,101,24,221,1 // vinsertf128 $0x1,%xmm5,%ymm3,%ymm3
16084 .byte 196,195,101,74,222,112 // vblendvps %ymm7,%ymm14,%ymm3,%ymm3
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]16085 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16086 .byte 197,252,16,100,36,128 // vmovups -0x80(%rsp),%ymm4
16087 .byte 197,252,16,108,36,160 // vmovups -0x60(%rsp),%ymm5
16088 .byte 197,252,16,116,36,192 // vmovups -0x40(%rsp),%ymm6
16089 .byte 197,252,16,124,36,224 // vmovups -0x20(%rsp),%ymm7
16090 .byte 72,131,196,24 // add $0x18,%rsp
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]16091 .byte 91 // pop %rbx
16092 .byte 65,92 // pop %r12
16093 .byte 65,94 // pop %r14
16094 .byte 65,95 // pop %r15
16095 .byte 255,224 // jmpq *%rax
16096
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16097HIDDEN _sk_store_f16_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16098.globl _sk_store_f16_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16099FUNCTION(_sk_store_f16_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16100_sk_store_f16_avx:
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16101 .byte 72,131,236,88 // sub $0x58,%rsp
16102 .byte 197,252,17,124,36,32 // vmovups %ymm7,0x20(%rsp)
16103 .byte 197,252,17,52,36 // vmovups %ymm6,(%rsp)
16104 .byte 197,252,17,108,36,224 // vmovups %ymm5,-0x20(%rsp)
16105 .byte 197,252,17,100,36,192 // vmovups %ymm4,-0x40(%rsp)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16106 .byte 196,98,125,24,13,182,22,0,0 // vbroadcastss 0x16b6(%rip),%ymm9 # 5d38 <_sk_callback_avx+0x20e>
16107 .byte 196,65,124,84,209 // vandps %ymm9,%ymm0,%ymm10
16108 .byte 197,252,17,68,36,128 // vmovups %ymm0,-0x80(%rsp)
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16109 .byte 196,65,124,87,218 // vxorps %ymm10,%ymm0,%ymm11
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16110 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16111 .byte 196,98,121,24,5,155,22,0,0 // vbroadcastss 0x169b(%rip),%xmm8 # 5d3c <_sk_callback_avx+0x212>
16112 .byte 196,65,57,102,236 // vpcmpgtd %xmm12,%xmm8,%xmm13
16113 .byte 196,65,57,102,243 // vpcmpgtd %xmm11,%xmm8,%xmm14
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16114 .byte 196,67,13,24,237,1 // vinsertf128 $0x1,%xmm13,%ymm14,%ymm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16115 .byte 196,193,9,114,210,16 // vpsrld $0x10,%xmm10,%xmm14
16116 .byte 196,67,125,25,210,1 // vextractf128 $0x1,%ymm10,%xmm10
16117 .byte 196,193,41,114,210,16 // vpsrld $0x10,%xmm10,%xmm10
16118 .byte 196,67,13,24,242,1 // vinsertf128 $0x1,%xmm10,%ymm14,%ymm14
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16119 .byte 196,193,33,114,211,13 // vpsrld $0xd,%xmm11,%xmm11
16120 .byte 196,193,25,114,212,13 // vpsrld $0xd,%xmm12,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16121 .byte 196,98,125,24,21,98,22,0,0 // vbroadcastss 0x1662(%rip),%ymm10 # 5d40 <_sk_callback_avx+0x216>
16122 .byte 196,65,12,86,242 // vorps %ymm10,%ymm14,%ymm14
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16123 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16124 .byte 196,65,1,254,228 // vpaddd %xmm12,%xmm15,%xmm12
16125 .byte 196,65,9,254,219 // vpaddd %xmm11,%xmm14,%xmm11
16126 .byte 196,67,37,24,228,1 // vinsertf128 $0x1,%xmm12,%ymm11,%ymm12
16127 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
16128 .byte 196,99,29,74,224,208 // vblendvps %ymm13,%ymm0,%ymm12,%ymm12
16129 .byte 196,65,116,84,233 // vandps %ymm9,%ymm1,%ymm13
16130 .byte 197,252,17,76,36,160 // vmovups %ymm1,-0x60(%rsp)
16131 .byte 196,65,116,87,245 // vxorps %ymm13,%ymm1,%ymm14
16132 .byte 196,67,125,25,247,1 // vextractf128 $0x1,%ymm14,%xmm15
16133 .byte 196,193,57,102,255 // vpcmpgtd %xmm15,%xmm8,%xmm7
16134 .byte 196,65,57,102,222 // vpcmpgtd %xmm14,%xmm8,%xmm11
16135 .byte 196,227,37,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm11,%ymm7
16136 .byte 196,193,33,114,213,16 // vpsrld $0x10,%xmm13,%xmm11
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16137 .byte 196,99,125,25,238,1 // vextractf128 $0x1,%ymm13,%xmm6
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16138 .byte 197,201,114,214,16 // vpsrld $0x10,%xmm6,%xmm6
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16139 .byte 196,227,37,24,246,1 // vinsertf128 $0x1,%xmm6,%ymm11,%ymm6
16140 .byte 196,193,33,114,215,13 // vpsrld $0xd,%xmm15,%xmm11
16141 .byte 196,193,76,86,242 // vorps %ymm10,%ymm6,%ymm6
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16142 .byte 196,227,125,25,245,1 // vextractf128 $0x1,%ymm6,%xmm5
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16143 .byte 196,193,81,254,235 // vpaddd %xmm11,%xmm5,%xmm5
16144 .byte 196,193,89,114,214,13 // vpsrld $0xd,%xmm14,%xmm4
16145 .byte 197,201,254,228 // vpaddd %xmm4,%xmm6,%xmm4
16146 .byte 196,227,93,24,229,1 // vinsertf128 $0x1,%xmm5,%ymm4,%ymm4
16147 .byte 196,99,93,74,232,112 // vblendvps %ymm7,%ymm0,%ymm4,%ymm13
16148 .byte 196,193,108,84,225 // vandps %ymm9,%ymm2,%ymm4
16149 .byte 197,236,87,236 // vxorps %ymm4,%ymm2,%ymm5
16150 .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6
16151 .byte 197,185,102,254 // vpcmpgtd %xmm6,%xmm8,%xmm7
16152 .byte 197,57,102,221 // vpcmpgtd %xmm5,%xmm8,%xmm11
16153 .byte 196,227,37,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm11,%ymm7
16154 .byte 197,161,114,212,16 // vpsrld $0x10,%xmm4,%xmm11
16155 .byte 196,227,125,25,228,1 // vextractf128 $0x1,%ymm4,%xmm4
16156 .byte 197,217,114,212,16 // vpsrld $0x10,%xmm4,%xmm4
16157 .byte 196,227,37,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm11,%ymm4
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16158 .byte 197,201,114,214,13 // vpsrld $0xd,%xmm6,%xmm6
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16159 .byte 196,193,92,86,226 // vorps %ymm10,%ymm4,%ymm4
16160 .byte 196,227,125,25,225,1 // vextractf128 $0x1,%ymm4,%xmm1
16161 .byte 197,241,254,206 // vpaddd %xmm6,%xmm1,%xmm1
16162 .byte 197,209,114,213,13 // vpsrld $0xd,%xmm5,%xmm5
16163 .byte 197,217,254,229 // vpaddd %xmm5,%xmm4,%xmm4
16164 .byte 196,227,93,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm4,%ymm1
16165 .byte 196,99,117,74,216,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm11
16166 .byte 196,193,100,84,225 // vandps %ymm9,%ymm3,%ymm4
16167 .byte 197,228,87,236 // vxorps %ymm4,%ymm3,%ymm5
16168 .byte 196,227,125,25,238,1 // vextractf128 $0x1,%ymm5,%xmm6
16169 .byte 197,185,102,254 // vpcmpgtd %xmm6,%xmm8,%xmm7
16170 .byte 197,57,102,197 // vpcmpgtd %xmm5,%xmm8,%xmm8
16171 .byte 196,227,61,24,255,1 // vinsertf128 $0x1,%xmm7,%ymm8,%ymm7
16172 .byte 197,185,114,212,16 // vpsrld $0x10,%xmm4,%xmm8
16173 .byte 196,227,125,25,228,1 // vextractf128 $0x1,%ymm4,%xmm4
16174 .byte 197,217,114,212,16 // vpsrld $0x10,%xmm4,%xmm4
16175 .byte 196,227,61,24,228,1 // vinsertf128 $0x1,%xmm4,%ymm8,%ymm4
16176 .byte 196,193,92,86,226 // vorps %ymm10,%ymm4,%ymm4
16177 .byte 197,201,114,214,13 // vpsrld $0xd,%xmm6,%xmm6
16178 .byte 196,227,125,25,225,1 // vextractf128 $0x1,%ymm4,%xmm1
16179 .byte 197,241,254,206 // vpaddd %xmm6,%xmm1,%xmm1
16180 .byte 197,209,114,213,13 // vpsrld $0xd,%xmm5,%xmm5
16181 .byte 197,217,254,229 // vpaddd %xmm5,%xmm4,%xmm4
16182 .byte 196,227,93,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm4,%ymm1
16183 .byte 196,227,117,74,200,112 // vblendvps %ymm7,%ymm0,%ymm1,%ymm1
16184 .byte 196,99,125,25,224,1 // vextractf128 $0x1,%ymm12,%xmm0
16185 .byte 196,226,25,43,192 // vpackusdw %xmm0,%xmm12,%xmm0
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]16186 .byte 196,99,125,25,236,1 // vextractf128 $0x1,%ymm13,%xmm4
16187 .byte 196,226,17,43,228 // vpackusdw %xmm4,%xmm13,%xmm4
16188 .byte 196,99,125,25,221,1 // vextractf128 $0x1,%ymm11,%xmm5
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16189 .byte 196,226,33,43,245 // vpackusdw %xmm5,%xmm11,%xmm6
16190 .byte 196,227,125,25,205,1 // vextractf128 $0x1,%ymm1,%xmm5
16191 .byte 196,226,113,43,205 // vpackusdw %xmm5,%xmm1,%xmm1
16192 .byte 197,249,97,236 // vpunpcklwd %xmm4,%xmm0,%xmm5
16193 .byte 197,249,105,196 // vpunpckhwd %xmm4,%xmm0,%xmm0
16194 .byte 197,201,97,225 // vpunpcklwd %xmm1,%xmm6,%xmm4
16195 .byte 197,201,105,201 // vpunpckhwd %xmm1,%xmm6,%xmm1
16196 .byte 197,81,98,220 // vpunpckldq %xmm4,%xmm5,%xmm11
16197 .byte 197,81,106,212 // vpunpckhdq %xmm4,%xmm5,%xmm10
16198 .byte 197,121,98,201 // vpunpckldq %xmm1,%xmm0,%xmm9
16199 .byte 197,121,106,193 // vpunpckhdq %xmm1,%xmm0,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16200 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16201 .byte 72,139,0 // mov (%rax),%rax
16202 .byte 72,133,201 // test %rcx,%rcx
16203 .byte 117,66 // jne 48bc <_sk_store_f16_avx+0x25e>
16204 .byte 197,120,17,28,248 // vmovups %xmm11,(%rax,%rdi,8)
16205 .byte 197,120,17,84,248,16 // vmovups %xmm10,0x10(%rax,%rdi,8)
16206 .byte 197,120,17,76,248,32 // vmovups %xmm9,0x20(%rax,%rdi,8)
16207 .byte 197,122,127,68,248,48 // vmovdqu %xmm8,0x30(%rax,%rdi,8)
16208 .byte 72,173 // lods %ds:(%rsi),%rax
16209 .byte 197,252,16,68,36,128 // vmovups -0x80(%rsp),%ymm0
16210 .byte 197,252,16,76,36,160 // vmovups -0x60(%rsp),%ymm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]16211 .byte 197,252,16,100,36,192 // vmovups -0x40(%rsp),%ymm4
16212 .byte 197,252,16,108,36,224 // vmovups -0x20(%rsp),%ymm5
16213 .byte 197,252,16,52,36 // vmovups (%rsp),%ymm6
16214 .byte 197,252,16,124,36,32 // vmovups 0x20(%rsp),%ymm7
16215 .byte 72,131,196,88 // add $0x58,%rsp
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16216 .byte 255,224 // jmpq *%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16217 .byte 197,121,214,28,248 // vmovq %xmm11,(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16218 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16219 .byte 116,202 // je 4891 <_sk_store_f16_avx+0x233>
16220 .byte 197,121,23,92,248,8 // vmovhpd %xmm11,0x8(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16221 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16222 .byte 114,190 // jb 4891 <_sk_store_f16_avx+0x233>
16223 .byte 197,121,214,84,248,16 // vmovq %xmm10,0x10(%rax,%rdi,8)
16224 .byte 116,182 // je 4891 <_sk_store_f16_avx+0x233>
16225 .byte 197,121,23,84,248,24 // vmovhpd %xmm10,0x18(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16226 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16227 .byte 114,170 // jb 4891 <_sk_store_f16_avx+0x233>
16228 .byte 197,121,214,76,248,32 // vmovq %xmm9,0x20(%rax,%rdi,8)
16229 .byte 116,162 // je 4891 <_sk_store_f16_avx+0x233>
16230 .byte 197,121,23,76,248,40 // vmovhpd %xmm9,0x28(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16231 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16232 .byte 114,150 // jb 4891 <_sk_store_f16_avx+0x233>
16233 .byte 197,121,214,68,248,48 // vmovq %xmm8,0x30(%rax,%rdi,8)
16234 .byte 235,142 // jmp 4891 <_sk_store_f16_avx+0x233>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16235
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16236HIDDEN _sk_load_u16_be_avx
16237.globl _sk_load_u16_be_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16238FUNCTION(_sk_load_u16_be_avx)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16239_sk_load_u16_be_avx:
16240 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16241 .byte 76,139,0 // mov (%rax),%r8
16242 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16243 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16244 .byte 15,133,5,1,0,0 // jne 4a1e <_sk_load_u16_be_avx+0x11b>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16245 .byte 196,65,121,16,4,64 // vmovupd (%r8,%rax,2),%xmm8
16246 .byte 196,193,121,16,84,64,16 // vmovupd 0x10(%r8,%rax,2),%xmm2
16247 .byte 196,193,121,16,92,64,32 // vmovupd 0x20(%r8,%rax,2),%xmm3
16248 .byte 196,65,122,111,76,64,48 // vmovdqu 0x30(%r8,%rax,2),%xmm9
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16249 .byte 197,185,97,194 // vpunpcklwd %xmm2,%xmm8,%xmm0
16250 .byte 197,185,105,210 // vpunpckhwd %xmm2,%xmm8,%xmm2
16251 .byte 196,193,97,97,201 // vpunpcklwd %xmm9,%xmm3,%xmm1
16252 .byte 196,193,97,105,217 // vpunpckhwd %xmm9,%xmm3,%xmm3
16253 .byte 197,121,97,210 // vpunpcklwd %xmm2,%xmm0,%xmm10
16254 .byte 197,121,105,194 // vpunpckhwd %xmm2,%xmm0,%xmm8
16255 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
16256 .byte 197,113,105,203 // vpunpckhwd %xmm3,%xmm1,%xmm9
16257 .byte 184,128,0,128,55 // mov $0x37800080,%eax
16258 .byte 197,249,110,192 // vmovd %eax,%xmm0
16259 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
16260 .byte 196,99,125,24,224,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm12
16261 .byte 197,169,108,194 // vpunpcklqdq %xmm2,%xmm10,%xmm0
16262 .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1
16263 .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0
16264 .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
16265 .byte 196,65,33,239,219 // vpxor %xmm11,%xmm11,%xmm11
16266 .byte 196,193,121,105,203 // vpunpckhwd %xmm11,%xmm0,%xmm1
16267 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
16268 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
16269 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
16270 .byte 197,156,89,192 // vmulps %ymm0,%ymm12,%ymm0
16271 .byte 197,169,109,202 // vpunpckhqdq %xmm2,%xmm10,%xmm1
16272 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
16273 .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1
16274 .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1
16275 .byte 196,193,113,105,211 // vpunpckhwd %xmm11,%xmm1,%xmm2
16276 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
16277 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
16278 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
16279 .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1
16280 .byte 196,193,57,108,209 // vpunpcklqdq %xmm9,%xmm8,%xmm2
16281 .byte 197,169,113,242,8 // vpsllw $0x8,%xmm2,%xmm10
16282 .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2
16283 .byte 197,169,235,210 // vpor %xmm2,%xmm10,%xmm2
16284 .byte 196,65,105,105,211 // vpunpckhwd %xmm11,%xmm2,%xmm10
16285 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
16286 .byte 196,195,109,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm2,%ymm2
16287 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
16288 .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2
16289 .byte 196,193,57,109,217 // vpunpckhqdq %xmm9,%xmm8,%xmm3
16290 .byte 197,185,113,243,8 // vpsllw $0x8,%xmm3,%xmm8
16291 .byte 197,225,113,211,8 // vpsrlw $0x8,%xmm3,%xmm3
16292 .byte 197,185,235,219 // vpor %xmm3,%xmm8,%xmm3
16293 .byte 196,65,97,105,195 // vpunpckhwd %xmm11,%xmm3,%xmm8
16294 .byte 196,226,121,51,219 // vpmovzxwd %xmm3,%xmm3
16295 .byte 196,195,101,24,216,1 // vinsertf128 $0x1,%xmm8,%ymm3,%ymm3
16296 .byte 197,252,91,219 // vcvtdq2ps %ymm3,%ymm3
16297 .byte 197,156,89,219 // vmulps %ymm3,%ymm12,%ymm3
16298 .byte 72,173 // lods %ds:(%rsi),%rax
16299 .byte 255,224 // jmpq *%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16300 .byte 196,65,123,16,4,64 // vmovsd (%r8,%rax,2),%xmm8
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16301 .byte 196,65,49,239,201 // vpxor %xmm9,%xmm9,%xmm9
16302 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16303 .byte 116,85 // je 4a84 <_sk_load_u16_be_avx+0x181>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16304 .byte 196,65,57,22,68,64,8 // vmovhpd 0x8(%r8,%rax,2),%xmm8,%xmm8
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16305 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16306 .byte 114,72 // jb 4a84 <_sk_load_u16_be_avx+0x181>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16307 .byte 196,193,123,16,84,64,16 // vmovsd 0x10(%r8,%rax,2),%xmm2
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16308 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16309 .byte 116,72 // je 4a91 <_sk_load_u16_be_avx+0x18e>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16310 .byte 196,193,105,22,84,64,24 // vmovhpd 0x18(%r8,%rax,2),%xmm2,%xmm2
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16311 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16312 .byte 114,59 // jb 4a91 <_sk_load_u16_be_avx+0x18e>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16313 .byte 196,193,123,16,92,64,32 // vmovsd 0x20(%r8,%rax,2),%xmm3
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16314 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16315 .byte 15,132,205,254,255,255 // je 4934 <_sk_load_u16_be_avx+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16316 .byte 196,193,97,22,92,64,40 // vmovhpd 0x28(%r8,%rax,2),%xmm3,%xmm3
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16317 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16318 .byte 15,130,188,254,255,255 // jb 4934 <_sk_load_u16_be_avx+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16319 .byte 196,65,122,126,76,64,48 // vmovq 0x30(%r8,%rax,2),%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16320 .byte 233,176,254,255,255 // jmpq 4934 <_sk_load_u16_be_avx+0x31>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16321 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
16322 .byte 197,233,87,210 // vxorpd %xmm2,%xmm2,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16323 .byte 233,163,254,255,255 // jmpq 4934 <_sk_load_u16_be_avx+0x31>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16324 .byte 197,225,87,219 // vxorpd %xmm3,%xmm3,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16325 .byte 233,154,254,255,255 // jmpq 4934 <_sk_load_u16_be_avx+0x31>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16326
16327HIDDEN _sk_load_rgb_u16_be_avx
16328.globl _sk_load_rgb_u16_be_avx
16329FUNCTION(_sk_load_rgb_u16_be_avx)
16330_sk_load_rgb_u16_be_avx:
16331 .byte 72,173 // lods %ds:(%rsi),%rax
16332 .byte 76,139,0 // mov (%rax),%r8
16333 .byte 72,141,4,127 // lea (%rdi,%rdi,2),%rax
16334 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16335 .byte 15,133,8,1,0,0 // jne 4bb4 <_sk_load_rgb_u16_be_avx+0x11a>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16336 .byte 196,193,122,111,4,64 // vmovdqu (%r8,%rax,2),%xmm0
16337 .byte 196,193,122,111,84,64,12 // vmovdqu 0xc(%r8,%rax,2),%xmm2
16338 .byte 196,193,122,111,76,64,24 // vmovdqu 0x18(%r8,%rax,2),%xmm1
16339 .byte 196,193,122,111,92,64,32 // vmovdqu 0x20(%r8,%rax,2),%xmm3
16340 .byte 197,225,115,219,4 // vpsrldq $0x4,%xmm3,%xmm3
16341 .byte 197,185,115,216,6 // vpsrldq $0x6,%xmm0,%xmm8
16342 .byte 197,177,115,218,6 // vpsrldq $0x6,%xmm2,%xmm9
16343 .byte 197,161,115,217,6 // vpsrldq $0x6,%xmm1,%xmm11
16344 .byte 197,169,115,219,6 // vpsrldq $0x6,%xmm3,%xmm10
16345 .byte 197,249,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm0
16346 .byte 196,193,57,97,209 // vpunpcklwd %xmm9,%xmm8,%xmm2
16347 .byte 197,241,97,203 // vpunpcklwd %xmm3,%xmm1,%xmm1
16348 .byte 196,193,33,97,218 // vpunpcklwd %xmm10,%xmm11,%xmm3
16349 .byte 197,121,97,194 // vpunpcklwd %xmm2,%xmm0,%xmm8
16350 .byte 197,121,105,202 // vpunpckhwd %xmm2,%xmm0,%xmm9
16351 .byte 197,241,97,211 // vpunpcklwd %xmm3,%xmm1,%xmm2
16352 .byte 197,113,105,219 // vpunpckhwd %xmm3,%xmm1,%xmm11
16353 .byte 184,128,0,128,55 // mov $0x37800080,%eax
16354 .byte 197,249,110,192 // vmovd %eax,%xmm0
16355 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
16356 .byte 196,99,125,24,208,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm10
16357 .byte 197,185,108,194 // vpunpcklqdq %xmm2,%xmm8,%xmm0
16358 .byte 197,241,113,240,8 // vpsllw $0x8,%xmm0,%xmm1
16359 .byte 197,249,113,208,8 // vpsrlw $0x8,%xmm0,%xmm0
16360 .byte 197,241,235,192 // vpor %xmm0,%xmm1,%xmm0
16361 .byte 196,65,25,239,228 // vpxor %xmm12,%xmm12,%xmm12
16362 .byte 196,193,121,105,204 // vpunpckhwd %xmm12,%xmm0,%xmm1
16363 .byte 196,226,121,51,192 // vpmovzxwd %xmm0,%xmm0
16364 .byte 196,227,125,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm0,%ymm0
16365 .byte 197,252,91,192 // vcvtdq2ps %ymm0,%ymm0
16366 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
16367 .byte 197,185,109,202 // vpunpckhqdq %xmm2,%xmm8,%xmm1
16368 .byte 197,233,113,241,8 // vpsllw $0x8,%xmm1,%xmm2
16369 .byte 197,241,113,209,8 // vpsrlw $0x8,%xmm1,%xmm1
16370 .byte 197,233,235,201 // vpor %xmm1,%xmm2,%xmm1
16371 .byte 196,193,113,105,212 // vpunpckhwd %xmm12,%xmm1,%xmm2
16372 .byte 196,226,121,51,201 // vpmovzxwd %xmm1,%xmm1
16373 .byte 196,227,117,24,202,1 // vinsertf128 $0x1,%xmm2,%ymm1,%ymm1
16374 .byte 197,252,91,201 // vcvtdq2ps %ymm1,%ymm1
16375 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
16376 .byte 196,193,49,108,211 // vpunpcklqdq %xmm11,%xmm9,%xmm2
16377 .byte 197,225,113,242,8 // vpsllw $0x8,%xmm2,%xmm3
16378 .byte 197,233,113,210,8 // vpsrlw $0x8,%xmm2,%xmm2
16379 .byte 197,225,235,210 // vpor %xmm2,%xmm3,%xmm2
16380 .byte 196,193,105,105,220 // vpunpckhwd %xmm12,%xmm2,%xmm3
16381 .byte 196,226,121,51,210 // vpmovzxwd %xmm2,%xmm2
16382 .byte 196,227,109,24,211,1 // vinsertf128 $0x1,%xmm3,%ymm2,%ymm2
16383 .byte 197,252,91,210 // vcvtdq2ps %ymm2,%ymm2
16384 .byte 197,172,89,210 // vmulps %ymm2,%ymm10,%ymm2
16385 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
16386 .byte 197,249,110,216 // vmovd %eax,%xmm3
16387 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
16388 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
16389 .byte 72,173 // lods %ds:(%rsi),%rax
16390 .byte 255,224 // jmpq *%rax
16391 .byte 196,193,121,110,4,64 // vmovd (%r8,%rax,2),%xmm0
16392 .byte 196,193,121,196,68,64,4,2 // vpinsrw $0x2,0x4(%r8,%rax,2),%xmm0,%xmm0
16393 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16394 .byte 117,5 // jne 4bcd <_sk_load_rgb_u16_be_avx+0x133>
16395 .byte 233,19,255,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16396 .byte 196,193,121,110,76,64,6 // vmovd 0x6(%r8,%rax,2),%xmm1
16397 .byte 196,65,113,196,68,64,10,2 // vpinsrw $0x2,0xa(%r8,%rax,2),%xmm1,%xmm8
16398 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16399 .byte 114,26 // jb 4bfc <_sk_load_rgb_u16_be_avx+0x162>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16400 .byte 196,193,121,110,76,64,12 // vmovd 0xc(%r8,%rax,2),%xmm1
16401 .byte 196,193,113,196,84,64,16,2 // vpinsrw $0x2,0x10(%r8,%rax,2),%xmm1,%xmm2
16402 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16403 .byte 117,10 // jne 4c01 <_sk_load_rgb_u16_be_avx+0x167>
16404 .byte 233,228,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
16405 .byte 233,223,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16406 .byte 196,193,121,110,76,64,18 // vmovd 0x12(%r8,%rax,2),%xmm1
16407 .byte 196,65,113,196,76,64,22,2 // vpinsrw $0x2,0x16(%r8,%rax,2),%xmm1,%xmm9
16408 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16409 .byte 114,26 // jb 4c30 <_sk_load_rgb_u16_be_avx+0x196>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16410 .byte 196,193,121,110,76,64,24 // vmovd 0x18(%r8,%rax,2),%xmm1
16411 .byte 196,193,113,196,76,64,28,2 // vpinsrw $0x2,0x1c(%r8,%rax,2),%xmm1,%xmm1
16412 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16413 .byte 117,10 // jne 4c35 <_sk_load_rgb_u16_be_avx+0x19b>
16414 .byte 233,176,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
16415 .byte 233,171,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16416 .byte 196,193,121,110,92,64,30 // vmovd 0x1e(%r8,%rax,2),%xmm3
16417 .byte 196,65,97,196,92,64,34,2 // vpinsrw $0x2,0x22(%r8,%rax,2),%xmm3,%xmm11
16418 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16419 .byte 114,20 // jb 4c5e <_sk_load_rgb_u16_be_avx+0x1c4>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16420 .byte 196,193,121,110,92,64,36 // vmovd 0x24(%r8,%rax,2),%xmm3
16421 .byte 196,193,97,196,92,64,40,2 // vpinsrw $0x2,0x28(%r8,%rax,2),%xmm3,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16422 .byte 233,130,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
16423 .byte 233,125,254,255,255 // jmpq 4ae0 <_sk_load_rgb_u16_be_avx+0x46>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16424
16425HIDDEN _sk_store_u16_be_avx
16426.globl _sk_store_u16_be_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16427FUNCTION(_sk_store_u16_be_avx)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16428_sk_store_u16_be_avx:
16429 .byte 72,173 // lods %ds:(%rsi),%rax
16430 .byte 76,139,0 // mov (%rax),%r8
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16431 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16432 .byte 184,0,255,127,71 // mov $0x477fff00,%eax
16433 .byte 197,121,110,192 // vmovd %eax,%xmm8
16434 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
16435 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
16436 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
16437 .byte 196,65,125,91,201 // vcvtps2dq %ymm9,%ymm9
16438 .byte 196,67,125,25,202,1 // vextractf128 $0x1,%ymm9,%xmm10
16439 .byte 196,66,49,43,202 // vpackusdw %xmm10,%xmm9,%xmm9
16440 .byte 196,193,41,113,241,8 // vpsllw $0x8,%xmm9,%xmm10
16441 .byte 196,193,49,113,209,8 // vpsrlw $0x8,%xmm9,%xmm9
16442 .byte 196,65,41,235,201 // vpor %xmm9,%xmm10,%xmm9
16443 .byte 197,60,89,209 // vmulps %ymm1,%ymm8,%ymm10
16444 .byte 196,65,125,91,210 // vcvtps2dq %ymm10,%ymm10
16445 .byte 196,67,125,25,211,1 // vextractf128 $0x1,%ymm10,%xmm11
16446 .byte 196,66,41,43,211 // vpackusdw %xmm11,%xmm10,%xmm10
16447 .byte 196,193,33,113,242,8 // vpsllw $0x8,%xmm10,%xmm11
16448 .byte 196,193,41,113,210,8 // vpsrlw $0x8,%xmm10,%xmm10
16449 .byte 196,65,33,235,210 // vpor %xmm10,%xmm11,%xmm10
16450 .byte 197,60,89,218 // vmulps %ymm2,%ymm8,%ymm11
16451 .byte 196,65,125,91,219 // vcvtps2dq %ymm11,%ymm11
16452 .byte 196,67,125,25,220,1 // vextractf128 $0x1,%ymm11,%xmm12
16453 .byte 196,66,33,43,220 // vpackusdw %xmm12,%xmm11,%xmm11
16454 .byte 196,193,25,113,243,8 // vpsllw $0x8,%xmm11,%xmm12
16455 .byte 196,193,33,113,211,8 // vpsrlw $0x8,%xmm11,%xmm11
16456 .byte 196,65,25,235,219 // vpor %xmm11,%xmm12,%xmm11
16457 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
16458 .byte 196,65,125,91,192 // vcvtps2dq %ymm8,%ymm8
16459 .byte 196,67,125,25,196,1 // vextractf128 $0x1,%ymm8,%xmm12
16460 .byte 196,66,57,43,196 // vpackusdw %xmm12,%xmm8,%xmm8
16461 .byte 196,193,25,113,240,8 // vpsllw $0x8,%xmm8,%xmm12
16462 .byte 196,193,57,113,208,8 // vpsrlw $0x8,%xmm8,%xmm8
16463 .byte 196,65,25,235,192 // vpor %xmm8,%xmm12,%xmm8
16464 .byte 196,65,49,97,226 // vpunpcklwd %xmm10,%xmm9,%xmm12
16465 .byte 196,65,49,105,234 // vpunpckhwd %xmm10,%xmm9,%xmm13
16466 .byte 196,65,33,97,200 // vpunpcklwd %xmm8,%xmm11,%xmm9
16467 .byte 196,65,33,105,192 // vpunpckhwd %xmm8,%xmm11,%xmm8
16468 .byte 196,65,25,98,217 // vpunpckldq %xmm9,%xmm12,%xmm11
16469 .byte 196,65,25,106,209 // vpunpckhdq %xmm9,%xmm12,%xmm10
16470 .byte 196,65,17,98,200 // vpunpckldq %xmm8,%xmm13,%xmm9
16471 .byte 196,65,17,106,192 // vpunpckhdq %xmm8,%xmm13,%xmm8
16472 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16473 .byte 117,31 // jne 4d65 <_sk_store_u16_be_avx+0x102>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16474 .byte 196,1,120,17,28,72 // vmovups %xmm11,(%r8,%r9,2)
16475 .byte 196,1,120,17,84,72,16 // vmovups %xmm10,0x10(%r8,%r9,2)
16476 .byte 196,1,120,17,76,72,32 // vmovups %xmm9,0x20(%r8,%r9,2)
16477 .byte 196,1,122,127,68,72,48 // vmovdqu %xmm8,0x30(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16478 .byte 72,173 // lods %ds:(%rsi),%rax
16479 .byte 255,224 // jmpq *%rax
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16480 .byte 196,1,121,214,28,72 // vmovq %xmm11,(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16481 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16482 .byte 116,240 // je 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16483 .byte 196,1,121,23,92,72,8 // vmovhpd %xmm11,0x8(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16484 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16485 .byte 114,227 // jb 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16486 .byte 196,1,121,214,84,72,16 // vmovq %xmm10,0x10(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16487 .byte 116,218 // je 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16488 .byte 196,1,121,23,84,72,24 // vmovhpd %xmm10,0x18(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16489 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16490 .byte 114,205 // jb 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16491 .byte 196,1,121,214,76,72,32 // vmovq %xmm9,0x20(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16492 .byte 116,196 // je 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16493 .byte 196,1,121,23,76,72,40 // vmovhpd %xmm9,0x28(%r8,%r9,2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16494 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16495 .byte 114,183 // jb 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]16496 .byte 196,1,121,214,68,72,48 // vmovq %xmm8,0x30(%r8,%r9,2)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16497 .byte 235,174 // jmp 4d61 <_sk_store_u16_be_avx+0xfe>
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]16498
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16499HIDDEN _sk_load_f32_avx
16500.globl _sk_load_f32_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16501FUNCTION(_sk_load_f32_avx)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16502_sk_load_f32_avx:
16503 .byte 72,173 // lods %ds:(%rsi),%rax
16504 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16505 .byte 119,110 // ja 4e29 <_sk_load_f32_avx+0x76>
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16506 .byte 76,139,0 // mov (%rax),%r8
16507 .byte 76,141,12,189,0,0,0,0 // lea 0x0(,%rdi,4),%r9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16508 .byte 76,141,21,135,0,0,0 // lea 0x87(%rip),%r10 # 4e54 <_sk_load_f32_avx+0xa1>
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16509 .byte 73,99,4,138 // movslq (%r10,%rcx,4),%rax
16510 .byte 76,1,208 // add %r10,%rax
16511 .byte 255,224 // jmpq *%rax
16512 .byte 196,3,125,24,68,136,112,1 // vinsertf128 $0x1,0x70(%r8,%r9,4),%ymm0,%ymm8
16513 .byte 196,131,125,24,92,136,96,1 // vinsertf128 $0x1,0x60(%r8,%r9,4),%ymm0,%ymm3
16514 .byte 196,131,125,24,76,136,80,1 // vinsertf128 $0x1,0x50(%r8,%r9,4),%ymm0,%ymm1
16515 .byte 196,131,125,24,84,136,64,1 // vinsertf128 $0x1,0x40(%r8,%r9,4),%ymm0,%ymm2
16516 .byte 196,129,121,16,68,136,48 // vmovupd 0x30(%r8,%r9,4),%xmm0
16517 .byte 196,195,125,13,192,12 // vblendpd $0xc,%ymm8,%ymm0,%ymm0
16518 .byte 196,1,121,16,68,136,32 // vmovupd 0x20(%r8,%r9,4),%xmm8
16519 .byte 196,99,61,13,203,12 // vblendpd $0xc,%ymm3,%ymm8,%ymm9
16520 .byte 196,129,121,16,92,136,16 // vmovupd 0x10(%r8,%r9,4),%xmm3
16521 .byte 196,99,101,13,209,12 // vblendpd $0xc,%ymm1,%ymm3,%ymm10
16522 .byte 196,129,121,16,12,136 // vmovupd (%r8,%r9,4),%xmm1
16523 .byte 196,227,117,13,202,12 // vblendpd $0xc,%ymm2,%ymm1,%ymm1
16524 .byte 196,193,116,20,210 // vunpcklps %ymm10,%ymm1,%ymm2
16525 .byte 196,193,116,21,218 // vunpckhps %ymm10,%ymm1,%ymm3
16526 .byte 197,180,20,200 // vunpcklps %ymm0,%ymm9,%ymm1
16527 .byte 197,52,21,192 // vunpckhps %ymm0,%ymm9,%ymm8
16528 .byte 197,237,20,193 // vunpcklpd %ymm1,%ymm2,%ymm0
16529 .byte 197,237,21,201 // vunpckhpd %ymm1,%ymm2,%ymm1
16530 .byte 196,193,101,20,208 // vunpcklpd %ymm8,%ymm3,%ymm2
16531 .byte 196,193,101,21,216 // vunpckhpd %ymm8,%ymm3,%ymm3
16532 .byte 72,173 // lods %ds:(%rsi),%rax
16533 .byte 255,224 // jmpq *%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16534 .byte 15,31,0 // nopl (%rax)
16535 .byte 130 // (bad)
16536 .byte 255 // (bad)
16537 .byte 255 // (bad)
16538 .byte 255,201 // dec %ecx
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16539 .byte 255 // (bad)
16540 .byte 255 // (bad)
16541 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16542 .byte 188,255,255,255,175 // mov $0xafffffff,%esp
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16543 .byte 255 // (bad)
16544 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16545 .byte 255,162,255,255,255,154 // jmpq *-0x65000001(%rdx)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16546 .byte 255 // (bad)
16547 .byte 255 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16548 .byte 255,146,255,255,255,138 // callq *-0x75000001(%rdx)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]16549 .byte 255 // (bad)
16550 .byte 255 // (bad)
16551 .byte 255 // .byte 0xff
16552
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16553HIDDEN _sk_store_f32_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16554.globl _sk_store_f32_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16555FUNCTION(_sk_store_f32_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16556_sk_store_f32_avx:
16557 .byte 72,173 // lods %ds:(%rsi),%rax
16558 .byte 76,139,0 // mov (%rax),%r8
16559 .byte 72,141,4,189,0,0,0,0 // lea 0x0(,%rdi,4),%rax
16560 .byte 197,124,20,193 // vunpcklps %ymm1,%ymm0,%ymm8
16561 .byte 197,124,21,217 // vunpckhps %ymm1,%ymm0,%ymm11
16562 .byte 197,108,20,203 // vunpcklps %ymm3,%ymm2,%ymm9
16563 .byte 197,108,21,227 // vunpckhps %ymm3,%ymm2,%ymm12
16564 .byte 196,65,61,20,209 // vunpcklpd %ymm9,%ymm8,%ymm10
16565 .byte 196,65,61,21,201 // vunpckhpd %ymm9,%ymm8,%ymm9
16566 .byte 196,65,37,20,196 // vunpcklpd %ymm12,%ymm11,%ymm8
16567 .byte 196,65,37,21,220 // vunpckhpd %ymm12,%ymm11,%ymm11
16568 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16569 .byte 117,55 // jne 4ee1 <_sk_store_f32_avx+0x6d>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16570 .byte 196,67,45,24,225,1 // vinsertf128 $0x1,%xmm9,%ymm10,%ymm12
16571 .byte 196,67,61,24,235,1 // vinsertf128 $0x1,%xmm11,%ymm8,%ymm13
16572 .byte 196,67,45,6,201,49 // vperm2f128 $0x31,%ymm9,%ymm10,%ymm9
16573 .byte 196,67,61,6,195,49 // vperm2f128 $0x31,%ymm11,%ymm8,%ymm8
16574 .byte 196,65,125,17,36,128 // vmovupd %ymm12,(%r8,%rax,4)
16575 .byte 196,65,125,17,108,128,32 // vmovupd %ymm13,0x20(%r8,%rax,4)
16576 .byte 196,65,125,17,76,128,64 // vmovupd %ymm9,0x40(%r8,%rax,4)
16577 .byte 196,65,125,17,68,128,96 // vmovupd %ymm8,0x60(%r8,%rax,4)
16578 .byte 72,173 // lods %ds:(%rsi),%rax
16579 .byte 255,224 // jmpq *%rax
16580 .byte 196,65,121,17,20,128 // vmovupd %xmm10,(%r8,%rax,4)
16581 .byte 72,131,249,1 // cmp $0x1,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16582 .byte 116,240 // je 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16583 .byte 196,65,121,17,76,128,16 // vmovupd %xmm9,0x10(%r8,%rax,4)
16584 .byte 72,131,249,3 // cmp $0x3,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16585 .byte 114,227 // jb 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16586 .byte 196,65,121,17,68,128,32 // vmovupd %xmm8,0x20(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16587 .byte 116,218 // je 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16588 .byte 196,65,121,17,92,128,48 // vmovupd %xmm11,0x30(%r8,%rax,4)
16589 .byte 72,131,249,5 // cmp $0x5,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16590 .byte 114,205 // jb 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16591 .byte 196,67,125,25,84,128,64,1 // vextractf128 $0x1,%ymm10,0x40(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16592 .byte 116,195 // je 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16593 .byte 196,67,125,25,76,128,80,1 // vextractf128 $0x1,%ymm9,0x50(%r8,%rax,4)
16594 .byte 72,131,249,7 // cmp $0x7,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16595 .byte 114,181 // jb 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16596 .byte 196,67,125,25,68,128,96,1 // vextractf128 $0x1,%ymm8,0x60(%r8,%rax,4)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16597 .byte 235,171 // jmp 4edd <_sk_store_f32_avx+0x69>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16598
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16599HIDDEN _sk_clamp_x_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16600.globl _sk_clamp_x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16601FUNCTION(_sk_clamp_x_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16602_sk_clamp_x_avx:
16603 .byte 72,173 // lods %ds:(%rsi),%rax
16604 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
16605 .byte 197,60,95,200 // vmaxps %ymm0,%ymm8,%ymm9
16606 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16607 .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0
16608 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16609 .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0
16610 .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8
16611 .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0
16612 .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0
16613 .byte 72,173 // lods %ds:(%rsi),%rax
16614 .byte 255,224 // jmpq *%rax
16615
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16616HIDDEN _sk_clamp_y_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16617.globl _sk_clamp_y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16618FUNCTION(_sk_clamp_y_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16619_sk_clamp_y_avx:
16620 .byte 72,173 // lods %ds:(%rsi),%rax
16621 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
16622 .byte 197,60,95,201 // vmaxps %ymm1,%ymm8,%ymm9
16623 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16624 .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1
16625 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16626 .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1
16627 .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8
16628 .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1
16629 .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1
16630 .byte 72,173 // lods %ds:(%rsi),%rax
16631 .byte 255,224 // jmpq *%rax
16632
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16633HIDDEN _sk_repeat_x_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16634.globl _sk_repeat_x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16635FUNCTION(_sk_repeat_x_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16636_sk_repeat_x_avx:
16637 .byte 72,173 // lods %ds:(%rsi),%rax
16638 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16639 .byte 196,65,124,94,200 // vdivps %ymm8,%ymm0,%ymm9
16640 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
16641 .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9
16642 .byte 196,65,124,92,201 // vsubps %ymm9,%ymm0,%ymm9
16643 .byte 196,99,125,25,192,1 // vextractf128 $0x1,%ymm8,%xmm0
16644 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16645 .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0
16646 .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8
16647 .byte 196,227,61,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm8,%ymm0
16648 .byte 197,180,93,192 // vminps %ymm0,%ymm9,%ymm0
16649 .byte 72,173 // lods %ds:(%rsi),%rax
16650 .byte 255,224 // jmpq *%rax
16651
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16652HIDDEN _sk_repeat_y_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16653.globl _sk_repeat_y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16654FUNCTION(_sk_repeat_y_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16655_sk_repeat_y_avx:
16656 .byte 72,173 // lods %ds:(%rsi),%rax
16657 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16658 .byte 196,65,116,94,200 // vdivps %ymm8,%ymm1,%ymm9
16659 .byte 196,67,125,8,201,1 // vroundps $0x1,%ymm9,%ymm9
16660 .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9
16661 .byte 196,65,116,92,201 // vsubps %ymm9,%ymm1,%ymm9
16662 .byte 196,99,125,25,193,1 // vextractf128 $0x1,%ymm8,%xmm1
16663 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16664 .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1
16665 .byte 196,65,57,254,194 // vpaddd %xmm10,%xmm8,%xmm8
16666 .byte 196,227,61,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm8,%ymm1
16667 .byte 197,180,93,201 // vminps %ymm1,%ymm9,%ymm1
16668 .byte 72,173 // lods %ds:(%rsi),%rax
16669 .byte 255,224 // jmpq *%rax
16670
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16671HIDDEN _sk_mirror_x_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16672.globl _sk_mirror_x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16673FUNCTION(_sk_mirror_x_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16674_sk_mirror_x_avx:
16675 .byte 72,173 // lods %ds:(%rsi),%rax
16676 .byte 197,121,110,0 // vmovd (%rax),%xmm8
16677 .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9
16678 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
16679 .byte 196,65,124,92,209 // vsubps %ymm9,%ymm0,%ymm10
16680 .byte 196,193,58,88,192 // vaddss %xmm8,%xmm8,%xmm0
16681 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
16682 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
16683 .byte 197,44,94,192 // vdivps %ymm0,%ymm10,%ymm8
16684 .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8
16685 .byte 197,188,89,192 // vmulps %ymm0,%ymm8,%ymm0
16686 .byte 197,172,92,192 // vsubps %ymm0,%ymm10,%ymm0
16687 .byte 196,193,124,92,193 // vsubps %ymm9,%ymm0,%ymm0
16688 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
16689 .byte 197,60,92,192 // vsubps %ymm0,%ymm8,%ymm8
16690 .byte 197,60,84,192 // vandps %ymm0,%ymm8,%ymm8
16691 .byte 196,99,125,25,200,1 // vextractf128 $0x1,%ymm9,%xmm0
16692 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16693 .byte 196,193,121,254,194 // vpaddd %xmm10,%xmm0,%xmm0
16694 .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9
16695 .byte 196,227,53,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm9,%ymm0
16696 .byte 197,188,93,192 // vminps %ymm0,%ymm8,%ymm0
16697 .byte 72,173 // lods %ds:(%rsi),%rax
16698 .byte 255,224 // jmpq *%rax
16699
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16700HIDDEN _sk_mirror_y_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16701.globl _sk_mirror_y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16702FUNCTION(_sk_mirror_y_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16703_sk_mirror_y_avx:
16704 .byte 72,173 // lods %ds:(%rsi),%rax
16705 .byte 197,121,110,0 // vmovd (%rax),%xmm8
16706 .byte 196,65,121,112,200,0 // vpshufd $0x0,%xmm8,%xmm9
16707 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
16708 .byte 196,65,116,92,209 // vsubps %ymm9,%ymm1,%ymm10
16709 .byte 196,193,58,88,200 // vaddss %xmm8,%xmm8,%xmm1
16710 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
16711 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
16712 .byte 197,44,94,193 // vdivps %ymm1,%ymm10,%ymm8
16713 .byte 196,67,125,8,192,1 // vroundps $0x1,%ymm8,%ymm8
16714 .byte 197,188,89,201 // vmulps %ymm1,%ymm8,%ymm1
16715 .byte 197,172,92,201 // vsubps %ymm1,%ymm10,%ymm1
16716 .byte 196,193,116,92,201 // vsubps %ymm9,%ymm1,%ymm1
16717 .byte 196,65,60,87,192 // vxorps %ymm8,%ymm8,%ymm8
16718 .byte 197,60,92,193 // vsubps %ymm1,%ymm8,%ymm8
16719 .byte 197,60,84,193 // vandps %ymm1,%ymm8,%ymm8
16720 .byte 196,99,125,25,201,1 // vextractf128 $0x1,%ymm9,%xmm1
16721 .byte 196,65,41,118,210 // vpcmpeqd %xmm10,%xmm10,%xmm10
16722 .byte 196,193,113,254,202 // vpaddd %xmm10,%xmm1,%xmm1
16723 .byte 196,65,49,254,202 // vpaddd %xmm10,%xmm9,%xmm9
16724 .byte 196,227,53,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm9,%ymm1
16725 .byte 197,188,93,201 // vminps %ymm1,%ymm8,%ymm1
16726 .byte 72,173 // lods %ds:(%rsi),%rax
16727 .byte 255,224 // jmpq *%rax
16728
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16729HIDDEN _sk_luminance_to_alpha_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16730.globl _sk_luminance_to_alpha_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16731FUNCTION(_sk_luminance_to_alpha_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16732_sk_luminance_to_alpha_avx:
16733 .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax
16734 .byte 197,249,110,216 // vmovd %eax,%xmm3
16735 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
16736 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
16737 .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0
16738 .byte 184,89,23,55,63 // mov $0x3f371759,%eax
16739 .byte 197,249,110,216 // vmovd %eax,%xmm3
16740 .byte 196,227,121,4,219,0 // vpermilps $0x0,%xmm3,%xmm3
16741 .byte 196,227,101,24,219,1 // vinsertf128 $0x1,%xmm3,%ymm3,%ymm3
16742 .byte 197,228,89,201 // vmulps %ymm1,%ymm3,%ymm1
16743 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
16744 .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax
16745 .byte 197,249,110,200 // vmovd %eax,%xmm1
16746 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
16747 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
16748 .byte 197,244,89,202 // vmulps %ymm2,%ymm1,%ymm1
16749 .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3
16750 .byte 72,173 // lods %ds:(%rsi),%rax
16751 .byte 197,252,87,192 // vxorps %ymm0,%ymm0,%ymm0
16752 .byte 197,244,87,201 // vxorps %ymm1,%ymm1,%ymm1
16753 .byte 197,236,87,210 // vxorps %ymm2,%ymm2,%ymm2
16754 .byte 255,224 // jmpq *%rax
16755
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16756HIDDEN _sk_matrix_2x3_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16757.globl _sk_matrix_2x3_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16758FUNCTION(_sk_matrix_2x3_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16759_sk_matrix_2x3_avx:
16760 .byte 72,173 // lods %ds:(%rsi),%rax
16761 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16762 .byte 196,98,125,24,72,8 // vbroadcastss 0x8(%rax),%ymm9
16763 .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10
16764 .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9
16765 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16766 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
16767 .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
16768 .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9
16769 .byte 196,98,125,24,80,12 // vbroadcastss 0xc(%rax),%ymm10
16770 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
16771 .byte 197,172,89,201 // vmulps %ymm1,%ymm10,%ymm1
16772 .byte 196,193,116,88,203 // vaddps %ymm11,%ymm1,%ymm1
16773 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
16774 .byte 197,252,88,201 // vaddps %ymm1,%ymm0,%ymm1
16775 .byte 72,173 // lods %ds:(%rsi),%rax
16776 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
16777 .byte 255,224 // jmpq *%rax
16778
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16779HIDDEN _sk_matrix_3x4_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16780.globl _sk_matrix_3x4_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16781FUNCTION(_sk_matrix_3x4_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16782_sk_matrix_3x4_avx:
16783 .byte 72,173 // lods %ds:(%rsi),%rax
16784 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16785 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
16786 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
16787 .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11
16788 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
16789 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16790 .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9
16791 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16792 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
16793 .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
16794 .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9
16795 .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10
16796 .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11
16797 .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12
16798 .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11
16799 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
16800 .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10
16801 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16802 .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9
16803 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16804 .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10
16805 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
16806 .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12
16807 .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13
16808 .byte 197,156,89,210 // vmulps %ymm2,%ymm12,%ymm2
16809 .byte 196,193,108,88,213 // vaddps %ymm13,%ymm2,%ymm2
16810 .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1
16811 .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
16812 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
16813 .byte 197,252,88,209 // vaddps %ymm1,%ymm0,%ymm2
16814 .byte 72,173 // lods %ds:(%rsi),%rax
16815 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
16816 .byte 197,124,41,201 // vmovaps %ymm9,%ymm1
16817 .byte 255,224 // jmpq *%rax
16818
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16819HIDDEN _sk_matrix_4x5_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16820.globl _sk_matrix_4x5_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16821FUNCTION(_sk_matrix_4x5_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16822_sk_matrix_4x5_avx:
16823 .byte 72,173 // lods %ds:(%rsi),%rax
16824 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16825 .byte 196,98,125,24,72,16 // vbroadcastss 0x10(%rax),%ymm9
16826 .byte 196,98,125,24,80,32 // vbroadcastss 0x20(%rax),%ymm10
16827 .byte 196,98,125,24,88,48 // vbroadcastss 0x30(%rax),%ymm11
16828 .byte 196,98,125,24,96,64 // vbroadcastss 0x40(%rax),%ymm12
16829 .byte 197,36,89,219 // vmulps %ymm3,%ymm11,%ymm11
16830 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
16831 .byte 197,44,89,210 // vmulps %ymm2,%ymm10,%ymm10
16832 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16833 .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9
16834 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16835 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
16836 .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
16837 .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9
16838 .byte 196,98,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm10
16839 .byte 196,98,125,24,88,36 // vbroadcastss 0x24(%rax),%ymm11
16840 .byte 196,98,125,24,96,52 // vbroadcastss 0x34(%rax),%ymm12
16841 .byte 196,98,125,24,104,68 // vbroadcastss 0x44(%rax),%ymm13
16842 .byte 197,28,89,227 // vmulps %ymm3,%ymm12,%ymm12
16843 .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
16844 .byte 197,36,89,218 // vmulps %ymm2,%ymm11,%ymm11
16845 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
16846 .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10
16847 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16848 .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9
16849 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16850 .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10
16851 .byte 196,98,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm11
16852 .byte 196,98,125,24,96,40 // vbroadcastss 0x28(%rax),%ymm12
16853 .byte 196,98,125,24,104,56 // vbroadcastss 0x38(%rax),%ymm13
16854 .byte 196,98,125,24,112,72 // vbroadcastss 0x48(%rax),%ymm14
16855 .byte 197,20,89,235 // vmulps %ymm3,%ymm13,%ymm13
16856 .byte 196,65,20,88,238 // vaddps %ymm14,%ymm13,%ymm13
16857 .byte 197,28,89,226 // vmulps %ymm2,%ymm12,%ymm12
16858 .byte 196,65,28,88,229 // vaddps %ymm13,%ymm12,%ymm12
16859 .byte 197,36,89,217 // vmulps %ymm1,%ymm11,%ymm11
16860 .byte 196,65,36,88,220 // vaddps %ymm12,%ymm11,%ymm11
16861 .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10
16862 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16863 .byte 196,98,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm11
16864 .byte 196,98,125,24,96,28 // vbroadcastss 0x1c(%rax),%ymm12
16865 .byte 196,98,125,24,104,44 // vbroadcastss 0x2c(%rax),%ymm13
16866 .byte 196,98,125,24,112,60 // vbroadcastss 0x3c(%rax),%ymm14
16867 .byte 196,98,125,24,120,76 // vbroadcastss 0x4c(%rax),%ymm15
16868 .byte 197,140,89,219 // vmulps %ymm3,%ymm14,%ymm3
16869 .byte 196,193,100,88,223 // vaddps %ymm15,%ymm3,%ymm3
16870 .byte 197,148,89,210 // vmulps %ymm2,%ymm13,%ymm2
16871 .byte 197,236,88,211 // vaddps %ymm3,%ymm2,%ymm2
16872 .byte 197,156,89,201 // vmulps %ymm1,%ymm12,%ymm1
16873 .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
16874 .byte 197,164,89,192 // vmulps %ymm0,%ymm11,%ymm0
16875 .byte 197,252,88,217 // vaddps %ymm1,%ymm0,%ymm3
16876 .byte 72,173 // lods %ds:(%rsi),%rax
16877 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
16878 .byte 197,124,41,201 // vmovaps %ymm9,%ymm1
16879 .byte 197,124,41,210 // vmovaps %ymm10,%ymm2
16880 .byte 255,224 // jmpq *%rax
16881
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16882HIDDEN _sk_matrix_perspective_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16883.globl _sk_matrix_perspective_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16884FUNCTION(_sk_matrix_perspective_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16885_sk_matrix_perspective_avx:
16886 .byte 72,173 // lods %ds:(%rsi),%rax
16887 .byte 196,98,125,24,0 // vbroadcastss (%rax),%ymm8
16888 .byte 196,98,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm9
16889 .byte 196,98,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm10
16890 .byte 197,52,89,201 // vmulps %ymm1,%ymm9,%ymm9
16891 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16892 .byte 197,60,89,192 // vmulps %ymm0,%ymm8,%ymm8
16893 .byte 196,65,60,88,193 // vaddps %ymm9,%ymm8,%ymm8
16894 .byte 196,98,125,24,72,12 // vbroadcastss 0xc(%rax),%ymm9
16895 .byte 196,98,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm10
16896 .byte 196,98,125,24,88,20 // vbroadcastss 0x14(%rax),%ymm11
16897 .byte 197,44,89,209 // vmulps %ymm1,%ymm10,%ymm10
16898 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
16899 .byte 197,52,89,200 // vmulps %ymm0,%ymm9,%ymm9
16900 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
16901 .byte 196,98,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm10
16902 .byte 196,98,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm11
16903 .byte 196,98,125,24,96,32 // vbroadcastss 0x20(%rax),%ymm12
16904 .byte 197,164,89,201 // vmulps %ymm1,%ymm11,%ymm1
16905 .byte 196,193,116,88,204 // vaddps %ymm12,%ymm1,%ymm1
16906 .byte 197,172,89,192 // vmulps %ymm0,%ymm10,%ymm0
16907 .byte 197,252,88,193 // vaddps %ymm1,%ymm0,%ymm0
16908 .byte 197,252,83,200 // vrcpps %ymm0,%ymm1
16909 .byte 197,188,89,193 // vmulps %ymm1,%ymm8,%ymm0
16910 .byte 197,180,89,201 // vmulps %ymm1,%ymm9,%ymm1
16911 .byte 72,173 // lods %ds:(%rsi),%rax
16912 .byte 255,224 // jmpq *%rax
16913
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16914HIDDEN _sk_linear_gradient_avx
16915.globl _sk_linear_gradient_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16916FUNCTION(_sk_linear_gradient_avx)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16917_sk_linear_gradient_avx:
16918 .byte 72,173 // lods %ds:(%rsi),%rax
16919 .byte 196,98,125,24,64,16 // vbroadcastss 0x10(%rax),%ymm8
16920 .byte 196,226,125,24,72,20 // vbroadcastss 0x14(%rax),%ymm1
16921 .byte 196,226,125,24,80,24 // vbroadcastss 0x18(%rax),%ymm2
16922 .byte 196,226,125,24,88,28 // vbroadcastss 0x1c(%rax),%ymm3
16923 .byte 76,139,0 // mov (%rax),%r8
16924 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16925 .byte 15,132,146,0,0,0 // je 5495 <_sk_linear_gradient_avx+0xb8>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16926 .byte 72,139,64,8 // mov 0x8(%rax),%rax
16927 .byte 72,131,192,32 // add $0x20,%rax
16928 .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
16929 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
16930 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
16931 .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11
16932 .byte 196,98,125,24,104,224 // vbroadcastss -0x20(%rax),%ymm13
16933 .byte 196,65,124,194,237,1 // vcmpltps %ymm13,%ymm0,%ymm13
16934 .byte 196,98,125,24,112,228 // vbroadcastss -0x1c(%rax),%ymm14
16935 .byte 196,67,13,74,228,208 // vblendvps %ymm13,%ymm12,%ymm14,%ymm12
16936 .byte 196,98,125,24,112,232 // vbroadcastss -0x18(%rax),%ymm14
16937 .byte 196,67,13,74,219,208 // vblendvps %ymm13,%ymm11,%ymm14,%ymm11
16938 .byte 196,98,125,24,112,236 // vbroadcastss -0x14(%rax),%ymm14
16939 .byte 196,67,13,74,210,208 // vblendvps %ymm13,%ymm10,%ymm14,%ymm10
16940 .byte 196,98,125,24,112,240 // vbroadcastss -0x10(%rax),%ymm14
16941 .byte 196,67,13,74,201,208 // vblendvps %ymm13,%ymm9,%ymm14,%ymm9
16942 .byte 196,98,125,24,112,244 // vbroadcastss -0xc(%rax),%ymm14
16943 .byte 196,67,13,74,192,208 // vblendvps %ymm13,%ymm8,%ymm14,%ymm8
16944 .byte 196,98,125,24,112,248 // vbroadcastss -0x8(%rax),%ymm14
16945 .byte 196,227,13,74,201,208 // vblendvps %ymm13,%ymm1,%ymm14,%ymm1
16946 .byte 196,98,125,24,112,252 // vbroadcastss -0x4(%rax),%ymm14
16947 .byte 196,227,13,74,210,208 // vblendvps %ymm13,%ymm2,%ymm14,%ymm2
16948 .byte 196,98,125,24,48 // vbroadcastss (%rax),%ymm14
16949 .byte 196,227,13,74,219,208 // vblendvps %ymm13,%ymm3,%ymm14,%ymm3
16950 .byte 72,131,192,36 // add $0x24,%rax
16951 .byte 73,255,200 // dec %r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]16952 .byte 117,140 // jne 541f <_sk_linear_gradient_avx+0x42>
16953 .byte 235,20 // jmp 54a9 <_sk_linear_gradient_avx+0xcc>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16954 .byte 196,65,36,87,219 // vxorps %ymm11,%ymm11,%ymm11
16955 .byte 196,65,44,87,210 // vxorps %ymm10,%ymm10,%ymm10
16956 .byte 196,65,52,87,201 // vxorps %ymm9,%ymm9,%ymm9
16957 .byte 196,65,28,87,228 // vxorps %ymm12,%ymm12,%ymm12
16958 .byte 197,28,89,224 // vmulps %ymm0,%ymm12,%ymm12
16959 .byte 196,65,60,88,196 // vaddps %ymm12,%ymm8,%ymm8
16960 .byte 197,36,89,216 // vmulps %ymm0,%ymm11,%ymm11
16961 .byte 197,164,88,201 // vaddps %ymm1,%ymm11,%ymm1
16962 .byte 197,44,89,208 // vmulps %ymm0,%ymm10,%ymm10
16963 .byte 197,172,88,210 // vaddps %ymm2,%ymm10,%ymm2
16964 .byte 197,180,89,192 // vmulps %ymm0,%ymm9,%ymm0
16965 .byte 197,252,88,219 // vaddps %ymm3,%ymm0,%ymm3
16966 .byte 72,173 // lods %ds:(%rsi),%rax
16967 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
16968 .byte 255,224 // jmpq *%rax
16969
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]16970HIDDEN _sk_linear_gradient_2stops_avx
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16971.globl _sk_linear_gradient_2stops_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16972FUNCTION(_sk_linear_gradient_2stops_avx)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16973_sk_linear_gradient_2stops_avx:
16974 .byte 72,173 // lods %ds:(%rsi),%rax
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16975 .byte 196,226,125,24,8 // vbroadcastss (%rax),%ymm1
16976 .byte 196,226,125,24,80,16 // vbroadcastss 0x10(%rax),%ymm2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16977 .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]16978 .byte 197,116,88,194 // vaddps %ymm2,%ymm1,%ymm8
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16979 .byte 196,226,125,24,72,4 // vbroadcastss 0x4(%rax),%ymm1
16980 .byte 196,226,125,24,80,20 // vbroadcastss 0x14(%rax),%ymm2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16981 .byte 197,244,89,200 // vmulps %ymm0,%ymm1,%ymm1
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]16982 .byte 197,244,88,202 // vaddps %ymm2,%ymm1,%ymm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16983 .byte 196,226,125,24,80,8 // vbroadcastss 0x8(%rax),%ymm2
16984 .byte 196,226,125,24,88,24 // vbroadcastss 0x18(%rax),%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16985 .byte 197,236,89,208 // vmulps %ymm0,%ymm2,%ymm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]16986 .byte 197,236,88,211 // vaddps %ymm3,%ymm2,%ymm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]16987 .byte 196,226,125,24,88,12 // vbroadcastss 0xc(%rax),%ymm3
16988 .byte 196,98,125,24,72,28 // vbroadcastss 0x1c(%rax),%ymm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16989 .byte 197,228,89,192 // vmulps %ymm0,%ymm3,%ymm0
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]16990 .byte 196,193,124,88,217 // vaddps %ymm9,%ymm0,%ymm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]16991 .byte 72,173 // lods %ds:(%rsi),%rax
16992 .byte 197,124,41,192 // vmovaps %ymm8,%ymm0
16993 .byte 255,224 // jmpq *%rax
16994
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]16995HIDDEN _sk_save_xy_avx
16996.globl _sk_save_xy_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]16997FUNCTION(_sk_save_xy_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]16998_sk_save_xy_avx:
16999 .byte 72,173 // lods %ds:(%rsi),%rax
17000 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17001 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17002 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17003 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17004 .byte 197,60,88,200 // vaddps %ymm0,%ymm8,%ymm9
17005 .byte 196,67,125,8,209,1 // vroundps $0x1,%ymm9,%ymm10
17006 .byte 196,65,52,92,202 // vsubps %ymm10,%ymm9,%ymm9
17007 .byte 197,60,88,193 // vaddps %ymm1,%ymm8,%ymm8
17008 .byte 196,67,125,8,208,1 // vroundps $0x1,%ymm8,%ymm10
17009 .byte 196,65,60,92,194 // vsubps %ymm10,%ymm8,%ymm8
17010 .byte 197,252,17,0 // vmovups %ymm0,(%rax)
17011 .byte 197,252,17,72,32 // vmovups %ymm1,0x20(%rax)
17012 .byte 197,124,17,72,64 // vmovups %ymm9,0x40(%rax)
17013 .byte 197,124,17,64,96 // vmovups %ymm8,0x60(%rax)
17014 .byte 72,173 // lods %ds:(%rsi),%rax
17015 .byte 255,224 // jmpq *%rax
17016
17017HIDDEN _sk_accumulate_avx
17018.globl _sk_accumulate_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17019FUNCTION(_sk_accumulate_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17020_sk_accumulate_avx:
17021 .byte 72,173 // lods %ds:(%rsi),%rax
17022 .byte 197,124,16,128,128,0,0,0 // vmovups 0x80(%rax),%ymm8
17023 .byte 197,60,89,128,160,0,0,0 // vmulps 0xa0(%rax),%ymm8,%ymm8
17024 .byte 197,60,89,200 // vmulps %ymm0,%ymm8,%ymm9
17025 .byte 197,180,88,228 // vaddps %ymm4,%ymm9,%ymm4
17026 .byte 197,60,89,201 // vmulps %ymm1,%ymm8,%ymm9
17027 .byte 197,180,88,237 // vaddps %ymm5,%ymm9,%ymm5
17028 .byte 197,60,89,202 // vmulps %ymm2,%ymm8,%ymm9
17029 .byte 197,180,88,246 // vaddps %ymm6,%ymm9,%ymm6
17030 .byte 197,60,89,195 // vmulps %ymm3,%ymm8,%ymm8
17031 .byte 197,188,88,255 // vaddps %ymm7,%ymm8,%ymm7
17032 .byte 72,173 // lods %ds:(%rsi),%rax
17033 .byte 255,224 // jmpq *%rax
17034
17035HIDDEN _sk_bilinear_nx_avx
17036.globl _sk_bilinear_nx_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17037FUNCTION(_sk_bilinear_nx_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17038_sk_bilinear_nx_avx:
17039 .byte 72,173 // lods %ds:(%rsi),%rax
17040 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
17041 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17042 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17043 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
17044 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
17045 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17046 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17047 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17048 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17049 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
17050 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17051 .byte 72,173 // lods %ds:(%rsi),%rax
17052 .byte 255,224 // jmpq *%rax
17053
17054HIDDEN _sk_bilinear_px_avx
17055.globl _sk_bilinear_px_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17056FUNCTION(_sk_bilinear_px_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17057_sk_bilinear_px_avx:
17058 .byte 72,173 // lods %ds:(%rsi),%rax
17059 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17060 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17061 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17062 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
17063 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
17064 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
17065 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17066 .byte 72,173 // lods %ds:(%rsi),%rax
17067 .byte 255,224 // jmpq *%rax
17068
17069HIDDEN _sk_bilinear_ny_avx
17070.globl _sk_bilinear_ny_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17071FUNCTION(_sk_bilinear_ny_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17072_sk_bilinear_ny_avx:
17073 .byte 72,173 // lods %ds:(%rsi),%rax
17074 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
17075 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17076 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17077 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
17078 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
17079 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17080 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17081 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17082 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17083 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
17084 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17085 .byte 72,173 // lods %ds:(%rsi),%rax
17086 .byte 255,224 // jmpq *%rax
17087
17088HIDDEN _sk_bilinear_py_avx
17089.globl _sk_bilinear_py_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17090FUNCTION(_sk_bilinear_py_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17091_sk_bilinear_py_avx:
17092 .byte 72,173 // lods %ds:(%rsi),%rax
17093 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17094 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17095 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17096 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
17097 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
17098 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
17099 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17100 .byte 72,173 // lods %ds:(%rsi),%rax
17101 .byte 255,224 // jmpq *%rax
17102
17103HIDDEN _sk_bicubic_n3x_avx
17104.globl _sk_bicubic_n3x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17105FUNCTION(_sk_bicubic_n3x_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17106_sk_bicubic_n3x_avx:
17107 .byte 72,173 // lods %ds:(%rsi),%rax
17108 .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d
17109 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17110 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17111 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
17112 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
17113 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17114 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17115 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17116 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17117 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
17118 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
17119 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
17120 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17121 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17122 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17123 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
17124 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17125 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17126 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17127 .byte 196,65,44,89,192 // vmulps %ymm8,%ymm10,%ymm8
17128 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
17129 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17130 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17131 .byte 72,173 // lods %ds:(%rsi),%rax
17132 .byte 255,224 // jmpq *%rax
17133
17134HIDDEN _sk_bicubic_n1x_avx
17135.globl _sk_bicubic_n1x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17136FUNCTION(_sk_bicubic_n1x_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17137_sk_bicubic_n1x_avx:
17138 .byte 72,173 // lods %ds:(%rsi),%rax
17139 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
17140 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17141 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17142 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
17143 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
17144 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17145 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17146 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17147 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17148 .byte 197,60,92,64,64 // vsubps 0x40(%rax),%ymm8,%ymm8
17149 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
17150 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
17151 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
17152 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
17153 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17154 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17155 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17156 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17157 .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9
17158 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
17159 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17160 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17161 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17162 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17163 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
17164 .byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9
17165 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
17166 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17167 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17168 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17169 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
17170 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
17171 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17172 .byte 72,173 // lods %ds:(%rsi),%rax
17173 .byte 255,224 // jmpq *%rax
17174
17175HIDDEN _sk_bicubic_p1x_avx
17176.globl _sk_bicubic_p1x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17177FUNCTION(_sk_bicubic_p1x_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17178_sk_bicubic_p1x_avx:
17179 .byte 72,173 // lods %ds:(%rsi),%rax
17180 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17181 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17182 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17183 .byte 196,99,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm8
17184 .byte 197,188,88,0 // vaddps (%rax),%ymm8,%ymm0
17185 .byte 197,124,16,72,64 // vmovups 0x40(%rax),%ymm9
17186 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
17187 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17188 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17189 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17190 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17191 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17192 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17193 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17194 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
17195 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
17196 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
17197 .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
17198 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
17199 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17200 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17201 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17202 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17203 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
17204 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17205 .byte 72,173 // lods %ds:(%rsi),%rax
17206 .byte 255,224 // jmpq *%rax
17207
17208HIDDEN _sk_bicubic_p3x_avx
17209.globl _sk_bicubic_p3x_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17210FUNCTION(_sk_bicubic_p3x_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17211_sk_bicubic_p3x_avx:
17212 .byte 72,173 // lods %ds:(%rsi),%rax
17213 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17214 .byte 196,193,121,110,192 // vmovd %r8d,%xmm0
17215 .byte 196,227,121,4,192,0 // vpermilps $0x0,%xmm0,%xmm0
17216 .byte 196,227,125,24,192,1 // vinsertf128 $0x1,%xmm0,%ymm0,%ymm0
17217 .byte 197,252,88,0 // vaddps (%rax),%ymm0,%ymm0
17218 .byte 197,124,16,64,64 // vmovups 0x40(%rax),%ymm8
17219 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
17220 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
17221 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17222 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17223 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17224 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
17225 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17226 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17227 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17228 .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
17229 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
17230 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17231 .byte 197,124,17,128,128,0,0,0 // vmovups %ymm8,0x80(%rax)
17232 .byte 72,173 // lods %ds:(%rsi),%rax
17233 .byte 255,224 // jmpq *%rax
17234
17235HIDDEN _sk_bicubic_n3y_avx
17236.globl _sk_bicubic_n3y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17237FUNCTION(_sk_bicubic_n3y_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17238_sk_bicubic_n3y_avx:
17239 .byte 72,173 // lods %ds:(%rsi),%rax
17240 .byte 65,184,0,0,192,191 // mov $0xbfc00000,%r8d
17241 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17242 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17243 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
17244 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
17245 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17246 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17247 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17248 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17249 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
17250 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
17251 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
17252 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17253 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17254 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17255 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
17256 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17257 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17258 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17259 .byte 196,65,44,89,192 // vmulps %ymm8,%ymm10,%ymm8
17260 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
17261 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17262 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17263 .byte 72,173 // lods %ds:(%rsi),%rax
17264 .byte 255,224 // jmpq *%rax
17265
17266HIDDEN _sk_bicubic_n1y_avx
17267.globl _sk_bicubic_n1y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17268FUNCTION(_sk_bicubic_n1y_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17269_sk_bicubic_n1y_avx:
17270 .byte 72,173 // lods %ds:(%rsi),%rax
17271 .byte 65,184,0,0,0,191 // mov $0xbf000000,%r8d
17272 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17273 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17274 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
17275 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
17276 .byte 65,184,0,0,128,63 // mov $0x3f800000,%r8d
17277 .byte 196,65,121,110,192 // vmovd %r8d,%xmm8
17278 .byte 196,67,121,4,192,0 // vpermilps $0x0,%xmm8,%xmm8
17279 .byte 196,67,61,24,192,1 // vinsertf128 $0x1,%xmm8,%ymm8,%ymm8
17280 .byte 197,60,92,64,96 // vsubps 0x60(%rax),%ymm8,%ymm8
17281 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
17282 .byte 196,65,121,110,200 // vmovd %r8d,%xmm9
17283 .byte 196,67,121,4,201,0 // vpermilps $0x0,%xmm9,%xmm9
17284 .byte 196,67,53,24,201,1 // vinsertf128 $0x1,%xmm9,%ymm9,%ymm9
17285 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17286 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17287 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17288 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17289 .byte 196,65,52,89,200 // vmulps %ymm8,%ymm9,%ymm9
17290 .byte 196,65,52,88,202 // vaddps %ymm10,%ymm9,%ymm9
17291 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17292 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17293 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17294 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17295 .byte 196,65,60,89,201 // vmulps %ymm9,%ymm8,%ymm9
17296 .byte 196,65,44,88,201 // vaddps %ymm9,%ymm10,%ymm9
17297 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
17298 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17299 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17300 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17301 .byte 196,65,60,89,193 // vmulps %ymm9,%ymm8,%ymm8
17302 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
17303 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17304 .byte 72,173 // lods %ds:(%rsi),%rax
17305 .byte 255,224 // jmpq *%rax
17306
17307HIDDEN _sk_bicubic_p1y_avx
17308.globl _sk_bicubic_p1y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17309FUNCTION(_sk_bicubic_p1y_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17310_sk_bicubic_p1y_avx:
17311 .byte 72,173 // lods %ds:(%rsi),%rax
17312 .byte 65,184,0,0,0,63 // mov $0x3f000000,%r8d
17313 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17314 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17315 .byte 196,99,117,24,193,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm8
17316 .byte 197,188,88,72,32 // vaddps 0x20(%rax),%ymm8,%ymm1
17317 .byte 197,124,16,72,96 // vmovups 0x60(%rax),%ymm9
17318 .byte 65,184,85,85,149,191 // mov $0xbf955555,%r8d
17319 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17320 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17321 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17322 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17323 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17324 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17325 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17326 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
17327 .byte 196,65,44,88,211 // vaddps %ymm11,%ymm10,%ymm10
17328 .byte 196,65,52,89,210 // vmulps %ymm10,%ymm9,%ymm10
17329 .byte 196,65,60,88,194 // vaddps %ymm10,%ymm8,%ymm8
17330 .byte 65,184,57,142,99,61 // mov $0x3d638e39,%r8d
17331 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17332 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17333 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17334 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17335 .byte 196,65,44,88,192 // vaddps %ymm8,%ymm10,%ymm8
17336 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17337 .byte 72,173 // lods %ds:(%rsi),%rax
17338 .byte 255,224 // jmpq *%rax
17339
17340HIDDEN _sk_bicubic_p3y_avx
17341.globl _sk_bicubic_p3y_avx
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17342FUNCTION(_sk_bicubic_p3y_avx)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]17343_sk_bicubic_p3y_avx:
17344 .byte 72,173 // lods %ds:(%rsi),%rax
17345 .byte 65,184,0,0,192,63 // mov $0x3fc00000,%r8d
17346 .byte 196,193,121,110,200 // vmovd %r8d,%xmm1
17347 .byte 196,227,121,4,201,0 // vpermilps $0x0,%xmm1,%xmm1
17348 .byte 196,227,117,24,201,1 // vinsertf128 $0x1,%xmm1,%ymm1,%ymm1
17349 .byte 197,244,88,72,32 // vaddps 0x20(%rax),%ymm1,%ymm1
17350 .byte 197,124,16,64,96 // vmovups 0x60(%rax),%ymm8
17351 .byte 196,65,60,89,200 // vmulps %ymm8,%ymm8,%ymm9
17352 .byte 65,184,114,28,199,62 // mov $0x3ec71c72,%r8d
17353 .byte 196,65,121,110,208 // vmovd %r8d,%xmm10
17354 .byte 196,67,121,4,210,0 // vpermilps $0x0,%xmm10,%xmm10
17355 .byte 196,67,45,24,210,1 // vinsertf128 $0x1,%xmm10,%ymm10,%ymm10
17356 .byte 65,184,171,170,170,190 // mov $0xbeaaaaab,%r8d
17357 .byte 196,65,121,110,216 // vmovd %r8d,%xmm11
17358 .byte 196,67,121,4,219,0 // vpermilps $0x0,%xmm11,%xmm11
17359 .byte 196,67,37,24,219,1 // vinsertf128 $0x1,%xmm11,%ymm11,%ymm11
17360 .byte 196,65,60,89,194 // vmulps %ymm10,%ymm8,%ymm8
17361 .byte 196,65,60,88,195 // vaddps %ymm11,%ymm8,%ymm8
17362 .byte 196,65,52,89,192 // vmulps %ymm8,%ymm9,%ymm8
17363 .byte 197,124,17,128,160,0,0,0 // vmovups %ymm8,0xa0(%rax)
17364 .byte 72,173 // lods %ds:(%rsi),%rax
17365 .byte 255,224 // jmpq *%rax
17366
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17367HIDDEN _sk_callback_avx
17368.globl _sk_callback_avx
17369FUNCTION(_sk_callback_avx)
17370_sk_callback_avx:
17371 .byte 65,87 // push %r15
17372 .byte 65,86 // push %r14
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17373 .byte 65,85 // push %r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17374 .byte 65,84 // push %r12
17375 .byte 83 // push %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17376 .byte 72,129,236,144,0,0,0 // sub $0x90,%rsp
17377 .byte 197,252,17,124,36,96 // vmovups %ymm7,0x60(%rsp)
17378 .byte 197,252,17,116,36,64 // vmovups %ymm6,0x40(%rsp)
17379 .byte 197,252,17,108,36,32 // vmovups %ymm5,0x20(%rsp)
17380 .byte 197,252,17,36,36 // vmovups %ymm4,(%rsp)
17381 .byte 73,137,205 // mov %rcx,%r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17382 .byte 73,137,214 // mov %rdx,%r14
17383 .byte 73,137,255 // mov %rdi,%r15
17384 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17385 .byte 72,137,195 // mov %rax,%rbx
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17386 .byte 73,137,244 // mov %rsi,%r12
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17387 .byte 197,252,20,225 // vunpcklps %ymm1,%ymm0,%ymm4
17388 .byte 197,252,21,193 // vunpckhps %ymm1,%ymm0,%ymm0
17389 .byte 197,236,20,203 // vunpcklps %ymm3,%ymm2,%ymm1
17390 .byte 197,236,21,211 // vunpckhps %ymm3,%ymm2,%ymm2
17391 .byte 197,221,20,217 // vunpcklpd %ymm1,%ymm4,%ymm3
17392 .byte 197,221,21,201 // vunpckhpd %ymm1,%ymm4,%ymm1
17393 .byte 197,253,20,226 // vunpcklpd %ymm2,%ymm0,%ymm4
17394 .byte 197,253,21,194 // vunpckhpd %ymm2,%ymm0,%ymm0
17395 .byte 196,227,101,24,209,1 // vinsertf128 $0x1,%xmm1,%ymm3,%ymm2
17396 .byte 196,227,93,24,232,1 // vinsertf128 $0x1,%xmm0,%ymm4,%ymm5
17397 .byte 196,227,101,6,201,49 // vperm2f128 $0x31,%ymm1,%ymm3,%ymm1
17398 .byte 196,227,93,6,192,49 // vperm2f128 $0x31,%ymm0,%ymm4,%ymm0
17399 .byte 197,253,17,83,8 // vmovupd %ymm2,0x8(%rbx)
17400 .byte 197,253,17,107,40 // vmovupd %ymm5,0x28(%rbx)
17401 .byte 197,253,17,75,72 // vmovupd %ymm1,0x48(%rbx)
17402 .byte 197,253,17,67,104 // vmovupd %ymm0,0x68(%rbx)
17403 .byte 77,133,237 // test %r13,%r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17404 .byte 190,8,0,0,0 // mov $0x8,%esi
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17405 .byte 65,15,69,245 // cmovne %r13d,%esi
17406 .byte 72,137,223 // mov %rbx,%rdi
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17407 .byte 197,248,119 // vzeroupper
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17408 .byte 255,19 // callq *(%rbx)
17409 .byte 72,139,131,136,0,0,0 // mov 0x88(%rbx),%rax
17410 .byte 197,248,16,0 // vmovups (%rax),%xmm0
17411 .byte 197,248,16,72,16 // vmovups 0x10(%rax),%xmm1
17412 .byte 197,248,16,80,32 // vmovups 0x20(%rax),%xmm2
17413 .byte 197,248,16,88,48 // vmovups 0x30(%rax),%xmm3
17414 .byte 196,227,101,24,88,112,1 // vinsertf128 $0x1,0x70(%rax),%ymm3,%ymm3
17415 .byte 196,227,109,24,80,96,1 // vinsertf128 $0x1,0x60(%rax),%ymm2,%ymm2
17416 .byte 196,227,117,24,72,80,1 // vinsertf128 $0x1,0x50(%rax),%ymm1,%ymm1
17417 .byte 196,227,125,24,64,64,1 // vinsertf128 $0x1,0x40(%rax),%ymm0,%ymm0
17418 .byte 197,252,20,225 // vunpcklps %ymm1,%ymm0,%ymm4
17419 .byte 197,252,21,233 // vunpckhps %ymm1,%ymm0,%ymm5
17420 .byte 197,236,20,203 // vunpcklps %ymm3,%ymm2,%ymm1
17421 .byte 197,236,21,219 // vunpckhps %ymm3,%ymm2,%ymm3
17422 .byte 197,221,20,193 // vunpcklpd %ymm1,%ymm4,%ymm0
17423 .byte 197,221,21,201 // vunpckhpd %ymm1,%ymm4,%ymm1
17424 .byte 197,213,20,211 // vunpcklpd %ymm3,%ymm5,%ymm2
17425 .byte 197,213,21,219 // vunpckhpd %ymm3,%ymm5,%ymm3
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17426 .byte 76,137,230 // mov %r12,%rsi
17427 .byte 72,173 // lods %ds:(%rsi),%rax
17428 .byte 76,137,255 // mov %r15,%rdi
17429 .byte 76,137,242 // mov %r14,%rdx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17430 .byte 76,137,233 // mov %r13,%rcx
17431 .byte 197,252,16,36,36 // vmovups (%rsp),%ymm4
17432 .byte 197,252,16,108,36,32 // vmovups 0x20(%rsp),%ymm5
17433 .byte 197,252,16,116,36,64 // vmovups 0x40(%rsp),%ymm6
17434 .byte 197,252,16,124,36,96 // vmovups 0x60(%rsp),%ymm7
17435 .byte 72,129,196,144,0,0,0 // add $0x90,%rsp
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17436 .byte 91 // pop %rbx
17437 .byte 65,92 // pop %r12
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]17438 .byte 65,93 // pop %r13
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]17439 .byte 65,94 // pop %r14
17440 .byte 65,95 // pop %r15
17441 .byte 255,224 // jmpq *%rax
17442
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]17443BALIGN4
17444 .byte 0,0 // add %al,(%rax)
17445 .byte 0,63 // add %bh,(%rdi)
17446 .byte 0,0 // add %al,(%rax)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]17447 .byte 128,63,0 // cmpb $0x0,(%rdi)
17448 .byte 0,0 // add %al,(%rax)
17449 .byte 52,255 // xor $0xff,%al
17450 .byte 255 // (bad)
17451 .byte 127,0 // jg 5c60 <.literal4+0x10>
17452 .byte 0,0 // add %al,(%rax)
17453 .byte 0,63 // add %bh,(%rdi)
17454 .byte 119,115 // ja 5cd9 <.literal4+0x89>
17455 .byte 248 // clc
17456 .byte 194,117,191 // retq $0xbf75
17457 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
17458 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
17459 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
17460 .byte 154 // (bad)
17461 .byte 64,254 // rex (bad)
17462 .byte 210,221 // rcr %cl,%ch
17463 .byte 65,0,0 // add %al,(%r8)
17464 .byte 0,75,0 // add %cl,0x0(%rbx)
17465 .byte 0,0 // add %al,(%rax)
17466 .byte 52,255 // xor $0xff,%al
17467 .byte 255 // (bad)
17468 .byte 127,0 // jg 5c90 <.literal4+0x40>
17469 .byte 0,0 // add %al,(%rax)
17470 .byte 0,63 // add %bh,(%rdi)
17471 .byte 119,115 // ja 5d09 <.literal4+0xb9>
17472 .byte 248 // clc
17473 .byte 194,117,191 // retq $0xbf75
17474 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
17475 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
17476 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
17477 .byte 154 // (bad)
17478 .byte 64,254 // rex (bad)
17479 .byte 210,221 // rcr %cl,%ch
17480 .byte 65,0,0 // add %al,(%r8)
17481 .byte 0,75,0 // add %cl,0x0(%rbx)
17482 .byte 0,0 // add %al,(%rax)
17483 .byte 52,255 // xor $0xff,%al
17484 .byte 255 // (bad)
17485 .byte 127,0 // jg 5cc0 <.literal4+0x70>
17486 .byte 0,0 // add %al,(%rax)
17487 .byte 0,63 // add %bh,(%rdi)
17488 .byte 119,115 // ja 5d39 <.literal4+0xe9>
17489 .byte 248 // clc
17490 .byte 194,117,191 // retq $0xbf75
17491 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
17492 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
17493 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
17494 .byte 154 // (bad)
17495 .byte 64,254 // rex (bad)
17496 .byte 210,221 // rcr %cl,%ch
17497 .byte 65,0,0 // add %al,(%r8)
17498 .byte 0,75,0 // add %cl,0x0(%rbx)
17499 .byte 0,0 // add %al,(%rax)
17500 .byte 52,255 // xor $0xff,%al
17501 .byte 255 // (bad)
17502 .byte 127,0 // jg 5cf0 <.literal4+0xa0>
17503 .byte 0,0 // add %al,(%rax)
17504 .byte 0,63 // add %bh,(%rdi)
17505 .byte 119,115 // ja 5d69 <_sk_callback_avx+0x23f>
17506 .byte 248 // clc
17507 .byte 194,117,191 // retq $0xbf75
17508 .byte 191,63,249,68,180 // mov $0xb444f93f,%edi
17509 .byte 62,163,233,220,63,81,140,242,66,141 // movabs %eax,%ds:0x8d42f28c513fdce9
17510 .byte 188,190,63,248,245 // mov $0xf5f83fbe,%esp
17511 .byte 154 // (bad)
17512 .byte 64,254 // rex (bad)
17513 .byte 210,221 // rcr %cl,%ch
17514 .byte 65,0,0 // add %al,(%r8)
17515 .byte 0,75,0 // add %cl,0x0(%rbx)
17516 .byte 128,0,0 // addb $0x0,(%rax)
17517 .byte 0,0 // add %al,(%rax)
17518 .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax)
17519 .byte 0,0 // add %al,(%rax)
17520 .byte 0,56 // add %bh,(%rax)
17521 .byte 0,128,0,0,0,0 // add %al,0x0(%rax)
17522 .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax)
17523 .byte 0,0 // add %al,(%rax)
17524 .byte 0,56 // add %bh,(%rax)
17525 .byte 0,0 // add %al,(%rax)
17526 .byte 0,128,0,0,128,56 // add %al,0x38800000(%rax)
17527 .byte 0,64,254 // add %al,-0x2(%rax)
17528 .byte 255 // .byte 0xff
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]17529BALIGN32
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]17530
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17531HIDDEN _sk_start_pipeline_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17532.globl _sk_start_pipeline_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17533FUNCTION(_sk_start_pipeline_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17534_sk_start_pipeline_sse41:
17535 .byte 65,87 // push %r15
17536 .byte 65,86 // push %r14
17537 .byte 65,85 // push %r13
17538 .byte 65,84 // push %r12
17539 .byte 83 // push %rbx
17540 .byte 73,137,207 // mov %rcx,%r15
17541 .byte 73,137,214 // mov %rdx,%r14
17542 .byte 72,137,251 // mov %rdi,%rbx
17543 .byte 72,173 // lods %ds:(%rsi),%rax
17544 .byte 73,137,196 // mov %rax,%r12
17545 .byte 73,137,245 // mov %rsi,%r13
17546 .byte 72,141,67,4 // lea 0x4(%rbx),%rax
17547 .byte 76,57,248 // cmp %r15,%rax
17548 .byte 118,5 // jbe 28 <_sk_start_pipeline_sse41+0x28>
17549 .byte 72,137,216 // mov %rbx,%rax
17550 .byte 235,52 // jmp 5c <_sk_start_pipeline_sse41+0x5c>
17551 .byte 15,87,192 // xorps %xmm0,%xmm0
17552 .byte 15,87,201 // xorps %xmm1,%xmm1
17553 .byte 15,87,210 // xorps %xmm2,%xmm2
17554 .byte 15,87,219 // xorps %xmm3,%xmm3
17555 .byte 15,87,228 // xorps %xmm4,%xmm4
17556 .byte 15,87,237 // xorps %xmm5,%xmm5
17557 .byte 15,87,246 // xorps %xmm6,%xmm6
17558 .byte 15,87,255 // xorps %xmm7,%xmm7
17559 .byte 72,137,223 // mov %rbx,%rdi
17560 .byte 76,137,238 // mov %r13,%rsi
17561 .byte 76,137,242 // mov %r14,%rdx
17562 .byte 65,255,212 // callq *%r12
17563 .byte 72,141,67,4 // lea 0x4(%rbx),%rax
17564 .byte 72,131,195,8 // add $0x8,%rbx
17565 .byte 76,57,251 // cmp %r15,%rbx
17566 .byte 72,137,195 // mov %rax,%rbx
17567 .byte 118,204 // jbe 28 <_sk_start_pipeline_sse41+0x28>
17568 .byte 91 // pop %rbx
17569 .byte 65,92 // pop %r12
17570 .byte 65,93 // pop %r13
17571 .byte 65,94 // pop %r14
17572 .byte 65,95 // pop %r15
17573 .byte 195 // retq
17574
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17575HIDDEN _sk_just_return_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17576.globl _sk_just_return_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17577FUNCTION(_sk_just_return_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17578_sk_just_return_sse41:
17579 .byte 195 // retq
17580
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17581HIDDEN _sk_seed_shader_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17582.globl _sk_seed_shader_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17583FUNCTION(_sk_seed_shader_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17584_sk_seed_shader_sse41:
17585 .byte 72,173 // lods %ds:(%rsi),%rax
17586 .byte 102,15,110,199 // movd %edi,%xmm0
17587 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
17588 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]17589 .byte 15,40,21,196,62,0,0 // movaps 0x3ec4(%rip),%xmm2 # 3f40 <_sk_callback_sse41+0xe0>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17590 .byte 15,88,202 // addps %xmm2,%xmm1
17591 .byte 15,16,2 // movups (%rdx),%xmm0
17592 .byte 15,88,193 // addps %xmm1,%xmm0
17593 .byte 102,15,110,8 // movd (%rax),%xmm1
17594 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
17595 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
17596 .byte 15,88,202 // addps %xmm2,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17597 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]17598 .byte 15,40,21,179,62,0,0 // movaps 0x3eb3(%rip),%xmm2 # 3f50 <_sk_callback_sse41+0xf0>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17599 .byte 15,87,219 // xorps %xmm3,%xmm3
17600 .byte 15,87,228 // xorps %xmm4,%xmm4
17601 .byte 15,87,237 // xorps %xmm5,%xmm5
17602 .byte 15,87,246 // xorps %xmm6,%xmm6
17603 .byte 15,87,255 // xorps %xmm7,%xmm7
17604 .byte 255,224 // jmpq *%rax
17605
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17606HIDDEN _sk_constant_color_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17607.globl _sk_constant_color_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17608FUNCTION(_sk_constant_color_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17609_sk_constant_color_sse41:
17610 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]17611 .byte 243,15,16,0 // movss (%rax),%xmm0
17612 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17613 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]17614 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
17615 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
17616 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
17617 .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3
17618 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17619 .byte 72,173 // lods %ds:(%rsi),%rax
17620 .byte 255,224 // jmpq *%rax
17621
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17622HIDDEN _sk_clear_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17623.globl _sk_clear_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17624FUNCTION(_sk_clear_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17625_sk_clear_sse41:
17626 .byte 72,173 // lods %ds:(%rsi),%rax
17627 .byte 15,87,192 // xorps %xmm0,%xmm0
17628 .byte 15,87,201 // xorps %xmm1,%xmm1
17629 .byte 15,87,210 // xorps %xmm2,%xmm2
17630 .byte 15,87,219 // xorps %xmm3,%xmm3
17631 .byte 255,224 // jmpq *%rax
17632
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17633HIDDEN _sk_srcatop_sse41
17634.globl _sk_srcatop_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17635FUNCTION(_sk_srcatop_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17636_sk_srcatop_sse41:
17637 .byte 15,89,199 // mulps %xmm7,%xmm0
17638 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17639 .byte 102,68,15,110,192 // movd %eax,%xmm8
17640 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17641 .byte 68,15,92,195 // subps %xmm3,%xmm8
17642 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17643 .byte 68,15,89,204 // mulps %xmm4,%xmm9
17644 .byte 65,15,88,193 // addps %xmm9,%xmm0
17645 .byte 15,89,207 // mulps %xmm7,%xmm1
17646 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17647 .byte 68,15,89,205 // mulps %xmm5,%xmm9
17648 .byte 65,15,88,201 // addps %xmm9,%xmm1
17649 .byte 15,89,215 // mulps %xmm7,%xmm2
17650 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17651 .byte 68,15,89,206 // mulps %xmm6,%xmm9
17652 .byte 65,15,88,209 // addps %xmm9,%xmm2
17653 .byte 15,89,223 // mulps %xmm7,%xmm3
17654 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17655 .byte 65,15,88,216 // addps %xmm8,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17656 .byte 72,173 // lods %ds:(%rsi),%rax
17657 .byte 255,224 // jmpq *%rax
17658
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17659HIDDEN _sk_dstatop_sse41
17660.globl _sk_dstatop_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17661FUNCTION(_sk_dstatop_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17662_sk_dstatop_sse41:
17663 .byte 68,15,40,195 // movaps %xmm3,%xmm8
17664 .byte 68,15,89,196 // mulps %xmm4,%xmm8
17665 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17666 .byte 102,68,15,110,200 // movd %eax,%xmm9
17667 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
17668 .byte 68,15,92,207 // subps %xmm7,%xmm9
17669 .byte 65,15,89,193 // mulps %xmm9,%xmm0
17670 .byte 65,15,88,192 // addps %xmm8,%xmm0
17671 .byte 68,15,40,195 // movaps %xmm3,%xmm8
17672 .byte 68,15,89,197 // mulps %xmm5,%xmm8
17673 .byte 65,15,89,201 // mulps %xmm9,%xmm1
17674 .byte 65,15,88,200 // addps %xmm8,%xmm1
17675 .byte 68,15,40,195 // movaps %xmm3,%xmm8
17676 .byte 68,15,89,198 // mulps %xmm6,%xmm8
17677 .byte 65,15,89,209 // mulps %xmm9,%xmm2
17678 .byte 65,15,88,208 // addps %xmm8,%xmm2
17679 .byte 68,15,89,203 // mulps %xmm3,%xmm9
17680 .byte 15,89,223 // mulps %xmm7,%xmm3
17681 .byte 65,15,88,217 // addps %xmm9,%xmm3
17682 .byte 72,173 // lods %ds:(%rsi),%rax
17683 .byte 255,224 // jmpq *%rax
17684
17685HIDDEN _sk_srcin_sse41
17686.globl _sk_srcin_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17687FUNCTION(_sk_srcin_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17688_sk_srcin_sse41:
17689 .byte 15,89,199 // mulps %xmm7,%xmm0
17690 .byte 15,89,207 // mulps %xmm7,%xmm1
17691 .byte 15,89,215 // mulps %xmm7,%xmm2
17692 .byte 15,89,223 // mulps %xmm7,%xmm3
17693 .byte 72,173 // lods %ds:(%rsi),%rax
17694 .byte 255,224 // jmpq *%rax
17695
17696HIDDEN _sk_dstin_sse41
17697.globl _sk_dstin_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17698FUNCTION(_sk_dstin_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17699_sk_dstin_sse41:
17700 .byte 15,40,195 // movaps %xmm3,%xmm0
17701 .byte 15,89,196 // mulps %xmm4,%xmm0
17702 .byte 15,40,203 // movaps %xmm3,%xmm1
17703 .byte 15,89,205 // mulps %xmm5,%xmm1
17704 .byte 15,40,211 // movaps %xmm3,%xmm2
17705 .byte 15,89,214 // mulps %xmm6,%xmm2
17706 .byte 15,89,223 // mulps %xmm7,%xmm3
17707 .byte 72,173 // lods %ds:(%rsi),%rax
17708 .byte 255,224 // jmpq *%rax
17709
17710HIDDEN _sk_srcout_sse41
17711.globl _sk_srcout_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17712FUNCTION(_sk_srcout_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17713_sk_srcout_sse41:
17714 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17715 .byte 102,68,15,110,192 // movd %eax,%xmm8
17716 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17717 .byte 68,15,92,199 // subps %xmm7,%xmm8
17718 .byte 65,15,89,192 // mulps %xmm8,%xmm0
17719 .byte 65,15,89,200 // mulps %xmm8,%xmm1
17720 .byte 65,15,89,208 // mulps %xmm8,%xmm2
17721 .byte 65,15,89,216 // mulps %xmm8,%xmm3
17722 .byte 72,173 // lods %ds:(%rsi),%rax
17723 .byte 255,224 // jmpq *%rax
17724
17725HIDDEN _sk_dstout_sse41
17726.globl _sk_dstout_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17727FUNCTION(_sk_dstout_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17728_sk_dstout_sse41:
17729 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17730 .byte 102,68,15,110,192 // movd %eax,%xmm8
17731 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17732 .byte 68,15,92,195 // subps %xmm3,%xmm8
17733 .byte 65,15,40,192 // movaps %xmm8,%xmm0
17734 .byte 15,89,196 // mulps %xmm4,%xmm0
17735 .byte 65,15,40,200 // movaps %xmm8,%xmm1
17736 .byte 15,89,205 // mulps %xmm5,%xmm1
17737 .byte 65,15,40,208 // movaps %xmm8,%xmm2
17738 .byte 15,89,214 // mulps %xmm6,%xmm2
17739 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17740 .byte 72,173 // lods %ds:(%rsi),%rax
17741 .byte 65,15,40,216 // movaps %xmm8,%xmm3
17742 .byte 255,224 // jmpq *%rax
17743
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17744HIDDEN _sk_srcover_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17745.globl _sk_srcover_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17746FUNCTION(_sk_srcover_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17747_sk_srcover_sse41:
17748 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17749 .byte 102,68,15,110,192 // movd %eax,%xmm8
17750 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17751 .byte 68,15,92,195 // subps %xmm3,%xmm8
17752 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17753 .byte 68,15,89,204 // mulps %xmm4,%xmm9
17754 .byte 65,15,88,193 // addps %xmm9,%xmm0
17755 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17756 .byte 68,15,89,205 // mulps %xmm5,%xmm9
17757 .byte 65,15,88,201 // addps %xmm9,%xmm1
17758 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17759 .byte 68,15,89,206 // mulps %xmm6,%xmm9
17760 .byte 65,15,88,209 // addps %xmm9,%xmm2
17761 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17762 .byte 65,15,88,216 // addps %xmm8,%xmm3
17763 .byte 72,173 // lods %ds:(%rsi),%rax
17764 .byte 255,224 // jmpq *%rax
17765
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]17766HIDDEN _sk_dstover_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17767.globl _sk_dstover_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17768FUNCTION(_sk_dstover_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]17769_sk_dstover_sse41:
17770 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17771 .byte 102,68,15,110,192 // movd %eax,%xmm8
17772 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17773 .byte 68,15,92,199 // subps %xmm7,%xmm8
17774 .byte 65,15,89,192 // mulps %xmm8,%xmm0
17775 .byte 15,88,196 // addps %xmm4,%xmm0
17776 .byte 65,15,89,200 // mulps %xmm8,%xmm1
17777 .byte 15,88,205 // addps %xmm5,%xmm1
17778 .byte 65,15,89,208 // mulps %xmm8,%xmm2
17779 .byte 15,88,214 // addps %xmm6,%xmm2
17780 .byte 65,15,89,216 // mulps %xmm8,%xmm3
17781 .byte 15,88,223 // addps %xmm7,%xmm3
17782 .byte 72,173 // lods %ds:(%rsi),%rax
17783 .byte 255,224 // jmpq *%rax
17784
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17785HIDDEN _sk_modulate_sse41
17786.globl _sk_modulate_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17787FUNCTION(_sk_modulate_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17788_sk_modulate_sse41:
17789 .byte 15,89,196 // mulps %xmm4,%xmm0
17790 .byte 15,89,205 // mulps %xmm5,%xmm1
17791 .byte 15,89,214 // mulps %xmm6,%xmm2
17792 .byte 15,89,223 // mulps %xmm7,%xmm3
17793 .byte 72,173 // lods %ds:(%rsi),%rax
17794 .byte 255,224 // jmpq *%rax
17795
17796HIDDEN _sk_multiply_sse41
17797.globl _sk_multiply_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17798FUNCTION(_sk_multiply_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17799_sk_multiply_sse41:
17800 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17801 .byte 102,68,15,110,192 // movd %eax,%xmm8
17802 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
17803 .byte 69,15,40,200 // movaps %xmm8,%xmm9
17804 .byte 68,15,92,207 // subps %xmm7,%xmm9
17805 .byte 69,15,40,209 // movaps %xmm9,%xmm10
17806 .byte 68,15,89,208 // mulps %xmm0,%xmm10
17807 .byte 68,15,92,195 // subps %xmm3,%xmm8
17808 .byte 69,15,40,216 // movaps %xmm8,%xmm11
17809 .byte 68,15,89,220 // mulps %xmm4,%xmm11
17810 .byte 69,15,88,218 // addps %xmm10,%xmm11
17811 .byte 15,89,196 // mulps %xmm4,%xmm0
17812 .byte 65,15,88,195 // addps %xmm11,%xmm0
17813 .byte 69,15,40,209 // movaps %xmm9,%xmm10
17814 .byte 68,15,89,209 // mulps %xmm1,%xmm10
17815 .byte 69,15,40,216 // movaps %xmm8,%xmm11
17816 .byte 68,15,89,221 // mulps %xmm5,%xmm11
17817 .byte 69,15,88,218 // addps %xmm10,%xmm11
17818 .byte 15,89,205 // mulps %xmm5,%xmm1
17819 .byte 65,15,88,203 // addps %xmm11,%xmm1
17820 .byte 69,15,40,209 // movaps %xmm9,%xmm10
17821 .byte 68,15,89,210 // mulps %xmm2,%xmm10
17822 .byte 69,15,40,216 // movaps %xmm8,%xmm11
17823 .byte 68,15,89,222 // mulps %xmm6,%xmm11
17824 .byte 69,15,88,218 // addps %xmm10,%xmm11
17825 .byte 15,89,214 // mulps %xmm6,%xmm2
17826 .byte 65,15,88,211 // addps %xmm11,%xmm2
17827 .byte 68,15,89,203 // mulps %xmm3,%xmm9
17828 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17829 .byte 69,15,88,193 // addps %xmm9,%xmm8
17830 .byte 15,89,223 // mulps %xmm7,%xmm3
17831 .byte 65,15,88,216 // addps %xmm8,%xmm3
17832 .byte 72,173 // lods %ds:(%rsi),%rax
17833 .byte 255,224 // jmpq *%rax
17834
17835HIDDEN _sk_plus__sse41
17836.globl _sk_plus__sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17837FUNCTION(_sk_plus__sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17838_sk_plus__sse41:
17839 .byte 15,88,196 // addps %xmm4,%xmm0
17840 .byte 15,88,205 // addps %xmm5,%xmm1
17841 .byte 15,88,214 // addps %xmm6,%xmm2
17842 .byte 15,88,223 // addps %xmm7,%xmm3
17843 .byte 72,173 // lods %ds:(%rsi),%rax
17844 .byte 255,224 // jmpq *%rax
17845
17846HIDDEN _sk_screen_sse41
17847.globl _sk_screen_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17848FUNCTION(_sk_screen_sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17849_sk_screen_sse41:
17850 .byte 68,15,40,192 // movaps %xmm0,%xmm8
17851 .byte 68,15,88,196 // addps %xmm4,%xmm8
17852 .byte 15,89,196 // mulps %xmm4,%xmm0
17853 .byte 68,15,92,192 // subps %xmm0,%xmm8
17854 .byte 68,15,40,201 // movaps %xmm1,%xmm9
17855 .byte 68,15,88,205 // addps %xmm5,%xmm9
17856 .byte 15,89,205 // mulps %xmm5,%xmm1
17857 .byte 68,15,92,201 // subps %xmm1,%xmm9
17858 .byte 68,15,40,210 // movaps %xmm2,%xmm10
17859 .byte 68,15,88,214 // addps %xmm6,%xmm10
17860 .byte 15,89,214 // mulps %xmm6,%xmm2
17861 .byte 68,15,92,210 // subps %xmm2,%xmm10
17862 .byte 68,15,40,219 // movaps %xmm3,%xmm11
17863 .byte 68,15,88,223 // addps %xmm7,%xmm11
17864 .byte 15,89,223 // mulps %xmm7,%xmm3
17865 .byte 68,15,92,219 // subps %xmm3,%xmm11
17866 .byte 72,173 // lods %ds:(%rsi),%rax
17867 .byte 65,15,40,192 // movaps %xmm8,%xmm0
17868 .byte 65,15,40,201 // movaps %xmm9,%xmm1
17869 .byte 65,15,40,210 // movaps %xmm10,%xmm2
17870 .byte 65,15,40,219 // movaps %xmm11,%xmm3
17871 .byte 255,224 // jmpq *%rax
17872
17873HIDDEN _sk_xor__sse41
17874.globl _sk_xor__sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17875FUNCTION(_sk_xor__sse41)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]17876_sk_xor__sse41:
17877 .byte 68,15,40,195 // movaps %xmm3,%xmm8
17878 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17879 .byte 102,15,110,216 // movd %eax,%xmm3
17880 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
17881 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17882 .byte 68,15,92,207 // subps %xmm7,%xmm9
17883 .byte 65,15,89,193 // mulps %xmm9,%xmm0
17884 .byte 65,15,92,216 // subps %xmm8,%xmm3
17885 .byte 68,15,40,211 // movaps %xmm3,%xmm10
17886 .byte 68,15,89,212 // mulps %xmm4,%xmm10
17887 .byte 65,15,88,194 // addps %xmm10,%xmm0
17888 .byte 65,15,89,201 // mulps %xmm9,%xmm1
17889 .byte 68,15,40,211 // movaps %xmm3,%xmm10
17890 .byte 68,15,89,213 // mulps %xmm5,%xmm10
17891 .byte 65,15,88,202 // addps %xmm10,%xmm1
17892 .byte 65,15,89,209 // mulps %xmm9,%xmm2
17893 .byte 68,15,40,211 // movaps %xmm3,%xmm10
17894 .byte 68,15,89,214 // mulps %xmm6,%xmm10
17895 .byte 65,15,88,210 // addps %xmm10,%xmm2
17896 .byte 69,15,89,200 // mulps %xmm8,%xmm9
17897 .byte 15,89,223 // mulps %xmm7,%xmm3
17898 .byte 65,15,88,217 // addps %xmm9,%xmm3
17899 .byte 72,173 // lods %ds:(%rsi),%rax
17900 .byte 255,224 // jmpq *%rax
17901
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]17902HIDDEN _sk_darken_sse41
17903.globl _sk_darken_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17904FUNCTION(_sk_darken_sse41)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]17905_sk_darken_sse41:
17906 .byte 68,15,40,193 // movaps %xmm1,%xmm8
17907 .byte 68,15,40,200 // movaps %xmm0,%xmm9
17908 .byte 15,88,196 // addps %xmm4,%xmm0
17909 .byte 68,15,89,207 // mulps %xmm7,%xmm9
17910 .byte 15,40,203 // movaps %xmm3,%xmm1
17911 .byte 15,89,204 // mulps %xmm4,%xmm1
17912 .byte 68,15,95,201 // maxps %xmm1,%xmm9
17913 .byte 65,15,92,193 // subps %xmm9,%xmm0
17914 .byte 65,15,40,200 // movaps %xmm8,%xmm1
17915 .byte 15,88,205 // addps %xmm5,%xmm1
17916 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17917 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17918 .byte 68,15,89,205 // mulps %xmm5,%xmm9
17919 .byte 69,15,95,193 // maxps %xmm9,%xmm8
17920 .byte 65,15,92,200 // subps %xmm8,%xmm1
17921 .byte 68,15,40,194 // movaps %xmm2,%xmm8
17922 .byte 68,15,88,198 // addps %xmm6,%xmm8
17923 .byte 15,89,215 // mulps %xmm7,%xmm2
17924 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17925 .byte 68,15,89,206 // mulps %xmm6,%xmm9
17926 .byte 65,15,95,209 // maxps %xmm9,%xmm2
17927 .byte 68,15,92,194 // subps %xmm2,%xmm8
17928 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17929 .byte 102,15,110,208 // movd %eax,%xmm2
17930 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
17931 .byte 15,92,211 // subps %xmm3,%xmm2
17932 .byte 15,89,215 // mulps %xmm7,%xmm2
17933 .byte 15,88,218 // addps %xmm2,%xmm3
17934 .byte 72,173 // lods %ds:(%rsi),%rax
17935 .byte 65,15,40,208 // movaps %xmm8,%xmm2
17936 .byte 255,224 // jmpq *%rax
17937
17938HIDDEN _sk_lighten_sse41
17939.globl _sk_lighten_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17940FUNCTION(_sk_lighten_sse41)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]17941_sk_lighten_sse41:
17942 .byte 68,15,40,193 // movaps %xmm1,%xmm8
17943 .byte 68,15,40,200 // movaps %xmm0,%xmm9
17944 .byte 15,88,196 // addps %xmm4,%xmm0
17945 .byte 68,15,89,207 // mulps %xmm7,%xmm9
17946 .byte 15,40,203 // movaps %xmm3,%xmm1
17947 .byte 15,89,204 // mulps %xmm4,%xmm1
17948 .byte 68,15,93,201 // minps %xmm1,%xmm9
17949 .byte 65,15,92,193 // subps %xmm9,%xmm0
17950 .byte 65,15,40,200 // movaps %xmm8,%xmm1
17951 .byte 15,88,205 // addps %xmm5,%xmm1
17952 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17953 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17954 .byte 68,15,89,205 // mulps %xmm5,%xmm9
17955 .byte 69,15,93,193 // minps %xmm9,%xmm8
17956 .byte 65,15,92,200 // subps %xmm8,%xmm1
17957 .byte 68,15,40,194 // movaps %xmm2,%xmm8
17958 .byte 68,15,88,198 // addps %xmm6,%xmm8
17959 .byte 15,89,215 // mulps %xmm7,%xmm2
17960 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17961 .byte 68,15,89,206 // mulps %xmm6,%xmm9
17962 .byte 65,15,93,209 // minps %xmm9,%xmm2
17963 .byte 68,15,92,194 // subps %xmm2,%xmm8
17964 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
17965 .byte 102,15,110,208 // movd %eax,%xmm2
17966 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
17967 .byte 15,92,211 // subps %xmm3,%xmm2
17968 .byte 15,89,215 // mulps %xmm7,%xmm2
17969 .byte 15,88,218 // addps %xmm2,%xmm3
17970 .byte 72,173 // lods %ds:(%rsi),%rax
17971 .byte 65,15,40,208 // movaps %xmm8,%xmm2
17972 .byte 255,224 // jmpq *%rax
17973
17974HIDDEN _sk_difference_sse41
17975.globl _sk_difference_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]17976FUNCTION(_sk_difference_sse41)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]17977_sk_difference_sse41:
17978 .byte 68,15,40,193 // movaps %xmm1,%xmm8
17979 .byte 68,15,40,200 // movaps %xmm0,%xmm9
17980 .byte 15,88,196 // addps %xmm4,%xmm0
17981 .byte 68,15,89,207 // mulps %xmm7,%xmm9
17982 .byte 15,40,203 // movaps %xmm3,%xmm1
17983 .byte 15,89,204 // mulps %xmm4,%xmm1
17984 .byte 68,15,93,201 // minps %xmm1,%xmm9
17985 .byte 69,15,88,201 // addps %xmm9,%xmm9
17986 .byte 65,15,92,193 // subps %xmm9,%xmm0
17987 .byte 65,15,40,200 // movaps %xmm8,%xmm1
17988 .byte 15,88,205 // addps %xmm5,%xmm1
17989 .byte 68,15,89,199 // mulps %xmm7,%xmm8
17990 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17991 .byte 68,15,89,205 // mulps %xmm5,%xmm9
17992 .byte 69,15,93,193 // minps %xmm9,%xmm8
17993 .byte 69,15,88,192 // addps %xmm8,%xmm8
17994 .byte 65,15,92,200 // subps %xmm8,%xmm1
17995 .byte 68,15,40,194 // movaps %xmm2,%xmm8
17996 .byte 68,15,88,198 // addps %xmm6,%xmm8
17997 .byte 15,89,215 // mulps %xmm7,%xmm2
17998 .byte 68,15,40,203 // movaps %xmm3,%xmm9
17999 .byte 68,15,89,206 // mulps %xmm6,%xmm9
18000 .byte 65,15,93,209 // minps %xmm9,%xmm2
18001 .byte 15,88,210 // addps %xmm2,%xmm2
18002 .byte 68,15,92,194 // subps %xmm2,%xmm8
18003 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18004 .byte 102,15,110,208 // movd %eax,%xmm2
18005 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
18006 .byte 15,92,211 // subps %xmm3,%xmm2
18007 .byte 15,89,215 // mulps %xmm7,%xmm2
18008 .byte 15,88,218 // addps %xmm2,%xmm3
18009 .byte 72,173 // lods %ds:(%rsi),%rax
18010 .byte 65,15,40,208 // movaps %xmm8,%xmm2
18011 .byte 255,224 // jmpq *%rax
18012
18013HIDDEN _sk_exclusion_sse41
18014.globl _sk_exclusion_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18015FUNCTION(_sk_exclusion_sse41)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]18016_sk_exclusion_sse41:
18017 .byte 68,15,40,193 // movaps %xmm1,%xmm8
18018 .byte 15,40,200 // movaps %xmm0,%xmm1
18019 .byte 15,88,196 // addps %xmm4,%xmm0
18020 .byte 15,89,204 // mulps %xmm4,%xmm1
18021 .byte 15,88,201 // addps %xmm1,%xmm1
18022 .byte 15,92,193 // subps %xmm1,%xmm0
18023 .byte 65,15,40,200 // movaps %xmm8,%xmm1
18024 .byte 15,88,205 // addps %xmm5,%xmm1
18025 .byte 68,15,89,197 // mulps %xmm5,%xmm8
18026 .byte 69,15,88,192 // addps %xmm8,%xmm8
18027 .byte 65,15,92,200 // subps %xmm8,%xmm1
18028 .byte 68,15,40,194 // movaps %xmm2,%xmm8
18029 .byte 68,15,88,198 // addps %xmm6,%xmm8
18030 .byte 15,89,214 // mulps %xmm6,%xmm2
18031 .byte 15,88,210 // addps %xmm2,%xmm2
18032 .byte 68,15,92,194 // subps %xmm2,%xmm8
18033 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18034 .byte 102,15,110,208 // movd %eax,%xmm2
18035 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
18036 .byte 15,92,211 // subps %xmm3,%xmm2
18037 .byte 15,89,215 // mulps %xmm7,%xmm2
18038 .byte 15,88,218 // addps %xmm2,%xmm3
18039 .byte 72,173 // lods %ds:(%rsi),%rax
18040 .byte 65,15,40,208 // movaps %xmm8,%xmm2
18041 .byte 255,224 // jmpq *%rax
18042
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18043HIDDEN _sk_colorburn_sse41
18044.globl _sk_colorburn_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18045FUNCTION(_sk_colorburn_sse41)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18046_sk_colorburn_sse41:
18047 .byte 68,15,40,200 // movaps %xmm0,%xmm9
18048 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18049 .byte 102,68,15,110,208 // movd %eax,%xmm10
18050 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
18051 .byte 69,15,40,218 // movaps %xmm10,%xmm11
18052 .byte 68,15,92,223 // subps %xmm7,%xmm11
18053 .byte 69,15,40,227 // movaps %xmm11,%xmm12
18054 .byte 69,15,89,225 // mulps %xmm9,%xmm12
18055 .byte 69,15,87,192 // xorps %xmm8,%xmm8
18056 .byte 15,40,199 // movaps %xmm7,%xmm0
18057 .byte 15,92,196 // subps %xmm4,%xmm0
18058 .byte 15,89,195 // mulps %xmm3,%xmm0
18059 .byte 65,15,94,193 // divps %xmm9,%xmm0
18060 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18061 .byte 68,15,93,232 // minps %xmm0,%xmm13
18062 .byte 68,15,40,247 // movaps %xmm7,%xmm14
18063 .byte 69,15,92,245 // subps %xmm13,%xmm14
18064 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18065 .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0
18066 .byte 68,15,92,211 // subps %xmm3,%xmm10
18067 .byte 68,15,89,243 // mulps %xmm3,%xmm14
18068 .byte 69,15,88,244 // addps %xmm12,%xmm14
18069 .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14
18070 .byte 69,15,40,202 // movaps %xmm10,%xmm9
18071 .byte 68,15,89,204 // mulps %xmm4,%xmm9
18072 .byte 68,15,88,228 // addps %xmm4,%xmm12
18073 .byte 15,40,196 // movaps %xmm4,%xmm0
18074 .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0
18075 .byte 69,15,88,206 // addps %xmm14,%xmm9
18076 .byte 102,69,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm9
18077 .byte 69,15,40,227 // movaps %xmm11,%xmm12
18078 .byte 68,15,89,225 // mulps %xmm1,%xmm12
18079 .byte 15,40,199 // movaps %xmm7,%xmm0
18080 .byte 15,92,197 // subps %xmm5,%xmm0
18081 .byte 15,89,195 // mulps %xmm3,%xmm0
18082 .byte 15,94,193 // divps %xmm1,%xmm0
18083 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18084 .byte 68,15,93,232 // minps %xmm0,%xmm13
18085 .byte 68,15,40,247 // movaps %xmm7,%xmm14
18086 .byte 69,15,92,245 // subps %xmm13,%xmm14
18087 .byte 15,40,193 // movaps %xmm1,%xmm0
18088 .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0
18089 .byte 68,15,89,243 // mulps %xmm3,%xmm14
18090 .byte 69,15,88,244 // addps %xmm12,%xmm14
18091 .byte 102,68,15,56,20,241 // blendvps %xmm0,%xmm1,%xmm14
18092 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18093 .byte 15,89,205 // mulps %xmm5,%xmm1
18094 .byte 68,15,88,229 // addps %xmm5,%xmm12
18095 .byte 65,15,88,206 // addps %xmm14,%xmm1
18096 .byte 15,40,197 // movaps %xmm5,%xmm0
18097 .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0
18098 .byte 102,65,15,56,20,204 // blendvps %xmm0,%xmm12,%xmm1
18099 .byte 15,40,199 // movaps %xmm7,%xmm0
18100 .byte 15,92,198 // subps %xmm6,%xmm0
18101 .byte 15,89,195 // mulps %xmm3,%xmm0
18102 .byte 15,94,194 // divps %xmm2,%xmm0
18103 .byte 68,15,40,231 // movaps %xmm7,%xmm12
18104 .byte 68,15,93,224 // minps %xmm0,%xmm12
18105 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18106 .byte 69,15,92,236 // subps %xmm12,%xmm13
18107 .byte 68,15,89,218 // mulps %xmm2,%xmm11
18108 .byte 68,15,194,194,0 // cmpeqps %xmm2,%xmm8
18109 .byte 68,15,89,235 // mulps %xmm3,%xmm13
18110 .byte 69,15,88,235 // addps %xmm11,%xmm13
18111 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18112 .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13
18113 .byte 68,15,88,222 // addps %xmm6,%xmm11
18114 .byte 65,15,40,210 // movaps %xmm10,%xmm2
18115 .byte 15,89,214 // mulps %xmm6,%xmm2
18116 .byte 65,15,88,213 // addps %xmm13,%xmm2
18117 .byte 15,40,198 // movaps %xmm6,%xmm0
18118 .byte 15,194,199,0 // cmpeqps %xmm7,%xmm0
18119 .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2
18120 .byte 68,15,89,215 // mulps %xmm7,%xmm10
18121 .byte 65,15,88,218 // addps %xmm10,%xmm3
18122 .byte 72,173 // lods %ds:(%rsi),%rax
18123 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18124 .byte 255,224 // jmpq *%rax
18125
18126HIDDEN _sk_colordodge_sse41
18127.globl _sk_colordodge_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18128FUNCTION(_sk_colordodge_sse41)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18129_sk_colordodge_sse41:
18130 .byte 68,15,40,192 // movaps %xmm0,%xmm8
18131 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18132 .byte 102,68,15,110,208 // movd %eax,%xmm10
18133 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
18134 .byte 69,15,40,218 // movaps %xmm10,%xmm11
18135 .byte 68,15,92,223 // subps %xmm7,%xmm11
18136 .byte 69,15,40,227 // movaps %xmm11,%xmm12
18137 .byte 69,15,89,224 // mulps %xmm8,%xmm12
18138 .byte 68,15,40,203 // movaps %xmm3,%xmm9
18139 .byte 68,15,89,204 // mulps %xmm4,%xmm9
18140 .byte 15,40,195 // movaps %xmm3,%xmm0
18141 .byte 65,15,92,192 // subps %xmm8,%xmm0
18142 .byte 68,15,94,200 // divps %xmm0,%xmm9
18143 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18144 .byte 68,15,40,247 // movaps %xmm7,%xmm14
18145 .byte 69,15,93,241 // minps %xmm9,%xmm14
18146 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18147 .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0
18148 .byte 68,15,89,243 // mulps %xmm3,%xmm14
18149 .byte 69,15,88,244 // addps %xmm12,%xmm14
18150 .byte 102,69,15,56,20,240 // blendvps %xmm0,%xmm8,%xmm14
18151 .byte 69,15,87,201 // xorps %xmm9,%xmm9
18152 .byte 68,15,92,211 // subps %xmm3,%xmm10
18153 .byte 69,15,40,194 // movaps %xmm10,%xmm8
18154 .byte 68,15,89,196 // mulps %xmm4,%xmm8
18155 .byte 68,15,88,228 // addps %xmm4,%xmm12
18156 .byte 69,15,88,198 // addps %xmm14,%xmm8
18157 .byte 15,40,196 // movaps %xmm4,%xmm0
18158 .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0
18159 .byte 102,69,15,56,20,196 // blendvps %xmm0,%xmm12,%xmm8
18160 .byte 68,15,40,227 // movaps %xmm3,%xmm12
18161 .byte 68,15,89,229 // mulps %xmm5,%xmm12
18162 .byte 15,40,195 // movaps %xmm3,%xmm0
18163 .byte 15,92,193 // subps %xmm1,%xmm0
18164 .byte 68,15,94,224 // divps %xmm0,%xmm12
18165 .byte 69,15,40,243 // movaps %xmm11,%xmm14
18166 .byte 68,15,89,241 // mulps %xmm1,%xmm14
18167 .byte 69,15,93,236 // minps %xmm12,%xmm13
18168 .byte 15,40,193 // movaps %xmm1,%xmm0
18169 .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0
18170 .byte 68,15,89,235 // mulps %xmm3,%xmm13
18171 .byte 69,15,88,238 // addps %xmm14,%xmm13
18172 .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13
18173 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18174 .byte 15,89,205 // mulps %xmm5,%xmm1
18175 .byte 68,15,88,245 // addps %xmm5,%xmm14
18176 .byte 65,15,88,205 // addps %xmm13,%xmm1
18177 .byte 15,40,197 // movaps %xmm5,%xmm0
18178 .byte 65,15,194,193,0 // cmpeqps %xmm9,%xmm0
18179 .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1
18180 .byte 68,15,40,227 // movaps %xmm3,%xmm12
18181 .byte 68,15,89,230 // mulps %xmm6,%xmm12
18182 .byte 15,40,195 // movaps %xmm3,%xmm0
18183 .byte 15,92,194 // subps %xmm2,%xmm0
18184 .byte 68,15,94,224 // divps %xmm0,%xmm12
18185 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18186 .byte 68,15,89,218 // mulps %xmm2,%xmm11
18187 .byte 69,15,93,236 // minps %xmm12,%xmm13
18188 .byte 15,40,194 // movaps %xmm2,%xmm0
18189 .byte 15,194,195,0 // cmpeqps %xmm3,%xmm0
18190 .byte 68,15,89,235 // mulps %xmm3,%xmm13
18191 .byte 69,15,88,235 // addps %xmm11,%xmm13
18192 .byte 102,68,15,56,20,234 // blendvps %xmm0,%xmm2,%xmm13
18193 .byte 65,15,40,210 // movaps %xmm10,%xmm2
18194 .byte 15,89,214 // mulps %xmm6,%xmm2
18195 .byte 65,15,88,213 // addps %xmm13,%xmm2
18196 .byte 68,15,194,206,0 // cmpeqps %xmm6,%xmm9
18197 .byte 68,15,88,222 // addps %xmm6,%xmm11
18198 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18199 .byte 102,65,15,56,20,211 // blendvps %xmm0,%xmm11,%xmm2
18200 .byte 68,15,89,215 // mulps %xmm7,%xmm10
18201 .byte 65,15,88,218 // addps %xmm10,%xmm3
18202 .byte 72,173 // lods %ds:(%rsi),%rax
18203 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18204 .byte 255,224 // jmpq *%rax
18205
18206HIDDEN _sk_hardlight_sse41
18207.globl _sk_hardlight_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18208FUNCTION(_sk_hardlight_sse41)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18209_sk_hardlight_sse41:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18210 .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18211 .byte 68,15,40,229 // movaps %xmm5,%xmm12
18212 .byte 15,40,244 // movaps %xmm4,%xmm6
18213 .byte 15,40,227 // movaps %xmm3,%xmm4
18214 .byte 15,40,234 // movaps %xmm2,%xmm5
18215 .byte 68,15,40,200 // movaps %xmm0,%xmm9
18216 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18217 .byte 102,68,15,110,208 // movd %eax,%xmm10
18218 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
18219 .byte 65,15,40,210 // movaps %xmm10,%xmm2
18220 .byte 15,92,215 // subps %xmm7,%xmm2
18221 .byte 15,40,194 // movaps %xmm2,%xmm0
18222 .byte 65,15,89,193 // mulps %xmm9,%xmm0
18223 .byte 68,15,92,212 // subps %xmm4,%xmm10
18224 .byte 69,15,40,194 // movaps %xmm10,%xmm8
18225 .byte 68,15,89,198 // mulps %xmm6,%xmm8
18226 .byte 68,15,88,192 // addps %xmm0,%xmm8
18227 .byte 68,15,40,252 // movaps %xmm4,%xmm15
18228 .byte 69,15,92,249 // subps %xmm9,%xmm15
18229 .byte 15,89,223 // mulps %xmm7,%xmm3
18230 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18231 .byte 68,15,40,247 // movaps %xmm7,%xmm14
18232 .byte 15,40,199 // movaps %xmm7,%xmm0
18233 .byte 15,92,198 // subps %xmm6,%xmm0
18234 .byte 65,15,89,199 // mulps %xmm15,%xmm0
18235 .byte 15,88,192 // addps %xmm0,%xmm0
18236 .byte 68,15,40,251 // movaps %xmm3,%xmm15
18237 .byte 68,15,92,248 // subps %xmm0,%xmm15
18238 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18239 .byte 15,88,192 // addps %xmm0,%xmm0
18240 .byte 15,194,196,2 // cmpleps %xmm4,%xmm0
18241 .byte 68,15,89,206 // mulps %xmm6,%xmm9
18242 .byte 69,15,88,201 // addps %xmm9,%xmm9
18243 .byte 102,69,15,56,20,249 // blendvps %xmm0,%xmm9,%xmm15
18244 .byte 68,15,40,218 // movaps %xmm2,%xmm11
18245 .byte 68,15,89,217 // mulps %xmm1,%xmm11
18246 .byte 15,40,196 // movaps %xmm4,%xmm0
18247 .byte 15,92,193 // subps %xmm1,%xmm0
18248 .byte 69,15,40,204 // movaps %xmm12,%xmm9
18249 .byte 69,15,92,233 // subps %xmm9,%xmm13
18250 .byte 68,15,89,232 // mulps %xmm0,%xmm13
18251 .byte 69,15,88,237 // addps %xmm13,%xmm13
18252 .byte 68,15,40,227 // movaps %xmm3,%xmm12
18253 .byte 69,15,92,229 // subps %xmm13,%xmm12
18254 .byte 15,40,193 // movaps %xmm1,%xmm0
18255 .byte 15,88,192 // addps %xmm0,%xmm0
18256 .byte 15,194,196,2 // cmpleps %xmm4,%xmm0
18257 .byte 65,15,89,201 // mulps %xmm9,%xmm1
18258 .byte 69,15,40,233 // movaps %xmm9,%xmm13
18259 .byte 15,88,201 // addps %xmm1,%xmm1
18260 .byte 102,68,15,56,20,225 // blendvps %xmm0,%xmm1,%xmm12
18261 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18262 .byte 69,15,40,202 // movaps %xmm10,%xmm9
18263 .byte 68,15,89,215 // mulps %xmm7,%xmm10
18264 .byte 69,15,88,199 // addps %xmm15,%xmm8
18265 .byte 65,15,89,205 // mulps %xmm13,%xmm1
18266 .byte 65,15,88,203 // addps %xmm11,%xmm1
18267 .byte 65,15,88,204 // addps %xmm12,%xmm1
18268 .byte 15,89,213 // mulps %xmm5,%xmm2
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18269 .byte 68,15,40,92,36,232 // movaps -0x18(%rsp),%xmm11
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18270 .byte 69,15,89,203 // mulps %xmm11,%xmm9
18271 .byte 68,15,88,202 // addps %xmm2,%xmm9
18272 .byte 15,40,197 // movaps %xmm5,%xmm0
18273 .byte 15,88,192 // addps %xmm0,%xmm0
18274 .byte 15,194,196,2 // cmpleps %xmm4,%xmm0
18275 .byte 15,40,212 // movaps %xmm4,%xmm2
18276 .byte 15,92,213 // subps %xmm5,%xmm2
18277 .byte 65,15,89,235 // mulps %xmm11,%xmm5
18278 .byte 15,88,237 // addps %xmm5,%xmm5
18279 .byte 69,15,92,243 // subps %xmm11,%xmm14
18280 .byte 68,15,89,242 // mulps %xmm2,%xmm14
18281 .byte 69,15,88,246 // addps %xmm14,%xmm14
18282 .byte 65,15,92,222 // subps %xmm14,%xmm3
18283 .byte 102,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm3
18284 .byte 68,15,88,203 // addps %xmm3,%xmm9
18285 .byte 65,15,88,226 // addps %xmm10,%xmm4
18286 .byte 72,173 // lods %ds:(%rsi),%rax
18287 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18288 .byte 65,15,40,209 // movaps %xmm9,%xmm2
18289 .byte 15,40,220 // movaps %xmm4,%xmm3
18290 .byte 15,40,230 // movaps %xmm6,%xmm4
18291 .byte 65,15,40,237 // movaps %xmm13,%xmm5
18292 .byte 65,15,40,243 // movaps %xmm11,%xmm6
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18293 .byte 255,224 // jmpq *%rax
18294
18295HIDDEN _sk_overlay_sse41
18296.globl _sk_overlay_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18297FUNCTION(_sk_overlay_sse41)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18298_sk_overlay_sse41:
18299 .byte 68,15,40,201 // movaps %xmm1,%xmm9
18300 .byte 68,15,40,240 // movaps %xmm0,%xmm14
18301 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18302 .byte 102,68,15,110,208 // movd %eax,%xmm10
18303 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
18304 .byte 69,15,40,218 // movaps %xmm10,%xmm11
18305 .byte 68,15,92,223 // subps %xmm7,%xmm11
18306 .byte 65,15,40,195 // movaps %xmm11,%xmm0
18307 .byte 65,15,89,198 // mulps %xmm14,%xmm0
18308 .byte 68,15,92,211 // subps %xmm3,%xmm10
18309 .byte 69,15,40,194 // movaps %xmm10,%xmm8
18310 .byte 68,15,89,196 // mulps %xmm4,%xmm8
18311 .byte 68,15,88,192 // addps %xmm0,%xmm8
18312 .byte 68,15,40,235 // movaps %xmm3,%xmm13
18313 .byte 69,15,92,238 // subps %xmm14,%xmm13
18314 .byte 68,15,89,244 // mulps %xmm4,%xmm14
18315 .byte 15,40,207 // movaps %xmm7,%xmm1
18316 .byte 15,92,204 // subps %xmm4,%xmm1
18317 .byte 15,40,196 // movaps %xmm4,%xmm0
18318 .byte 15,88,192 // addps %xmm0,%xmm0
18319 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18320 .byte 69,15,88,246 // addps %xmm14,%xmm14
18321 .byte 68,15,40,227 // movaps %xmm3,%xmm12
18322 .byte 68,15,89,231 // mulps %xmm7,%xmm12
18323 .byte 65,15,89,205 // mulps %xmm13,%xmm1
18324 .byte 15,88,201 // addps %xmm1,%xmm1
18325 .byte 69,15,40,236 // movaps %xmm12,%xmm13
18326 .byte 68,15,92,233 // subps %xmm1,%xmm13
18327 .byte 102,69,15,56,20,238 // blendvps %xmm0,%xmm14,%xmm13
18328 .byte 69,15,88,197 // addps %xmm13,%xmm8
18329 .byte 65,15,40,195 // movaps %xmm11,%xmm0
18330 .byte 65,15,89,193 // mulps %xmm9,%xmm0
18331 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18332 .byte 15,89,205 // mulps %xmm5,%xmm1
18333 .byte 15,88,200 // addps %xmm0,%xmm1
18334 .byte 68,15,40,235 // movaps %xmm3,%xmm13
18335 .byte 69,15,92,233 // subps %xmm9,%xmm13
18336 .byte 68,15,89,205 // mulps %xmm5,%xmm9
18337 .byte 68,15,40,247 // movaps %xmm7,%xmm14
18338 .byte 68,15,92,245 // subps %xmm5,%xmm14
18339 .byte 15,40,197 // movaps %xmm5,%xmm0
18340 .byte 15,88,192 // addps %xmm0,%xmm0
18341 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18342 .byte 69,15,88,201 // addps %xmm9,%xmm9
18343 .byte 69,15,89,245 // mulps %xmm13,%xmm14
18344 .byte 69,15,88,246 // addps %xmm14,%xmm14
18345 .byte 69,15,40,236 // movaps %xmm12,%xmm13
18346 .byte 69,15,92,238 // subps %xmm14,%xmm13
18347 .byte 102,69,15,56,20,233 // blendvps %xmm0,%xmm9,%xmm13
18348 .byte 65,15,88,205 // addps %xmm13,%xmm1
18349 .byte 68,15,89,218 // mulps %xmm2,%xmm11
18350 .byte 69,15,40,202 // movaps %xmm10,%xmm9
18351 .byte 68,15,89,206 // mulps %xmm6,%xmm9
18352 .byte 69,15,88,203 // addps %xmm11,%xmm9
18353 .byte 68,15,40,219 // movaps %xmm3,%xmm11
18354 .byte 68,15,92,218 // subps %xmm2,%xmm11
18355 .byte 15,89,214 // mulps %xmm6,%xmm2
18356 .byte 68,15,40,239 // movaps %xmm7,%xmm13
18357 .byte 68,15,92,238 // subps %xmm6,%xmm13
18358 .byte 15,40,198 // movaps %xmm6,%xmm0
18359 .byte 15,88,192 // addps %xmm0,%xmm0
18360 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18361 .byte 15,88,210 // addps %xmm2,%xmm2
18362 .byte 69,15,89,235 // mulps %xmm11,%xmm13
18363 .byte 69,15,88,237 // addps %xmm13,%xmm13
18364 .byte 69,15,92,229 // subps %xmm13,%xmm12
18365 .byte 102,68,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm12
18366 .byte 69,15,88,204 // addps %xmm12,%xmm9
18367 .byte 68,15,89,215 // mulps %xmm7,%xmm10
18368 .byte 65,15,88,218 // addps %xmm10,%xmm3
18369 .byte 72,173 // lods %ds:(%rsi),%rax
18370 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18371 .byte 65,15,40,209 // movaps %xmm9,%xmm2
18372 .byte 255,224 // jmpq *%rax
18373
18374HIDDEN _sk_softlight_sse41
18375.globl _sk_softlight_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18376FUNCTION(_sk_softlight_sse41)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18377_sk_softlight_sse41:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18378 .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18379 .byte 15,40,244 // movaps %xmm4,%xmm6
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18380 .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18381 .byte 68,15,40,225 // movaps %xmm1,%xmm12
18382 .byte 68,15,40,192 // movaps %xmm0,%xmm8
18383 .byte 15,87,228 // xorps %xmm4,%xmm4
18384 .byte 15,194,231,1 // cmpltps %xmm7,%xmm4
18385 .byte 15,40,198 // movaps %xmm6,%xmm0
18386 .byte 15,94,199 // divps %xmm7,%xmm0
18387 .byte 15,84,196 // andps %xmm4,%xmm0
18388 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18389 .byte 102,15,110,200 // movd %eax,%xmm1
18390 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
18391 .byte 68,15,40,209 // movaps %xmm1,%xmm10
18392 .byte 68,15,92,208 // subps %xmm0,%xmm10
18393 .byte 68,15,40,240 // movaps %xmm0,%xmm14
18394 .byte 68,15,40,248 // movaps %xmm0,%xmm15
18395 .byte 15,82,208 // rsqrtps %xmm0,%xmm2
18396 .byte 68,15,83,218 // rcpps %xmm2,%xmm11
18397 .byte 68,15,92,216 // subps %xmm0,%xmm11
18398 .byte 15,88,192 // addps %xmm0,%xmm0
18399 .byte 15,88,192 // addps %xmm0,%xmm0
18400 .byte 15,40,208 // movaps %xmm0,%xmm2
18401 .byte 15,89,210 // mulps %xmm2,%xmm2
18402 .byte 15,88,208 // addps %xmm0,%xmm2
18403 .byte 69,15,40,200 // movaps %xmm8,%xmm9
18404 .byte 69,15,88,201 // addps %xmm9,%xmm9
18405 .byte 68,15,92,241 // subps %xmm1,%xmm14
18406 .byte 68,15,89,242 // mulps %xmm2,%xmm14
18407 .byte 184,0,0,224,64 // mov $0x40e00000,%eax
18408 .byte 102,68,15,110,232 // movd %eax,%xmm13
18409 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
18410 .byte 69,15,89,253 // mulps %xmm13,%xmm15
18411 .byte 69,15,88,254 // addps %xmm14,%xmm15
18412 .byte 15,40,198 // movaps %xmm6,%xmm0
18413 .byte 15,88,192 // addps %xmm0,%xmm0
18414 .byte 15,88,192 // addps %xmm0,%xmm0
18415 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18416 .byte 102,69,15,56,20,223 // blendvps %xmm0,%xmm15,%xmm11
18417 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18418 .byte 15,92,195 // subps %xmm3,%xmm0
18419 .byte 68,15,89,208 // mulps %xmm0,%xmm10
18420 .byte 68,15,88,211 // addps %xmm3,%xmm10
18421 .byte 68,15,89,214 // mulps %xmm6,%xmm10
18422 .byte 15,40,211 // movaps %xmm3,%xmm2
18423 .byte 15,89,214 // mulps %xmm6,%xmm2
18424 .byte 15,89,199 // mulps %xmm7,%xmm0
18425 .byte 68,15,89,216 // mulps %xmm0,%xmm11
18426 .byte 68,15,88,218 // addps %xmm2,%xmm11
18427 .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9
18428 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18429 .byte 102,69,15,56,20,218 // blendvps %xmm0,%xmm10,%xmm11
18430 .byte 68,15,40,213 // movaps %xmm5,%xmm10
18431 .byte 68,15,94,215 // divps %xmm7,%xmm10
18432 .byte 68,15,84,212 // andps %xmm4,%xmm10
18433 .byte 68,15,40,201 // movaps %xmm1,%xmm9
18434 .byte 69,15,92,202 // subps %xmm10,%xmm9
18435 .byte 65,15,40,210 // movaps %xmm10,%xmm2
18436 .byte 15,88,210 // addps %xmm2,%xmm2
18437 .byte 15,88,210 // addps %xmm2,%xmm2
18438 .byte 15,40,194 // movaps %xmm2,%xmm0
18439 .byte 15,89,192 // mulps %xmm0,%xmm0
18440 .byte 15,88,194 // addps %xmm2,%xmm0
18441 .byte 65,15,40,210 // movaps %xmm10,%xmm2
18442 .byte 15,92,209 // subps %xmm1,%xmm2
18443 .byte 15,89,208 // mulps %xmm0,%xmm2
18444 .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0
18445 .byte 68,15,83,240 // rcpps %xmm0,%xmm14
18446 .byte 69,15,92,242 // subps %xmm10,%xmm14
18447 .byte 69,15,89,213 // mulps %xmm13,%xmm10
18448 .byte 68,15,88,210 // addps %xmm2,%xmm10
18449 .byte 15,40,197 // movaps %xmm5,%xmm0
18450 .byte 15,88,192 // addps %xmm0,%xmm0
18451 .byte 15,88,192 // addps %xmm0,%xmm0
18452 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18453 .byte 102,69,15,56,20,242 // blendvps %xmm0,%xmm10,%xmm14
18454 .byte 65,15,40,196 // movaps %xmm12,%xmm0
18455 .byte 15,88,192 // addps %xmm0,%xmm0
18456 .byte 15,40,208 // movaps %xmm0,%xmm2
18457 .byte 15,92,211 // subps %xmm3,%xmm2
18458 .byte 68,15,89,202 // mulps %xmm2,%xmm9
18459 .byte 68,15,88,203 // addps %xmm3,%xmm9
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18460 .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18461 .byte 68,15,89,205 // mulps %xmm5,%xmm9
18462 .byte 15,89,215 // mulps %xmm7,%xmm2
18463 .byte 68,15,89,242 // mulps %xmm2,%xmm14
18464 .byte 15,40,211 // movaps %xmm3,%xmm2
18465 .byte 15,89,213 // mulps %xmm5,%xmm2
18466 .byte 68,15,88,242 // addps %xmm2,%xmm14
18467 .byte 68,15,40,249 // movaps %xmm1,%xmm15
18468 .byte 15,194,195,2 // cmpleps %xmm3,%xmm0
18469 .byte 102,69,15,56,20,241 // blendvps %xmm0,%xmm9,%xmm14
18470 .byte 68,15,40,209 // movaps %xmm1,%xmm10
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18471 .byte 15,40,108,36,216 // movaps -0x28(%rsp),%xmm5
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18472 .byte 15,40,197 // movaps %xmm5,%xmm0
18473 .byte 15,94,199 // divps %xmm7,%xmm0
18474 .byte 15,84,196 // andps %xmm4,%xmm0
18475 .byte 15,40,208 // movaps %xmm0,%xmm2
18476 .byte 15,92,209 // subps %xmm1,%xmm2
18477 .byte 15,92,200 // subps %xmm0,%xmm1
18478 .byte 68,15,89,232 // mulps %xmm0,%xmm13
18479 .byte 15,82,224 // rsqrtps %xmm0,%xmm4
18480 .byte 68,15,83,204 // rcpps %xmm4,%xmm9
18481 .byte 68,15,92,200 // subps %xmm0,%xmm9
18482 .byte 15,88,192 // addps %xmm0,%xmm0
18483 .byte 15,88,192 // addps %xmm0,%xmm0
18484 .byte 15,40,224 // movaps %xmm0,%xmm4
18485 .byte 15,89,228 // mulps %xmm4,%xmm4
18486 .byte 15,88,224 // addps %xmm0,%xmm4
18487 .byte 15,89,226 // mulps %xmm2,%xmm4
18488 .byte 68,15,88,236 // addps %xmm4,%xmm13
18489 .byte 15,40,197 // movaps %xmm5,%xmm0
18490 .byte 15,88,192 // addps %xmm0,%xmm0
18491 .byte 15,88,192 // addps %xmm0,%xmm0
18492 .byte 15,194,199,2 // cmpleps %xmm7,%xmm0
18493 .byte 102,69,15,56,20,205 // blendvps %xmm0,%xmm13,%xmm9
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18494 .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18495 .byte 65,15,40,197 // movaps %xmm13,%xmm0
18496 .byte 15,88,192 // addps %xmm0,%xmm0
18497 .byte 15,40,208 // movaps %xmm0,%xmm2
18498 .byte 15,92,211 // subps %xmm3,%xmm2
18499 .byte 15,89,202 // mulps %xmm2,%xmm1
18500 .byte 15,89,215 // mulps %xmm7,%xmm2
18501 .byte 68,15,89,202 // mulps %xmm2,%xmm9
18502 .byte 15,40,211 // movaps %xmm3,%xmm2
18503 .byte 15,89,213 // mulps %xmm5,%xmm2
18504 .byte 68,15,88,202 // addps %xmm2,%xmm9
18505 .byte 15,88,203 // addps %xmm3,%xmm1
18506 .byte 15,89,205 // mulps %xmm5,%xmm1
18507 .byte 15,40,213 // movaps %xmm5,%xmm2
18508 .byte 15,194,195,2 // cmpleps %xmm3,%xmm0
18509 .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9
18510 .byte 68,15,92,255 // subps %xmm7,%xmm15
18511 .byte 69,15,89,199 // mulps %xmm15,%xmm8
18512 .byte 69,15,89,231 // mulps %xmm15,%xmm12
18513 .byte 69,15,89,253 // mulps %xmm13,%xmm15
18514 .byte 68,15,92,211 // subps %xmm3,%xmm10
18515 .byte 65,15,40,194 // movaps %xmm10,%xmm0
18516 .byte 15,89,198 // mulps %xmm6,%xmm0
18517 .byte 68,15,88,192 // addps %xmm0,%xmm8
18518 .byte 69,15,88,195 // addps %xmm11,%xmm8
18519 .byte 65,15,40,194 // movaps %xmm10,%xmm0
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18520 .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18521 .byte 15,89,197 // mulps %xmm5,%xmm0
18522 .byte 68,15,88,224 // addps %xmm0,%xmm12
18523 .byte 69,15,88,230 // addps %xmm14,%xmm12
18524 .byte 65,15,40,194 // movaps %xmm10,%xmm0
18525 .byte 15,89,194 // mulps %xmm2,%xmm0
18526 .byte 65,15,88,199 // addps %xmm15,%xmm0
18527 .byte 68,15,88,200 // addps %xmm0,%xmm9
18528 .byte 68,15,89,215 // mulps %xmm7,%xmm10
18529 .byte 65,15,88,218 // addps %xmm10,%xmm3
18530 .byte 72,173 // lods %ds:(%rsi),%rax
18531 .byte 15,40,230 // movaps %xmm6,%xmm4
18532 .byte 15,40,242 // movaps %xmm2,%xmm6
18533 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18534 .byte 65,15,40,204 // movaps %xmm12,%xmm1
18535 .byte 65,15,40,209 // movaps %xmm9,%xmm2
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]18536 .byte 255,224 // jmpq *%rax
18537
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18538HIDDEN _sk_clamp_0_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18539.globl _sk_clamp_0_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18540FUNCTION(_sk_clamp_0_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18541_sk_clamp_0_sse41:
18542 .byte 69,15,87,192 // xorps %xmm8,%xmm8
18543 .byte 65,15,95,192 // maxps %xmm8,%xmm0
18544 .byte 65,15,95,200 // maxps %xmm8,%xmm1
18545 .byte 65,15,95,208 // maxps %xmm8,%xmm2
18546 .byte 65,15,95,216 // maxps %xmm8,%xmm3
18547 .byte 72,173 // lods %ds:(%rsi),%rax
18548 .byte 255,224 // jmpq *%rax
18549
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18550HIDDEN _sk_clamp_1_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18551.globl _sk_clamp_1_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18552FUNCTION(_sk_clamp_1_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18553_sk_clamp_1_sse41:
18554 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18555 .byte 102,68,15,110,192 // movd %eax,%xmm8
18556 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18557 .byte 65,15,93,192 // minps %xmm8,%xmm0
18558 .byte 65,15,93,200 // minps %xmm8,%xmm1
18559 .byte 65,15,93,208 // minps %xmm8,%xmm2
18560 .byte 65,15,93,216 // minps %xmm8,%xmm3
18561 .byte 72,173 // lods %ds:(%rsi),%rax
18562 .byte 255,224 // jmpq *%rax
18563
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18564HIDDEN _sk_clamp_a_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18565.globl _sk_clamp_a_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18566FUNCTION(_sk_clamp_a_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18567_sk_clamp_a_sse41:
18568 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18569 .byte 102,68,15,110,192 // movd %eax,%xmm8
18570 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18571 .byte 65,15,93,216 // minps %xmm8,%xmm3
18572 .byte 15,93,195 // minps %xmm3,%xmm0
18573 .byte 15,93,203 // minps %xmm3,%xmm1
18574 .byte 15,93,211 // minps %xmm3,%xmm2
18575 .byte 72,173 // lods %ds:(%rsi),%rax
18576 .byte 255,224 // jmpq *%rax
18577
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18578HIDDEN _sk_set_rgb_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18579.globl _sk_set_rgb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18580FUNCTION(_sk_set_rgb_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18581_sk_set_rgb_sse41:
18582 .byte 72,173 // lods %ds:(%rsi),%rax
18583 .byte 243,15,16,0 // movss (%rax),%xmm0
18584 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
18585 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
18586 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
18587 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
18588 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
18589 .byte 72,173 // lods %ds:(%rsi),%rax
18590 .byte 255,224 // jmpq *%rax
18591
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18592HIDDEN _sk_swap_rb_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18593.globl _sk_swap_rb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18594FUNCTION(_sk_swap_rb_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18595_sk_swap_rb_sse41:
18596 .byte 68,15,40,192 // movaps %xmm0,%xmm8
18597 .byte 72,173 // lods %ds:(%rsi),%rax
18598 .byte 15,40,194 // movaps %xmm2,%xmm0
18599 .byte 65,15,40,208 // movaps %xmm8,%xmm2
18600 .byte 255,224 // jmpq *%rax
18601
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18602HIDDEN _sk_swap_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18603.globl _sk_swap_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18604FUNCTION(_sk_swap_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18605_sk_swap_sse41:
18606 .byte 68,15,40,195 // movaps %xmm3,%xmm8
18607 .byte 68,15,40,202 // movaps %xmm2,%xmm9
18608 .byte 68,15,40,209 // movaps %xmm1,%xmm10
18609 .byte 68,15,40,216 // movaps %xmm0,%xmm11
18610 .byte 72,173 // lods %ds:(%rsi),%rax
18611 .byte 15,40,196 // movaps %xmm4,%xmm0
18612 .byte 15,40,205 // movaps %xmm5,%xmm1
18613 .byte 15,40,214 // movaps %xmm6,%xmm2
18614 .byte 15,40,223 // movaps %xmm7,%xmm3
18615 .byte 65,15,40,227 // movaps %xmm11,%xmm4
18616 .byte 65,15,40,234 // movaps %xmm10,%xmm5
18617 .byte 65,15,40,241 // movaps %xmm9,%xmm6
18618 .byte 65,15,40,248 // movaps %xmm8,%xmm7
18619 .byte 255,224 // jmpq *%rax
18620
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18621HIDDEN _sk_move_src_dst_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18622.globl _sk_move_src_dst_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18623FUNCTION(_sk_move_src_dst_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18624_sk_move_src_dst_sse41:
18625 .byte 72,173 // lods %ds:(%rsi),%rax
18626 .byte 15,40,224 // movaps %xmm0,%xmm4
18627 .byte 15,40,233 // movaps %xmm1,%xmm5
18628 .byte 15,40,242 // movaps %xmm2,%xmm6
18629 .byte 15,40,251 // movaps %xmm3,%xmm7
18630 .byte 255,224 // jmpq *%rax
18631
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18632HIDDEN _sk_move_dst_src_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18633.globl _sk_move_dst_src_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18634FUNCTION(_sk_move_dst_src_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18635_sk_move_dst_src_sse41:
18636 .byte 72,173 // lods %ds:(%rsi),%rax
18637 .byte 15,40,196 // movaps %xmm4,%xmm0
18638 .byte 15,40,205 // movaps %xmm5,%xmm1
18639 .byte 15,40,214 // movaps %xmm6,%xmm2
18640 .byte 15,40,223 // movaps %xmm7,%xmm3
18641 .byte 255,224 // jmpq *%rax
18642
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18643HIDDEN _sk_premul_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18644.globl _sk_premul_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18645FUNCTION(_sk_premul_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18646_sk_premul_sse41:
18647 .byte 15,89,195 // mulps %xmm3,%xmm0
18648 .byte 15,89,203 // mulps %xmm3,%xmm1
18649 .byte 15,89,211 // mulps %xmm3,%xmm2
18650 .byte 72,173 // lods %ds:(%rsi),%rax
18651 .byte 255,224 // jmpq *%rax
18652
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18653HIDDEN _sk_unpremul_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18654.globl _sk_unpremul_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18655FUNCTION(_sk_unpremul_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18656_sk_unpremul_sse41:
18657 .byte 69,15,87,192 // xorps %xmm8,%xmm8
18658 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18659 .byte 102,68,15,110,200 // movd %eax,%xmm9
18660 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
18661 .byte 68,15,94,203 // divps %xmm3,%xmm9
18662 .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8
18663 .byte 69,15,84,193 // andps %xmm9,%xmm8
18664 .byte 65,15,89,192 // mulps %xmm8,%xmm0
18665 .byte 65,15,89,200 // mulps %xmm8,%xmm1
18666 .byte 65,15,89,208 // mulps %xmm8,%xmm2
18667 .byte 72,173 // lods %ds:(%rsi),%rax
18668 .byte 255,224 // jmpq *%rax
18669
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18670HIDDEN _sk_from_srgb_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18671.globl _sk_from_srgb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18672FUNCTION(_sk_from_srgb_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18673_sk_from_srgb_sse41:
18674 .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax
18675 .byte 102,68,15,110,216 // movd %eax,%xmm11
18676 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
18677 .byte 69,15,40,211 // movaps %xmm11,%xmm10
18678 .byte 68,15,89,208 // mulps %xmm0,%xmm10
18679 .byte 68,15,40,240 // movaps %xmm0,%xmm14
18680 .byte 69,15,89,246 // mulps %xmm14,%xmm14
18681 .byte 184,154,153,153,62 // mov $0x3e99999a,%eax
18682 .byte 102,68,15,110,192 // movd %eax,%xmm8
18683 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18684 .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax
18685 .byte 102,68,15,110,224 // movd %eax,%xmm12
18686 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
18687 .byte 69,15,40,200 // movaps %xmm8,%xmm9
18688 .byte 68,15,89,200 // mulps %xmm0,%xmm9
18689 .byte 69,15,88,204 // addps %xmm12,%xmm9
18690 .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax
18691 .byte 102,68,15,110,232 // movd %eax,%xmm13
18692 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
18693 .byte 69,15,89,206 // mulps %xmm14,%xmm9
18694 .byte 69,15,88,205 // addps %xmm13,%xmm9
18695 .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax
18696 .byte 102,68,15,110,240 // movd %eax,%xmm14
18697 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
18698 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
18699 .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9
18700 .byte 69,15,40,251 // movaps %xmm11,%xmm15
18701 .byte 68,15,89,249 // mulps %xmm1,%xmm15
18702 .byte 15,40,193 // movaps %xmm1,%xmm0
18703 .byte 15,89,192 // mulps %xmm0,%xmm0
18704 .byte 69,15,40,208 // movaps %xmm8,%xmm10
18705 .byte 68,15,89,209 // mulps %xmm1,%xmm10
18706 .byte 69,15,88,212 // addps %xmm12,%xmm10
18707 .byte 68,15,89,208 // mulps %xmm0,%xmm10
18708 .byte 69,15,88,213 // addps %xmm13,%xmm10
18709 .byte 65,15,194,206,1 // cmpltps %xmm14,%xmm1
18710 .byte 15,40,193 // movaps %xmm1,%xmm0
18711 .byte 102,69,15,56,20,215 // blendvps %xmm0,%xmm15,%xmm10
18712 .byte 68,15,89,218 // mulps %xmm2,%xmm11
18713 .byte 15,40,194 // movaps %xmm2,%xmm0
18714 .byte 15,89,192 // mulps %xmm0,%xmm0
18715 .byte 68,15,89,194 // mulps %xmm2,%xmm8
18716 .byte 69,15,88,196 // addps %xmm12,%xmm8
18717 .byte 68,15,89,192 // mulps %xmm0,%xmm8
18718 .byte 69,15,88,197 // addps %xmm13,%xmm8
18719 .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2
18720 .byte 15,40,194 // movaps %xmm2,%xmm0
18721 .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8
18722 .byte 72,173 // lods %ds:(%rsi),%rax
18723 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18724 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18725 .byte 65,15,40,208 // movaps %xmm8,%xmm2
18726 .byte 255,224 // jmpq *%rax
18727
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]18728HIDDEN _sk_to_srgb_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18729.globl _sk_to_srgb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18730FUNCTION(_sk_to_srgb_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18731_sk_to_srgb_sse41:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18732 .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18733 .byte 15,40,254 // movaps %xmm6,%xmm7
18734 .byte 15,40,245 // movaps %xmm5,%xmm6
18735 .byte 15,40,236 // movaps %xmm4,%xmm5
18736 .byte 15,40,227 // movaps %xmm3,%xmm4
18737 .byte 15,40,218 // movaps %xmm2,%xmm3
18738 .byte 15,40,209 // movaps %xmm1,%xmm2
18739 .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8
18740 .byte 69,15,83,200 // rcpps %xmm8,%xmm9
18741 .byte 69,15,82,248 // rsqrtps %xmm8,%xmm15
18742 .byte 184,41,92,71,65 // mov $0x41475c29,%eax
18743 .byte 102,68,15,110,216 // movd %eax,%xmm11
18744 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
18745 .byte 69,15,40,211 // movaps %xmm11,%xmm10
18746 .byte 68,15,89,208 // mulps %xmm0,%xmm10
18747 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18748 .byte 102,68,15,110,192 // movd %eax,%xmm8
18749 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18750 .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax
18751 .byte 102,68,15,110,224 // movd %eax,%xmm12
18752 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
18753 .byte 184,206,111,48,63 // mov $0x3f306fce,%eax
18754 .byte 102,68,15,110,232 // movd %eax,%xmm13
18755 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
18756 .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax
18757 .byte 53,0,0,0,128 // xor $0x80000000,%eax
18758 .byte 102,68,15,110,240 // movd %eax,%xmm14
18759 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
18760 .byte 69,15,89,205 // mulps %xmm13,%xmm9
18761 .byte 69,15,88,206 // addps %xmm14,%xmm9
18762 .byte 69,15,89,252 // mulps %xmm12,%xmm15
18763 .byte 69,15,88,249 // addps %xmm9,%xmm15
18764 .byte 69,15,40,200 // movaps %xmm8,%xmm9
18765 .byte 69,15,93,207 // minps %xmm15,%xmm9
18766 .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax
18767 .byte 102,68,15,110,248 // movd %eax,%xmm15
18768 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15
18769 .byte 65,15,194,199,1 // cmpltps %xmm15,%xmm0
18770 .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9
18771 .byte 68,15,82,210 // rsqrtps %xmm2,%xmm10
18772 .byte 65,15,83,194 // rcpps %xmm10,%xmm0
18773 .byte 69,15,82,210 // rsqrtps %xmm10,%xmm10
18774 .byte 65,15,89,197 // mulps %xmm13,%xmm0
18775 .byte 65,15,88,198 // addps %xmm14,%xmm0
18776 .byte 69,15,89,212 // mulps %xmm12,%xmm10
18777 .byte 68,15,88,208 // addps %xmm0,%xmm10
18778 .byte 65,15,40,200 // movaps %xmm8,%xmm1
18779 .byte 65,15,93,202 // minps %xmm10,%xmm1
18780 .byte 69,15,40,211 // movaps %xmm11,%xmm10
18781 .byte 68,15,89,210 // mulps %xmm2,%xmm10
18782 .byte 65,15,194,215,1 // cmpltps %xmm15,%xmm2
18783 .byte 15,40,194 // movaps %xmm2,%xmm0
18784 .byte 102,65,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm1
18785 .byte 15,82,195 // rsqrtps %xmm3,%xmm0
18786 .byte 15,83,208 // rcpps %xmm0,%xmm2
18787 .byte 65,15,89,213 // mulps %xmm13,%xmm2
18788 .byte 65,15,88,214 // addps %xmm14,%xmm2
18789 .byte 15,82,192 // rsqrtps %xmm0,%xmm0
18790 .byte 65,15,89,196 // mulps %xmm12,%xmm0
18791 .byte 15,88,194 // addps %xmm2,%xmm0
18792 .byte 68,15,93,192 // minps %xmm0,%xmm8
18793 .byte 68,15,89,219 // mulps %xmm3,%xmm11
18794 .byte 65,15,194,223,1 // cmpltps %xmm15,%xmm3
18795 .byte 15,40,195 // movaps %xmm3,%xmm0
18796 .byte 102,69,15,56,20,195 // blendvps %xmm0,%xmm11,%xmm8
18797 .byte 72,173 // lods %ds:(%rsi),%rax
18798 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18799 .byte 65,15,40,208 // movaps %xmm8,%xmm2
18800 .byte 15,40,220 // movaps %xmm4,%xmm3
18801 .byte 15,40,229 // movaps %xmm5,%xmm4
18802 .byte 15,40,238 // movaps %xmm6,%xmm5
18803 .byte 15,40,247 // movaps %xmm7,%xmm6
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]18804 .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]18805 .byte 255,224 // jmpq *%rax
18806
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]18807HIDDEN _sk_rgb_to_hsl_sse41
18808.globl _sk_rgb_to_hsl_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18809FUNCTION(_sk_rgb_to_hsl_sse41)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]18810_sk_rgb_to_hsl_sse41:
18811 .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp)
18812 .byte 15,40,254 // movaps %xmm6,%xmm7
18813 .byte 15,40,245 // movaps %xmm5,%xmm6
18814 .byte 15,40,236 // movaps %xmm4,%xmm5
18815 .byte 15,40,227 // movaps %xmm3,%xmm4
18816 .byte 15,40,218 // movaps %xmm2,%xmm3
18817 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18818 .byte 102,68,15,110,216 // movd %eax,%xmm11
18819 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
18820 .byte 65,185,0,0,192,64 // mov $0x40c00000,%r9d
18821 .byte 184,0,0,0,64 // mov $0x40000000,%eax
18822 .byte 185,0,0,128,64 // mov $0x40800000,%ecx
18823 .byte 102,68,15,110,193 // movd %ecx,%xmm8
18824 .byte 68,15,40,224 // movaps %xmm0,%xmm12
18825 .byte 68,15,95,225 // maxps %xmm1,%xmm12
18826 .byte 68,15,95,227 // maxps %xmm3,%xmm12
18827 .byte 68,15,40,232 // movaps %xmm0,%xmm13
18828 .byte 68,15,93,233 // minps %xmm1,%xmm13
18829 .byte 68,15,93,235 // minps %xmm3,%xmm13
18830 .byte 69,15,40,204 // movaps %xmm12,%xmm9
18831 .byte 68,15,194,200,0 // cmpeqps %xmm0,%xmm9
18832 .byte 68,15,40,241 // movaps %xmm1,%xmm14
18833 .byte 68,15,92,243 // subps %xmm3,%xmm14
18834 .byte 68,15,40,249 // movaps %xmm1,%xmm15
18835 .byte 68,15,194,251,1 // cmpltps %xmm3,%xmm15
18836 .byte 69,15,40,212 // movaps %xmm12,%xmm10
18837 .byte 68,15,194,209,0 // cmpeqps %xmm1,%xmm10
18838 .byte 15,92,216 // subps %xmm0,%xmm3
18839 .byte 15,92,193 // subps %xmm1,%xmm0
18840 .byte 65,15,40,212 // movaps %xmm12,%xmm2
18841 .byte 65,15,92,213 // subps %xmm13,%xmm2
18842 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
18843 .byte 68,15,94,218 // divps %xmm2,%xmm11
18844 .byte 65,15,89,195 // mulps %xmm11,%xmm0
18845 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18846 .byte 68,15,88,192 // addps %xmm0,%xmm8
18847 .byte 102,15,110,200 // movd %eax,%xmm1
18848 .byte 65,15,89,219 // mulps %xmm11,%xmm3
18849 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
18850 .byte 15,88,217 // addps %xmm1,%xmm3
18851 .byte 65,15,40,194 // movaps %xmm10,%xmm0
18852 .byte 102,68,15,56,20,195 // blendvps %xmm0,%xmm3,%xmm8
18853 .byte 69,15,89,243 // mulps %xmm11,%xmm14
18854 .byte 102,65,15,110,217 // movd %r9d,%xmm3
18855 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
18856 .byte 65,15,84,223 // andps %xmm15,%xmm3
18857 .byte 65,15,88,222 // addps %xmm14,%xmm3
18858 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
18859 .byte 102,68,15,110,208 // movd %eax,%xmm10
18860 .byte 65,15,40,193 // movaps %xmm9,%xmm0
18861 .byte 102,68,15,56,20,195 // blendvps %xmm0,%xmm3,%xmm8
18862 .byte 65,15,40,220 // movaps %xmm12,%xmm3
18863 .byte 65,15,92,204 // subps %xmm12,%xmm1
18864 .byte 69,15,88,229 // addps %xmm13,%xmm12
18865 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
18866 .byte 69,15,40,204 // movaps %xmm12,%xmm9
18867 .byte 69,15,89,202 // mulps %xmm10,%xmm9
18868 .byte 69,15,194,209,1 // cmpltps %xmm9,%xmm10
18869 .byte 65,15,92,205 // subps %xmm13,%xmm1
18870 .byte 65,15,40,194 // movaps %xmm10,%xmm0
18871 .byte 102,68,15,56,20,225 // blendvps %xmm0,%xmm1,%xmm12
18872 .byte 65,15,194,221,4 // cmpneqps %xmm13,%xmm3
18873 .byte 102,65,15,110,192 // movd %r8d,%xmm0
18874 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
18875 .byte 68,15,84,195 // andps %xmm3,%xmm8
18876 .byte 68,15,89,192 // mulps %xmm0,%xmm8
18877 .byte 65,15,94,212 // divps %xmm12,%xmm2
18878 .byte 15,84,211 // andps %xmm3,%xmm2
18879 .byte 72,173 // lods %ds:(%rsi),%rax
18880 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18881 .byte 15,40,202 // movaps %xmm2,%xmm1
18882 .byte 65,15,40,209 // movaps %xmm9,%xmm2
18883 .byte 15,40,220 // movaps %xmm4,%xmm3
18884 .byte 15,40,229 // movaps %xmm5,%xmm4
18885 .byte 15,40,238 // movaps %xmm6,%xmm5
18886 .byte 15,40,247 // movaps %xmm7,%xmm6
18887 .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7
18888 .byte 255,224 // jmpq *%rax
18889
18890HIDDEN _sk_hsl_to_rgb_sse41
18891.globl _sk_hsl_to_rgb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]18892FUNCTION(_sk_hsl_to_rgb_sse41)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]18893_sk_hsl_to_rgb_sse41:
18894 .byte 72,131,236,24 // sub $0x18,%rsp
18895 .byte 15,41,60,36 // movaps %xmm7,(%rsp)
18896 .byte 15,41,116,36,240 // movaps %xmm6,-0x10(%rsp)
18897 .byte 15,41,108,36,224 // movaps %xmm5,-0x20(%rsp)
18898 .byte 15,41,100,36,208 // movaps %xmm4,-0x30(%rsp)
18899 .byte 15,41,92,36,192 // movaps %xmm3,-0x40(%rsp)
18900 .byte 68,15,40,208 // movaps %xmm0,%xmm10
18901 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
18902 .byte 102,15,110,216 // movd %eax,%xmm3
18903 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
18904 .byte 15,41,92,36,128 // movaps %xmm3,-0x80(%rsp)
18905 .byte 15,40,194 // movaps %xmm2,%xmm0
18906 .byte 15,194,195,1 // cmpltps %xmm3,%xmm0
18907 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
18908 .byte 102,68,15,110,192 // movd %eax,%xmm8
18909 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
18910 .byte 65,15,40,224 // movaps %xmm8,%xmm4
18911 .byte 15,88,225 // addps %xmm1,%xmm4
18912 .byte 15,89,226 // mulps %xmm2,%xmm4
18913 .byte 15,40,217 // movaps %xmm1,%xmm3
18914 .byte 15,40,249 // movaps %xmm1,%xmm7
18915 .byte 15,88,250 // addps %xmm2,%xmm7
18916 .byte 15,89,218 // mulps %xmm2,%xmm3
18917 .byte 15,40,234 // movaps %xmm2,%xmm5
18918 .byte 15,92,251 // subps %xmm3,%xmm7
18919 .byte 102,15,56,20,252 // blendvps %xmm0,%xmm4,%xmm7
18920 .byte 184,0,0,0,64 // mov $0x40000000,%eax
18921 .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx
18922 .byte 102,15,110,209 // movd %ecx,%xmm2
18923 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
18924 .byte 15,41,84,36,176 // movaps %xmm2,-0x50(%rsp)
18925 .byte 65,15,88,210 // addps %xmm10,%xmm2
18926 .byte 185,0,0,0,0 // mov $0x0,%ecx
18927 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18928 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0
18929 .byte 15,40,218 // movaps %xmm2,%xmm3
18930 .byte 65,15,92,216 // subps %xmm8,%xmm3
18931 .byte 68,15,40,226 // movaps %xmm2,%xmm12
18932 .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12
18933 .byte 102,68,15,110,241 // movd %ecx,%xmm14
18934 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
18935 .byte 15,40,194 // movaps %xmm2,%xmm0
18936 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
18937 .byte 68,15,41,116,36,160 // movaps %xmm14,-0x60(%rsp)
18938 .byte 65,15,40,216 // movaps %xmm8,%xmm3
18939 .byte 15,88,218 // addps %xmm2,%xmm3
18940 .byte 102,68,15,56,20,227 // blendvps %xmm0,%xmm3,%xmm12
18941 .byte 102,68,15,110,200 // movd %eax,%xmm9
18942 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
18943 .byte 15,40,229 // movaps %xmm5,%xmm4
18944 .byte 15,41,100,36,144 // movaps %xmm4,-0x70(%rsp)
18945 .byte 68,15,89,204 // mulps %xmm4,%xmm9
18946 .byte 68,15,92,207 // subps %xmm7,%xmm9
18947 .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax
18948 .byte 15,40,199 // movaps %xmm7,%xmm0
18949 .byte 65,15,92,193 // subps %xmm9,%xmm0
18950 .byte 185,0,0,192,64 // mov $0x40c00000,%ecx
18951 .byte 102,15,110,241 // movd %ecx,%xmm6
18952 .byte 15,198,246,0 // shufps $0x0,%xmm6,%xmm6
18953 .byte 15,89,240 // mulps %xmm0,%xmm6
18954 .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx
18955 .byte 102,15,110,217 // movd %ecx,%xmm3
18956 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
18957 .byte 15,40,235 // movaps %xmm3,%xmm5
18958 .byte 65,15,92,236 // subps %xmm12,%xmm5
18959 .byte 69,15,40,236 // movaps %xmm12,%xmm13
18960 .byte 69,15,40,252 // movaps %xmm12,%xmm15
18961 .byte 68,15,194,227,1 // cmpltps %xmm3,%xmm12
18962 .byte 15,89,238 // mulps %xmm6,%xmm5
18963 .byte 65,15,88,233 // addps %xmm9,%xmm5
18964 .byte 69,15,40,217 // movaps %xmm9,%xmm11
18965 .byte 65,15,40,196 // movaps %xmm12,%xmm0
18966 .byte 102,68,15,56,20,221 // blendvps %xmm0,%xmm5,%xmm11
18967 .byte 68,15,194,124,36,128,1 // cmpltps -0x80(%rsp),%xmm15
18968 .byte 65,15,40,199 // movaps %xmm15,%xmm0
18969 .byte 102,68,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm11
18970 .byte 102,15,110,232 // movd %eax,%xmm5
18971 .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5
18972 .byte 68,15,194,237,1 // cmpltps %xmm5,%xmm13
18973 .byte 15,89,214 // mulps %xmm6,%xmm2
18974 .byte 65,15,88,209 // addps %xmm9,%xmm2
18975 .byte 65,15,40,197 // movaps %xmm13,%xmm0
18976 .byte 102,68,15,56,20,218 // blendvps %xmm0,%xmm2,%xmm11
18977 .byte 69,15,87,228 // xorps %xmm12,%xmm12
18978 .byte 68,15,194,225,0 // cmpeqps %xmm1,%xmm12
18979 .byte 65,15,40,196 // movaps %xmm12,%xmm0
18980 .byte 102,68,15,56,20,220 // blendvps %xmm0,%xmm4,%xmm11
18981 .byte 65,15,40,192 // movaps %xmm8,%xmm0
18982 .byte 65,15,194,194,1 // cmpltps %xmm10,%xmm0
18983 .byte 65,15,40,202 // movaps %xmm10,%xmm1
18984 .byte 65,15,92,200 // subps %xmm8,%xmm1
18985 .byte 69,15,40,234 // movaps %xmm10,%xmm13
18986 .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13
18987 .byte 65,15,40,194 // movaps %xmm10,%xmm0
18988 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
18989 .byte 65,15,40,200 // movaps %xmm8,%xmm1
18990 .byte 65,15,88,202 // addps %xmm10,%xmm1
18991 .byte 102,68,15,56,20,233 // blendvps %xmm0,%xmm1,%xmm13
18992 .byte 68,15,40,243 // movaps %xmm3,%xmm14
18993 .byte 69,15,92,245 // subps %xmm13,%xmm14
18994 .byte 65,15,40,229 // movaps %xmm13,%xmm4
18995 .byte 69,15,40,253 // movaps %xmm13,%xmm15
18996 .byte 68,15,194,235,1 // cmpltps %xmm3,%xmm13
18997 .byte 68,15,89,246 // mulps %xmm6,%xmm14
18998 .byte 69,15,88,241 // addps %xmm9,%xmm14
18999 .byte 65,15,40,201 // movaps %xmm9,%xmm1
19000 .byte 65,15,40,197 // movaps %xmm13,%xmm0
19001 .byte 102,65,15,56,20,206 // blendvps %xmm0,%xmm14,%xmm1
19002 .byte 68,15,40,116,36,128 // movaps -0x80(%rsp),%xmm14
19003 .byte 69,15,194,254,1 // cmpltps %xmm14,%xmm15
19004 .byte 65,15,40,199 // movaps %xmm15,%xmm0
19005 .byte 102,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm1
19006 .byte 15,194,229,1 // cmpltps %xmm5,%xmm4
19007 .byte 15,40,214 // movaps %xmm6,%xmm2
19008 .byte 65,15,89,210 // mulps %xmm10,%xmm2
19009 .byte 65,15,88,209 // addps %xmm9,%xmm2
19010 .byte 15,40,196 // movaps %xmm4,%xmm0
19011 .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1
19012 .byte 65,15,40,196 // movaps %xmm12,%xmm0
19013 .byte 68,15,40,124,36,144 // movaps -0x70(%rsp),%xmm15
19014 .byte 102,65,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm1
19015 .byte 68,15,92,84,36,176 // subps -0x50(%rsp),%xmm10
19016 .byte 65,15,40,210 // movaps %xmm10,%xmm2
19017 .byte 65,15,92,208 // subps %xmm8,%xmm2
19018 .byte 69,15,40,232 // movaps %xmm8,%xmm13
19019 .byte 69,15,194,194,1 // cmpltps %xmm10,%xmm8
19020 .byte 65,15,40,226 // movaps %xmm10,%xmm4
19021 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19022 .byte 102,15,56,20,226 // blendvps %xmm0,%xmm2,%xmm4
19023 .byte 65,15,40,194 // movaps %xmm10,%xmm0
19024 .byte 15,194,68,36,160,1 // cmpltps -0x60(%rsp),%xmm0
19025 .byte 69,15,88,234 // addps %xmm10,%xmm13
19026 .byte 102,65,15,56,20,229 // blendvps %xmm0,%xmm13,%xmm4
19027 .byte 68,15,89,214 // mulps %xmm6,%xmm10
19028 .byte 15,40,196 // movaps %xmm4,%xmm0
19029 .byte 15,194,195,1 // cmpltps %xmm3,%xmm0
19030 .byte 15,92,220 // subps %xmm4,%xmm3
19031 .byte 15,89,222 // mulps %xmm6,%xmm3
19032 .byte 69,15,88,209 // addps %xmm9,%xmm10
19033 .byte 65,15,88,217 // addps %xmm9,%xmm3
19034 .byte 102,68,15,56,20,203 // blendvps %xmm0,%xmm3,%xmm9
19035 .byte 15,40,196 // movaps %xmm4,%xmm0
19036 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
19037 .byte 102,68,15,56,20,207 // blendvps %xmm0,%xmm7,%xmm9
19038 .byte 15,40,196 // movaps %xmm4,%xmm0
19039 .byte 15,194,197,1 // cmpltps %xmm5,%xmm0
19040 .byte 102,69,15,56,20,202 // blendvps %xmm0,%xmm10,%xmm9
19041 .byte 65,15,40,196 // movaps %xmm12,%xmm0
19042 .byte 102,69,15,56,20,207 // blendvps %xmm0,%xmm15,%xmm9
19043 .byte 72,173 // lods %ds:(%rsi),%rax
19044 .byte 65,15,40,195 // movaps %xmm11,%xmm0
19045 .byte 65,15,40,209 // movaps %xmm9,%xmm2
19046 .byte 15,40,92,36,192 // movaps -0x40(%rsp),%xmm3
19047 .byte 15,40,100,36,208 // movaps -0x30(%rsp),%xmm4
19048 .byte 15,40,108,36,224 // movaps -0x20(%rsp),%xmm5
19049 .byte 15,40,116,36,240 // movaps -0x10(%rsp),%xmm6
19050 .byte 15,40,60,36 // movaps (%rsp),%xmm7
19051 .byte 72,131,196,24 // add $0x18,%rsp
19052 .byte 255,224 // jmpq *%rax
19053
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19054HIDDEN _sk_scale_1_float_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19055.globl _sk_scale_1_float_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19056FUNCTION(_sk_scale_1_float_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19057_sk_scale_1_float_sse41:
19058 .byte 72,173 // lods %ds:(%rsi),%rax
19059 .byte 243,68,15,16,0 // movss (%rax),%xmm8
19060 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19061 .byte 65,15,89,192 // mulps %xmm8,%xmm0
19062 .byte 65,15,89,200 // mulps %xmm8,%xmm1
19063 .byte 65,15,89,208 // mulps %xmm8,%xmm2
19064 .byte 65,15,89,216 // mulps %xmm8,%xmm3
19065 .byte 72,173 // lods %ds:(%rsi),%rax
19066 .byte 255,224 // jmpq *%rax
19067
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19068HIDDEN _sk_scale_u8_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19069.globl _sk_scale_u8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19070FUNCTION(_sk_scale_u8_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19071_sk_scale_u8_sse41:
19072 .byte 72,173 // lods %ds:(%rsi),%rax
19073 .byte 72,139,0 // mov (%rax),%rax
19074 .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8
19075 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19076 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
19077 .byte 102,68,15,110,200 // movd %eax,%xmm9
19078 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19079 .byte 69,15,89,200 // mulps %xmm8,%xmm9
19080 .byte 65,15,89,193 // mulps %xmm9,%xmm0
19081 .byte 65,15,89,201 // mulps %xmm9,%xmm1
19082 .byte 65,15,89,209 // mulps %xmm9,%xmm2
19083 .byte 65,15,89,217 // mulps %xmm9,%xmm3
19084 .byte 72,173 // lods %ds:(%rsi),%rax
19085 .byte 255,224 // jmpq *%rax
19086
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19087HIDDEN _sk_lerp_1_float_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19088.globl _sk_lerp_1_float_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19089FUNCTION(_sk_lerp_1_float_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19090_sk_lerp_1_float_sse41:
19091 .byte 72,173 // lods %ds:(%rsi),%rax
19092 .byte 243,68,15,16,0 // movss (%rax),%xmm8
19093 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19094 .byte 15,92,196 // subps %xmm4,%xmm0
19095 .byte 65,15,89,192 // mulps %xmm8,%xmm0
19096 .byte 15,88,196 // addps %xmm4,%xmm0
19097 .byte 15,92,205 // subps %xmm5,%xmm1
19098 .byte 65,15,89,200 // mulps %xmm8,%xmm1
19099 .byte 15,88,205 // addps %xmm5,%xmm1
19100 .byte 15,92,214 // subps %xmm6,%xmm2
19101 .byte 65,15,89,208 // mulps %xmm8,%xmm2
19102 .byte 15,88,214 // addps %xmm6,%xmm2
19103 .byte 15,92,223 // subps %xmm7,%xmm3
19104 .byte 65,15,89,216 // mulps %xmm8,%xmm3
19105 .byte 15,88,223 // addps %xmm7,%xmm3
19106 .byte 72,173 // lods %ds:(%rsi),%rax
19107 .byte 255,224 // jmpq *%rax
19108
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19109HIDDEN _sk_lerp_u8_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19110.globl _sk_lerp_u8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19111FUNCTION(_sk_lerp_u8_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19112_sk_lerp_u8_sse41:
19113 .byte 72,173 // lods %ds:(%rsi),%rax
19114 .byte 72,139,0 // mov (%rax),%rax
19115 .byte 102,68,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm8
19116 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19117 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
19118 .byte 102,68,15,110,200 // movd %eax,%xmm9
19119 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19120 .byte 69,15,89,200 // mulps %xmm8,%xmm9
19121 .byte 15,92,196 // subps %xmm4,%xmm0
19122 .byte 65,15,89,193 // mulps %xmm9,%xmm0
19123 .byte 15,88,196 // addps %xmm4,%xmm0
19124 .byte 15,92,205 // subps %xmm5,%xmm1
19125 .byte 65,15,89,201 // mulps %xmm9,%xmm1
19126 .byte 15,88,205 // addps %xmm5,%xmm1
19127 .byte 15,92,214 // subps %xmm6,%xmm2
19128 .byte 65,15,89,209 // mulps %xmm9,%xmm2
19129 .byte 15,88,214 // addps %xmm6,%xmm2
19130 .byte 15,92,223 // subps %xmm7,%xmm3
19131 .byte 65,15,89,217 // mulps %xmm9,%xmm3
19132 .byte 15,88,223 // addps %xmm7,%xmm3
19133 .byte 72,173 // lods %ds:(%rsi),%rax
19134 .byte 255,224 // jmpq *%rax
19135
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19136HIDDEN _sk_lerp_565_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19137.globl _sk_lerp_565_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19138FUNCTION(_sk_lerp_565_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19139_sk_lerp_565_sse41:
19140 .byte 72,173 // lods %ds:(%rsi),%rax
19141 .byte 72,139,0 // mov (%rax),%rax
19142 .byte 102,68,15,56,51,4,120 // pmovzxwd (%rax,%rdi,2),%xmm8
19143 .byte 184,0,248,0,0 // mov $0xf800,%eax
19144 .byte 102,15,110,216 // movd %eax,%xmm3
19145 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
19146 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
19147 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
19148 .byte 184,8,33,132,55 // mov $0x37842108,%eax
19149 .byte 102,68,15,110,208 // movd %eax,%xmm10
19150 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19151 .byte 69,15,89,209 // mulps %xmm9,%xmm10
19152 .byte 184,224,7,0,0 // mov $0x7e0,%eax
19153 .byte 102,15,110,216 // movd %eax,%xmm3
19154 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
19155 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
19156 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
19157 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
19158 .byte 102,68,15,110,216 // movd %eax,%xmm11
19159 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
19160 .byte 69,15,89,217 // mulps %xmm9,%xmm11
19161 .byte 184,31,0,0,0 // mov $0x1f,%eax
19162 .byte 102,15,110,216 // movd %eax,%xmm3
19163 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
19164 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
19165 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
19166 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
19167 .byte 102,15,110,216 // movd %eax,%xmm3
19168 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19169 .byte 65,15,89,216 // mulps %xmm8,%xmm3
19170 .byte 15,92,196 // subps %xmm4,%xmm0
19171 .byte 65,15,89,194 // mulps %xmm10,%xmm0
19172 .byte 15,88,196 // addps %xmm4,%xmm0
19173 .byte 15,92,205 // subps %xmm5,%xmm1
19174 .byte 65,15,89,203 // mulps %xmm11,%xmm1
19175 .byte 15,88,205 // addps %xmm5,%xmm1
19176 .byte 15,92,214 // subps %xmm6,%xmm2
19177 .byte 15,89,211 // mulps %xmm3,%xmm2
19178 .byte 15,88,214 // addps %xmm6,%xmm2
19179 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19180 .byte 102,15,110,216 // movd %eax,%xmm3
19181 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19182 .byte 72,173 // lods %ds:(%rsi),%rax
19183 .byte 255,224 // jmpq *%rax
19184
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]19185HIDDEN _sk_load_tables_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19186.globl _sk_load_tables_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19187FUNCTION(_sk_load_tables_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]19188_sk_load_tables_sse41:
19189 .byte 72,173 // lods %ds:(%rsi),%rax
19190 .byte 72,139,8 // mov (%rax),%rcx
19191 .byte 76,139,64,8 // mov 0x8(%rax),%r8
19192 .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8
19193 .byte 185,255,0,0,0 // mov $0xff,%ecx
19194 .byte 102,15,110,193 // movd %ecx,%xmm0
19195 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
19196 .byte 102,65,15,111,200 // movdqa %xmm8,%xmm1
19197 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
19198 .byte 102,15,219,200 // pand %xmm0,%xmm1
19199 .byte 102,65,15,111,208 // movdqa %xmm8,%xmm2
19200 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
19201 .byte 102,15,219,208 // pand %xmm0,%xmm2
19202 .byte 102,65,15,219,192 // pand %xmm8,%xmm0
19203 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx
19204 .byte 65,137,201 // mov %ecx,%r9d
19205 .byte 72,193,233,32 // shr $0x20,%rcx
19206 .byte 102,73,15,126,194 // movq %xmm0,%r10
19207 .byte 69,137,211 // mov %r10d,%r11d
19208 .byte 73,193,234,32 // shr $0x20,%r10
19209 .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0
19210 .byte 102,67,15,58,33,4,144,16 // insertps $0x10,(%r8,%r10,4),%xmm0
19211 .byte 102,67,15,58,33,4,136,32 // insertps $0x20,(%r8,%r9,4),%xmm0
19212 .byte 102,65,15,58,33,4,136,48 // insertps $0x30,(%r8,%rcx,4),%xmm0
19213 .byte 76,139,64,16 // mov 0x10(%rax),%r8
19214 .byte 102,73,15,58,22,202,1 // pextrq $0x1,%xmm1,%r10
19215 .byte 77,137,209 // mov %r10,%r9
19216 .byte 73,193,233,32 // shr $0x20,%r9
19217 .byte 102,72,15,126,201 // movq %xmm1,%rcx
19218 .byte 65,137,203 // mov %ecx,%r11d
19219 .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d
19220 .byte 72,193,233,30 // shr $0x1e,%rcx
19221 .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d
19222 .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1
19223 .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1
19224 .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3
19225 .byte 102,15,58,33,203,32 // insertps $0x20,%xmm3,%xmm1
19226 .byte 243,67,15,16,28,136 // movss (%r8,%r9,4),%xmm3
19227 .byte 102,15,58,33,203,48 // insertps $0x30,%xmm3,%xmm1
19228 .byte 76,139,72,24 // mov 0x18(%rax),%r9
19229 .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx
19230 .byte 68,15,183,193 // movzwl %cx,%r8d
19231 .byte 72,193,233,32 // shr $0x20,%rcx
19232 .byte 102,72,15,126,208 // movq %xmm2,%rax
19233 .byte 68,15,183,208 // movzwl %ax,%r10d
19234 .byte 72,193,232,30 // shr $0x1e,%rax
19235 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
19236 .byte 102,65,15,58,33,20,1,16 // insertps $0x10,(%r9,%rax,1),%xmm2
19237 .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3
19238 .byte 102,15,58,33,211,32 // insertps $0x20,%xmm3,%xmm2
19239 .byte 243,65,15,16,28,137 // movss (%r9,%rcx,4),%xmm3
19240 .byte 102,15,58,33,211,48 // insertps $0x30,%xmm3,%xmm2
19241 .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8
19242 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19243 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
19244 .byte 102,15,110,216 // movd %eax,%xmm3
19245 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19246 .byte 65,15,89,216 // mulps %xmm8,%xmm3
19247 .byte 72,173 // lods %ds:(%rsi),%rax
19248 .byte 255,224 // jmpq *%rax
19249
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]19250HIDDEN _sk_load_tables_u16_be_sse41
19251.globl _sk_load_tables_u16_be_sse41
19252FUNCTION(_sk_load_tables_u16_be_sse41)
19253_sk_load_tables_u16_be_sse41:
19254 .byte 72,173 // lods %ds:(%rsi),%rax
19255 .byte 72,139,8 // mov (%rax),%rcx
19256 .byte 76,139,64,8 // mov 0x8(%rax),%r8
19257 .byte 243,15,111,4,249 // movdqu (%rcx,%rdi,8),%xmm0
19258 .byte 243,15,111,76,249,16 // movdqu 0x10(%rcx,%rdi,8),%xmm1
19259 .byte 102,15,111,216 // movdqa %xmm0,%xmm3
19260 .byte 102,15,97,217 // punpcklwd %xmm1,%xmm3
19261 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
19262 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
19263 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
19264 .byte 102,15,105,216 // punpckhwd %xmm0,%xmm3
19265 .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0
19266 .byte 185,255,0,0,0 // mov $0xff,%ecx
19267 .byte 102,15,110,209 // movd %ecx,%xmm2
19268 .byte 102,68,15,112,194,0 // pshufd $0x0,%xmm2,%xmm8
19269 .byte 102,65,15,219,192 // pand %xmm8,%xmm0
19270 .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9
19271 .byte 102,72,15,126,193 // movq %xmm0,%rcx
19272 .byte 65,137,202 // mov %ecx,%r10d
19273 .byte 72,193,233,30 // shr $0x1e,%rcx
19274 .byte 69,137,203 // mov %r9d,%r11d
19275 .byte 73,193,233,30 // shr $0x1e,%r9
19276 .byte 243,67,15,16,4,144 // movss (%r8,%r10,4),%xmm0
19277 .byte 102,65,15,58,33,4,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm0
19278 .byte 102,67,15,58,33,4,152,32 // insertps $0x20,(%r8,%r11,4),%xmm0
19279 .byte 102,67,15,58,33,4,8,48 // insertps $0x30,(%r8,%r9,1),%xmm0
19280 .byte 76,139,64,16 // mov 0x10(%rax),%r8
19281 .byte 102,15,115,217,8 // psrldq $0x8,%xmm1
19282 .byte 102,15,56,51,201 // pmovzxwd %xmm1,%xmm1
19283 .byte 102,65,15,219,200 // pand %xmm8,%xmm1
19284 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9
19285 .byte 102,72,15,126,201 // movq %xmm1,%rcx
19286 .byte 65,137,202 // mov %ecx,%r10d
19287 .byte 72,193,233,30 // shr $0x1e,%rcx
19288 .byte 69,137,203 // mov %r9d,%r11d
19289 .byte 73,193,233,30 // shr $0x1e,%r9
19290 .byte 243,67,15,16,12,144 // movss (%r8,%r10,4),%xmm1
19291 .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1
19292 .byte 243,67,15,16,20,152 // movss (%r8,%r11,4),%xmm2
19293 .byte 102,15,58,33,202,32 // insertps $0x20,%xmm2,%xmm1
19294 .byte 243,67,15,16,20,8 // movss (%r8,%r9,1),%xmm2
19295 .byte 102,15,58,33,202,48 // insertps $0x30,%xmm2,%xmm1
19296 .byte 76,139,64,24 // mov 0x18(%rax),%r8
19297 .byte 102,15,56,51,211 // pmovzxwd %xmm3,%xmm2
19298 .byte 102,65,15,219,208 // pand %xmm8,%xmm2
19299 .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx
19300 .byte 102,72,15,126,208 // movq %xmm2,%rax
19301 .byte 65,137,193 // mov %eax,%r9d
19302 .byte 72,193,232,30 // shr $0x1e,%rax
19303 .byte 65,137,202 // mov %ecx,%r10d
19304 .byte 72,193,233,30 // shr $0x1e,%rcx
19305 .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2
19306 .byte 102,65,15,58,33,20,0,16 // insertps $0x10,(%r8,%rax,1),%xmm2
19307 .byte 102,67,15,58,33,20,144,32 // insertps $0x20,(%r8,%r10,4),%xmm2
19308 .byte 102,65,15,58,33,20,8,48 // insertps $0x30,(%r8,%rcx,1),%xmm2
19309 .byte 184,128,0,128,55 // mov $0x37800080,%eax
19310 .byte 102,68,15,110,192 // movd %eax,%xmm8
19311 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19312 .byte 102,15,112,219,78 // pshufd $0x4e,%xmm3,%xmm3
19313 .byte 102,68,15,111,203 // movdqa %xmm3,%xmm9
19314 .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9
19315 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
19316 .byte 102,65,15,235,217 // por %xmm9,%xmm3
19317 .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3
19318 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
19319 .byte 65,15,89,216 // mulps %xmm8,%xmm3
19320 .byte 72,173 // lods %ds:(%rsi),%rax
19321 .byte 255,224 // jmpq *%rax
19322
19323HIDDEN _sk_load_tables_rgb_u16_be_sse41
19324.globl _sk_load_tables_rgb_u16_be_sse41
19325FUNCTION(_sk_load_tables_rgb_u16_be_sse41)
19326_sk_load_tables_rgb_u16_be_sse41:
19327 .byte 72,173 // lods %ds:(%rsi),%rax
19328 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
19329 .byte 72,139,8 // mov (%rax),%rcx
19330 .byte 76,139,64,8 // mov 0x8(%rax),%r8
19331 .byte 243,66,15,111,20,73 // movdqu (%rcx,%r9,2),%xmm2
19332 .byte 243,66,15,111,68,73,8 // movdqu 0x8(%rcx,%r9,2),%xmm0
19333 .byte 102,15,115,216,4 // psrldq $0x4,%xmm0
19334 .byte 102,15,111,202 // movdqa %xmm2,%xmm1
19335 .byte 102,15,115,217,6 // psrldq $0x6,%xmm1
19336 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2
19337 .byte 102,15,115,216,6 // psrldq $0x6,%xmm0
19338 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
19339 .byte 102,15,111,194 // movdqa %xmm2,%xmm0
19340 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
19341 .byte 102,68,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm8
19342 .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2
19343 .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0
19344 .byte 185,255,0,0,0 // mov $0xff,%ecx
19345 .byte 102,15,110,201 // movd %ecx,%xmm1
19346 .byte 102,68,15,112,201,0 // pshufd $0x0,%xmm1,%xmm9
19347 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
19348 .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9
19349 .byte 102,72,15,126,193 // movq %xmm0,%rcx
19350 .byte 65,137,202 // mov %ecx,%r10d
19351 .byte 72,193,233,30 // shr $0x1e,%rcx
19352 .byte 69,137,203 // mov %r9d,%r11d
19353 .byte 73,193,233,30 // shr $0x1e,%r9
19354 .byte 243,67,15,16,4,144 // movss (%r8,%r10,4),%xmm0
19355 .byte 102,65,15,58,33,4,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm0
19356 .byte 102,67,15,58,33,4,152,32 // insertps $0x20,(%r8,%r11,4),%xmm0
19357 .byte 102,67,15,58,33,4,8,48 // insertps $0x30,(%r8,%r9,1),%xmm0
19358 .byte 76,139,64,16 // mov 0x10(%rax),%r8
19359 .byte 102,65,15,56,51,200 // pmovzxwd %xmm8,%xmm1
19360 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
19361 .byte 102,73,15,58,22,201,1 // pextrq $0x1,%xmm1,%r9
19362 .byte 102,72,15,126,201 // movq %xmm1,%rcx
19363 .byte 65,137,202 // mov %ecx,%r10d
19364 .byte 72,193,233,30 // shr $0x1e,%rcx
19365 .byte 69,137,203 // mov %r9d,%r11d
19366 .byte 73,193,233,30 // shr $0x1e,%r9
19367 .byte 243,67,15,16,12,144 // movss (%r8,%r10,4),%xmm1
19368 .byte 102,65,15,58,33,12,8,16 // insertps $0x10,(%r8,%rcx,1),%xmm1
19369 .byte 243,67,15,16,28,152 // movss (%r8,%r11,4),%xmm3
19370 .byte 102,15,58,33,203,32 // insertps $0x20,%xmm3,%xmm1
19371 .byte 243,67,15,16,28,8 // movss (%r8,%r9,1),%xmm3
19372 .byte 102,15,58,33,203,48 // insertps $0x30,%xmm3,%xmm1
19373 .byte 76,139,64,24 // mov 0x18(%rax),%r8
19374 .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2
19375 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
19376 .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx
19377 .byte 102,72,15,126,208 // movq %xmm2,%rax
19378 .byte 65,137,193 // mov %eax,%r9d
19379 .byte 72,193,232,30 // shr $0x1e,%rax
19380 .byte 65,137,202 // mov %ecx,%r10d
19381 .byte 72,193,233,30 // shr $0x1e,%rcx
19382 .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2
19383 .byte 102,65,15,58,33,20,0,16 // insertps $0x10,(%r8,%rax,1),%xmm2
19384 .byte 102,67,15,58,33,20,144,32 // insertps $0x20,(%r8,%r10,4),%xmm2
19385 .byte 102,65,15,58,33,20,8,48 // insertps $0x30,(%r8,%rcx,1),%xmm2
19386 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19387 .byte 102,15,110,216 // movd %eax,%xmm3
19388 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19389 .byte 72,173 // lods %ds:(%rsi),%rax
19390 .byte 255,224 // jmpq *%rax
19391
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]19392HIDDEN _sk_byte_tables_sse41
19393.globl _sk_byte_tables_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19394FUNCTION(_sk_byte_tables_sse41)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]19395_sk_byte_tables_sse41:
19396 .byte 65,86 // push %r14
19397 .byte 83 // push %rbx
19398 .byte 72,173 // lods %ds:(%rsi),%rax
19399 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
19400 .byte 102,68,15,110,193 // movd %ecx,%xmm8
19401 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19402 .byte 65,15,89,192 // mulps %xmm8,%xmm0
19403 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0
19404 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx
19405 .byte 65,137,200 // mov %ecx,%r8d
19406 .byte 72,193,233,32 // shr $0x20,%rcx
19407 .byte 102,73,15,126,193 // movq %xmm0,%r9
19408 .byte 69,137,202 // mov %r9d,%r10d
19409 .byte 77,137,203 // mov %r9,%r11
19410 .byte 73,193,235,32 // shr $0x20,%r11
19411 .byte 76,139,48 // mov (%rax),%r14
19412 .byte 76,139,72,8 // mov 0x8(%rax),%r9
19413 .byte 102,67,15,58,32,4,22,0 // pinsrb $0x0,(%r14,%r10,1),%xmm0
19414 .byte 102,67,15,58,32,4,30,1 // pinsrb $0x1,(%r14,%r11,1),%xmm0
19415 .byte 67,15,182,28,6 // movzbl (%r14,%r8,1),%ebx
19416 .byte 102,15,58,32,195,2 // pinsrb $0x2,%ebx,%xmm0
19417 .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx
19418 .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0
19419 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0
19420 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
19421 .byte 185,129,128,128,59 // mov $0x3b808081,%ecx
19422 .byte 102,68,15,110,201 // movd %ecx,%xmm9
19423 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19424 .byte 65,15,89,193 // mulps %xmm9,%xmm0
19425 .byte 65,15,89,200 // mulps %xmm8,%xmm1
19426 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
19427 .byte 102,72,15,58,22,201,1 // pextrq $0x1,%xmm1,%rcx
19428 .byte 65,137,200 // mov %ecx,%r8d
19429 .byte 72,193,233,32 // shr $0x20,%rcx
19430 .byte 102,72,15,126,203 // movq %xmm1,%rbx
19431 .byte 65,137,218 // mov %ebx,%r10d
19432 .byte 72,193,235,32 // shr $0x20,%rbx
19433 .byte 102,67,15,58,32,12,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm1
19434 .byte 102,65,15,58,32,12,25,1 // pinsrb $0x1,(%r9,%rbx,1),%xmm1
19435 .byte 67,15,182,28,1 // movzbl (%r9,%r8,1),%ebx
19436 .byte 102,15,58,32,203,2 // pinsrb $0x2,%ebx,%xmm1
19437 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
19438 .byte 102,15,58,32,201,3 // pinsrb $0x3,%ecx,%xmm1
19439 .byte 102,15,56,49,201 // pmovzxbd %xmm1,%xmm1
19440 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
19441 .byte 65,15,89,201 // mulps %xmm9,%xmm1
19442 .byte 76,139,72,16 // mov 0x10(%rax),%r9
19443 .byte 65,15,89,208 // mulps %xmm8,%xmm2
19444 .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2
19445 .byte 102,72,15,58,22,211,1 // pextrq $0x1,%xmm2,%rbx
19446 .byte 65,137,216 // mov %ebx,%r8d
19447 .byte 72,193,235,32 // shr $0x20,%rbx
19448 .byte 102,72,15,126,209 // movq %xmm2,%rcx
19449 .byte 65,137,202 // mov %ecx,%r10d
19450 .byte 72,193,233,32 // shr $0x20,%rcx
19451 .byte 102,67,15,58,32,20,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm2
19452 .byte 102,65,15,58,32,20,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm2
19453 .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx
19454 .byte 102,15,58,32,209,2 // pinsrb $0x2,%ecx,%xmm2
19455 .byte 65,15,182,12,25 // movzbl (%r9,%rbx,1),%ecx
19456 .byte 102,15,58,32,209,3 // pinsrb $0x3,%ecx,%xmm2
19457 .byte 102,15,56,49,210 // pmovzxbd %xmm2,%xmm2
19458 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
19459 .byte 65,15,89,209 // mulps %xmm9,%xmm2
19460 .byte 72,139,64,24 // mov 0x18(%rax),%rax
19461 .byte 68,15,89,195 // mulps %xmm3,%xmm8
19462 .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3
19463 .byte 102,72,15,58,22,217,1 // pextrq $0x1,%xmm3,%rcx
19464 .byte 65,137,200 // mov %ecx,%r8d
19465 .byte 72,193,233,32 // shr $0x20,%rcx
19466 .byte 102,72,15,126,219 // movq %xmm3,%rbx
19467 .byte 65,137,217 // mov %ebx,%r9d
19468 .byte 72,193,235,32 // shr $0x20,%rbx
19469 .byte 102,66,15,58,32,28,8,0 // pinsrb $0x0,(%rax,%r9,1),%xmm3
19470 .byte 102,15,58,32,28,24,1 // pinsrb $0x1,(%rax,%rbx,1),%xmm3
19471 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
19472 .byte 102,15,58,32,219,2 // pinsrb $0x2,%ebx,%xmm3
19473 .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax
19474 .byte 102,15,58,32,216,3 // pinsrb $0x3,%eax,%xmm3
19475 .byte 102,15,56,49,219 // pmovzxbd %xmm3,%xmm3
19476 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
19477 .byte 65,15,89,217 // mulps %xmm9,%xmm3
19478 .byte 72,173 // lods %ds:(%rsi),%rax
19479 .byte 91 // pop %rbx
19480 .byte 65,94 // pop %r14
19481 .byte 255,224 // jmpq *%rax
19482
19483HIDDEN _sk_byte_tables_rgb_sse41
19484.globl _sk_byte_tables_rgb_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]19485FUNCTION(_sk_byte_tables_rgb_sse41)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]19486_sk_byte_tables_rgb_sse41:
19487 .byte 65,86 // push %r14
19488 .byte 83 // push %rbx
19489 .byte 72,173 // lods %ds:(%rsi),%rax
19490 .byte 139,72,24 // mov 0x18(%rax),%ecx
19491 .byte 255,201 // dec %ecx
19492 .byte 102,68,15,110,193 // movd %ecx,%xmm8
19493 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
19494 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19495 .byte 65,15,89,192 // mulps %xmm8,%xmm0
19496 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0
19497 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx
19498 .byte 65,137,200 // mov %ecx,%r8d
19499 .byte 72,193,233,32 // shr $0x20,%rcx
19500 .byte 102,73,15,126,193 // movq %xmm0,%r9
19501 .byte 69,137,202 // mov %r9d,%r10d
19502 .byte 77,137,203 // mov %r9,%r11
19503 .byte 73,193,235,32 // shr $0x20,%r11
19504 .byte 76,139,48 // mov (%rax),%r14
19505 .byte 76,139,72,8 // mov 0x8(%rax),%r9
19506 .byte 102,67,15,58,32,4,22,0 // pinsrb $0x0,(%r14,%r10,1),%xmm0
19507 .byte 102,67,15,58,32,4,30,1 // pinsrb $0x1,(%r14,%r11,1),%xmm0
19508 .byte 67,15,182,28,6 // movzbl (%r14,%r8,1),%ebx
19509 .byte 102,15,58,32,195,2 // pinsrb $0x2,%ebx,%xmm0
19510 .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx
19511 .byte 102,15,58,32,193,3 // pinsrb $0x3,%ecx,%xmm0
19512 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0
19513 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
19514 .byte 185,129,128,128,59 // mov $0x3b808081,%ecx
19515 .byte 102,68,15,110,201 // movd %ecx,%xmm9
19516 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19517 .byte 65,15,89,193 // mulps %xmm9,%xmm0
19518 .byte 65,15,89,200 // mulps %xmm8,%xmm1
19519 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
19520 .byte 102,72,15,58,22,201,1 // pextrq $0x1,%xmm1,%rcx
19521 .byte 65,137,200 // mov %ecx,%r8d
19522 .byte 72,193,233,32 // shr $0x20,%rcx
19523 .byte 102,72,15,126,203 // movq %xmm1,%rbx
19524 .byte 65,137,218 // mov %ebx,%r10d
19525 .byte 72,193,235,32 // shr $0x20,%rbx
19526 .byte 102,67,15,58,32,12,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm1
19527 .byte 102,65,15,58,32,12,25,1 // pinsrb $0x1,(%r9,%rbx,1),%xmm1
19528 .byte 67,15,182,28,1 // movzbl (%r9,%r8,1),%ebx
19529 .byte 102,15,58,32,203,2 // pinsrb $0x2,%ebx,%xmm1
19530 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
19531 .byte 102,15,58,32,201,3 // pinsrb $0x3,%ecx,%xmm1
19532 .byte 102,15,56,49,201 // pmovzxbd %xmm1,%xmm1
19533 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
19534 .byte 65,15,89,201 // mulps %xmm9,%xmm1
19535 .byte 72,139,64,16 // mov 0x10(%rax),%rax
19536 .byte 65,15,89,208 // mulps %xmm8,%xmm2
19537 .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2
19538 .byte 102,72,15,58,22,209,1 // pextrq $0x1,%xmm2,%rcx
19539 .byte 65,137,200 // mov %ecx,%r8d
19540 .byte 72,193,233,32 // shr $0x20,%rcx
19541 .byte 102,72,15,126,211 // movq %xmm2,%rbx
19542 .byte 65,137,217 // mov %ebx,%r9d
19543 .byte 72,193,235,32 // shr $0x20,%rbx
19544 .byte 102,66,15,58,32,20,8,0 // pinsrb $0x0,(%rax,%r9,1),%xmm2
19545 .byte 102,15,58,32,20,24,1 // pinsrb $0x1,(%rax,%rbx,1),%xmm2
19546 .byte 66,15,182,28,0 // movzbl (%rax,%r8,1),%ebx
19547 .byte 102,15,58,32,211,2 // pinsrb $0x2,%ebx,%xmm2
19548 .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax
19549 .byte 102,15,58,32,208,3 // pinsrb $0x3,%eax,%xmm2
19550 .byte 102,15,56,49,210 // pmovzxbd %xmm2,%xmm2
19551 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
19552 .byte 65,15,89,209 // mulps %xmm9,%xmm2
19553 .byte 72,173 // lods %ds:(%rsi),%rax
19554 .byte 91 // pop %rbx
19555 .byte 65,94 // pop %r14
19556 .byte 255,224 // jmpq *%rax
19557
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]19558HIDDEN _sk_table_r_sse41
19559.globl _sk_table_r_sse41
19560FUNCTION(_sk_table_r_sse41)
19561_sk_table_r_sse41:
19562 .byte 72,173 // lods %ds:(%rsi),%rax
19563 .byte 76,139,8 // mov (%rax),%r9
19564 .byte 139,64,8 // mov 0x8(%rax),%eax
19565 .byte 255,200 // dec %eax
19566 .byte 102,68,15,110,192 // movd %eax,%xmm8
19567 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
19568 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19569 .byte 68,15,89,192 // mulps %xmm0,%xmm8
19570 .byte 102,65,15,91,192 // cvtps2dq %xmm8,%xmm0
19571 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
19572 .byte 65,137,192 // mov %eax,%r8d
19573 .byte 72,193,232,32 // shr $0x20,%rax
19574 .byte 102,72,15,126,193 // movq %xmm0,%rcx
19575 .byte 65,137,202 // mov %ecx,%r10d
19576 .byte 72,193,233,32 // shr $0x20,%rcx
19577 .byte 243,67,15,16,4,145 // movss (%r9,%r10,4),%xmm0
19578 .byte 102,65,15,58,33,4,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm0
19579 .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8
19580 .byte 102,65,15,58,33,192,32 // insertps $0x20,%xmm8,%xmm0
19581 .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8
19582 .byte 102,65,15,58,33,192,48 // insertps $0x30,%xmm8,%xmm0
19583 .byte 72,173 // lods %ds:(%rsi),%rax
19584 .byte 255,224 // jmpq *%rax
19585
19586HIDDEN _sk_table_g_sse41
19587.globl _sk_table_g_sse41
19588FUNCTION(_sk_table_g_sse41)
19589_sk_table_g_sse41:
19590 .byte 72,173 // lods %ds:(%rsi),%rax
19591 .byte 76,139,8 // mov (%rax),%r9
19592 .byte 139,64,8 // mov 0x8(%rax),%eax
19593 .byte 255,200 // dec %eax
19594 .byte 102,68,15,110,192 // movd %eax,%xmm8
19595 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
19596 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19597 .byte 68,15,89,193 // mulps %xmm1,%xmm8
19598 .byte 102,65,15,91,200 // cvtps2dq %xmm8,%xmm1
19599 .byte 102,72,15,58,22,200,1 // pextrq $0x1,%xmm1,%rax
19600 .byte 65,137,192 // mov %eax,%r8d
19601 .byte 72,193,232,32 // shr $0x20,%rax
19602 .byte 102,72,15,126,201 // movq %xmm1,%rcx
19603 .byte 65,137,202 // mov %ecx,%r10d
19604 .byte 72,193,233,32 // shr $0x20,%rcx
19605 .byte 243,67,15,16,12,145 // movss (%r9,%r10,4),%xmm1
19606 .byte 102,65,15,58,33,12,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm1
19607 .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8
19608 .byte 102,65,15,58,33,200,32 // insertps $0x20,%xmm8,%xmm1
19609 .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8
19610 .byte 102,65,15,58,33,200,48 // insertps $0x30,%xmm8,%xmm1
19611 .byte 72,173 // lods %ds:(%rsi),%rax
19612 .byte 255,224 // jmpq *%rax
19613
19614HIDDEN _sk_table_b_sse41
19615.globl _sk_table_b_sse41
19616FUNCTION(_sk_table_b_sse41)
19617_sk_table_b_sse41:
19618 .byte 72,173 // lods %ds:(%rsi),%rax
19619 .byte 76,139,8 // mov (%rax),%r9
19620 .byte 139,64,8 // mov 0x8(%rax),%eax
19621 .byte 255,200 // dec %eax
19622 .byte 102,68,15,110,192 // movd %eax,%xmm8
19623 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
19624 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19625 .byte 68,15,89,194 // mulps %xmm2,%xmm8
19626 .byte 102,65,15,91,208 // cvtps2dq %xmm8,%xmm2
19627 .byte 102,72,15,58,22,208,1 // pextrq $0x1,%xmm2,%rax
19628 .byte 65,137,192 // mov %eax,%r8d
19629 .byte 72,193,232,32 // shr $0x20,%rax
19630 .byte 102,72,15,126,209 // movq %xmm2,%rcx
19631 .byte 65,137,202 // mov %ecx,%r10d
19632 .byte 72,193,233,32 // shr $0x20,%rcx
19633 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
19634 .byte 102,65,15,58,33,20,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm2
19635 .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8
19636 .byte 102,65,15,58,33,208,32 // insertps $0x20,%xmm8,%xmm2
19637 .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8
19638 .byte 102,65,15,58,33,208,48 // insertps $0x30,%xmm8,%xmm2
19639 .byte 72,173 // lods %ds:(%rsi),%rax
19640 .byte 255,224 // jmpq *%rax
19641
19642HIDDEN _sk_table_a_sse41
19643.globl _sk_table_a_sse41
19644FUNCTION(_sk_table_a_sse41)
19645_sk_table_a_sse41:
19646 .byte 72,173 // lods %ds:(%rsi),%rax
19647 .byte 76,139,8 // mov (%rax),%r9
19648 .byte 139,64,8 // mov 0x8(%rax),%eax
19649 .byte 255,200 // dec %eax
19650 .byte 102,68,15,110,192 // movd %eax,%xmm8
19651 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
19652 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
19653 .byte 68,15,89,195 // mulps %xmm3,%xmm8
19654 .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3
19655 .byte 102,72,15,58,22,216,1 // pextrq $0x1,%xmm3,%rax
19656 .byte 65,137,192 // mov %eax,%r8d
19657 .byte 72,193,232,32 // shr $0x20,%rax
19658 .byte 102,72,15,126,217 // movq %xmm3,%rcx
19659 .byte 65,137,202 // mov %ecx,%r10d
19660 .byte 72,193,233,32 // shr $0x20,%rcx
19661 .byte 243,67,15,16,28,145 // movss (%r9,%r10,4),%xmm3
19662 .byte 102,65,15,58,33,28,137,16 // insertps $0x10,(%r9,%rcx,4),%xmm3
19663 .byte 243,71,15,16,4,129 // movss (%r9,%r8,4),%xmm8
19664 .byte 102,65,15,58,33,216,32 // insertps $0x20,%xmm8,%xmm3
19665 .byte 243,69,15,16,4,129 // movss (%r9,%rax,4),%xmm8
19666 .byte 102,65,15,58,33,216,48 // insertps $0x30,%xmm8,%xmm3
19667 .byte 72,173 // lods %ds:(%rsi),%rax
19668 .byte 255,224 // jmpq *%rax
19669
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19670HIDDEN _sk_parametric_r_sse41
19671.globl _sk_parametric_r_sse41
19672FUNCTION(_sk_parametric_r_sse41)
19673_sk_parametric_r_sse41:
19674 .byte 72,173 // lods %ds:(%rsi),%rax
19675 .byte 243,68,15,16,64,16 // movss 0x10(%rax),%xmm8
19676 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19677 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
19678 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19679 .byte 68,15,89,200 // mulps %xmm0,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19680 .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10
19681 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19682 .byte 68,15,89,208 // mulps %xmm0,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19683 .byte 65,15,194,192,2 // cmpleps %xmm8,%xmm0
19684 .byte 243,68,15,16,64,24 // movss 0x18(%rax),%xmm8
19685 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19686 .byte 69,15,88,200 // addps %xmm8,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19687 .byte 243,68,15,16,24 // movss (%rax),%xmm11
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19688 .byte 243,68,15,16,64,8 // movss 0x8(%rax),%xmm8
19689 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19690 .byte 69,15,88,208 // addps %xmm8,%xmm10
19691 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
19692 .byte 69,15,91,194 // cvtdq2ps %xmm10,%xmm8
19693 .byte 68,15,89,5,149,33,0,0 // mulps 0x2195(%rip),%xmm8 # 3f60 <_sk_callback_sse41+0x100>
19694 .byte 68,15,84,21,157,33,0,0 // andps 0x219d(%rip),%xmm10 # 3f70 <_sk_callback_sse41+0x110>
19695 .byte 68,15,86,21,165,33,0,0 // orps 0x21a5(%rip),%xmm10 # 3f80 <_sk_callback_sse41+0x120>
19696 .byte 68,15,88,5,173,33,0,0 // addps 0x21ad(%rip),%xmm8 # 3f90 <_sk_callback_sse41+0x130>
19697 .byte 68,15,40,37,181,33,0,0 // movaps 0x21b5(%rip),%xmm12 # 3fa0 <_sk_callback_sse41+0x140>
19698 .byte 69,15,89,226 // mulps %xmm10,%xmm12
19699 .byte 69,15,92,196 // subps %xmm12,%xmm8
19700 .byte 68,15,88,21,181,33,0,0 // addps 0x21b5(%rip),%xmm10 # 3fb0 <_sk_callback_sse41+0x150>
19701 .byte 68,15,40,37,189,33,0,0 // movaps 0x21bd(%rip),%xmm12 # 3fc0 <_sk_callback_sse41+0x160>
19702 .byte 69,15,94,226 // divps %xmm10,%xmm12
19703 .byte 69,15,92,196 // subps %xmm12,%xmm8
19704 .byte 69,15,89,195 // mulps %xmm11,%xmm8
19705 .byte 102,69,15,58,8,208,1 // roundps $0x1,%xmm8,%xmm10
19706 .byte 69,15,40,216 // movaps %xmm8,%xmm11
19707 .byte 69,15,92,218 // subps %xmm10,%xmm11
19708 .byte 68,15,88,5,170,33,0,0 // addps 0x21aa(%rip),%xmm8 # 3fd0 <_sk_callback_sse41+0x170>
19709 .byte 68,15,40,21,178,33,0,0 // movaps 0x21b2(%rip),%xmm10 # 3fe0 <_sk_callback_sse41+0x180>
19710 .byte 69,15,89,211 // mulps %xmm11,%xmm10
19711 .byte 69,15,92,194 // subps %xmm10,%xmm8
19712 .byte 68,15,40,21,178,33,0,0 // movaps 0x21b2(%rip),%xmm10 # 3ff0 <_sk_callback_sse41+0x190>
19713 .byte 69,15,92,211 // subps %xmm11,%xmm10
19714 .byte 68,15,40,29,182,33,0,0 // movaps 0x21b6(%rip),%xmm11 # 4000 <_sk_callback_sse41+0x1a0>
19715 .byte 69,15,94,218 // divps %xmm10,%xmm11
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19716 .byte 69,15,88,216 // addps %xmm8,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19717 .byte 68,15,89,29,182,33,0,0 // mulps 0x21b6(%rip),%xmm11 # 4010 <_sk_callback_sse41+0x1b0>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19718 .byte 102,69,15,91,211 // cvtps2dq %xmm11,%xmm10
19719 .byte 243,68,15,16,64,20 // movss 0x14(%rax),%xmm8
19720 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19721 .byte 69,15,88,194 // addps %xmm10,%xmm8
19722 .byte 102,69,15,56,20,193 // blendvps %xmm0,%xmm9,%xmm8
19723 .byte 15,87,192 // xorps %xmm0,%xmm0
19724 .byte 68,15,95,192 // maxps %xmm0,%xmm8
19725 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19726 .byte 102,15,110,192 // movd %eax,%xmm0
19727 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
19728 .byte 68,15,93,192 // minps %xmm0,%xmm8
19729 .byte 72,173 // lods %ds:(%rsi),%rax
19730 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19731 .byte 255,224 // jmpq *%rax
19732
19733HIDDEN _sk_parametric_g_sse41
19734.globl _sk_parametric_g_sse41
19735FUNCTION(_sk_parametric_g_sse41)
19736_sk_parametric_g_sse41:
19737 .byte 68,15,40,192 // movaps %xmm0,%xmm8
19738 .byte 72,173 // lods %ds:(%rsi),%rax
19739 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
19740 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19741 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
19742 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19743 .byte 68,15,89,201 // mulps %xmm1,%xmm9
19744 .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11
19745 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
19746 .byte 68,15,89,217 // mulps %xmm1,%xmm11
19747 .byte 15,40,193 // movaps %xmm1,%xmm0
19748 .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0
19749 .byte 243,15,16,72,24 // movss 0x18(%rax),%xmm1
19750 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
19751 .byte 68,15,88,201 // addps %xmm1,%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19752 .byte 243,68,15,16,16 // movss (%rax),%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19753 .byte 243,15,16,72,8 // movss 0x8(%rax),%xmm1
19754 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
19755 .byte 68,15,88,217 // addps %xmm1,%xmm11
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19756 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19757 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19758 .byte 68,15,89,37,37,33,0,0 // mulps 0x2125(%rip),%xmm12 # 4020 <_sk_callback_sse41+0x1c0>
19759 .byte 68,15,84,29,45,33,0,0 // andps 0x212d(%rip),%xmm11 # 4030 <_sk_callback_sse41+0x1d0>
19760 .byte 68,15,86,29,53,33,0,0 // orps 0x2135(%rip),%xmm11 # 4040 <_sk_callback_sse41+0x1e0>
19761 .byte 68,15,88,37,61,33,0,0 // addps 0x213d(%rip),%xmm12 # 4050 <_sk_callback_sse41+0x1f0>
19762 .byte 15,40,13,70,33,0,0 // movaps 0x2146(%rip),%xmm1 # 4060 <_sk_callback_sse41+0x200>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19763 .byte 65,15,89,203 // mulps %xmm11,%xmm1
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19764 .byte 68,15,92,225 // subps %xmm1,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19765 .byte 68,15,88,29,70,33,0,0 // addps 0x2146(%rip),%xmm11 # 4070 <_sk_callback_sse41+0x210>
19766 .byte 15,40,13,79,33,0,0 // movaps 0x214f(%rip),%xmm1 # 4080 <_sk_callback_sse41+0x220>
19767 .byte 65,15,94,203 // divps %xmm11,%xmm1
19768 .byte 68,15,92,225 // subps %xmm1,%xmm12
19769 .byte 69,15,89,226 // mulps %xmm10,%xmm12
19770 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10
19771 .byte 69,15,40,220 // movaps %xmm12,%xmm11
19772 .byte 69,15,92,218 // subps %xmm10,%xmm11
19773 .byte 68,15,88,37,60,33,0,0 // addps 0x213c(%rip),%xmm12 # 4090 <_sk_callback_sse41+0x230>
19774 .byte 15,40,13,69,33,0,0 // movaps 0x2145(%rip),%xmm1 # 40a0 <_sk_callback_sse41+0x240>
19775 .byte 65,15,89,203 // mulps %xmm11,%xmm1
19776 .byte 68,15,92,225 // subps %xmm1,%xmm12
19777 .byte 68,15,40,21,69,33,0,0 // movaps 0x2145(%rip),%xmm10 # 40b0 <_sk_callback_sse41+0x250>
19778 .byte 69,15,92,211 // subps %xmm11,%xmm10
19779 .byte 15,40,13,74,33,0,0 // movaps 0x214a(%rip),%xmm1 # 40c0 <_sk_callback_sse41+0x260>
19780 .byte 65,15,94,202 // divps %xmm10,%xmm1
19781 .byte 65,15,88,204 // addps %xmm12,%xmm1
19782 .byte 15,89,13,75,33,0,0 // mulps 0x214b(%rip),%xmm1 # 40d0 <_sk_callback_sse41+0x270>
19783 .byte 102,68,15,91,209 // cvtps2dq %xmm1,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19784 .byte 243,15,16,72,20 // movss 0x14(%rax),%xmm1
19785 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
19786 .byte 65,15,88,202 // addps %xmm10,%xmm1
19787 .byte 102,65,15,56,20,201 // blendvps %xmm0,%xmm9,%xmm1
19788 .byte 15,87,192 // xorps %xmm0,%xmm0
19789 .byte 15,95,200 // maxps %xmm0,%xmm1
19790 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19791 .byte 102,15,110,192 // movd %eax,%xmm0
19792 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
19793 .byte 15,93,200 // minps %xmm0,%xmm1
19794 .byte 72,173 // lods %ds:(%rsi),%rax
19795 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19796 .byte 255,224 // jmpq *%rax
19797
19798HIDDEN _sk_parametric_b_sse41
19799.globl _sk_parametric_b_sse41
19800FUNCTION(_sk_parametric_b_sse41)
19801_sk_parametric_b_sse41:
19802 .byte 68,15,40,192 // movaps %xmm0,%xmm8
19803 .byte 72,173 // lods %ds:(%rsi),%rax
19804 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
19805 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19806 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
19807 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19808 .byte 68,15,89,202 // mulps %xmm2,%xmm9
19809 .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11
19810 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
19811 .byte 68,15,89,218 // mulps %xmm2,%xmm11
19812 .byte 15,40,194 // movaps %xmm2,%xmm0
19813 .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0
19814 .byte 243,15,16,80,24 // movss 0x18(%rax),%xmm2
19815 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
19816 .byte 68,15,88,202 // addps %xmm2,%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19817 .byte 243,68,15,16,16 // movss (%rax),%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19818 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
19819 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
19820 .byte 68,15,88,218 // addps %xmm2,%xmm11
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19821 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19822 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19823 .byte 68,15,89,37,190,32,0,0 // mulps 0x20be(%rip),%xmm12 # 40e0 <_sk_callback_sse41+0x280>
19824 .byte 68,15,84,29,198,32,0,0 // andps 0x20c6(%rip),%xmm11 # 40f0 <_sk_callback_sse41+0x290>
19825 .byte 68,15,86,29,206,32,0,0 // orps 0x20ce(%rip),%xmm11 # 4100 <_sk_callback_sse41+0x2a0>
19826 .byte 68,15,88,37,214,32,0,0 // addps 0x20d6(%rip),%xmm12 # 4110 <_sk_callback_sse41+0x2b0>
19827 .byte 15,40,21,223,32,0,0 // movaps 0x20df(%rip),%xmm2 # 4120 <_sk_callback_sse41+0x2c0>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19828 .byte 65,15,89,211 // mulps %xmm11,%xmm2
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19829 .byte 68,15,92,226 // subps %xmm2,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19830 .byte 68,15,88,29,223,32,0,0 // addps 0x20df(%rip),%xmm11 # 4130 <_sk_callback_sse41+0x2d0>
19831 .byte 15,40,21,232,32,0,0 // movaps 0x20e8(%rip),%xmm2 # 4140 <_sk_callback_sse41+0x2e0>
19832 .byte 65,15,94,211 // divps %xmm11,%xmm2
19833 .byte 68,15,92,226 // subps %xmm2,%xmm12
19834 .byte 69,15,89,226 // mulps %xmm10,%xmm12
19835 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10
19836 .byte 69,15,40,220 // movaps %xmm12,%xmm11
19837 .byte 69,15,92,218 // subps %xmm10,%xmm11
19838 .byte 68,15,88,37,213,32,0,0 // addps 0x20d5(%rip),%xmm12 # 4150 <_sk_callback_sse41+0x2f0>
19839 .byte 15,40,21,222,32,0,0 // movaps 0x20de(%rip),%xmm2 # 4160 <_sk_callback_sse41+0x300>
19840 .byte 65,15,89,211 // mulps %xmm11,%xmm2
19841 .byte 68,15,92,226 // subps %xmm2,%xmm12
19842 .byte 68,15,40,21,222,32,0,0 // movaps 0x20de(%rip),%xmm10 # 4170 <_sk_callback_sse41+0x310>
19843 .byte 69,15,92,211 // subps %xmm11,%xmm10
19844 .byte 15,40,21,227,32,0,0 // movaps 0x20e3(%rip),%xmm2 # 4180 <_sk_callback_sse41+0x320>
19845 .byte 65,15,94,210 // divps %xmm10,%xmm2
19846 .byte 65,15,88,212 // addps %xmm12,%xmm2
19847 .byte 15,89,21,228,32,0,0 // mulps 0x20e4(%rip),%xmm2 # 4190 <_sk_callback_sse41+0x330>
19848 .byte 102,68,15,91,210 // cvtps2dq %xmm2,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19849 .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2
19850 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
19851 .byte 65,15,88,210 // addps %xmm10,%xmm2
19852 .byte 102,65,15,56,20,209 // blendvps %xmm0,%xmm9,%xmm2
19853 .byte 15,87,192 // xorps %xmm0,%xmm0
19854 .byte 15,95,208 // maxps %xmm0,%xmm2
19855 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19856 .byte 102,15,110,192 // movd %eax,%xmm0
19857 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
19858 .byte 15,93,208 // minps %xmm0,%xmm2
19859 .byte 72,173 // lods %ds:(%rsi),%rax
19860 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19861 .byte 255,224 // jmpq *%rax
19862
19863HIDDEN _sk_parametric_a_sse41
19864.globl _sk_parametric_a_sse41
19865FUNCTION(_sk_parametric_a_sse41)
19866_sk_parametric_a_sse41:
19867 .byte 68,15,40,192 // movaps %xmm0,%xmm8
19868 .byte 72,173 // lods %ds:(%rsi),%rax
19869 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
19870 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19871 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
19872 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19873 .byte 68,15,89,203 // mulps %xmm3,%xmm9
19874 .byte 243,68,15,16,88,4 // movss 0x4(%rax),%xmm11
19875 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
19876 .byte 68,15,89,219 // mulps %xmm3,%xmm11
19877 .byte 15,40,195 // movaps %xmm3,%xmm0
19878 .byte 65,15,194,194,2 // cmpleps %xmm10,%xmm0
19879 .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3
19880 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19881 .byte 68,15,88,203 // addps %xmm3,%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19882 .byte 243,68,15,16,16 // movss (%rax),%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19883 .byte 243,15,16,88,8 // movss 0x8(%rax),%xmm3
19884 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19885 .byte 68,15,88,219 // addps %xmm3,%xmm11
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19886 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19887 .byte 69,15,91,227 // cvtdq2ps %xmm11,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19888 .byte 68,15,89,37,87,32,0,0 // mulps 0x2057(%rip),%xmm12 # 41a0 <_sk_callback_sse41+0x340>
19889 .byte 68,15,84,29,95,32,0,0 // andps 0x205f(%rip),%xmm11 # 41b0 <_sk_callback_sse41+0x350>
19890 .byte 68,15,86,29,103,32,0,0 // orps 0x2067(%rip),%xmm11 # 41c0 <_sk_callback_sse41+0x360>
19891 .byte 68,15,88,37,111,32,0,0 // addps 0x206f(%rip),%xmm12 # 41d0 <_sk_callback_sse41+0x370>
19892 .byte 15,40,29,120,32,0,0 // movaps 0x2078(%rip),%xmm3 # 41e0 <_sk_callback_sse41+0x380>
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19893 .byte 65,15,89,219 // mulps %xmm11,%xmm3
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]19894 .byte 68,15,92,227 // subps %xmm3,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]19895 .byte 68,15,88,29,120,32,0,0 // addps 0x2078(%rip),%xmm11 # 41f0 <_sk_callback_sse41+0x390>
19896 .byte 15,40,29,129,32,0,0 // movaps 0x2081(%rip),%xmm3 # 4200 <_sk_callback_sse41+0x3a0>
19897 .byte 65,15,94,219 // divps %xmm11,%xmm3
19898 .byte 68,15,92,227 // subps %xmm3,%xmm12
19899 .byte 69,15,89,226 // mulps %xmm10,%xmm12
19900 .byte 102,69,15,58,8,212,1 // roundps $0x1,%xmm12,%xmm10
19901 .byte 69,15,40,220 // movaps %xmm12,%xmm11
19902 .byte 69,15,92,218 // subps %xmm10,%xmm11
19903 .byte 68,15,88,37,110,32,0,0 // addps 0x206e(%rip),%xmm12 # 4210 <_sk_callback_sse41+0x3b0>
19904 .byte 15,40,29,119,32,0,0 // movaps 0x2077(%rip),%xmm3 # 4220 <_sk_callback_sse41+0x3c0>
19905 .byte 65,15,89,219 // mulps %xmm11,%xmm3
19906 .byte 68,15,92,227 // subps %xmm3,%xmm12
19907 .byte 68,15,40,21,119,32,0,0 // movaps 0x2077(%rip),%xmm10 # 4230 <_sk_callback_sse41+0x3d0>
19908 .byte 69,15,92,211 // subps %xmm11,%xmm10
19909 .byte 15,40,29,124,32,0,0 // movaps 0x207c(%rip),%xmm3 # 4240 <_sk_callback_sse41+0x3e0>
19910 .byte 65,15,94,218 // divps %xmm10,%xmm3
19911 .byte 65,15,88,220 // addps %xmm12,%xmm3
19912 .byte 15,89,29,125,32,0,0 // mulps 0x207d(%rip),%xmm3 # 4250 <_sk_callback_sse41+0x3f0>
19913 .byte 102,68,15,91,211 // cvtps2dq %xmm3,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]19914 .byte 243,15,16,88,20 // movss 0x14(%rax),%xmm3
19915 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
19916 .byte 65,15,88,218 // addps %xmm10,%xmm3
19917 .byte 102,65,15,56,20,217 // blendvps %xmm0,%xmm9,%xmm3
19918 .byte 15,87,192 // xorps %xmm0,%xmm0
19919 .byte 15,95,216 // maxps %xmm0,%xmm3
19920 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
19921 .byte 102,15,110,192 // movd %eax,%xmm0
19922 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
19923 .byte 15,93,216 // minps %xmm0,%xmm3
19924 .byte 72,173 // lods %ds:(%rsi),%rax
19925 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19926 .byte 255,224 // jmpq *%rax
19927
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]19928HIDDEN _sk_lab_to_xyz_sse41
19929.globl _sk_lab_to_xyz_sse41
19930FUNCTION(_sk_lab_to_xyz_sse41)
19931_sk_lab_to_xyz_sse41:
19932 .byte 68,15,40,193 // movaps %xmm1,%xmm8
19933 .byte 184,0,0,200,66 // mov $0x42c80000,%eax
19934 .byte 102,68,15,110,200 // movd %eax,%xmm9
19935 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
19936 .byte 68,15,89,200 // mulps %xmm0,%xmm9
19937 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
19938 .byte 102,68,15,110,208 // movd %eax,%xmm10
19939 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19940 .byte 69,15,89,194 // mulps %xmm10,%xmm8
19941 .byte 184,0,0,0,67 // mov $0x43000000,%eax
19942 .byte 102,15,110,200 // movd %eax,%xmm1
19943 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
19944 .byte 68,15,92,193 // subps %xmm1,%xmm8
19945 .byte 68,15,89,210 // mulps %xmm2,%xmm10
19946 .byte 68,15,92,209 // subps %xmm1,%xmm10
19947 .byte 184,0,0,128,65 // mov $0x41800000,%eax
19948 .byte 102,15,110,208 // movd %eax,%xmm2
19949 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
19950 .byte 65,15,88,209 // addps %xmm9,%xmm2
19951 .byte 184,203,61,13,60 // mov $0x3c0d3dcb,%eax
19952 .byte 102,15,110,200 // movd %eax,%xmm1
19953 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
19954 .byte 15,89,202 // mulps %xmm2,%xmm1
19955 .byte 184,111,18,3,59 // mov $0x3b03126f,%eax
19956 .byte 102,68,15,110,232 // movd %eax,%xmm13
19957 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
19958 .byte 69,15,89,232 // mulps %xmm8,%xmm13
19959 .byte 68,15,88,233 // addps %xmm1,%xmm13
19960 .byte 184,10,215,163,59 // mov $0x3ba3d70a,%eax
19961 .byte 102,15,110,192 // movd %eax,%xmm0
19962 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
19963 .byte 65,15,89,194 // mulps %xmm10,%xmm0
19964 .byte 68,15,40,201 // movaps %xmm1,%xmm9
19965 .byte 68,15,92,200 // subps %xmm0,%xmm9
19966 .byte 69,15,40,221 // movaps %xmm13,%xmm11
19967 .byte 69,15,89,219 // mulps %xmm11,%xmm11
19968 .byte 69,15,89,221 // mulps %xmm13,%xmm11
19969 .byte 184,194,24,17,60 // mov $0x3c1118c2,%eax
19970 .byte 102,68,15,110,192 // movd %eax,%xmm8
19971 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
19972 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19973 .byte 65,15,194,195,1 // cmpltps %xmm11,%xmm0
19974 .byte 184,203,61,13,62 // mov $0x3e0d3dcb,%eax
19975 .byte 102,68,15,110,208 // movd %eax,%xmm10
19976 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
19977 .byte 69,15,92,234 // subps %xmm10,%xmm13
19978 .byte 184,80,128,3,62 // mov $0x3e038050,%eax
19979 .byte 102,68,15,110,224 // movd %eax,%xmm12
19980 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
19981 .byte 69,15,89,236 // mulps %xmm12,%xmm13
19982 .byte 102,69,15,56,20,235 // blendvps %xmm0,%xmm11,%xmm13
19983 .byte 15,40,209 // movaps %xmm1,%xmm2
19984 .byte 15,89,210 // mulps %xmm2,%xmm2
19985 .byte 15,89,209 // mulps %xmm1,%xmm2
19986 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19987 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0
19988 .byte 65,15,92,202 // subps %xmm10,%xmm1
19989 .byte 65,15,89,204 // mulps %xmm12,%xmm1
19990 .byte 102,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm1
19991 .byte 65,15,40,209 // movaps %xmm9,%xmm2
19992 .byte 15,89,210 // mulps %xmm2,%xmm2
19993 .byte 65,15,89,209 // mulps %xmm9,%xmm2
19994 .byte 68,15,194,194,1 // cmpltps %xmm2,%xmm8
19995 .byte 69,15,92,202 // subps %xmm10,%xmm9
19996 .byte 69,15,89,204 // mulps %xmm12,%xmm9
19997 .byte 65,15,40,192 // movaps %xmm8,%xmm0
19998 .byte 102,68,15,56,20,202 // blendvps %xmm0,%xmm2,%xmm9
19999 .byte 184,31,215,118,63 // mov $0x3f76d71f,%eax
20000 .byte 102,15,110,192 // movd %eax,%xmm0
20001 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20002 .byte 65,15,89,197 // mulps %xmm13,%xmm0
20003 .byte 184,246,64,83,63 // mov $0x3f5340f6,%eax
20004 .byte 102,15,110,208 // movd %eax,%xmm2
20005 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
20006 .byte 65,15,89,209 // mulps %xmm9,%xmm2
20007 .byte 72,173 // lods %ds:(%rsi),%rax
20008 .byte 255,224 // jmpq *%rax
20009
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20010HIDDEN _sk_load_a8_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20011.globl _sk_load_a8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20012FUNCTION(_sk_load_a8_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20013_sk_load_a8_sse41:
20014 .byte 72,173 // lods %ds:(%rsi),%rax
20015 .byte 72,139,0 // mov (%rax),%rax
20016 .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0
20017 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20018 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20019 .byte 102,15,110,216 // movd %eax,%xmm3
20020 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20021 .byte 15,89,216 // mulps %xmm0,%xmm3
20022 .byte 72,173 // lods %ds:(%rsi),%rax
20023 .byte 15,87,192 // xorps %xmm0,%xmm0
20024 .byte 15,87,201 // xorps %xmm1,%xmm1
20025 .byte 15,87,210 // xorps %xmm2,%xmm2
20026 .byte 255,224 // jmpq *%rax
20027
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20028HIDDEN _sk_gather_a8_sse41
20029.globl _sk_gather_a8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20030FUNCTION(_sk_gather_a8_sse41)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20031_sk_gather_a8_sse41:
20032 .byte 72,173 // lods %ds:(%rsi),%rax
20033 .byte 76,139,8 // mov (%rax),%r9
20034 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20035 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20036 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20037 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20038 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20039 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20040 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
20041 .byte 65,137,192 // mov %eax,%r8d
20042 .byte 72,193,232,32 // shr $0x20,%rax
20043 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20044 .byte 65,137,202 // mov %ecx,%r10d
20045 .byte 72,193,233,32 // shr $0x20,%rcx
20046 .byte 102,67,15,58,32,4,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm0
20047 .byte 102,65,15,58,32,4,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm0
20048 .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx
20049 .byte 102,15,58,32,193,2 // pinsrb $0x2,%ecx,%xmm0
20050 .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax
20051 .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0
20052 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0
20053 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20054 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20055 .byte 102,15,110,216 // movd %eax,%xmm3
20056 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20057 .byte 15,89,216 // mulps %xmm0,%xmm3
20058 .byte 72,173 // lods %ds:(%rsi),%rax
20059 .byte 15,87,192 // xorps %xmm0,%xmm0
20060 .byte 102,15,239,201 // pxor %xmm1,%xmm1
20061 .byte 102,15,239,210 // pxor %xmm2,%xmm2
20062 .byte 255,224 // jmpq *%rax
20063
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20064HIDDEN _sk_store_a8_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20065.globl _sk_store_a8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20066FUNCTION(_sk_store_a8_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20067_sk_store_a8_sse41:
20068 .byte 72,173 // lods %ds:(%rsi),%rax
20069 .byte 72,139,0 // mov (%rax),%rax
20070 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
20071 .byte 102,68,15,110,193 // movd %ecx,%xmm8
20072 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20073 .byte 68,15,89,195 // mulps %xmm3,%xmm8
20074 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
20075 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
20076 .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8
20077 .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1)
20078 .byte 72,173 // lods %ds:(%rsi),%rax
20079 .byte 255,224 // jmpq *%rax
20080
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20081HIDDEN _sk_load_g8_sse41
20082.globl _sk_load_g8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20083FUNCTION(_sk_load_g8_sse41)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20084_sk_load_g8_sse41:
20085 .byte 72,173 // lods %ds:(%rsi),%rax
20086 .byte 72,139,0 // mov (%rax),%rax
20087 .byte 102,15,56,49,4,56 // pmovzxbd (%rax,%rdi,1),%xmm0
20088 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20089 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20090 .byte 102,15,110,192 // movd %eax,%xmm0
20091 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20092 .byte 15,89,193 // mulps %xmm1,%xmm0
20093 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
20094 .byte 102,15,110,216 // movd %eax,%xmm3
20095 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20096 .byte 72,173 // lods %ds:(%rsi),%rax
20097 .byte 15,40,200 // movaps %xmm0,%xmm1
20098 .byte 15,40,208 // movaps %xmm0,%xmm2
20099 .byte 255,224 // jmpq *%rax
20100
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20101HIDDEN _sk_gather_g8_sse41
20102.globl _sk_gather_g8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20103FUNCTION(_sk_gather_g8_sse41)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20104_sk_gather_g8_sse41:
20105 .byte 72,173 // lods %ds:(%rsi),%rax
20106 .byte 76,139,8 // mov (%rax),%r9
20107 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20108 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20109 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20110 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20111 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20112 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20113 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
20114 .byte 65,137,192 // mov %eax,%r8d
20115 .byte 72,193,232,32 // shr $0x20,%rax
20116 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20117 .byte 65,137,202 // mov %ecx,%r10d
20118 .byte 72,193,233,32 // shr $0x20,%rcx
20119 .byte 102,67,15,58,32,4,17,0 // pinsrb $0x0,(%r9,%r10,1),%xmm0
20120 .byte 102,65,15,58,32,4,9,1 // pinsrb $0x1,(%r9,%rcx,1),%xmm0
20121 .byte 67,15,182,12,1 // movzbl (%r9,%r8,1),%ecx
20122 .byte 102,15,58,32,193,2 // pinsrb $0x2,%ecx,%xmm0
20123 .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax
20124 .byte 102,15,58,32,192,3 // pinsrb $0x3,%eax,%xmm0
20125 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0
20126 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20127 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20128 .byte 102,15,110,192 // movd %eax,%xmm0
20129 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20130 .byte 15,89,193 // mulps %xmm1,%xmm0
20131 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
20132 .byte 102,15,110,216 // movd %eax,%xmm3
20133 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20134 .byte 72,173 // lods %ds:(%rsi),%rax
20135 .byte 15,40,200 // movaps %xmm0,%xmm1
20136 .byte 15,40,208 // movaps %xmm0,%xmm2
20137 .byte 255,224 // jmpq *%rax
20138
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]20139HIDDEN _sk_gather_i8_sse41
20140.globl _sk_gather_i8_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20141FUNCTION(_sk_gather_i8_sse41)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]20142_sk_gather_i8_sse41:
20143 .byte 72,173 // lods %ds:(%rsi),%rax
20144 .byte 73,137,192 // mov %rax,%r8
20145 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20146 .byte 116,5 // je 250a <_sk_gather_i8_sse41+0xf>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]20147 .byte 76,137,192 // mov %r8,%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20148 .byte 235,2 // jmp 250c <_sk_gather_i8_sse41+0x11>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]20149 .byte 72,173 // lods %ds:(%rsi),%rax
20150 .byte 76,139,16 // mov (%rax),%r10
20151 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20152 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20153 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20154 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20155 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20156 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20157 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
20158 .byte 65,137,193 // mov %eax,%r9d
20159 .byte 72,193,232,32 // shr $0x20,%rax
20160 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20161 .byte 65,137,203 // mov %ecx,%r11d
20162 .byte 72,193,233,32 // shr $0x20,%rcx
20163 .byte 102,67,15,58,32,4,26,0 // pinsrb $0x0,(%r10,%r11,1),%xmm0
20164 .byte 102,65,15,58,32,4,10,1 // pinsrb $0x1,(%r10,%rcx,1),%xmm0
20165 .byte 102,67,15,58,32,4,10,2 // pinsrb $0x2,(%r10,%r9,1),%xmm0
20166 .byte 102,65,15,58,32,4,2,3 // pinsrb $0x3,(%r10,%rax,1),%xmm0
20167 .byte 102,15,56,49,192 // pmovzxbd %xmm0,%xmm0
20168 .byte 102,73,15,58,22,193,1 // pextrq $0x1,%xmm0,%r9
20169 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20170 .byte 73,139,64,8 // mov 0x8(%r8),%rax
20171 .byte 65,137,200 // mov %ecx,%r8d
20172 .byte 72,193,233,30 // shr $0x1e,%rcx
20173 .byte 69,137,202 // mov %r9d,%r10d
20174 .byte 73,193,233,30 // shr $0x1e,%r9
20175 .byte 102,66,15,110,28,128 // movd (%rax,%r8,4),%xmm3
20176 .byte 102,15,58,34,28,8,1 // pinsrd $0x1,(%rax,%rcx,1),%xmm3
20177 .byte 102,66,15,58,34,28,144,2 // pinsrd $0x2,(%rax,%r10,4),%xmm3
20178 .byte 102,66,15,58,34,28,8,3 // pinsrd $0x3,(%rax,%r9,1),%xmm3
20179 .byte 184,255,0,0,0 // mov $0xff,%eax
20180 .byte 102,15,110,192 // movd %eax,%xmm0
20181 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20182 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
20183 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
20184 .byte 102,15,219,200 // pand %xmm0,%xmm1
20185 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
20186 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
20187 .byte 102,15,219,208 // pand %xmm0,%xmm2
20188 .byte 102,15,219,195 // pand %xmm3,%xmm0
20189 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20190 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20191 .byte 102,68,15,110,192 // movd %eax,%xmm8
20192 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20193 .byte 65,15,89,192 // mulps %xmm8,%xmm0
20194 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
20195 .byte 65,15,89,200 // mulps %xmm8,%xmm1
20196 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
20197 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20198 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
20199 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20200 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20201 .byte 72,173 // lods %ds:(%rsi),%rax
20202 .byte 255,224 // jmpq *%rax
20203
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20204HIDDEN _sk_load_565_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20205.globl _sk_load_565_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20206FUNCTION(_sk_load_565_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20207_sk_load_565_sse41:
20208 .byte 72,173 // lods %ds:(%rsi),%rax
20209 .byte 72,139,0 // mov (%rax),%rax
20210 .byte 102,15,56,51,20,120 // pmovzxwd (%rax,%rdi,2),%xmm2
20211 .byte 184,0,248,0,0 // mov $0xf800,%eax
20212 .byte 102,15,110,192 // movd %eax,%xmm0
20213 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20214 .byte 102,15,219,194 // pand %xmm2,%xmm0
20215 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20216 .byte 184,8,33,132,55 // mov $0x37842108,%eax
20217 .byte 102,15,110,192 // movd %eax,%xmm0
20218 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20219 .byte 15,89,193 // mulps %xmm1,%xmm0
20220 .byte 184,224,7,0,0 // mov $0x7e0,%eax
20221 .byte 102,15,110,200 // movd %eax,%xmm1
20222 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
20223 .byte 102,15,219,202 // pand %xmm2,%xmm1
20224 .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3
20225 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
20226 .byte 102,15,110,200 // movd %eax,%xmm1
20227 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
20228 .byte 15,89,203 // mulps %xmm3,%xmm1
20229 .byte 184,31,0,0,0 // mov $0x1f,%eax
20230 .byte 102,15,110,216 // movd %eax,%xmm3
20231 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
20232 .byte 102,15,219,218 // pand %xmm2,%xmm3
20233 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20234 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
20235 .byte 102,15,110,208 // movd %eax,%xmm2
20236 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
20237 .byte 15,89,211 // mulps %xmm3,%xmm2
20238 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
20239 .byte 102,15,110,216 // movd %eax,%xmm3
20240 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20241 .byte 72,173 // lods %ds:(%rsi),%rax
20242 .byte 255,224 // jmpq *%rax
20243
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20244HIDDEN _sk_gather_565_sse41
20245.globl _sk_gather_565_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20246FUNCTION(_sk_gather_565_sse41)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20247_sk_gather_565_sse41:
20248 .byte 72,173 // lods %ds:(%rsi),%rax
20249 .byte 76,139,8 // mov (%rax),%r9
20250 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20251 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20252 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20253 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20254 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20255 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20256 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
20257 .byte 65,137,192 // mov %eax,%r8d
20258 .byte 72,193,232,32 // shr $0x20,%rax
20259 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20260 .byte 65,137,202 // mov %ecx,%r10d
20261 .byte 72,193,233,32 // shr $0x20,%rcx
20262 .byte 102,67,15,196,4,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm0
20263 .byte 102,65,15,196,4,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm0
20264 .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx
20265 .byte 102,15,196,193,2 // pinsrw $0x2,%ecx,%xmm0
20266 .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax
20267 .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0
20268 .byte 102,15,56,51,208 // pmovzxwd %xmm0,%xmm2
20269 .byte 184,0,248,0,0 // mov $0xf800,%eax
20270 .byte 102,15,110,192 // movd %eax,%xmm0
20271 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20272 .byte 102,15,219,194 // pand %xmm2,%xmm0
20273 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20274 .byte 184,8,33,132,55 // mov $0x37842108,%eax
20275 .byte 102,15,110,192 // movd %eax,%xmm0
20276 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20277 .byte 15,89,193 // mulps %xmm1,%xmm0
20278 .byte 184,224,7,0,0 // mov $0x7e0,%eax
20279 .byte 102,15,110,200 // movd %eax,%xmm1
20280 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
20281 .byte 102,15,219,202 // pand %xmm2,%xmm1
20282 .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3
20283 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
20284 .byte 102,15,110,200 // movd %eax,%xmm1
20285 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
20286 .byte 15,89,203 // mulps %xmm3,%xmm1
20287 .byte 184,31,0,0,0 // mov $0x1f,%eax
20288 .byte 102,15,110,216 // movd %eax,%xmm3
20289 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
20290 .byte 102,15,219,218 // pand %xmm2,%xmm3
20291 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20292 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
20293 .byte 102,15,110,208 // movd %eax,%xmm2
20294 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
20295 .byte 15,89,211 // mulps %xmm3,%xmm2
20296 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
20297 .byte 102,15,110,216 // movd %eax,%xmm3
20298 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20299 .byte 72,173 // lods %ds:(%rsi),%rax
20300 .byte 255,224 // jmpq *%rax
20301
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20302HIDDEN _sk_store_565_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20303.globl _sk_store_565_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20304FUNCTION(_sk_store_565_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20305_sk_store_565_sse41:
20306 .byte 72,173 // lods %ds:(%rsi),%rax
20307 .byte 72,139,0 // mov (%rax),%rax
20308 .byte 185,0,0,248,65 // mov $0x41f80000,%ecx
20309 .byte 102,68,15,110,193 // movd %ecx,%xmm8
20310 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20311 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20312 .byte 68,15,89,200 // mulps %xmm0,%xmm9
20313 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20314 .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9
20315 .byte 185,0,0,124,66 // mov $0x427c0000,%ecx
20316 .byte 102,68,15,110,209 // movd %ecx,%xmm10
20317 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
20318 .byte 68,15,89,209 // mulps %xmm1,%xmm10
20319 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
20320 .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10
20321 .byte 102,69,15,235,209 // por %xmm9,%xmm10
20322 .byte 68,15,89,194 // mulps %xmm2,%xmm8
20323 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
20324 .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
20325 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
20326 .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
20327 .byte 72,173 // lods %ds:(%rsi),%rax
20328 .byte 255,224 // jmpq *%rax
20329
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20330HIDDEN _sk_load_4444_sse41
20331.globl _sk_load_4444_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20332FUNCTION(_sk_load_4444_sse41)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20333_sk_load_4444_sse41:
20334 .byte 72,173 // lods %ds:(%rsi),%rax
20335 .byte 72,139,0 // mov (%rax),%rax
20336 .byte 102,68,15,56,51,12,120 // pmovzxwd (%rax,%rdi,2),%xmm9
20337 .byte 184,0,240,0,0 // mov $0xf000,%eax
20338 .byte 102,15,110,192 // movd %eax,%xmm0
20339 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20340 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
20341 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20342 .byte 184,137,136,136,55 // mov $0x37888889,%eax
20343 .byte 102,15,110,192 // movd %eax,%xmm0
20344 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20345 .byte 15,89,193 // mulps %xmm1,%xmm0
20346 .byte 184,0,15,0,0 // mov $0xf00,%eax
20347 .byte 102,15,110,200 // movd %eax,%xmm1
20348 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
20349 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
20350 .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
20351 .byte 184,137,136,136,57 // mov $0x39888889,%eax
20352 .byte 102,15,110,200 // movd %eax,%xmm1
20353 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
20354 .byte 15,89,202 // mulps %xmm2,%xmm1
20355 .byte 184,240,0,0,0 // mov $0xf0,%eax
20356 .byte 102,15,110,208 // movd %eax,%xmm2
20357 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20358 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
20359 .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
20360 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
20361 .byte 102,15,110,208 // movd %eax,%xmm2
20362 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
20363 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20364 .byte 184,15,0,0,0 // mov $0xf,%eax
20365 .byte 102,15,110,216 // movd %eax,%xmm3
20366 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
20367 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
20368 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
20369 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
20370 .byte 102,15,110,216 // movd %eax,%xmm3
20371 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20372 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20373 .byte 72,173 // lods %ds:(%rsi),%rax
20374 .byte 255,224 // jmpq *%rax
20375
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20376HIDDEN _sk_gather_4444_sse41
20377.globl _sk_gather_4444_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20378FUNCTION(_sk_gather_4444_sse41)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]20379_sk_gather_4444_sse41:
20380 .byte 72,173 // lods %ds:(%rsi),%rax
20381 .byte 76,139,8 // mov (%rax),%r9
20382 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20383 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20384 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20385 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20386 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20387 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20388 .byte 102,72,15,58,22,192,1 // pextrq $0x1,%xmm0,%rax
20389 .byte 65,137,192 // mov %eax,%r8d
20390 .byte 72,193,232,32 // shr $0x20,%rax
20391 .byte 102,72,15,126,193 // movq %xmm0,%rcx
20392 .byte 65,137,202 // mov %ecx,%r10d
20393 .byte 72,193,233,32 // shr $0x20,%rcx
20394 .byte 102,67,15,196,4,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm0
20395 .byte 102,65,15,196,4,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm0
20396 .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx
20397 .byte 102,15,196,193,2 // pinsrw $0x2,%ecx,%xmm0
20398 .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax
20399 .byte 102,15,196,192,3 // pinsrw $0x3,%eax,%xmm0
20400 .byte 102,68,15,56,51,200 // pmovzxwd %xmm0,%xmm9
20401 .byte 184,0,240,0,0 // mov $0xf000,%eax
20402 .byte 102,15,110,192 // movd %eax,%xmm0
20403 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20404 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
20405 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
20406 .byte 184,137,136,136,55 // mov $0x37888889,%eax
20407 .byte 102,15,110,192 // movd %eax,%xmm0
20408 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
20409 .byte 15,89,193 // mulps %xmm1,%xmm0
20410 .byte 184,0,15,0,0 // mov $0xf00,%eax
20411 .byte 102,15,110,200 // movd %eax,%xmm1
20412 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
20413 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
20414 .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
20415 .byte 184,137,136,136,57 // mov $0x39888889,%eax
20416 .byte 102,15,110,200 // movd %eax,%xmm1
20417 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
20418 .byte 15,89,202 // mulps %xmm2,%xmm1
20419 .byte 184,240,0,0,0 // mov $0xf0,%eax
20420 .byte 102,15,110,208 // movd %eax,%xmm2
20421 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20422 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
20423 .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
20424 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
20425 .byte 102,15,110,208 // movd %eax,%xmm2
20426 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
20427 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20428 .byte 184,15,0,0,0 // mov $0xf,%eax
20429 .byte 102,15,110,216 // movd %eax,%xmm3
20430 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
20431 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
20432 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
20433 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
20434 .byte 102,15,110,216 // movd %eax,%xmm3
20435 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20436 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20437 .byte 72,173 // lods %ds:(%rsi),%rax
20438 .byte 255,224 // jmpq *%rax
20439
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20440HIDDEN _sk_store_4444_sse41
20441.globl _sk_store_4444_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20442FUNCTION(_sk_store_4444_sse41)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]20443_sk_store_4444_sse41:
20444 .byte 72,173 // lods %ds:(%rsi),%rax
20445 .byte 72,139,0 // mov (%rax),%rax
20446 .byte 185,0,0,112,65 // mov $0x41700000,%ecx
20447 .byte 102,68,15,110,193 // movd %ecx,%xmm8
20448 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20449 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20450 .byte 68,15,89,200 // mulps %xmm0,%xmm9
20451 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20452 .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9
20453 .byte 69,15,40,208 // movaps %xmm8,%xmm10
20454 .byte 68,15,89,209 // mulps %xmm1,%xmm10
20455 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
20456 .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
20457 .byte 102,69,15,235,209 // por %xmm9,%xmm10
20458 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20459 .byte 68,15,89,202 // mulps %xmm2,%xmm9
20460 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20461 .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9
20462 .byte 68,15,89,195 // mulps %xmm3,%xmm8
20463 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
20464 .byte 102,69,15,86,193 // orpd %xmm9,%xmm8
20465 .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
20466 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
20467 .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
20468 .byte 72,173 // lods %ds:(%rsi),%rax
20469 .byte 255,224 // jmpq *%rax
20470
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20471HIDDEN _sk_load_8888_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20472.globl _sk_load_8888_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20473FUNCTION(_sk_load_8888_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20474_sk_load_8888_sse41:
20475 .byte 72,173 // lods %ds:(%rsi),%rax
20476 .byte 72,139,0 // mov (%rax),%rax
20477 .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3
20478 .byte 184,255,0,0,0 // mov $0xff,%eax
20479 .byte 102,15,110,192 // movd %eax,%xmm0
20480 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20481 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
20482 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
20483 .byte 102,15,219,200 // pand %xmm0,%xmm1
20484 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
20485 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
20486 .byte 102,15,219,208 // pand %xmm0,%xmm2
20487 .byte 102,15,219,195 // pand %xmm3,%xmm0
20488 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20489 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20490 .byte 102,68,15,110,192 // movd %eax,%xmm8
20491 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20492 .byte 65,15,89,192 // mulps %xmm8,%xmm0
20493 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
20494 .byte 65,15,89,200 // mulps %xmm8,%xmm1
20495 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
20496 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20497 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
20498 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20499 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20500 .byte 72,173 // lods %ds:(%rsi),%rax
20501 .byte 255,224 // jmpq *%rax
20502
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]20503HIDDEN _sk_gather_8888_sse41
20504.globl _sk_gather_8888_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20505FUNCTION(_sk_gather_8888_sse41)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]20506_sk_gather_8888_sse41:
20507 .byte 72,173 // lods %ds:(%rsi),%rax
20508 .byte 76,139,8 // mov (%rax),%r9
20509 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20510 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20511 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20512 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20513 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20514 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20515 .byte 102,72,15,126,192 // movq %xmm0,%rax
20516 .byte 65,137,192 // mov %eax,%r8d
20517 .byte 72,193,232,32 // shr $0x20,%rax
20518 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx
20519 .byte 65,137,202 // mov %ecx,%r10d
20520 .byte 72,193,233,32 // shr $0x20,%rcx
20521 .byte 102,67,15,110,28,129 // movd (%r9,%r8,4),%xmm3
20522 .byte 102,65,15,58,34,28,129,1 // pinsrd $0x1,(%r9,%rax,4),%xmm3
20523 .byte 102,67,15,58,34,28,145,2 // pinsrd $0x2,(%r9,%r10,4),%xmm3
20524 .byte 102,65,15,58,34,28,137,3 // pinsrd $0x3,(%r9,%rcx,4),%xmm3
20525 .byte 184,255,0,0,0 // mov $0xff,%eax
20526 .byte 102,15,110,192 // movd %eax,%xmm0
20527 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
20528 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
20529 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
20530 .byte 102,15,219,200 // pand %xmm0,%xmm1
20531 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
20532 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
20533 .byte 102,15,219,208 // pand %xmm0,%xmm2
20534 .byte 102,15,219,195 // pand %xmm3,%xmm0
20535 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20536 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
20537 .byte 102,68,15,110,192 // movd %eax,%xmm8
20538 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20539 .byte 65,15,89,192 // mulps %xmm8,%xmm0
20540 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
20541 .byte 65,15,89,200 // mulps %xmm8,%xmm1
20542 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
20543 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20544 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
20545 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20546 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20547 .byte 72,173 // lods %ds:(%rsi),%rax
20548 .byte 255,224 // jmpq *%rax
20549
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20550HIDDEN _sk_store_8888_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20551.globl _sk_store_8888_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20552FUNCTION(_sk_store_8888_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20553_sk_store_8888_sse41:
20554 .byte 72,173 // lods %ds:(%rsi),%rax
20555 .byte 72,139,0 // mov (%rax),%rax
20556 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
20557 .byte 102,68,15,110,193 // movd %ecx,%xmm8
20558 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20559 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20560 .byte 68,15,89,200 // mulps %xmm0,%xmm9
20561 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20562 .byte 69,15,40,208 // movaps %xmm8,%xmm10
20563 .byte 68,15,89,209 // mulps %xmm1,%xmm10
20564 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
20565 .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
20566 .byte 102,69,15,235,209 // por %xmm9,%xmm10
20567 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20568 .byte 68,15,89,202 // mulps %xmm2,%xmm9
20569 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20570 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
20571 .byte 68,15,89,195 // mulps %xmm3,%xmm8
20572 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
20573 .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8
20574 .byte 102,69,15,235,193 // por %xmm9,%xmm8
20575 .byte 102,69,15,235,194 // por %xmm10,%xmm8
20576 .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4)
20577 .byte 72,173 // lods %ds:(%rsi),%rax
20578 .byte 255,224 // jmpq *%rax
20579
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20580HIDDEN _sk_load_f16_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20581.globl _sk_load_f16_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20582FUNCTION(_sk_load_f16_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20583_sk_load_f16_sse41:
20584 .byte 72,173 // lods %ds:(%rsi),%rax
20585 .byte 72,139,0 // mov (%rax),%rax
20586 .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0
20587 .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20588 .byte 102,68,15,111,200 // movdqa %xmm0,%xmm9
20589 .byte 102,68,15,97,201 // punpcklwd %xmm1,%xmm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20590 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20591 .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11
20592 .byte 102,68,15,97,216 // punpcklwd %xmm0,%xmm11
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20593 .byte 102,68,15,105,200 // punpckhwd %xmm0,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20594 .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1
20595 .byte 102,68,15,111,5,152,22,0,0 // movdqa 0x1698(%rip),%xmm8 # 4260 <_sk_callback_sse41+0x400>
20596 .byte 102,15,111,209 // movdqa %xmm1,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20597 .byte 102,65,15,219,208 // pand %xmm8,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20598 .byte 102,15,239,202 // pxor %xmm2,%xmm1
20599 .byte 102,15,111,29,147,22,0,0 // movdqa 0x1693(%rip),%xmm3 # 4270 <_sk_callback_sse41+0x410>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20600 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20601 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
20602 .byte 102,15,56,63,195 // pmaxud %xmm3,%xmm0
20603 .byte 102,15,118,193 // pcmpeqd %xmm1,%xmm0
20604 .byte 102,15,114,241,13 // pslld $0xd,%xmm1
20605 .byte 102,15,235,202 // por %xmm2,%xmm1
20606 .byte 102,68,15,111,21,127,22,0,0 // movdqa 0x167f(%rip),%xmm10 # 4280 <_sk_callback_sse41+0x420>
20607 .byte 102,65,15,254,202 // paddd %xmm10,%xmm1
20608 .byte 102,15,219,193 // pand %xmm1,%xmm0
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20609 .byte 102,65,15,115,219,8 // psrldq $0x8,%xmm11
20610 .byte 102,69,15,56,51,219 // pmovzxwd %xmm11,%xmm11
20611 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2
20612 .byte 102,65,15,219,208 // pand %xmm8,%xmm2
20613 .byte 102,68,15,239,218 // pxor %xmm2,%xmm11
20614 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20615 .byte 102,65,15,111,203 // movdqa %xmm11,%xmm1
20616 .byte 102,15,56,63,203 // pmaxud %xmm3,%xmm1
20617 .byte 102,65,15,118,203 // pcmpeqd %xmm11,%xmm1
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20618 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20619 .byte 102,68,15,235,218 // por %xmm2,%xmm11
20620 .byte 102,69,15,254,218 // paddd %xmm10,%xmm11
20621 .byte 102,65,15,219,203 // pand %xmm11,%xmm1
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20622 .byte 102,69,15,56,51,217 // pmovzxwd %xmm9,%xmm11
20623 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12
20624 .byte 102,69,15,219,224 // pand %xmm8,%xmm12
20625 .byte 102,69,15,239,220 // pxor %xmm12,%xmm11
20626 .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20627 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2
20628 .byte 102,15,56,63,211 // pmaxud %xmm3,%xmm2
20629 .byte 102,65,15,118,211 // pcmpeqd %xmm11,%xmm2
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20630 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20631 .byte 102,69,15,235,220 // por %xmm12,%xmm11
20632 .byte 102,69,15,254,218 // paddd %xmm10,%xmm11
20633 .byte 102,65,15,219,211 // pand %xmm11,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20634 .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9
20635 .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9
20636 .byte 102,69,15,219,193 // pand %xmm9,%xmm8
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20637 .byte 102,69,15,239,200 // pxor %xmm8,%xmm9
20638 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20639 .byte 102,65,15,56,63,217 // pmaxud %xmm9,%xmm3
20640 .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20641 .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20642 .byte 102,69,15,235,200 // por %xmm8,%xmm9
20643 .byte 102,69,15,254,202 // paddd %xmm10,%xmm9
20644 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20645 .byte 72,173 // lods %ds:(%rsi),%rax
20646 .byte 255,224 // jmpq *%rax
20647
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]20648HIDDEN _sk_gather_f16_sse41
20649.globl _sk_gather_f16_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20650FUNCTION(_sk_gather_f16_sse41)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]20651_sk_gather_f16_sse41:
20652 .byte 72,173 // lods %ds:(%rsi),%rax
20653 .byte 76,139,8 // mov (%rax),%r9
20654 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
20655 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
20656 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
20657 .byte 102,15,56,64,209 // pmulld %xmm1,%xmm2
20658 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
20659 .byte 102,15,254,194 // paddd %xmm2,%xmm0
20660 .byte 102,72,15,126,192 // movq %xmm0,%rax
20661 .byte 65,137,192 // mov %eax,%r8d
20662 .byte 72,193,232,32 // shr $0x20,%rax
20663 .byte 102,72,15,58,22,193,1 // pextrq $0x1,%xmm0,%rcx
20664 .byte 65,137,202 // mov %ecx,%r10d
20665 .byte 72,193,233,32 // shr $0x20,%rcx
20666 .byte 243,65,15,126,4,201 // movq (%r9,%rcx,8),%xmm0
20667 .byte 243,67,15,126,12,209 // movq (%r9,%r10,8),%xmm1
20668 .byte 102,15,108,200 // punpcklqdq %xmm0,%xmm1
20669 .byte 243,65,15,126,4,193 // movq (%r9,%rax,8),%xmm0
20670 .byte 243,67,15,126,20,193 // movq (%r9,%r8,8),%xmm2
20671 .byte 102,15,108,208 // punpcklqdq %xmm0,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20672 .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9
20673 .byte 102,68,15,97,201 // punpcklwd %xmm1,%xmm9
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]20674 .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20675 .byte 102,69,15,111,217 // movdqa %xmm9,%xmm11
20676 .byte 102,68,15,97,218 // punpcklwd %xmm2,%xmm11
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20677 .byte 102,68,15,105,202 // punpckhwd %xmm2,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20678 .byte 102,65,15,56,51,203 // pmovzxwd %xmm11,%xmm1
20679 .byte 102,68,15,111,5,61,21,0,0 // movdqa 0x153d(%rip),%xmm8 # 4290 <_sk_callback_sse41+0x430>
20680 .byte 102,15,111,209 // movdqa %xmm1,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20681 .byte 102,65,15,219,208 // pand %xmm8,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20682 .byte 102,15,239,202 // pxor %xmm2,%xmm1
20683 .byte 102,15,111,29,56,21,0,0 // movdqa 0x1538(%rip),%xmm3 # 42a0 <_sk_callback_sse41+0x440>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20684 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20685 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
20686 .byte 102,15,56,63,195 // pmaxud %xmm3,%xmm0
20687 .byte 102,15,118,193 // pcmpeqd %xmm1,%xmm0
20688 .byte 102,15,114,241,13 // pslld $0xd,%xmm1
20689 .byte 102,15,235,202 // por %xmm2,%xmm1
20690 .byte 102,68,15,111,21,36,21,0,0 // movdqa 0x1524(%rip),%xmm10 # 42b0 <_sk_callback_sse41+0x450>
20691 .byte 102,65,15,254,202 // paddd %xmm10,%xmm1
20692 .byte 102,15,219,193 // pand %xmm1,%xmm0
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20693 .byte 102,65,15,115,219,8 // psrldq $0x8,%xmm11
20694 .byte 102,69,15,56,51,219 // pmovzxwd %xmm11,%xmm11
20695 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2
20696 .byte 102,65,15,219,208 // pand %xmm8,%xmm2
20697 .byte 102,68,15,239,218 // pxor %xmm2,%xmm11
20698 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20699 .byte 102,65,15,111,203 // movdqa %xmm11,%xmm1
20700 .byte 102,15,56,63,203 // pmaxud %xmm3,%xmm1
20701 .byte 102,65,15,118,203 // pcmpeqd %xmm11,%xmm1
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20702 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20703 .byte 102,68,15,235,218 // por %xmm2,%xmm11
20704 .byte 102,69,15,254,218 // paddd %xmm10,%xmm11
20705 .byte 102,65,15,219,203 // pand %xmm11,%xmm1
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20706 .byte 102,69,15,56,51,217 // pmovzxwd %xmm9,%xmm11
20707 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12
20708 .byte 102,69,15,219,224 // pand %xmm8,%xmm12
20709 .byte 102,69,15,239,220 // pxor %xmm12,%xmm11
20710 .byte 102,65,15,114,244,16 // pslld $0x10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20711 .byte 102,65,15,111,211 // movdqa %xmm11,%xmm2
20712 .byte 102,15,56,63,211 // pmaxud %xmm3,%xmm2
20713 .byte 102,65,15,118,211 // pcmpeqd %xmm11,%xmm2
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20714 .byte 102,65,15,114,243,13 // pslld $0xd,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20715 .byte 102,69,15,235,220 // por %xmm12,%xmm11
20716 .byte 102,69,15,254,218 // paddd %xmm10,%xmm11
20717 .byte 102,65,15,219,211 // pand %xmm11,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20718 .byte 102,65,15,115,217,8 // psrldq $0x8,%xmm9
20719 .byte 102,69,15,56,51,201 // pmovzxwd %xmm9,%xmm9
20720 .byte 102,69,15,219,193 // pand %xmm9,%xmm8
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20721 .byte 102,69,15,239,200 // pxor %xmm8,%xmm9
20722 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20723 .byte 102,65,15,56,63,217 // pmaxud %xmm9,%xmm3
20724 .byte 102,65,15,118,217 // pcmpeqd %xmm9,%xmm3
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20725 .byte 102,65,15,114,241,13 // pslld $0xd,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20726 .byte 102,69,15,235,200 // por %xmm8,%xmm9
20727 .byte 102,69,15,254,202 // paddd %xmm10,%xmm9
20728 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]20729 .byte 72,173 // lods %ds:(%rsi),%rax
20730 .byte 255,224 // jmpq *%rax
20731
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20732HIDDEN _sk_store_f16_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20733.globl _sk_store_f16_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20734FUNCTION(_sk_store_f16_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20735_sk_store_f16_sse41:
20736 .byte 72,173 // lods %ds:(%rsi),%rax
20737 .byte 72,139,0 // mov (%rax),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20738 .byte 102,68,15,111,21,90,20,0,0 // movdqa 0x145a(%rip),%xmm10 # 42c0 <_sk_callback_sse41+0x460>
20739 .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20740 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20741 .byte 102,69,15,219,234 // pand %xmm10,%xmm13
20742 .byte 102,69,15,239,229 // pxor %xmm13,%xmm12
20743 .byte 102,68,15,111,13,77,20,0,0 // movdqa 0x144d(%rip),%xmm9 # 42d0 <_sk_callback_sse41+0x470>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20744 .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20745 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8
20746 .byte 102,69,15,102,196 // pcmpgtd %xmm12,%xmm8
20747 .byte 102,65,15,114,212,13 // psrld $0xd,%xmm12
20748 .byte 102,68,15,111,29,62,20,0,0 // movdqa 0x143e(%rip),%xmm11 # 42e0 <_sk_callback_sse41+0x480>
20749 .byte 102,69,15,235,235 // por %xmm11,%xmm13
20750 .byte 102,69,15,254,236 // paddd %xmm12,%xmm13
20751 .byte 102,69,15,223,197 // pandn %xmm13,%xmm8
20752 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
20753 .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13
20754 .byte 102,68,15,111,241 // movdqa %xmm1,%xmm14
20755 .byte 102,69,15,219,242 // pand %xmm10,%xmm14
20756 .byte 102,69,15,239,238 // pxor %xmm14,%xmm13
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20757 .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20758 .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12
20759 .byte 102,69,15,102,229 // pcmpgtd %xmm13,%xmm12
20760 .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13
20761 .byte 102,69,15,235,243 // por %xmm11,%xmm14
20762 .byte 102,69,15,254,245 // paddd %xmm13,%xmm14
20763 .byte 102,69,15,223,230 // pandn %xmm14,%xmm12
20764 .byte 102,69,15,56,43,228 // packusdw %xmm12,%xmm12
20765 .byte 102,68,15,111,242 // movdqa %xmm2,%xmm14
20766 .byte 102,68,15,111,250 // movdqa %xmm2,%xmm15
20767 .byte 102,69,15,219,250 // pand %xmm10,%xmm15
20768 .byte 102,69,15,239,247 // pxor %xmm15,%xmm14
20769 .byte 102,65,15,114,215,16 // psrld $0x10,%xmm15
20770 .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13
20771 .byte 102,69,15,102,238 // pcmpgtd %xmm14,%xmm13
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]20772 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20773 .byte 102,69,15,235,251 // por %xmm11,%xmm15
20774 .byte 102,69,15,254,254 // paddd %xmm14,%xmm15
20775 .byte 102,69,15,223,239 // pandn %xmm15,%xmm13
20776 .byte 102,69,15,56,43,237 // packusdw %xmm13,%xmm13
20777 .byte 102,68,15,219,211 // pand %xmm3,%xmm10
20778 .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14
20779 .byte 102,69,15,239,242 // pxor %xmm10,%xmm14
20780 .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10
20781 .byte 102,69,15,102,206 // pcmpgtd %xmm14,%xmm9
20782 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14
20783 .byte 102,69,15,235,211 // por %xmm11,%xmm10
20784 .byte 102,69,15,254,214 // paddd %xmm14,%xmm10
20785 .byte 102,69,15,223,202 // pandn %xmm10,%xmm9
20786 .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20787 .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]20788 .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]20789 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20790 .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20791 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]20792 .byte 102,69,15,106,197 // punpckhdq %xmm13,%xmm8
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]20793 .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20794 .byte 72,173 // lods %ds:(%rsi),%rax
20795 .byte 255,224 // jmpq *%rax
20796
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]20797HIDDEN _sk_load_u16_be_sse41
20798.globl _sk_load_u16_be_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20799FUNCTION(_sk_load_u16_be_sse41)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]20800_sk_load_u16_be_sse41:
20801 .byte 72,173 // lods %ds:(%rsi),%rax
20802 .byte 72,139,0 // mov (%rax),%rax
20803 .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0
20804 .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1
20805 .byte 102,15,111,208 // movdqa %xmm0,%xmm2
20806 .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2
20807 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
20808 .byte 102,15,111,202 // movdqa %xmm2,%xmm1
20809 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
20810 .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2
20811 .byte 184,128,0,128,55 // mov $0x37800080,%eax
20812 .byte 102,68,15,110,192 // movd %eax,%xmm8
20813 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20814 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
20815 .byte 102,15,113,240,8 // psllw $0x8,%xmm0
20816 .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3
20817 .byte 102,15,113,209,8 // psrlw $0x8,%xmm1
20818 .byte 102,15,235,200 // por %xmm0,%xmm1
20819 .byte 102,15,56,51,193 // pmovzxwd %xmm1,%xmm0
20820 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20821 .byte 65,15,89,192 // mulps %xmm8,%xmm0
20822 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
20823 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
20824 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
20825 .byte 102,15,235,217 // por %xmm1,%xmm3
20826 .byte 102,15,56,51,203 // pmovzxwd %xmm3,%xmm1
20827 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
20828 .byte 65,15,89,200 // mulps %xmm8,%xmm1
20829 .byte 102,68,15,111,202 // movdqa %xmm2,%xmm9
20830 .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9
20831 .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3
20832 .byte 102,15,113,210,8 // psrlw $0x8,%xmm2
20833 .byte 102,65,15,235,209 // por %xmm9,%xmm2
20834 .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2
20835 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
20836 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20837 .byte 102,68,15,111,203 // movdqa %xmm3,%xmm9
20838 .byte 102,65,15,113,241,8 // psllw $0x8,%xmm9
20839 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
20840 .byte 102,65,15,235,217 // por %xmm9,%xmm3
20841 .byte 102,15,56,51,219 // pmovzxwd %xmm3,%xmm3
20842 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
20843 .byte 65,15,89,216 // mulps %xmm8,%xmm3
20844 .byte 72,173 // lods %ds:(%rsi),%rax
20845 .byte 255,224 // jmpq *%rax
20846
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]20847HIDDEN _sk_load_rgb_u16_be_sse41
20848.globl _sk_load_rgb_u16_be_sse41
20849FUNCTION(_sk_load_rgb_u16_be_sse41)
20850_sk_load_rgb_u16_be_sse41:
20851 .byte 72,173 // lods %ds:(%rsi),%rax
20852 .byte 72,139,0 // mov (%rax),%rax
20853 .byte 72,141,12,127 // lea (%rdi,%rdi,2),%rcx
20854 .byte 243,15,111,20,72 // movdqu (%rax,%rcx,2),%xmm2
20855 .byte 243,15,111,68,72,8 // movdqu 0x8(%rax,%rcx,2),%xmm0
20856 .byte 102,15,115,216,4 // psrldq $0x4,%xmm0
20857 .byte 102,15,111,202 // movdqa %xmm2,%xmm1
20858 .byte 102,15,115,217,6 // psrldq $0x6,%xmm1
20859 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2
20860 .byte 102,15,115,216,6 // psrldq $0x6,%xmm0
20861 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
20862 .byte 102,15,111,194 // movdqa %xmm2,%xmm0
20863 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
20864 .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3
20865 .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2
20866 .byte 184,128,0,128,55 // mov $0x37800080,%eax
20867 .byte 102,68,15,110,192 // movd %eax,%xmm8
20868 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
20869 .byte 102,15,111,200 // movdqa %xmm0,%xmm1
20870 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
20871 .byte 102,15,113,208,8 // psrlw $0x8,%xmm0
20872 .byte 102,15,235,193 // por %xmm1,%xmm0
20873 .byte 102,15,56,51,192 // pmovzxwd %xmm0,%xmm0
20874 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
20875 .byte 65,15,89,192 // mulps %xmm8,%xmm0
20876 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
20877 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
20878 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
20879 .byte 102,15,235,217 // por %xmm1,%xmm3
20880 .byte 102,15,56,51,203 // pmovzxwd %xmm3,%xmm1
20881 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
20882 .byte 65,15,89,200 // mulps %xmm8,%xmm1
20883 .byte 102,15,111,218 // movdqa %xmm2,%xmm3
20884 .byte 102,15,113,243,8 // psllw $0x8,%xmm3
20885 .byte 102,15,113,210,8 // psrlw $0x8,%xmm2
20886 .byte 102,15,235,211 // por %xmm3,%xmm2
20887 .byte 102,15,56,51,210 // pmovzxwd %xmm2,%xmm2
20888 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
20889 .byte 65,15,89,208 // mulps %xmm8,%xmm2
20890 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
20891 .byte 102,15,110,216 // movd %eax,%xmm3
20892 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
20893 .byte 72,173 // lods %ds:(%rsi),%rax
20894 .byte 255,224 // jmpq *%rax
20895
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]20896HIDDEN _sk_store_u16_be_sse41
20897.globl _sk_store_u16_be_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20898FUNCTION(_sk_store_u16_be_sse41)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]20899_sk_store_u16_be_sse41:
20900 .byte 72,173 // lods %ds:(%rsi),%rax
20901 .byte 72,139,0 // mov (%rax),%rax
20902 .byte 185,0,255,127,71 // mov $0x477fff00,%ecx
20903 .byte 102,68,15,110,201 // movd %ecx,%xmm9
20904 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
20905 .byte 69,15,40,193 // movaps %xmm9,%xmm8
20906 .byte 68,15,89,192 // mulps %xmm0,%xmm8
20907 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
20908 .byte 102,69,15,56,43,192 // packusdw %xmm8,%xmm8
20909 .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
20910 .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10
20911 .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8
20912 .byte 102,69,15,235,194 // por %xmm10,%xmm8
20913 .byte 69,15,40,209 // movaps %xmm9,%xmm10
20914 .byte 68,15,89,209 // mulps %xmm1,%xmm10
20915 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
20916 .byte 102,69,15,56,43,210 // packusdw %xmm10,%xmm10
20917 .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11
20918 .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11
20919 .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10
20920 .byte 102,69,15,235,211 // por %xmm11,%xmm10
20921 .byte 69,15,40,217 // movaps %xmm9,%xmm11
20922 .byte 68,15,89,218 // mulps %xmm2,%xmm11
20923 .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11
20924 .byte 102,69,15,56,43,219 // packusdw %xmm11,%xmm11
20925 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12
20926 .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12
20927 .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11
20928 .byte 102,69,15,235,220 // por %xmm12,%xmm11
20929 .byte 68,15,89,203 // mulps %xmm3,%xmm9
20930 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
20931 .byte 102,69,15,56,43,201 // packusdw %xmm9,%xmm9
20932 .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12
20933 .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12
20934 .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9
20935 .byte 102,69,15,235,204 // por %xmm12,%xmm9
20936 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
20937 .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11
20938 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
20939 .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9
20940 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
20941 .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8
20942 .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
20943 .byte 72,173 // lods %ds:(%rsi),%rax
20944 .byte 255,224 // jmpq *%rax
20945
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]20946HIDDEN _sk_load_f32_sse41
20947.globl _sk_load_f32_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20948FUNCTION(_sk_load_f32_sse41)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]20949_sk_load_f32_sse41:
20950 .byte 72,173 // lods %ds:(%rsi),%rax
20951 .byte 72,139,0 // mov (%rax),%rax
20952 .byte 72,137,249 // mov %rdi,%rcx
20953 .byte 72,193,225,4 // shl $0x4,%rcx
20954 .byte 68,15,16,4,8 // movups (%rax,%rcx,1),%xmm8
20955 .byte 15,16,68,8,16 // movups 0x10(%rax,%rcx,1),%xmm0
20956 .byte 15,16,92,8,32 // movups 0x20(%rax,%rcx,1),%xmm3
20957 .byte 68,15,16,76,8,48 // movups 0x30(%rax,%rcx,1),%xmm9
20958 .byte 65,15,40,208 // movaps %xmm8,%xmm2
20959 .byte 15,20,208 // unpcklps %xmm0,%xmm2
20960 .byte 15,40,203 // movaps %xmm3,%xmm1
20961 .byte 65,15,20,201 // unpcklps %xmm9,%xmm1
20962 .byte 68,15,21,192 // unpckhps %xmm0,%xmm8
20963 .byte 65,15,21,217 // unpckhps %xmm9,%xmm3
20964 .byte 15,40,194 // movaps %xmm2,%xmm0
20965 .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0
20966 .byte 15,18,202 // movhlps %xmm2,%xmm1
20967 .byte 65,15,40,208 // movaps %xmm8,%xmm2
20968 .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2
20969 .byte 65,15,18,216 // movhlps %xmm8,%xmm3
20970 .byte 72,173 // lods %ds:(%rsi),%rax
20971 .byte 255,224 // jmpq *%rax
20972
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]20973HIDDEN _sk_store_f32_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20974.globl _sk_store_f32_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]20975FUNCTION(_sk_store_f32_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]20976_sk_store_f32_sse41:
20977 .byte 72,173 // lods %ds:(%rsi),%rax
20978 .byte 72,139,0 // mov (%rax),%rax
20979 .byte 72,137,249 // mov %rdi,%rcx
20980 .byte 72,193,225,4 // shl $0x4,%rcx
20981 .byte 68,15,40,192 // movaps %xmm0,%xmm8
20982 .byte 68,15,40,200 // movaps %xmm0,%xmm9
20983 .byte 68,15,20,201 // unpcklps %xmm1,%xmm9
20984 .byte 68,15,40,210 // movaps %xmm2,%xmm10
20985 .byte 68,15,40,218 // movaps %xmm2,%xmm11
20986 .byte 68,15,20,219 // unpcklps %xmm3,%xmm11
20987 .byte 68,15,21,193 // unpckhps %xmm1,%xmm8
20988 .byte 68,15,21,211 // unpckhps %xmm3,%xmm10
20989 .byte 69,15,40,225 // movaps %xmm9,%xmm12
20990 .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12
20991 .byte 69,15,18,217 // movhlps %xmm9,%xmm11
20992 .byte 69,15,40,200 // movaps %xmm8,%xmm9
20993 .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9
20994 .byte 69,15,18,208 // movhlps %xmm8,%xmm10
20995 .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1)
20996 .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1)
20997 .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1)
20998 .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1)
20999 .byte 72,173 // lods %ds:(%rsi),%rax
21000 .byte 255,224 // jmpq *%rax
21001
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21002HIDDEN _sk_clamp_x_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21003.globl _sk_clamp_x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21004FUNCTION(_sk_clamp_x_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21005_sk_clamp_x_sse41:
21006 .byte 72,173 // lods %ds:(%rsi),%rax
21007 .byte 69,15,87,192 // xorps %xmm8,%xmm8
21008 .byte 68,15,95,192 // maxps %xmm0,%xmm8
21009 .byte 243,68,15,16,8 // movss (%rax),%xmm9
21010 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21011 .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0
21012 .byte 102,65,15,254,193 // paddd %xmm9,%xmm0
21013 .byte 68,15,93,192 // minps %xmm0,%xmm8
21014 .byte 72,173 // lods %ds:(%rsi),%rax
21015 .byte 65,15,40,192 // movaps %xmm8,%xmm0
21016 .byte 255,224 // jmpq *%rax
21017
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21018HIDDEN _sk_clamp_y_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21019.globl _sk_clamp_y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21020FUNCTION(_sk_clamp_y_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21021_sk_clamp_y_sse41:
21022 .byte 72,173 // lods %ds:(%rsi),%rax
21023 .byte 69,15,87,192 // xorps %xmm8,%xmm8
21024 .byte 68,15,95,193 // maxps %xmm1,%xmm8
21025 .byte 243,68,15,16,8 // movss (%rax),%xmm9
21026 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21027 .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1
21028 .byte 102,65,15,254,201 // paddd %xmm9,%xmm1
21029 .byte 68,15,93,193 // minps %xmm1,%xmm8
21030 .byte 72,173 // lods %ds:(%rsi),%rax
21031 .byte 65,15,40,200 // movaps %xmm8,%xmm1
21032 .byte 255,224 // jmpq *%rax
21033
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21034HIDDEN _sk_repeat_x_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21035.globl _sk_repeat_x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21036FUNCTION(_sk_repeat_x_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21037_sk_repeat_x_sse41:
21038 .byte 72,173 // lods %ds:(%rsi),%rax
21039 .byte 243,68,15,16,0 // movss (%rax),%xmm8
21040 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21041 .byte 68,15,40,200 // movaps %xmm0,%xmm9
21042 .byte 69,15,94,200 // divps %xmm8,%xmm9
21043 .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9
21044 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21045 .byte 65,15,92,193 // subps %xmm9,%xmm0
21046 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9
21047 .byte 102,69,15,254,200 // paddd %xmm8,%xmm9
21048 .byte 65,15,93,193 // minps %xmm9,%xmm0
21049 .byte 72,173 // lods %ds:(%rsi),%rax
21050 .byte 255,224 // jmpq *%rax
21051
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21052HIDDEN _sk_repeat_y_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21053.globl _sk_repeat_y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21054FUNCTION(_sk_repeat_y_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21055_sk_repeat_y_sse41:
21056 .byte 72,173 // lods %ds:(%rsi),%rax
21057 .byte 243,68,15,16,0 // movss (%rax),%xmm8
21058 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21059 .byte 68,15,40,201 // movaps %xmm1,%xmm9
21060 .byte 69,15,94,200 // divps %xmm8,%xmm9
21061 .byte 102,69,15,58,8,201,1 // roundps $0x1,%xmm9,%xmm9
21062 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21063 .byte 65,15,92,201 // subps %xmm9,%xmm1
21064 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9
21065 .byte 102,69,15,254,200 // paddd %xmm8,%xmm9
21066 .byte 65,15,93,201 // minps %xmm9,%xmm1
21067 .byte 72,173 // lods %ds:(%rsi),%rax
21068 .byte 255,224 // jmpq *%rax
21069
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21070HIDDEN _sk_mirror_x_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21071.globl _sk_mirror_x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21072FUNCTION(_sk_mirror_x_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21073_sk_mirror_x_sse41:
21074 .byte 72,173 // lods %ds:(%rsi),%rax
21075 .byte 243,68,15,16,0 // movss (%rax),%xmm8
21076 .byte 69,15,40,200 // movaps %xmm8,%xmm9
21077 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21078 .byte 65,15,92,193 // subps %xmm9,%xmm0
21079 .byte 243,69,15,88,192 // addss %xmm8,%xmm8
21080 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21081 .byte 68,15,40,208 // movaps %xmm0,%xmm10
21082 .byte 69,15,94,208 // divps %xmm8,%xmm10
21083 .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10
21084 .byte 69,15,89,208 // mulps %xmm8,%xmm10
21085 .byte 65,15,92,194 // subps %xmm10,%xmm0
21086 .byte 65,15,92,193 // subps %xmm9,%xmm0
21087 .byte 69,15,87,192 // xorps %xmm8,%xmm8
21088 .byte 68,15,92,192 // subps %xmm0,%xmm8
21089 .byte 65,15,84,192 // andps %xmm8,%xmm0
21090 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8
21091 .byte 102,69,15,254,193 // paddd %xmm9,%xmm8
21092 .byte 65,15,93,192 // minps %xmm8,%xmm0
21093 .byte 72,173 // lods %ds:(%rsi),%rax
21094 .byte 255,224 // jmpq *%rax
21095
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21096HIDDEN _sk_mirror_y_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21097.globl _sk_mirror_y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21098FUNCTION(_sk_mirror_y_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21099_sk_mirror_y_sse41:
21100 .byte 72,173 // lods %ds:(%rsi),%rax
21101 .byte 243,68,15,16,0 // movss (%rax),%xmm8
21102 .byte 69,15,40,200 // movaps %xmm8,%xmm9
21103 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21104 .byte 65,15,92,201 // subps %xmm9,%xmm1
21105 .byte 243,69,15,88,192 // addss %xmm8,%xmm8
21106 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21107 .byte 68,15,40,209 // movaps %xmm1,%xmm10
21108 .byte 69,15,94,208 // divps %xmm8,%xmm10
21109 .byte 102,69,15,58,8,210,1 // roundps $0x1,%xmm10,%xmm10
21110 .byte 69,15,89,208 // mulps %xmm8,%xmm10
21111 .byte 65,15,92,202 // subps %xmm10,%xmm1
21112 .byte 65,15,92,201 // subps %xmm9,%xmm1
21113 .byte 69,15,87,192 // xorps %xmm8,%xmm8
21114 .byte 68,15,92,193 // subps %xmm1,%xmm8
21115 .byte 65,15,84,200 // andps %xmm8,%xmm1
21116 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8
21117 .byte 102,69,15,254,193 // paddd %xmm9,%xmm8
21118 .byte 65,15,93,200 // minps %xmm8,%xmm1
21119 .byte 72,173 // lods %ds:(%rsi),%rax
21120 .byte 255,224 // jmpq *%rax
21121
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21122HIDDEN _sk_luminance_to_alpha_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21123.globl _sk_luminance_to_alpha_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21124FUNCTION(_sk_luminance_to_alpha_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21125_sk_luminance_to_alpha_sse41:
21126 .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax
21127 .byte 102,15,110,216 // movd %eax,%xmm3
21128 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
21129 .byte 15,89,216 // mulps %xmm0,%xmm3
21130 .byte 184,89,23,55,63 // mov $0x3f371759,%eax
21131 .byte 102,15,110,192 // movd %eax,%xmm0
21132 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
21133 .byte 15,89,193 // mulps %xmm1,%xmm0
21134 .byte 15,88,195 // addps %xmm3,%xmm0
21135 .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax
21136 .byte 102,15,110,216 // movd %eax,%xmm3
21137 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
21138 .byte 15,89,218 // mulps %xmm2,%xmm3
21139 .byte 15,88,216 // addps %xmm0,%xmm3
21140 .byte 72,173 // lods %ds:(%rsi),%rax
21141 .byte 15,87,192 // xorps %xmm0,%xmm0
21142 .byte 15,87,201 // xorps %xmm1,%xmm1
21143 .byte 15,87,210 // xorps %xmm2,%xmm2
21144 .byte 255,224 // jmpq *%rax
21145
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21146HIDDEN _sk_matrix_2x3_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21147.globl _sk_matrix_2x3_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21148FUNCTION(_sk_matrix_2x3_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21149_sk_matrix_2x3_sse41:
21150 .byte 68,15,40,201 // movaps %xmm1,%xmm9
21151 .byte 68,15,40,192 // movaps %xmm0,%xmm8
21152 .byte 72,173 // lods %ds:(%rsi),%rax
21153 .byte 243,15,16,0 // movss (%rax),%xmm0
21154 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
21155 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
21156 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
21157 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21158 .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11
21159 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21160 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21161 .byte 69,15,88,211 // addps %xmm11,%xmm10
21162 .byte 65,15,89,192 // mulps %xmm8,%xmm0
21163 .byte 65,15,88,194 // addps %xmm10,%xmm0
21164 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
21165 .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10
21166 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21167 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
21168 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21169 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21170 .byte 69,15,88,211 // addps %xmm11,%xmm10
21171 .byte 65,15,89,200 // mulps %xmm8,%xmm1
21172 .byte 65,15,88,202 // addps %xmm10,%xmm1
21173 .byte 72,173 // lods %ds:(%rsi),%rax
21174 .byte 255,224 // jmpq *%rax
21175
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21176HIDDEN _sk_matrix_3x4_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21177.globl _sk_matrix_3x4_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21178FUNCTION(_sk_matrix_3x4_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21179_sk_matrix_3x4_sse41:
21180 .byte 68,15,40,201 // movaps %xmm1,%xmm9
21181 .byte 68,15,40,192 // movaps %xmm0,%xmm8
21182 .byte 72,173 // lods %ds:(%rsi),%rax
21183 .byte 243,15,16,0 // movss (%rax),%xmm0
21184 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
21185 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
21186 .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10
21187 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21188 .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11
21189 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21190 .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12
21191 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21192 .byte 68,15,89,218 // mulps %xmm2,%xmm11
21193 .byte 69,15,88,220 // addps %xmm12,%xmm11
21194 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21195 .byte 69,15,88,211 // addps %xmm11,%xmm10
21196 .byte 65,15,89,192 // mulps %xmm8,%xmm0
21197 .byte 65,15,88,194 // addps %xmm10,%xmm0
21198 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
21199 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
21200 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21201 .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11
21202 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21203 .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12
21204 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21205 .byte 68,15,89,218 // mulps %xmm2,%xmm11
21206 .byte 69,15,88,220 // addps %xmm12,%xmm11
21207 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21208 .byte 69,15,88,211 // addps %xmm11,%xmm10
21209 .byte 65,15,89,200 // mulps %xmm8,%xmm1
21210 .byte 65,15,88,202 // addps %xmm10,%xmm1
21211 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
21212 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21213 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
21214 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21215 .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12
21216 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21217 .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13
21218 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21219 .byte 68,15,89,226 // mulps %xmm2,%xmm12
21220 .byte 69,15,88,229 // addps %xmm13,%xmm12
21221 .byte 69,15,89,217 // mulps %xmm9,%xmm11
21222 .byte 69,15,88,220 // addps %xmm12,%xmm11
21223 .byte 69,15,89,208 // mulps %xmm8,%xmm10
21224 .byte 69,15,88,211 // addps %xmm11,%xmm10
21225 .byte 72,173 // lods %ds:(%rsi),%rax
21226 .byte 65,15,40,210 // movaps %xmm10,%xmm2
21227 .byte 255,224 // jmpq *%rax
21228
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21229HIDDEN _sk_matrix_4x5_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21230.globl _sk_matrix_4x5_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21231FUNCTION(_sk_matrix_4x5_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21232_sk_matrix_4x5_sse41:
21233 .byte 68,15,40,201 // movaps %xmm1,%xmm9
21234 .byte 68,15,40,192 // movaps %xmm0,%xmm8
21235 .byte 72,173 // lods %ds:(%rsi),%rax
21236 .byte 243,15,16,0 // movss (%rax),%xmm0
21237 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
21238 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
21239 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
21240 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21241 .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11
21242 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21243 .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12
21244 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21245 .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13
21246 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21247 .byte 68,15,89,227 // mulps %xmm3,%xmm12
21248 .byte 69,15,88,229 // addps %xmm13,%xmm12
21249 .byte 68,15,89,218 // mulps %xmm2,%xmm11
21250 .byte 69,15,88,220 // addps %xmm12,%xmm11
21251 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21252 .byte 69,15,88,211 // addps %xmm11,%xmm10
21253 .byte 65,15,89,192 // mulps %xmm8,%xmm0
21254 .byte 65,15,88,194 // addps %xmm10,%xmm0
21255 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
21256 .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10
21257 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21258 .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11
21259 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21260 .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12
21261 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21262 .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13
21263 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21264 .byte 68,15,89,227 // mulps %xmm3,%xmm12
21265 .byte 69,15,88,229 // addps %xmm13,%xmm12
21266 .byte 68,15,89,218 // mulps %xmm2,%xmm11
21267 .byte 69,15,88,220 // addps %xmm12,%xmm11
21268 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21269 .byte 69,15,88,211 // addps %xmm11,%xmm10
21270 .byte 65,15,89,200 // mulps %xmm8,%xmm1
21271 .byte 65,15,88,202 // addps %xmm10,%xmm1
21272 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
21273 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21274 .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11
21275 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21276 .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12
21277 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21278 .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13
21279 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21280 .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14
21281 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
21282 .byte 68,15,89,235 // mulps %xmm3,%xmm13
21283 .byte 69,15,88,238 // addps %xmm14,%xmm13
21284 .byte 68,15,89,226 // mulps %xmm2,%xmm12
21285 .byte 69,15,88,229 // addps %xmm13,%xmm12
21286 .byte 69,15,89,217 // mulps %xmm9,%xmm11
21287 .byte 69,15,88,220 // addps %xmm12,%xmm11
21288 .byte 69,15,89,208 // mulps %xmm8,%xmm10
21289 .byte 69,15,88,211 // addps %xmm11,%xmm10
21290 .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11
21291 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21292 .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12
21293 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21294 .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13
21295 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21296 .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14
21297 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
21298 .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15
21299 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15
21300 .byte 68,15,89,243 // mulps %xmm3,%xmm14
21301 .byte 69,15,88,247 // addps %xmm15,%xmm14
21302 .byte 68,15,89,234 // mulps %xmm2,%xmm13
21303 .byte 69,15,88,238 // addps %xmm14,%xmm13
21304 .byte 69,15,89,225 // mulps %xmm9,%xmm12
21305 .byte 69,15,88,229 // addps %xmm13,%xmm12
21306 .byte 69,15,89,216 // mulps %xmm8,%xmm11
21307 .byte 69,15,88,220 // addps %xmm12,%xmm11
21308 .byte 72,173 // lods %ds:(%rsi),%rax
21309 .byte 65,15,40,210 // movaps %xmm10,%xmm2
21310 .byte 65,15,40,219 // movaps %xmm11,%xmm3
21311 .byte 255,224 // jmpq *%rax
21312
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21313HIDDEN _sk_matrix_perspective_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21314.globl _sk_matrix_perspective_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21315FUNCTION(_sk_matrix_perspective_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21316_sk_matrix_perspective_sse41:
21317 .byte 68,15,40,192 // movaps %xmm0,%xmm8
21318 .byte 72,173 // lods %ds:(%rsi),%rax
21319 .byte 243,15,16,0 // movss (%rax),%xmm0
21320 .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9
21321 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
21322 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21323 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
21324 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21325 .byte 68,15,89,201 // mulps %xmm1,%xmm9
21326 .byte 69,15,88,202 // addps %xmm10,%xmm9
21327 .byte 65,15,89,192 // mulps %xmm8,%xmm0
21328 .byte 65,15,88,193 // addps %xmm9,%xmm0
21329 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
21330 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21331 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
21332 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21333 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
21334 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21335 .byte 68,15,89,209 // mulps %xmm1,%xmm10
21336 .byte 69,15,88,211 // addps %xmm11,%xmm10
21337 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21338 .byte 69,15,88,202 // addps %xmm10,%xmm9
21339 .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10
21340 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21341 .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11
21342 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21343 .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12
21344 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21345 .byte 68,15,89,217 // mulps %xmm1,%xmm11
21346 .byte 69,15,88,220 // addps %xmm12,%xmm11
21347 .byte 69,15,89,208 // mulps %xmm8,%xmm10
21348 .byte 69,15,88,211 // addps %xmm11,%xmm10
21349 .byte 65,15,83,202 // rcpps %xmm10,%xmm1
21350 .byte 15,89,193 // mulps %xmm1,%xmm0
21351 .byte 68,15,89,201 // mulps %xmm1,%xmm9
21352 .byte 72,173 // lods %ds:(%rsi),%rax
21353 .byte 65,15,40,201 // movaps %xmm9,%xmm1
21354 .byte 255,224 // jmpq *%rax
21355
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21356HIDDEN _sk_linear_gradient_sse41
21357.globl _sk_linear_gradient_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21358FUNCTION(_sk_linear_gradient_sse41)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21359_sk_linear_gradient_sse41:
21360 .byte 72,173 // lods %ds:(%rsi),%rax
21361 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
21362 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21363 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
21364 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21365 .byte 243,68,15,16,96,24 // movss 0x18(%rax),%xmm12
21366 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21367 .byte 243,68,15,16,104,28 // movss 0x1c(%rax),%xmm13
21368 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21369 .byte 72,139,8 // mov (%rax),%rcx
21370 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]21371 .byte 15,132,254,0,0,0 // je 38e2 <_sk_linear_gradient_sse41+0x138>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21372 .byte 15,41,100,36,168 // movaps %xmm4,-0x58(%rsp)
21373 .byte 15,41,108,36,184 // movaps %xmm5,-0x48(%rsp)
21374 .byte 15,41,116,36,200 // movaps %xmm6,-0x38(%rsp)
21375 .byte 15,41,124,36,216 // movaps %xmm7,-0x28(%rsp)
21376 .byte 72,139,64,8 // mov 0x8(%rax),%rax
21377 .byte 72,131,192,32 // add $0x20,%rax
21378 .byte 69,15,87,201 // xorps %xmm9,%xmm9
21379 .byte 15,87,219 // xorps %xmm3,%xmm3
21380 .byte 15,87,210 // xorps %xmm2,%xmm2
21381 .byte 15,87,201 // xorps %xmm1,%xmm1
21382 .byte 15,40,233 // movaps %xmm1,%xmm5
21383 .byte 15,40,242 // movaps %xmm2,%xmm6
21384 .byte 15,40,251 // movaps %xmm3,%xmm7
21385 .byte 69,15,40,194 // movaps %xmm10,%xmm8
21386 .byte 69,15,40,243 // movaps %xmm11,%xmm14
21387 .byte 69,15,40,252 // movaps %xmm12,%xmm15
21388 .byte 68,15,41,108,36,232 // movaps %xmm13,-0x18(%rsp)
21389 .byte 65,15,40,201 // movaps %xmm9,%xmm1
21390 .byte 243,15,16,80,224 // movss -0x20(%rax),%xmm2
21391 .byte 243,68,15,16,72,228 // movss -0x1c(%rax),%xmm9
21392 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
21393 .byte 15,40,224 // movaps %xmm0,%xmm4
21394 .byte 15,194,194,1 // cmpltps %xmm2,%xmm0
21395 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21396 .byte 102,68,15,56,20,201 // blendvps %xmm0,%xmm1,%xmm9
21397 .byte 243,15,16,72,232 // movss -0x18(%rax),%xmm1
21398 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
21399 .byte 102,15,56,20,205 // blendvps %xmm0,%xmm5,%xmm1
21400 .byte 243,15,16,80,236 // movss -0x14(%rax),%xmm2
21401 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
21402 .byte 102,15,56,20,214 // blendvps %xmm0,%xmm6,%xmm2
21403 .byte 243,15,16,88,240 // movss -0x10(%rax),%xmm3
21404 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
21405 .byte 102,15,56,20,223 // blendvps %xmm0,%xmm7,%xmm3
21406 .byte 243,68,15,16,80,244 // movss -0xc(%rax),%xmm10
21407 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21408 .byte 102,69,15,56,20,208 // blendvps %xmm0,%xmm8,%xmm10
21409 .byte 243,68,15,16,88,248 // movss -0x8(%rax),%xmm11
21410 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21411 .byte 102,69,15,56,20,222 // blendvps %xmm0,%xmm14,%xmm11
21412 .byte 243,68,15,16,96,252 // movss -0x4(%rax),%xmm12
21413 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
21414 .byte 102,69,15,56,20,231 // blendvps %xmm0,%xmm15,%xmm12
21415 .byte 243,68,15,16,40 // movss (%rax),%xmm13
21416 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
21417 .byte 102,68,15,56,20,108,36,232 // blendvps %xmm0,-0x18(%rsp),%xmm13
21418 .byte 15,40,196 // movaps %xmm4,%xmm0
21419 .byte 72,131,192,36 // add $0x24,%rax
21420 .byte 72,255,201 // dec %rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]21421 .byte 15,133,65,255,255,255 // jne 380d <_sk_linear_gradient_sse41+0x63>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21422 .byte 15,40,124,36,216 // movaps -0x28(%rsp),%xmm7
21423 .byte 15,40,116,36,200 // movaps -0x38(%rsp),%xmm6
21424 .byte 15,40,108,36,184 // movaps -0x48(%rsp),%xmm5
21425 .byte 15,40,100,36,168 // movaps -0x58(%rsp),%xmm4
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]21426 .byte 235,13 // jmp 38ef <_sk_linear_gradient_sse41+0x145>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21427 .byte 15,87,201 // xorps %xmm1,%xmm1
21428 .byte 15,87,210 // xorps %xmm2,%xmm2
21429 .byte 15,87,219 // xorps %xmm3,%xmm3
21430 .byte 69,15,87,201 // xorps %xmm9,%xmm9
21431 .byte 68,15,89,200 // mulps %xmm0,%xmm9
21432 .byte 69,15,88,202 // addps %xmm10,%xmm9
21433 .byte 15,89,200 // mulps %xmm0,%xmm1
21434 .byte 65,15,88,203 // addps %xmm11,%xmm1
21435 .byte 15,89,208 // mulps %xmm0,%xmm2
21436 .byte 65,15,88,212 // addps %xmm12,%xmm2
21437 .byte 15,89,216 // mulps %xmm0,%xmm3
21438 .byte 65,15,88,221 // addps %xmm13,%xmm3
21439 .byte 72,173 // lods %ds:(%rsi),%rax
21440 .byte 65,15,40,193 // movaps %xmm9,%xmm0
21441 .byte 255,224 // jmpq *%rax
21442
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]21443HIDDEN _sk_linear_gradient_2stops_sse41
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21444.globl _sk_linear_gradient_2stops_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21445FUNCTION(_sk_linear_gradient_2stops_sse41)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21446_sk_linear_gradient_2stops_sse41:
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21447 .byte 68,15,40,192 // movaps %xmm0,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21448 .byte 72,173 // lods %ds:(%rsi),%rax
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21449 .byte 243,15,16,0 // movss (%rax),%xmm0
21450 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21451 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21452 .byte 243,15,16,80,16 // movss 0x10(%rax),%xmm2
21453 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21454 .byte 65,15,89,192 // mulps %xmm8,%xmm0
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21455 .byte 15,88,194 // addps %xmm2,%xmm0
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21456 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21457 .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21458 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
21459 .byte 65,15,89,200 // mulps %xmm8,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21460 .byte 15,88,202 // addps %xmm2,%xmm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21461 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21462 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21463 .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21464 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
21465 .byte 65,15,89,208 // mulps %xmm8,%xmm2
21466 .byte 15,88,211 // addps %xmm3,%xmm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21467 .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21468 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]21469 .byte 243,68,15,16,72,28 // movss 0x1c(%rax),%xmm9
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]21470 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21471 .byte 65,15,89,216 // mulps %xmm8,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21472 .byte 65,15,88,217 // addps %xmm9,%xmm3
21473 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]21474 .byte 255,224 // jmpq *%rax
21475
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21476HIDDEN _sk_save_xy_sse41
21477.globl _sk_save_xy_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21478FUNCTION(_sk_save_xy_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21479_sk_save_xy_sse41:
21480 .byte 72,173 // lods %ds:(%rsi),%rax
21481 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21482 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21483 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21484 .byte 69,15,40,200 // movaps %xmm8,%xmm9
21485 .byte 68,15,88,200 // addps %xmm0,%xmm9
21486 .byte 102,69,15,58,8,209,1 // roundps $0x1,%xmm9,%xmm10
21487 .byte 69,15,92,202 // subps %xmm10,%xmm9
21488 .byte 68,15,88,193 // addps %xmm1,%xmm8
21489 .byte 102,69,15,58,8,208,1 // roundps $0x1,%xmm8,%xmm10
21490 .byte 69,15,92,194 // subps %xmm10,%xmm8
21491 .byte 15,17,0 // movups %xmm0,(%rax)
21492 .byte 15,17,72,32 // movups %xmm1,0x20(%rax)
21493 .byte 68,15,17,72,64 // movups %xmm9,0x40(%rax)
21494 .byte 68,15,17,64,96 // movups %xmm8,0x60(%rax)
21495 .byte 72,173 // lods %ds:(%rsi),%rax
21496 .byte 255,224 // jmpq *%rax
21497
21498HIDDEN _sk_accumulate_sse41
21499.globl _sk_accumulate_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21500FUNCTION(_sk_accumulate_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21501_sk_accumulate_sse41:
21502 .byte 72,173 // lods %ds:(%rsi),%rax
21503 .byte 68,15,16,128,128,0,0,0 // movups 0x80(%rax),%xmm8
21504 .byte 68,15,16,136,160,0,0,0 // movups 0xa0(%rax),%xmm9
21505 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21506 .byte 69,15,40,193 // movaps %xmm9,%xmm8
21507 .byte 68,15,89,192 // mulps %xmm0,%xmm8
21508 .byte 65,15,88,224 // addps %xmm8,%xmm4
21509 .byte 69,15,40,193 // movaps %xmm9,%xmm8
21510 .byte 68,15,89,193 // mulps %xmm1,%xmm8
21511 .byte 65,15,88,232 // addps %xmm8,%xmm5
21512 .byte 69,15,40,193 // movaps %xmm9,%xmm8
21513 .byte 68,15,89,194 // mulps %xmm2,%xmm8
21514 .byte 65,15,88,240 // addps %xmm8,%xmm6
21515 .byte 68,15,89,203 // mulps %xmm3,%xmm9
21516 .byte 65,15,88,249 // addps %xmm9,%xmm7
21517 .byte 72,173 // lods %ds:(%rsi),%rax
21518 .byte 255,224 // jmpq *%rax
21519
21520HIDDEN _sk_bilinear_nx_sse41
21521.globl _sk_bilinear_nx_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21522FUNCTION(_sk_bilinear_nx_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21523_sk_bilinear_nx_sse41:
21524 .byte 72,173 // lods %ds:(%rsi),%rax
21525 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
21526 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21527 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21528 .byte 15,16,0 // movups (%rax),%xmm0
21529 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21530 .byte 65,15,88,192 // addps %xmm8,%xmm0
21531 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21532 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21533 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21534 .byte 69,15,92,193 // subps %xmm9,%xmm8
21535 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax)
21536 .byte 72,173 // lods %ds:(%rsi),%rax
21537 .byte 255,224 // jmpq *%rax
21538
21539HIDDEN _sk_bilinear_px_sse41
21540.globl _sk_bilinear_px_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21541FUNCTION(_sk_bilinear_px_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21542_sk_bilinear_px_sse41:
21543 .byte 72,173 // lods %ds:(%rsi),%rax
21544 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21545 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21546 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21547 .byte 15,16,0 // movups (%rax),%xmm0
21548 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21549 .byte 65,15,88,192 // addps %xmm8,%xmm0
21550 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
21551 .byte 72,173 // lods %ds:(%rsi),%rax
21552 .byte 255,224 // jmpq *%rax
21553
21554HIDDEN _sk_bilinear_ny_sse41
21555.globl _sk_bilinear_ny_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21556FUNCTION(_sk_bilinear_ny_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21557_sk_bilinear_ny_sse41:
21558 .byte 72,173 // lods %ds:(%rsi),%rax
21559 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
21560 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21561 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21562 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21563 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21564 .byte 65,15,88,200 // addps %xmm8,%xmm1
21565 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21566 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21567 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21568 .byte 69,15,92,193 // subps %xmm9,%xmm8
21569 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax)
21570 .byte 72,173 // lods %ds:(%rsi),%rax
21571 .byte 255,224 // jmpq *%rax
21572
21573HIDDEN _sk_bilinear_py_sse41
21574.globl _sk_bilinear_py_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21575FUNCTION(_sk_bilinear_py_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21576_sk_bilinear_py_sse41:
21577 .byte 72,173 // lods %ds:(%rsi),%rax
21578 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21579 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21580 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21581 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21582 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21583 .byte 65,15,88,200 // addps %xmm8,%xmm1
21584 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
21585 .byte 72,173 // lods %ds:(%rsi),%rax
21586 .byte 255,224 // jmpq *%rax
21587
21588HIDDEN _sk_bicubic_n3x_sse41
21589.globl _sk_bicubic_n3x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21590FUNCTION(_sk_bicubic_n3x_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21591_sk_bicubic_n3x_sse41:
21592 .byte 72,173 // lods %ds:(%rsi),%rax
21593 .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx
21594 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21595 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21596 .byte 15,16,0 // movups (%rax),%xmm0
21597 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21598 .byte 65,15,88,192 // addps %xmm8,%xmm0
21599 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21600 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21601 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21602 .byte 69,15,92,193 // subps %xmm9,%xmm8
21603 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
21604 .byte 102,68,15,110,201 // movd %ecx,%xmm9
21605 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21606 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
21607 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21608 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21609 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21610 .byte 69,15,89,192 // mulps %xmm8,%xmm8
21611 .byte 69,15,88,202 // addps %xmm10,%xmm9
21612 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21613 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
21614 .byte 72,173 // lods %ds:(%rsi),%rax
21615 .byte 255,224 // jmpq *%rax
21616
21617HIDDEN _sk_bicubic_n1x_sse41
21618.globl _sk_bicubic_n1x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21619FUNCTION(_sk_bicubic_n1x_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21620_sk_bicubic_n1x_sse41:
21621 .byte 72,173 // lods %ds:(%rsi),%rax
21622 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
21623 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21624 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21625 .byte 15,16,0 // movups (%rax),%xmm0
21626 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21627 .byte 65,15,88,192 // addps %xmm8,%xmm0
21628 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21629 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21630 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21631 .byte 69,15,92,193 // subps %xmm9,%xmm8
21632 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
21633 .byte 102,68,15,110,201 // movd %ecx,%xmm9
21634 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21635 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21636 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21637 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21638 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21639 .byte 69,15,88,202 // addps %xmm10,%xmm9
21640 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21641 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21642 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21643 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21644 .byte 69,15,88,202 // addps %xmm10,%xmm9
21645 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
21646 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21647 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21648 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21649 .byte 69,15,88,202 // addps %xmm10,%xmm9
21650 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
21651 .byte 72,173 // lods %ds:(%rsi),%rax
21652 .byte 255,224 // jmpq *%rax
21653
21654HIDDEN _sk_bicubic_p1x_sse41
21655.globl _sk_bicubic_p1x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21656FUNCTION(_sk_bicubic_p1x_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21657_sk_bicubic_p1x_sse41:
21658 .byte 72,173 // lods %ds:(%rsi),%rax
21659 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21660 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21661 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21662 .byte 15,16,0 // movups (%rax),%xmm0
21663 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21664 .byte 65,15,88,192 // addps %xmm8,%xmm0
21665 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
21666 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21667 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21668 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21669 .byte 102,68,15,110,217 // movd %ecx,%xmm11
21670 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21671 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21672 .byte 69,15,88,211 // addps %xmm11,%xmm10
21673 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21674 .byte 69,15,88,208 // addps %xmm8,%xmm10
21675 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
21676 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21677 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21678 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21679 .byte 69,15,88,208 // addps %xmm8,%xmm10
21680 .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax)
21681 .byte 72,173 // lods %ds:(%rsi),%rax
21682 .byte 255,224 // jmpq *%rax
21683
21684HIDDEN _sk_bicubic_p3x_sse41
21685.globl _sk_bicubic_p3x_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21686FUNCTION(_sk_bicubic_p3x_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21687_sk_bicubic_p3x_sse41:
21688 .byte 72,173 // lods %ds:(%rsi),%rax
21689 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21690 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21691 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21692 .byte 15,16,0 // movups (%rax),%xmm0
21693 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
21694 .byte 65,15,88,192 // addps %xmm8,%xmm0
21695 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
21696 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21697 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21698 .byte 69,15,89,193 // mulps %xmm9,%xmm8
21699 .byte 69,15,89,201 // mulps %xmm9,%xmm9
21700 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
21701 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21702 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21703 .byte 69,15,88,194 // addps %xmm10,%xmm8
21704 .byte 69,15,89,193 // mulps %xmm9,%xmm8
21705 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax)
21706 .byte 72,173 // lods %ds:(%rsi),%rax
21707 .byte 255,224 // jmpq *%rax
21708
21709HIDDEN _sk_bicubic_n3y_sse41
21710.globl _sk_bicubic_n3y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21711FUNCTION(_sk_bicubic_n3y_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21712_sk_bicubic_n3y_sse41:
21713 .byte 72,173 // lods %ds:(%rsi),%rax
21714 .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx
21715 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21716 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21717 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21718 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21719 .byte 65,15,88,200 // addps %xmm8,%xmm1
21720 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21721 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21722 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21723 .byte 69,15,92,193 // subps %xmm9,%xmm8
21724 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
21725 .byte 102,68,15,110,201 // movd %ecx,%xmm9
21726 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21727 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
21728 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21729 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21730 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21731 .byte 69,15,89,192 // mulps %xmm8,%xmm8
21732 .byte 69,15,88,202 // addps %xmm10,%xmm9
21733 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21734 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
21735 .byte 72,173 // lods %ds:(%rsi),%rax
21736 .byte 255,224 // jmpq *%rax
21737
21738HIDDEN _sk_bicubic_n1y_sse41
21739.globl _sk_bicubic_n1y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21740FUNCTION(_sk_bicubic_n1y_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21741_sk_bicubic_n1y_sse41:
21742 .byte 72,173 // lods %ds:(%rsi),%rax
21743 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
21744 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21745 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21746 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21747 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21748 .byte 65,15,88,200 // addps %xmm8,%xmm1
21749 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
21750 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21751 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21752 .byte 69,15,92,193 // subps %xmm9,%xmm8
21753 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
21754 .byte 102,68,15,110,201 // movd %ecx,%xmm9
21755 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
21756 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21757 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21758 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21759 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21760 .byte 69,15,88,202 // addps %xmm10,%xmm9
21761 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21762 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21763 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21764 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21765 .byte 69,15,88,202 // addps %xmm10,%xmm9
21766 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
21767 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21768 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21769 .byte 69,15,89,200 // mulps %xmm8,%xmm9
21770 .byte 69,15,88,202 // addps %xmm10,%xmm9
21771 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
21772 .byte 72,173 // lods %ds:(%rsi),%rax
21773 .byte 255,224 // jmpq *%rax
21774
21775HIDDEN _sk_bicubic_p1y_sse41
21776.globl _sk_bicubic_p1y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21777FUNCTION(_sk_bicubic_p1y_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21778_sk_bicubic_p1y_sse41:
21779 .byte 72,173 // lods %ds:(%rsi),%rax
21780 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
21781 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21782 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21783 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21784 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21785 .byte 65,15,88,200 // addps %xmm8,%xmm1
21786 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
21787 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21788 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21789 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21790 .byte 102,68,15,110,217 // movd %ecx,%xmm11
21791 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
21792 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21793 .byte 69,15,88,211 // addps %xmm11,%xmm10
21794 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21795 .byte 69,15,88,208 // addps %xmm8,%xmm10
21796 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
21797 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21798 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21799 .byte 69,15,89,209 // mulps %xmm9,%xmm10
21800 .byte 69,15,88,208 // addps %xmm8,%xmm10
21801 .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax)
21802 .byte 72,173 // lods %ds:(%rsi),%rax
21803 .byte 255,224 // jmpq *%rax
21804
21805HIDDEN _sk_bicubic_p3y_sse41
21806.globl _sk_bicubic_p3y_sse41
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]21807FUNCTION(_sk_bicubic_p3y_sse41)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]21808_sk_bicubic_p3y_sse41:
21809 .byte 72,173 // lods %ds:(%rsi),%rax
21810 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
21811 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21812 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21813 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
21814 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
21815 .byte 65,15,88,200 // addps %xmm8,%xmm1
21816 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
21817 .byte 102,68,15,110,193 // movd %ecx,%xmm8
21818 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
21819 .byte 69,15,89,193 // mulps %xmm9,%xmm8
21820 .byte 69,15,89,201 // mulps %xmm9,%xmm9
21821 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
21822 .byte 102,68,15,110,209 // movd %ecx,%xmm10
21823 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
21824 .byte 69,15,88,194 // addps %xmm10,%xmm8
21825 .byte 69,15,89,193 // mulps %xmm9,%xmm8
21826 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax)
21827 .byte 72,173 // lods %ds:(%rsi),%rax
21828 .byte 255,224 // jmpq *%rax
21829
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21830HIDDEN _sk_callback_sse41
21831.globl _sk_callback_sse41
21832FUNCTION(_sk_callback_sse41)
21833_sk_callback_sse41:
21834 .byte 65,87 // push %r15
21835 .byte 65,86 // push %r14
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21836 .byte 65,84 // push %r12
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21837 .byte 83 // push %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21838 .byte 72,131,236,72 // sub $0x48,%rsp
21839 .byte 15,41,124,36,48 // movaps %xmm7,0x30(%rsp)
21840 .byte 15,41,116,36,32 // movaps %xmm6,0x20(%rsp)
21841 .byte 15,41,108,36,16 // movaps %xmm5,0x10(%rsp)
21842 .byte 15,41,36,36 // movaps %xmm4,(%rsp)
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21843 .byte 73,137,214 // mov %rdx,%r14
21844 .byte 73,137,255 // mov %rdi,%r15
21845 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21846 .byte 72,137,195 // mov %rax,%rbx
21847 .byte 73,137,244 // mov %rsi,%r12
21848 .byte 15,40,224 // movaps %xmm0,%xmm4
21849 .byte 15,20,225 // unpcklps %xmm1,%xmm4
21850 .byte 15,40,234 // movaps %xmm2,%xmm5
21851 .byte 15,20,235 // unpcklps %xmm3,%xmm5
21852 .byte 15,21,193 // unpckhps %xmm1,%xmm0
21853 .byte 15,21,211 // unpckhps %xmm3,%xmm2
21854 .byte 15,40,204 // movaps %xmm4,%xmm1
21855 .byte 102,15,20,205 // unpcklpd %xmm5,%xmm1
21856 .byte 15,18,236 // movhlps %xmm4,%xmm5
21857 .byte 15,40,216 // movaps %xmm0,%xmm3
21858 .byte 102,15,20,218 // unpcklpd %xmm2,%xmm3
21859 .byte 15,18,208 // movhlps %xmm0,%xmm2
21860 .byte 102,15,17,75,8 // movupd %xmm1,0x8(%rbx)
21861 .byte 15,17,107,24 // movups %xmm5,0x18(%rbx)
21862 .byte 102,15,17,91,40 // movupd %xmm3,0x28(%rbx)
21863 .byte 15,17,83,56 // movups %xmm2,0x38(%rbx)
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21864 .byte 190,4,0,0,0 // mov $0x4,%esi
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21865 .byte 72,137,223 // mov %rbx,%rdi
21866 .byte 255,19 // callq *(%rbx)
21867 .byte 72,139,131,136,0,0,0 // mov 0x88(%rbx),%rax
21868 .byte 15,16,32 // movups (%rax),%xmm4
21869 .byte 15,16,64,16 // movups 0x10(%rax),%xmm0
21870 .byte 15,16,88,32 // movups 0x20(%rax),%xmm3
21871 .byte 15,16,80,48 // movups 0x30(%rax),%xmm2
21872 .byte 15,40,236 // movaps %xmm4,%xmm5
21873 .byte 15,20,232 // unpcklps %xmm0,%xmm5
21874 .byte 15,40,203 // movaps %xmm3,%xmm1
21875 .byte 15,20,202 // unpcklps %xmm2,%xmm1
21876 .byte 15,21,224 // unpckhps %xmm0,%xmm4
21877 .byte 15,21,218 // unpckhps %xmm2,%xmm3
21878 .byte 15,40,197 // movaps %xmm5,%xmm0
21879 .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0
21880 .byte 15,18,205 // movhlps %xmm5,%xmm1
21881 .byte 15,40,212 // movaps %xmm4,%xmm2
21882 .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2
21883 .byte 15,18,220 // movhlps %xmm4,%xmm3
21884 .byte 76,137,230 // mov %r12,%rsi
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21885 .byte 72,173 // lods %ds:(%rsi),%rax
21886 .byte 76,137,255 // mov %r15,%rdi
21887 .byte 76,137,242 // mov %r14,%rdx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21888 .byte 15,40,36,36 // movaps (%rsp),%xmm4
21889 .byte 15,40,108,36,16 // movaps 0x10(%rsp),%xmm5
21890 .byte 15,40,116,36,32 // movaps 0x20(%rsp),%xmm6
21891 .byte 15,40,124,36,48 // movaps 0x30(%rsp),%xmm7
21892 .byte 72,131,196,72 // add $0x48,%rsp
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21893 .byte 91 // pop %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]21894 .byte 65,92 // pop %r12
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]21895 .byte 65,94 // pop %r14
21896 .byte 65,95 // pop %r15
21897 .byte 255,224 // jmpq *%rax
21898
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]21899BALIGN16
21900 .byte 0,0 // add %al,(%rax)
21901 .byte 0,63 // add %bh,(%rdi)
21902 .byte 0,0 // add %al,(%rax)
21903 .byte 0,63 // add %bh,(%rdi)
21904 .byte 0,0 // add %al,(%rax)
21905 .byte 0,63 // add %bh,(%rdi)
21906 .byte 0,0 // add %al,(%rax)
21907 .byte 0,63 // add %bh,(%rdi)
21908 .byte 0,0 // add %al,(%rax)
21909 .byte 128,63,0 // cmpb $0x0,(%rdi)
21910 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
21911 .byte 63 // (bad)
21912 .byte 0,0 // add %al,(%rax)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]21913 .byte 128,63,0 // cmpb $0x0,(%rdi)
21914 .byte 0,0 // add %al,(%rax)
21915 .byte 52,0 // xor $0x0,%al
21916 .byte 0,0 // add %al,(%rax)
21917 .byte 52,0 // xor $0x0,%al
21918 .byte 0,0 // add %al,(%rax)
21919 .byte 52,0 // xor $0x0,%al
21920 .byte 0,0 // add %al,(%rax)
21921 .byte 52,255 // xor $0xff,%al
21922 .byte 255 // (bad)
21923 .byte 127,0 // jg 3f74 <.literal16+0x34>
21924 .byte 255 // (bad)
21925 .byte 255 // (bad)
21926 .byte 127,0 // jg 3f78 <.literal16+0x38>
21927 .byte 255 // (bad)
21928 .byte 255 // (bad)
21929 .byte 127,0 // jg 3f7c <.literal16+0x3c>
21930 .byte 255 // (bad)
21931 .byte 255 // (bad)
21932 .byte 127,0 // jg 3f80 <.literal16+0x40>
21933 .byte 0,0 // add %al,(%rax)
21934 .byte 0,63 // add %bh,(%rdi)
21935 .byte 0,0 // add %al,(%rax)
21936 .byte 0,63 // add %bh,(%rdi)
21937 .byte 0,0 // add %al,(%rax)
21938 .byte 0,63 // add %bh,(%rdi)
21939 .byte 0,0 // add %al,(%rax)
21940 .byte 0,63 // add %bh,(%rdi)
21941 .byte 119,115 // ja 4005 <.literal16+0xc5>
21942 .byte 248 // clc
21943 .byte 194,119,115 // retq $0x7377
21944 .byte 248 // clc
21945 .byte 194,119,115 // retq $0x7377
21946 .byte 248 // clc
21947 .byte 194,119,115 // retq $0x7377
21948 .byte 248 // clc
21949 .byte 194,117,191 // retq $0xbf75
21950 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]21951 .byte 63 // (bad)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]21952 .byte 117,191 // jne 3f69 <.literal16+0x29>
21953 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
21954 .byte 63 // (bad)
21955 .byte 249 // stc
21956 .byte 68,180,62 // rex.R mov $0x3e,%spl
21957 .byte 249 // stc
21958 .byte 68,180,62 // rex.R mov $0x3e,%spl
21959 .byte 249 // stc
21960 .byte 68,180,62 // rex.R mov $0x3e,%spl
21961 .byte 249 // stc
21962 .byte 68,180,62 // rex.R mov $0x3e,%spl
21963 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
21964 .byte 233,220,63,163,233 // jmpq ffffffffe9a37faa <_sk_callback_sse41+0xffffffffe9a3414a>
21965 .byte 220,63 // fdivrl (%rdi)
21966 .byte 81 // push %rcx
21967 .byte 140,242 // mov %?,%edx
21968 .byte 66,81 // rex.X push %rcx
21969 .byte 140,242 // mov %?,%edx
21970 .byte 66,81 // rex.X push %rcx
21971 .byte 140,242 // mov %?,%edx
21972 .byte 66,81 // rex.X push %rcx
21973 .byte 140,242 // mov %?,%edx
21974 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
21975 .byte 63 // (bad)
21976 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
21977 .byte 63 // (bad)
21978 .byte 248 // clc
21979 .byte 245 // cmc
21980 .byte 154 // (bad)
21981 .byte 64,248 // rex clc
21982 .byte 245 // cmc
21983 .byte 154 // (bad)
21984 .byte 64,248 // rex clc
21985 .byte 245 // cmc
21986 .byte 154 // (bad)
21987 .byte 64,248 // rex clc
21988 .byte 245 // cmc
21989 .byte 154 // (bad)
21990 .byte 64,254 // rex (bad)
21991 .byte 210,221 // rcr %cl,%ch
21992 .byte 65,254 // rex.B (bad)
21993 .byte 210,221 // rcr %cl,%ch
21994 .byte 65,254 // rex.B (bad)
21995 .byte 210,221 // rcr %cl,%ch
21996 .byte 65,254 // rex.B (bad)
21997 .byte 210,221 // rcr %cl,%ch
21998 .byte 65,0,0 // add %al,(%r8)
21999 .byte 0,75,0 // add %cl,0x0(%rbx)
22000 .byte 0,0 // add %al,(%rax)
22001 .byte 75,0,0 // rex.WXB add %al,(%r8)
22002 .byte 0,75,0 // add %cl,0x0(%rbx)
22003 .byte 0,0 // add %al,(%rax)
22004 .byte 75,0,0 // rex.WXB add %al,(%r8)
22005 .byte 0,52,0 // add %dh,(%rax,%rax,1)
22006 .byte 0,0 // add %al,(%rax)
22007 .byte 52,0 // xor $0x0,%al
22008 .byte 0,0 // add %al,(%rax)
22009 .byte 52,0 // xor $0x0,%al
22010 .byte 0,0 // add %al,(%rax)
22011 .byte 52,255 // xor $0xff,%al
22012 .byte 255 // (bad)
22013 .byte 127,0 // jg 4034 <.literal16+0xf4>
22014 .byte 255 // (bad)
22015 .byte 255 // (bad)
22016 .byte 127,0 // jg 4038 <.literal16+0xf8>
22017 .byte 255 // (bad)
22018 .byte 255 // (bad)
22019 .byte 127,0 // jg 403c <.literal16+0xfc>
22020 .byte 255 // (bad)
22021 .byte 255 // (bad)
22022 .byte 127,0 // jg 4040 <.literal16+0x100>
22023 .byte 0,0 // add %al,(%rax)
22024 .byte 0,63 // add %bh,(%rdi)
22025 .byte 0,0 // add %al,(%rax)
22026 .byte 0,63 // add %bh,(%rdi)
22027 .byte 0,0 // add %al,(%rax)
22028 .byte 0,63 // add %bh,(%rdi)
22029 .byte 0,0 // add %al,(%rax)
22030 .byte 0,63 // add %bh,(%rdi)
22031 .byte 119,115 // ja 40c5 <.literal16+0x185>
22032 .byte 248 // clc
22033 .byte 194,119,115 // retq $0x7377
22034 .byte 248 // clc
22035 .byte 194,119,115 // retq $0x7377
22036 .byte 248 // clc
22037 .byte 194,119,115 // retq $0x7377
22038 .byte 248 // clc
22039 .byte 194,117,191 // retq $0xbf75
22040 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22041 .byte 63 // (bad)
22042 .byte 117,191 // jne 4029 <.literal16+0xe9>
22043 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22044 .byte 63 // (bad)
22045 .byte 249 // stc
22046 .byte 68,180,62 // rex.R mov $0x3e,%spl
22047 .byte 249 // stc
22048 .byte 68,180,62 // rex.R mov $0x3e,%spl
22049 .byte 249 // stc
22050 .byte 68,180,62 // rex.R mov $0x3e,%spl
22051 .byte 249 // stc
22052 .byte 68,180,62 // rex.R mov $0x3e,%spl
22053 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
22054 .byte 233,220,63,163,233 // jmpq ffffffffe9a3806a <_sk_callback_sse41+0xffffffffe9a3420a>
22055 .byte 220,63 // fdivrl (%rdi)
22056 .byte 81 // push %rcx
22057 .byte 140,242 // mov %?,%edx
22058 .byte 66,81 // rex.X push %rcx
22059 .byte 140,242 // mov %?,%edx
22060 .byte 66,81 // rex.X push %rcx
22061 .byte 140,242 // mov %?,%edx
22062 .byte 66,81 // rex.X push %rcx
22063 .byte 140,242 // mov %?,%edx
22064 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
22065 .byte 63 // (bad)
22066 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
22067 .byte 63 // (bad)
22068 .byte 248 // clc
22069 .byte 245 // cmc
22070 .byte 154 // (bad)
22071 .byte 64,248 // rex clc
22072 .byte 245 // cmc
22073 .byte 154 // (bad)
22074 .byte 64,248 // rex clc
22075 .byte 245 // cmc
22076 .byte 154 // (bad)
22077 .byte 64,248 // rex clc
22078 .byte 245 // cmc
22079 .byte 154 // (bad)
22080 .byte 64,254 // rex (bad)
22081 .byte 210,221 // rcr %cl,%ch
22082 .byte 65,254 // rex.B (bad)
22083 .byte 210,221 // rcr %cl,%ch
22084 .byte 65,254 // rex.B (bad)
22085 .byte 210,221 // rcr %cl,%ch
22086 .byte 65,254 // rex.B (bad)
22087 .byte 210,221 // rcr %cl,%ch
22088 .byte 65,0,0 // add %al,(%r8)
22089 .byte 0,75,0 // add %cl,0x0(%rbx)
22090 .byte 0,0 // add %al,(%rax)
22091 .byte 75,0,0 // rex.WXB add %al,(%r8)
22092 .byte 0,75,0 // add %cl,0x0(%rbx)
22093 .byte 0,0 // add %al,(%rax)
22094 .byte 75,0,0 // rex.WXB add %al,(%r8)
22095 .byte 0,52,0 // add %dh,(%rax,%rax,1)
22096 .byte 0,0 // add %al,(%rax)
22097 .byte 52,0 // xor $0x0,%al
22098 .byte 0,0 // add %al,(%rax)
22099 .byte 52,0 // xor $0x0,%al
22100 .byte 0,0 // add %al,(%rax)
22101 .byte 52,255 // xor $0xff,%al
22102 .byte 255 // (bad)
22103 .byte 127,0 // jg 40f4 <.literal16+0x1b4>
22104 .byte 255 // (bad)
22105 .byte 255 // (bad)
22106 .byte 127,0 // jg 40f8 <.literal16+0x1b8>
22107 .byte 255 // (bad)
22108 .byte 255 // (bad)
22109 .byte 127,0 // jg 40fc <.literal16+0x1bc>
22110 .byte 255 // (bad)
22111 .byte 255 // (bad)
22112 .byte 127,0 // jg 4100 <.literal16+0x1c0>
22113 .byte 0,0 // add %al,(%rax)
22114 .byte 0,63 // add %bh,(%rdi)
22115 .byte 0,0 // add %al,(%rax)
22116 .byte 0,63 // add %bh,(%rdi)
22117 .byte 0,0 // add %al,(%rax)
22118 .byte 0,63 // add %bh,(%rdi)
22119 .byte 0,0 // add %al,(%rax)
22120 .byte 0,63 // add %bh,(%rdi)
22121 .byte 119,115 // ja 4185 <.literal16+0x245>
22122 .byte 248 // clc
22123 .byte 194,119,115 // retq $0x7377
22124 .byte 248 // clc
22125 .byte 194,119,115 // retq $0x7377
22126 .byte 248 // clc
22127 .byte 194,119,115 // retq $0x7377
22128 .byte 248 // clc
22129 .byte 194,117,191 // retq $0xbf75
22130 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22131 .byte 63 // (bad)
22132 .byte 117,191 // jne 40e9 <.literal16+0x1a9>
22133 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22134 .byte 63 // (bad)
22135 .byte 249 // stc
22136 .byte 68,180,62 // rex.R mov $0x3e,%spl
22137 .byte 249 // stc
22138 .byte 68,180,62 // rex.R mov $0x3e,%spl
22139 .byte 249 // stc
22140 .byte 68,180,62 // rex.R mov $0x3e,%spl
22141 .byte 249 // stc
22142 .byte 68,180,62 // rex.R mov $0x3e,%spl
22143 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
22144 .byte 233,220,63,163,233 // jmpq ffffffffe9a3812a <_sk_callback_sse41+0xffffffffe9a342ca>
22145 .byte 220,63 // fdivrl (%rdi)
22146 .byte 81 // push %rcx
22147 .byte 140,242 // mov %?,%edx
22148 .byte 66,81 // rex.X push %rcx
22149 .byte 140,242 // mov %?,%edx
22150 .byte 66,81 // rex.X push %rcx
22151 .byte 140,242 // mov %?,%edx
22152 .byte 66,81 // rex.X push %rcx
22153 .byte 140,242 // mov %?,%edx
22154 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
22155 .byte 63 // (bad)
22156 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
22157 .byte 63 // (bad)
22158 .byte 248 // clc
22159 .byte 245 // cmc
22160 .byte 154 // (bad)
22161 .byte 64,248 // rex clc
22162 .byte 245 // cmc
22163 .byte 154 // (bad)
22164 .byte 64,248 // rex clc
22165 .byte 245 // cmc
22166 .byte 154 // (bad)
22167 .byte 64,248 // rex clc
22168 .byte 245 // cmc
22169 .byte 154 // (bad)
22170 .byte 64,254 // rex (bad)
22171 .byte 210,221 // rcr %cl,%ch
22172 .byte 65,254 // rex.B (bad)
22173 .byte 210,221 // rcr %cl,%ch
22174 .byte 65,254 // rex.B (bad)
22175 .byte 210,221 // rcr %cl,%ch
22176 .byte 65,254 // rex.B (bad)
22177 .byte 210,221 // rcr %cl,%ch
22178 .byte 65,0,0 // add %al,(%r8)
22179 .byte 0,75,0 // add %cl,0x0(%rbx)
22180 .byte 0,0 // add %al,(%rax)
22181 .byte 75,0,0 // rex.WXB add %al,(%r8)
22182 .byte 0,75,0 // add %cl,0x0(%rbx)
22183 .byte 0,0 // add %al,(%rax)
22184 .byte 75,0,0 // rex.WXB add %al,(%r8)
22185 .byte 0,52,0 // add %dh,(%rax,%rax,1)
22186 .byte 0,0 // add %al,(%rax)
22187 .byte 52,0 // xor $0x0,%al
22188 .byte 0,0 // add %al,(%rax)
22189 .byte 52,0 // xor $0x0,%al
22190 .byte 0,0 // add %al,(%rax)
22191 .byte 52,255 // xor $0xff,%al
22192 .byte 255 // (bad)
22193 .byte 127,0 // jg 41b4 <.literal16+0x274>
22194 .byte 255 // (bad)
22195 .byte 255 // (bad)
22196 .byte 127,0 // jg 41b8 <.literal16+0x278>
22197 .byte 255 // (bad)
22198 .byte 255 // (bad)
22199 .byte 127,0 // jg 41bc <.literal16+0x27c>
22200 .byte 255 // (bad)
22201 .byte 255 // (bad)
22202 .byte 127,0 // jg 41c0 <.literal16+0x280>
22203 .byte 0,0 // add %al,(%rax)
22204 .byte 0,63 // add %bh,(%rdi)
22205 .byte 0,0 // add %al,(%rax)
22206 .byte 0,63 // add %bh,(%rdi)
22207 .byte 0,0 // add %al,(%rax)
22208 .byte 0,63 // add %bh,(%rdi)
22209 .byte 0,0 // add %al,(%rax)
22210 .byte 0,63 // add %bh,(%rdi)
22211 .byte 119,115 // ja 4245 <.literal16+0x305>
22212 .byte 248 // clc
22213 .byte 194,119,115 // retq $0x7377
22214 .byte 248 // clc
22215 .byte 194,119,115 // retq $0x7377
22216 .byte 248 // clc
22217 .byte 194,119,115 // retq $0x7377
22218 .byte 248 // clc
22219 .byte 194,117,191 // retq $0xbf75
22220 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22221 .byte 63 // (bad)
22222 .byte 117,191 // jne 41a9 <.literal16+0x269>
22223 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
22224 .byte 63 // (bad)
22225 .byte 249 // stc
22226 .byte 68,180,62 // rex.R mov $0x3e,%spl
22227 .byte 249 // stc
22228 .byte 68,180,62 // rex.R mov $0x3e,%spl
22229 .byte 249 // stc
22230 .byte 68,180,62 // rex.R mov $0x3e,%spl
22231 .byte 249 // stc
22232 .byte 68,180,62 // rex.R mov $0x3e,%spl
22233 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
22234 .byte 233,220,63,163,233 // jmpq ffffffffe9a381ea <_sk_callback_sse41+0xffffffffe9a3438a>
22235 .byte 220,63 // fdivrl (%rdi)
22236 .byte 81 // push %rcx
22237 .byte 140,242 // mov %?,%edx
22238 .byte 66,81 // rex.X push %rcx
22239 .byte 140,242 // mov %?,%edx
22240 .byte 66,81 // rex.X push %rcx
22241 .byte 140,242 // mov %?,%edx
22242 .byte 66,81 // rex.X push %rcx
22243 .byte 140,242 // mov %?,%edx
22244 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
22245 .byte 63 // (bad)
22246 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
22247 .byte 63 // (bad)
22248 .byte 248 // clc
22249 .byte 245 // cmc
22250 .byte 154 // (bad)
22251 .byte 64,248 // rex clc
22252 .byte 245 // cmc
22253 .byte 154 // (bad)
22254 .byte 64,248 // rex clc
22255 .byte 245 // cmc
22256 .byte 154 // (bad)
22257 .byte 64,248 // rex clc
22258 .byte 245 // cmc
22259 .byte 154 // (bad)
22260 .byte 64,254 // rex (bad)
22261 .byte 210,221 // rcr %cl,%ch
22262 .byte 65,254 // rex.B (bad)
22263 .byte 210,221 // rcr %cl,%ch
22264 .byte 65,254 // rex.B (bad)
22265 .byte 210,221 // rcr %cl,%ch
22266 .byte 65,254 // rex.B (bad)
22267 .byte 210,221 // rcr %cl,%ch
22268 .byte 65,0,0 // add %al,(%r8)
22269 .byte 0,75,0 // add %cl,0x0(%rbx)
22270 .byte 0,0 // add %al,(%rax)
22271 .byte 75,0,0 // rex.WXB add %al,(%r8)
22272 .byte 0,75,0 // add %cl,0x0(%rbx)
22273 .byte 0,0 // add %al,(%rax)
22274 .byte 75,0,128,0,0,0,128 // rex.WXB add %al,-0x80000000(%r8)
22275 .byte 0,0 // add %al,(%rax)
22276 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
22277 .byte 0,0 // add %al,(%rax)
22278 .byte 0,4,0 // add %al,(%rax,%rax,1)
22279 .byte 0,0 // add %al,(%rax)
22280 .byte 4,0 // add $0x0,%al
22281 .byte 0,0 // add %al,(%rax)
22282 .byte 4,0 // add $0x0,%al
22283 .byte 0,0 // add %al,(%rax)
22284 .byte 4,0 // add $0x0,%al
22285 .byte 0,0 // add %al,(%rax)
22286 .byte 0,0 // add %al,(%rax)
22287 .byte 56,0 // cmp %al,(%rax)
22288 .byte 0,0 // add %al,(%rax)
22289 .byte 56,0 // cmp %al,(%rax)
22290 .byte 0,0 // add %al,(%rax)
22291 .byte 56,0 // cmp %al,(%rax)
22292 .byte 0,0 // add %al,(%rax)
22293 .byte 56,0 // cmp %al,(%rax)
22294 .byte 128,0,0 // addb $0x0,(%rax)
22295 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
22296 .byte 0,0 // add %al,(%rax)
22297 .byte 0,128,0,0,0,4 // add %al,0x4000000(%rax)
22298 .byte 0,0 // add %al,(%rax)
22299 .byte 0,4,0 // add %al,(%rax,%rax,1)
22300 .byte 0,0 // add %al,(%rax)
22301 .byte 4,0 // add $0x0,%al
22302 .byte 0,0 // add %al,(%rax)
22303 .byte 4,0 // add $0x0,%al
22304 .byte 0,0 // add %al,(%rax)
22305 .byte 0,0 // add %al,(%rax)
22306 .byte 56,0 // cmp %al,(%rax)
22307 .byte 0,0 // add %al,(%rax)
22308 .byte 56,0 // cmp %al,(%rax)
22309 .byte 0,0 // add %al,(%rax)
22310 .byte 56,0 // cmp %al,(%rax)
22311 .byte 0,0 // add %al,(%rax)
22312 .byte 56,0 // cmp %al,(%rax)
22313 .byte 0,0 // add %al,(%rax)
22314 .byte 128,0,0 // addb $0x0,(%rax)
22315 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
22316 .byte 0,0 // add %al,(%rax)
22317 .byte 0,128,0,0,128,56 // add %al,0x38800000(%rax)
22318 .byte 0,0 // add %al,(%rax)
22319 .byte 128,56,0 // cmpb $0x0,(%rax)
22320 .byte 0,128,56,0,0,128 // add %al,-0x7fffffc8(%rax)
22321 .byte 56,0 // cmp %al,(%rax)
22322 .byte 64,254 // rex (bad)
22323 .byte 255,0 // incl (%rax)
22324 .byte 64,254 // rex (bad)
22325 .byte 255,0 // incl (%rax)
22326 .byte 64,254 // rex (bad)
22327 .byte 255,0 // incl (%rax)
22328 .byte 64,254 // rex (bad)
22329 .byte 255 // .byte 0xff
Mike Klein67e61712017-04-26 09:52:09 -0400[diff] [blame]22330BALIGN32
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]22331
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22332HIDDEN _sk_start_pipeline_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22333.globl _sk_start_pipeline_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22334FUNCTION(_sk_start_pipeline_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22335_sk_start_pipeline_sse2:
22336 .byte 65,87 // push %r15
22337 .byte 65,86 // push %r14
22338 .byte 65,85 // push %r13
22339 .byte 65,84 // push %r12
22340 .byte 83 // push %rbx
22341 .byte 73,137,207 // mov %rcx,%r15
22342 .byte 73,137,214 // mov %rdx,%r14
22343 .byte 72,137,251 // mov %rdi,%rbx
22344 .byte 72,173 // lods %ds:(%rsi),%rax
22345 .byte 73,137,196 // mov %rax,%r12
22346 .byte 73,137,245 // mov %rsi,%r13
22347 .byte 72,141,67,4 // lea 0x4(%rbx),%rax
22348 .byte 76,57,248 // cmp %r15,%rax
22349 .byte 118,5 // jbe 28 <_sk_start_pipeline_sse2+0x28>
22350 .byte 72,137,216 // mov %rbx,%rax
22351 .byte 235,52 // jmp 5c <_sk_start_pipeline_sse2+0x5c>
22352 .byte 15,87,192 // xorps %xmm0,%xmm0
22353 .byte 15,87,201 // xorps %xmm1,%xmm1
22354 .byte 15,87,210 // xorps %xmm2,%xmm2
22355 .byte 15,87,219 // xorps %xmm3,%xmm3
22356 .byte 15,87,228 // xorps %xmm4,%xmm4
22357 .byte 15,87,237 // xorps %xmm5,%xmm5
22358 .byte 15,87,246 // xorps %xmm6,%xmm6
22359 .byte 15,87,255 // xorps %xmm7,%xmm7
22360 .byte 72,137,223 // mov %rbx,%rdi
22361 .byte 76,137,238 // mov %r13,%rsi
22362 .byte 76,137,242 // mov %r14,%rdx
22363 .byte 65,255,212 // callq *%r12
22364 .byte 72,141,67,4 // lea 0x4(%rbx),%rax
22365 .byte 72,131,195,8 // add $0x8,%rbx
22366 .byte 76,57,251 // cmp %r15,%rbx
22367 .byte 72,137,195 // mov %rax,%rbx
22368 .byte 118,204 // jbe 28 <_sk_start_pipeline_sse2+0x28>
22369 .byte 91 // pop %rbx
22370 .byte 65,92 // pop %r12
22371 .byte 65,93 // pop %r13
22372 .byte 65,94 // pop %r14
22373 .byte 65,95 // pop %r15
22374 .byte 195 // retq
22375
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22376HIDDEN _sk_just_return_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22377.globl _sk_just_return_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22378FUNCTION(_sk_just_return_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22379_sk_just_return_sse2:
22380 .byte 195 // retq
22381
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22382HIDDEN _sk_seed_shader_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22383.globl _sk_seed_shader_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22384FUNCTION(_sk_seed_shader_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22385_sk_seed_shader_sse2:
22386 .byte 72,173 // lods %ds:(%rsi),%rax
22387 .byte 102,15,110,199 // movd %edi,%xmm0
22388 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
22389 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]22390 .byte 15,40,21,212,66,0,0 // movaps 0x42d4(%rip),%xmm2 # 4350 <_sk_callback_sse2+0xe0>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22391 .byte 15,88,202 // addps %xmm2,%xmm1
22392 .byte 15,16,2 // movups (%rdx),%xmm0
22393 .byte 15,88,193 // addps %xmm1,%xmm0
22394 .byte 102,15,110,8 // movd (%rax),%xmm1
22395 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
22396 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
22397 .byte 15,88,202 // addps %xmm2,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22398 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]22399 .byte 15,40,21,195,66,0,0 // movaps 0x42c3(%rip),%xmm2 # 4360 <_sk_callback_sse2+0xf0>
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22400 .byte 15,87,219 // xorps %xmm3,%xmm3
22401 .byte 15,87,228 // xorps %xmm4,%xmm4
22402 .byte 15,87,237 // xorps %xmm5,%xmm5
22403 .byte 15,87,246 // xorps %xmm6,%xmm6
22404 .byte 15,87,255 // xorps %xmm7,%xmm7
22405 .byte 255,224 // jmpq *%rax
22406
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22407HIDDEN _sk_constant_color_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22408.globl _sk_constant_color_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22409FUNCTION(_sk_constant_color_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22410_sk_constant_color_sse2:
22411 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]22412 .byte 243,15,16,0 // movss (%rax),%xmm0
22413 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22414 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]22415 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
22416 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
22417 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
22418 .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3
22419 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22420 .byte 72,173 // lods %ds:(%rsi),%rax
22421 .byte 255,224 // jmpq *%rax
22422
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22423HIDDEN _sk_clear_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22424.globl _sk_clear_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22425FUNCTION(_sk_clear_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22426_sk_clear_sse2:
22427 .byte 72,173 // lods %ds:(%rsi),%rax
22428 .byte 15,87,192 // xorps %xmm0,%xmm0
22429 .byte 15,87,201 // xorps %xmm1,%xmm1
22430 .byte 15,87,210 // xorps %xmm2,%xmm2
22431 .byte 15,87,219 // xorps %xmm3,%xmm3
22432 .byte 255,224 // jmpq *%rax
22433
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22434HIDDEN _sk_srcatop_sse2
22435.globl _sk_srcatop_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22436FUNCTION(_sk_srcatop_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22437_sk_srcatop_sse2:
22438 .byte 15,89,199 // mulps %xmm7,%xmm0
22439 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22440 .byte 102,68,15,110,192 // movd %eax,%xmm8
22441 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22442 .byte 68,15,92,195 // subps %xmm3,%xmm8
22443 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22444 .byte 68,15,89,204 // mulps %xmm4,%xmm9
22445 .byte 65,15,88,193 // addps %xmm9,%xmm0
22446 .byte 15,89,207 // mulps %xmm7,%xmm1
22447 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22448 .byte 68,15,89,205 // mulps %xmm5,%xmm9
22449 .byte 65,15,88,201 // addps %xmm9,%xmm1
22450 .byte 15,89,215 // mulps %xmm7,%xmm2
22451 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22452 .byte 68,15,89,206 // mulps %xmm6,%xmm9
22453 .byte 65,15,88,209 // addps %xmm9,%xmm2
22454 .byte 15,89,223 // mulps %xmm7,%xmm3
22455 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22456 .byte 65,15,88,216 // addps %xmm8,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22457 .byte 72,173 // lods %ds:(%rsi),%rax
22458 .byte 255,224 // jmpq *%rax
22459
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22460HIDDEN _sk_dstatop_sse2
22461.globl _sk_dstatop_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22462FUNCTION(_sk_dstatop_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22463_sk_dstatop_sse2:
22464 .byte 68,15,40,195 // movaps %xmm3,%xmm8
22465 .byte 68,15,89,196 // mulps %xmm4,%xmm8
22466 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22467 .byte 102,68,15,110,200 // movd %eax,%xmm9
22468 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
22469 .byte 68,15,92,207 // subps %xmm7,%xmm9
22470 .byte 65,15,89,193 // mulps %xmm9,%xmm0
22471 .byte 65,15,88,192 // addps %xmm8,%xmm0
22472 .byte 68,15,40,195 // movaps %xmm3,%xmm8
22473 .byte 68,15,89,197 // mulps %xmm5,%xmm8
22474 .byte 65,15,89,201 // mulps %xmm9,%xmm1
22475 .byte 65,15,88,200 // addps %xmm8,%xmm1
22476 .byte 68,15,40,195 // movaps %xmm3,%xmm8
22477 .byte 68,15,89,198 // mulps %xmm6,%xmm8
22478 .byte 65,15,89,209 // mulps %xmm9,%xmm2
22479 .byte 65,15,88,208 // addps %xmm8,%xmm2
22480 .byte 68,15,89,203 // mulps %xmm3,%xmm9
22481 .byte 15,89,223 // mulps %xmm7,%xmm3
22482 .byte 65,15,88,217 // addps %xmm9,%xmm3
22483 .byte 72,173 // lods %ds:(%rsi),%rax
22484 .byte 255,224 // jmpq *%rax
22485
22486HIDDEN _sk_srcin_sse2
22487.globl _sk_srcin_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22488FUNCTION(_sk_srcin_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22489_sk_srcin_sse2:
22490 .byte 15,89,199 // mulps %xmm7,%xmm0
22491 .byte 15,89,207 // mulps %xmm7,%xmm1
22492 .byte 15,89,215 // mulps %xmm7,%xmm2
22493 .byte 15,89,223 // mulps %xmm7,%xmm3
22494 .byte 72,173 // lods %ds:(%rsi),%rax
22495 .byte 255,224 // jmpq *%rax
22496
22497HIDDEN _sk_dstin_sse2
22498.globl _sk_dstin_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22499FUNCTION(_sk_dstin_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22500_sk_dstin_sse2:
22501 .byte 15,40,195 // movaps %xmm3,%xmm0
22502 .byte 15,89,196 // mulps %xmm4,%xmm0
22503 .byte 15,40,203 // movaps %xmm3,%xmm1
22504 .byte 15,89,205 // mulps %xmm5,%xmm1
22505 .byte 15,40,211 // movaps %xmm3,%xmm2
22506 .byte 15,89,214 // mulps %xmm6,%xmm2
22507 .byte 15,89,223 // mulps %xmm7,%xmm3
22508 .byte 72,173 // lods %ds:(%rsi),%rax
22509 .byte 255,224 // jmpq *%rax
22510
22511HIDDEN _sk_srcout_sse2
22512.globl _sk_srcout_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22513FUNCTION(_sk_srcout_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22514_sk_srcout_sse2:
22515 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22516 .byte 102,68,15,110,192 // movd %eax,%xmm8
22517 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22518 .byte 68,15,92,199 // subps %xmm7,%xmm8
22519 .byte 65,15,89,192 // mulps %xmm8,%xmm0
22520 .byte 65,15,89,200 // mulps %xmm8,%xmm1
22521 .byte 65,15,89,208 // mulps %xmm8,%xmm2
22522 .byte 65,15,89,216 // mulps %xmm8,%xmm3
22523 .byte 72,173 // lods %ds:(%rsi),%rax
22524 .byte 255,224 // jmpq *%rax
22525
22526HIDDEN _sk_dstout_sse2
22527.globl _sk_dstout_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22528FUNCTION(_sk_dstout_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22529_sk_dstout_sse2:
22530 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22531 .byte 102,68,15,110,192 // movd %eax,%xmm8
22532 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22533 .byte 68,15,92,195 // subps %xmm3,%xmm8
22534 .byte 65,15,40,192 // movaps %xmm8,%xmm0
22535 .byte 15,89,196 // mulps %xmm4,%xmm0
22536 .byte 65,15,40,200 // movaps %xmm8,%xmm1
22537 .byte 15,89,205 // mulps %xmm5,%xmm1
22538 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22539 .byte 15,89,214 // mulps %xmm6,%xmm2
22540 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22541 .byte 72,173 // lods %ds:(%rsi),%rax
22542 .byte 65,15,40,216 // movaps %xmm8,%xmm3
22543 .byte 255,224 // jmpq *%rax
22544
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22545HIDDEN _sk_srcover_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22546.globl _sk_srcover_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22547FUNCTION(_sk_srcover_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22548_sk_srcover_sse2:
22549 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22550 .byte 102,68,15,110,192 // movd %eax,%xmm8
22551 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22552 .byte 68,15,92,195 // subps %xmm3,%xmm8
22553 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22554 .byte 68,15,89,204 // mulps %xmm4,%xmm9
22555 .byte 65,15,88,193 // addps %xmm9,%xmm0
22556 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22557 .byte 68,15,89,205 // mulps %xmm5,%xmm9
22558 .byte 65,15,88,201 // addps %xmm9,%xmm1
22559 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22560 .byte 68,15,89,206 // mulps %xmm6,%xmm9
22561 .byte 65,15,88,209 // addps %xmm9,%xmm2
22562 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22563 .byte 65,15,88,216 // addps %xmm8,%xmm3
22564 .byte 72,173 // lods %ds:(%rsi),%rax
22565 .byte 255,224 // jmpq *%rax
22566
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]22567HIDDEN _sk_dstover_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22568.globl _sk_dstover_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22569FUNCTION(_sk_dstover_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]22570_sk_dstover_sse2:
22571 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22572 .byte 102,68,15,110,192 // movd %eax,%xmm8
22573 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22574 .byte 68,15,92,199 // subps %xmm7,%xmm8
22575 .byte 65,15,89,192 // mulps %xmm8,%xmm0
22576 .byte 15,88,196 // addps %xmm4,%xmm0
22577 .byte 65,15,89,200 // mulps %xmm8,%xmm1
22578 .byte 15,88,205 // addps %xmm5,%xmm1
22579 .byte 65,15,89,208 // mulps %xmm8,%xmm2
22580 .byte 15,88,214 // addps %xmm6,%xmm2
22581 .byte 65,15,89,216 // mulps %xmm8,%xmm3
22582 .byte 15,88,223 // addps %xmm7,%xmm3
22583 .byte 72,173 // lods %ds:(%rsi),%rax
22584 .byte 255,224 // jmpq *%rax
22585
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22586HIDDEN _sk_modulate_sse2
22587.globl _sk_modulate_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22588FUNCTION(_sk_modulate_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22589_sk_modulate_sse2:
22590 .byte 15,89,196 // mulps %xmm4,%xmm0
22591 .byte 15,89,205 // mulps %xmm5,%xmm1
22592 .byte 15,89,214 // mulps %xmm6,%xmm2
22593 .byte 15,89,223 // mulps %xmm7,%xmm3
22594 .byte 72,173 // lods %ds:(%rsi),%rax
22595 .byte 255,224 // jmpq *%rax
22596
22597HIDDEN _sk_multiply_sse2
22598.globl _sk_multiply_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22599FUNCTION(_sk_multiply_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22600_sk_multiply_sse2:
22601 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22602 .byte 102,68,15,110,192 // movd %eax,%xmm8
22603 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
22604 .byte 69,15,40,200 // movaps %xmm8,%xmm9
22605 .byte 68,15,92,207 // subps %xmm7,%xmm9
22606 .byte 69,15,40,209 // movaps %xmm9,%xmm10
22607 .byte 68,15,89,208 // mulps %xmm0,%xmm10
22608 .byte 68,15,92,195 // subps %xmm3,%xmm8
22609 .byte 69,15,40,216 // movaps %xmm8,%xmm11
22610 .byte 68,15,89,220 // mulps %xmm4,%xmm11
22611 .byte 69,15,88,218 // addps %xmm10,%xmm11
22612 .byte 15,89,196 // mulps %xmm4,%xmm0
22613 .byte 65,15,88,195 // addps %xmm11,%xmm0
22614 .byte 69,15,40,209 // movaps %xmm9,%xmm10
22615 .byte 68,15,89,209 // mulps %xmm1,%xmm10
22616 .byte 69,15,40,216 // movaps %xmm8,%xmm11
22617 .byte 68,15,89,221 // mulps %xmm5,%xmm11
22618 .byte 69,15,88,218 // addps %xmm10,%xmm11
22619 .byte 15,89,205 // mulps %xmm5,%xmm1
22620 .byte 65,15,88,203 // addps %xmm11,%xmm1
22621 .byte 69,15,40,209 // movaps %xmm9,%xmm10
22622 .byte 68,15,89,210 // mulps %xmm2,%xmm10
22623 .byte 69,15,40,216 // movaps %xmm8,%xmm11
22624 .byte 68,15,89,222 // mulps %xmm6,%xmm11
22625 .byte 69,15,88,218 // addps %xmm10,%xmm11
22626 .byte 15,89,214 // mulps %xmm6,%xmm2
22627 .byte 65,15,88,211 // addps %xmm11,%xmm2
22628 .byte 68,15,89,203 // mulps %xmm3,%xmm9
22629 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22630 .byte 69,15,88,193 // addps %xmm9,%xmm8
22631 .byte 15,89,223 // mulps %xmm7,%xmm3
22632 .byte 65,15,88,216 // addps %xmm8,%xmm3
22633 .byte 72,173 // lods %ds:(%rsi),%rax
22634 .byte 255,224 // jmpq *%rax
22635
22636HIDDEN _sk_plus__sse2
22637.globl _sk_plus__sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22638FUNCTION(_sk_plus__sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22639_sk_plus__sse2:
22640 .byte 15,88,196 // addps %xmm4,%xmm0
22641 .byte 15,88,205 // addps %xmm5,%xmm1
22642 .byte 15,88,214 // addps %xmm6,%xmm2
22643 .byte 15,88,223 // addps %xmm7,%xmm3
22644 .byte 72,173 // lods %ds:(%rsi),%rax
22645 .byte 255,224 // jmpq *%rax
22646
22647HIDDEN _sk_screen_sse2
22648.globl _sk_screen_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22649FUNCTION(_sk_screen_sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22650_sk_screen_sse2:
22651 .byte 68,15,40,192 // movaps %xmm0,%xmm8
22652 .byte 68,15,88,196 // addps %xmm4,%xmm8
22653 .byte 15,89,196 // mulps %xmm4,%xmm0
22654 .byte 68,15,92,192 // subps %xmm0,%xmm8
22655 .byte 68,15,40,201 // movaps %xmm1,%xmm9
22656 .byte 68,15,88,205 // addps %xmm5,%xmm9
22657 .byte 15,89,205 // mulps %xmm5,%xmm1
22658 .byte 68,15,92,201 // subps %xmm1,%xmm9
22659 .byte 68,15,40,210 // movaps %xmm2,%xmm10
22660 .byte 68,15,88,214 // addps %xmm6,%xmm10
22661 .byte 15,89,214 // mulps %xmm6,%xmm2
22662 .byte 68,15,92,210 // subps %xmm2,%xmm10
22663 .byte 68,15,40,219 // movaps %xmm3,%xmm11
22664 .byte 68,15,88,223 // addps %xmm7,%xmm11
22665 .byte 15,89,223 // mulps %xmm7,%xmm3
22666 .byte 68,15,92,219 // subps %xmm3,%xmm11
22667 .byte 72,173 // lods %ds:(%rsi),%rax
22668 .byte 65,15,40,192 // movaps %xmm8,%xmm0
22669 .byte 65,15,40,201 // movaps %xmm9,%xmm1
22670 .byte 65,15,40,210 // movaps %xmm10,%xmm2
22671 .byte 65,15,40,219 // movaps %xmm11,%xmm3
22672 .byte 255,224 // jmpq *%rax
22673
22674HIDDEN _sk_xor__sse2
22675.globl _sk_xor__sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22676FUNCTION(_sk_xor__sse2)
Mike Kleinaaca1e42017-03-31 09:29:01 -0400[diff] [blame]22677_sk_xor__sse2:
22678 .byte 68,15,40,195 // movaps %xmm3,%xmm8
22679 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22680 .byte 102,15,110,216 // movd %eax,%xmm3
22681 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
22682 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22683 .byte 68,15,92,207 // subps %xmm7,%xmm9
22684 .byte 65,15,89,193 // mulps %xmm9,%xmm0
22685 .byte 65,15,92,216 // subps %xmm8,%xmm3
22686 .byte 68,15,40,211 // movaps %xmm3,%xmm10
22687 .byte 68,15,89,212 // mulps %xmm4,%xmm10
22688 .byte 65,15,88,194 // addps %xmm10,%xmm0
22689 .byte 65,15,89,201 // mulps %xmm9,%xmm1
22690 .byte 68,15,40,211 // movaps %xmm3,%xmm10
22691 .byte 68,15,89,213 // mulps %xmm5,%xmm10
22692 .byte 65,15,88,202 // addps %xmm10,%xmm1
22693 .byte 65,15,89,209 // mulps %xmm9,%xmm2
22694 .byte 68,15,40,211 // movaps %xmm3,%xmm10
22695 .byte 68,15,89,214 // mulps %xmm6,%xmm10
22696 .byte 65,15,88,210 // addps %xmm10,%xmm2
22697 .byte 69,15,89,200 // mulps %xmm8,%xmm9
22698 .byte 15,89,223 // mulps %xmm7,%xmm3
22699 .byte 65,15,88,217 // addps %xmm9,%xmm3
22700 .byte 72,173 // lods %ds:(%rsi),%rax
22701 .byte 255,224 // jmpq *%rax
22702
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]22703HIDDEN _sk_darken_sse2
22704.globl _sk_darken_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22705FUNCTION(_sk_darken_sse2)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]22706_sk_darken_sse2:
22707 .byte 68,15,40,193 // movaps %xmm1,%xmm8
22708 .byte 68,15,40,200 // movaps %xmm0,%xmm9
22709 .byte 15,88,196 // addps %xmm4,%xmm0
22710 .byte 68,15,89,207 // mulps %xmm7,%xmm9
22711 .byte 15,40,203 // movaps %xmm3,%xmm1
22712 .byte 15,89,204 // mulps %xmm4,%xmm1
22713 .byte 68,15,95,201 // maxps %xmm1,%xmm9
22714 .byte 65,15,92,193 // subps %xmm9,%xmm0
22715 .byte 65,15,40,200 // movaps %xmm8,%xmm1
22716 .byte 15,88,205 // addps %xmm5,%xmm1
22717 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22718 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22719 .byte 68,15,89,205 // mulps %xmm5,%xmm9
22720 .byte 69,15,95,193 // maxps %xmm9,%xmm8
22721 .byte 65,15,92,200 // subps %xmm8,%xmm1
22722 .byte 68,15,40,194 // movaps %xmm2,%xmm8
22723 .byte 68,15,88,198 // addps %xmm6,%xmm8
22724 .byte 15,89,215 // mulps %xmm7,%xmm2
22725 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22726 .byte 68,15,89,206 // mulps %xmm6,%xmm9
22727 .byte 65,15,95,209 // maxps %xmm9,%xmm2
22728 .byte 68,15,92,194 // subps %xmm2,%xmm8
22729 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22730 .byte 102,15,110,208 // movd %eax,%xmm2
22731 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
22732 .byte 15,92,211 // subps %xmm3,%xmm2
22733 .byte 15,89,215 // mulps %xmm7,%xmm2
22734 .byte 15,88,218 // addps %xmm2,%xmm3
22735 .byte 72,173 // lods %ds:(%rsi),%rax
22736 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22737 .byte 255,224 // jmpq *%rax
22738
22739HIDDEN _sk_lighten_sse2
22740.globl _sk_lighten_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22741FUNCTION(_sk_lighten_sse2)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]22742_sk_lighten_sse2:
22743 .byte 68,15,40,193 // movaps %xmm1,%xmm8
22744 .byte 68,15,40,200 // movaps %xmm0,%xmm9
22745 .byte 15,88,196 // addps %xmm4,%xmm0
22746 .byte 68,15,89,207 // mulps %xmm7,%xmm9
22747 .byte 15,40,203 // movaps %xmm3,%xmm1
22748 .byte 15,89,204 // mulps %xmm4,%xmm1
22749 .byte 68,15,93,201 // minps %xmm1,%xmm9
22750 .byte 65,15,92,193 // subps %xmm9,%xmm0
22751 .byte 65,15,40,200 // movaps %xmm8,%xmm1
22752 .byte 15,88,205 // addps %xmm5,%xmm1
22753 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22754 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22755 .byte 68,15,89,205 // mulps %xmm5,%xmm9
22756 .byte 69,15,93,193 // minps %xmm9,%xmm8
22757 .byte 65,15,92,200 // subps %xmm8,%xmm1
22758 .byte 68,15,40,194 // movaps %xmm2,%xmm8
22759 .byte 68,15,88,198 // addps %xmm6,%xmm8
22760 .byte 15,89,215 // mulps %xmm7,%xmm2
22761 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22762 .byte 68,15,89,206 // mulps %xmm6,%xmm9
22763 .byte 65,15,93,209 // minps %xmm9,%xmm2
22764 .byte 68,15,92,194 // subps %xmm2,%xmm8
22765 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22766 .byte 102,15,110,208 // movd %eax,%xmm2
22767 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
22768 .byte 15,92,211 // subps %xmm3,%xmm2
22769 .byte 15,89,215 // mulps %xmm7,%xmm2
22770 .byte 15,88,218 // addps %xmm2,%xmm3
22771 .byte 72,173 // lods %ds:(%rsi),%rax
22772 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22773 .byte 255,224 // jmpq *%rax
22774
22775HIDDEN _sk_difference_sse2
22776.globl _sk_difference_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22777FUNCTION(_sk_difference_sse2)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]22778_sk_difference_sse2:
22779 .byte 68,15,40,193 // movaps %xmm1,%xmm8
22780 .byte 68,15,40,200 // movaps %xmm0,%xmm9
22781 .byte 15,88,196 // addps %xmm4,%xmm0
22782 .byte 68,15,89,207 // mulps %xmm7,%xmm9
22783 .byte 15,40,203 // movaps %xmm3,%xmm1
22784 .byte 15,89,204 // mulps %xmm4,%xmm1
22785 .byte 68,15,93,201 // minps %xmm1,%xmm9
22786 .byte 69,15,88,201 // addps %xmm9,%xmm9
22787 .byte 65,15,92,193 // subps %xmm9,%xmm0
22788 .byte 65,15,40,200 // movaps %xmm8,%xmm1
22789 .byte 15,88,205 // addps %xmm5,%xmm1
22790 .byte 68,15,89,199 // mulps %xmm7,%xmm8
22791 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22792 .byte 68,15,89,205 // mulps %xmm5,%xmm9
22793 .byte 69,15,93,193 // minps %xmm9,%xmm8
22794 .byte 69,15,88,192 // addps %xmm8,%xmm8
22795 .byte 65,15,92,200 // subps %xmm8,%xmm1
22796 .byte 68,15,40,194 // movaps %xmm2,%xmm8
22797 .byte 68,15,88,198 // addps %xmm6,%xmm8
22798 .byte 15,89,215 // mulps %xmm7,%xmm2
22799 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22800 .byte 68,15,89,206 // mulps %xmm6,%xmm9
22801 .byte 65,15,93,209 // minps %xmm9,%xmm2
22802 .byte 15,88,210 // addps %xmm2,%xmm2
22803 .byte 68,15,92,194 // subps %xmm2,%xmm8
22804 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22805 .byte 102,15,110,208 // movd %eax,%xmm2
22806 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
22807 .byte 15,92,211 // subps %xmm3,%xmm2
22808 .byte 15,89,215 // mulps %xmm7,%xmm2
22809 .byte 15,88,218 // addps %xmm2,%xmm3
22810 .byte 72,173 // lods %ds:(%rsi),%rax
22811 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22812 .byte 255,224 // jmpq *%rax
22813
22814HIDDEN _sk_exclusion_sse2
22815.globl _sk_exclusion_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22816FUNCTION(_sk_exclusion_sse2)
Mike Klein66b09ab2017-03-31 10:29:40 -0400[diff] [blame]22817_sk_exclusion_sse2:
22818 .byte 68,15,40,193 // movaps %xmm1,%xmm8
22819 .byte 15,40,200 // movaps %xmm0,%xmm1
22820 .byte 15,88,196 // addps %xmm4,%xmm0
22821 .byte 15,89,204 // mulps %xmm4,%xmm1
22822 .byte 15,88,201 // addps %xmm1,%xmm1
22823 .byte 15,92,193 // subps %xmm1,%xmm0
22824 .byte 65,15,40,200 // movaps %xmm8,%xmm1
22825 .byte 15,88,205 // addps %xmm5,%xmm1
22826 .byte 68,15,89,197 // mulps %xmm5,%xmm8
22827 .byte 69,15,88,192 // addps %xmm8,%xmm8
22828 .byte 65,15,92,200 // subps %xmm8,%xmm1
22829 .byte 68,15,40,194 // movaps %xmm2,%xmm8
22830 .byte 68,15,88,198 // addps %xmm6,%xmm8
22831 .byte 15,89,214 // mulps %xmm6,%xmm2
22832 .byte 15,88,210 // addps %xmm2,%xmm2
22833 .byte 68,15,92,194 // subps %xmm2,%xmm8
22834 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22835 .byte 102,15,110,208 // movd %eax,%xmm2
22836 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
22837 .byte 15,92,211 // subps %xmm3,%xmm2
22838 .byte 15,89,215 // mulps %xmm7,%xmm2
22839 .byte 15,88,218 // addps %xmm2,%xmm3
22840 .byte 72,173 // lods %ds:(%rsi),%rax
22841 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22842 .byte 255,224 // jmpq *%rax
22843
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]22844HIDDEN _sk_colorburn_sse2
22845.globl _sk_colorburn_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22846FUNCTION(_sk_colorburn_sse2)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]22847_sk_colorburn_sse2:
22848 .byte 68,15,40,193 // movaps %xmm1,%xmm8
22849 .byte 68,15,40,224 // movaps %xmm0,%xmm12
22850 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22851 .byte 102,68,15,110,200 // movd %eax,%xmm9
22852 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
22853 .byte 69,15,40,217 // movaps %xmm9,%xmm11
22854 .byte 68,15,92,223 // subps %xmm7,%xmm11
22855 .byte 65,15,40,195 // movaps %xmm11,%xmm0
22856 .byte 65,15,89,196 // mulps %xmm12,%xmm0
22857 .byte 69,15,87,210 // xorps %xmm10,%xmm10
22858 .byte 15,40,207 // movaps %xmm7,%xmm1
22859 .byte 15,92,204 // subps %xmm4,%xmm1
22860 .byte 15,89,203 // mulps %xmm3,%xmm1
22861 .byte 65,15,94,204 // divps %xmm12,%xmm1
22862 .byte 68,15,40,239 // movaps %xmm7,%xmm13
22863 .byte 68,15,93,233 // minps %xmm1,%xmm13
22864 .byte 68,15,40,247 // movaps %xmm7,%xmm14
22865 .byte 69,15,92,245 // subps %xmm13,%xmm14
22866 .byte 65,15,40,204 // movaps %xmm12,%xmm1
22867 .byte 65,15,194,202,0 // cmpeqps %xmm10,%xmm1
22868 .byte 68,15,92,203 // subps %xmm3,%xmm9
22869 .byte 68,15,89,243 // mulps %xmm3,%xmm14
22870 .byte 68,15,88,240 // addps %xmm0,%xmm14
22871 .byte 68,15,84,225 // andps %xmm1,%xmm12
22872 .byte 65,15,85,206 // andnps %xmm14,%xmm1
22873 .byte 69,15,40,233 // movaps %xmm9,%xmm13
22874 .byte 68,15,89,236 // mulps %xmm4,%xmm13
22875 .byte 15,88,196 // addps %xmm4,%xmm0
22876 .byte 65,15,86,204 // orps %xmm12,%xmm1
22877 .byte 68,15,40,228 // movaps %xmm4,%xmm12
22878 .byte 68,15,194,231,0 // cmpeqps %xmm7,%xmm12
22879 .byte 65,15,88,205 // addps %xmm13,%xmm1
22880 .byte 65,15,84,196 // andps %xmm12,%xmm0
22881 .byte 68,15,85,225 // andnps %xmm1,%xmm12
22882 .byte 65,15,86,196 // orps %xmm12,%xmm0
22883 .byte 65,15,40,203 // movaps %xmm11,%xmm1
22884 .byte 65,15,89,200 // mulps %xmm8,%xmm1
22885 .byte 68,15,40,231 // movaps %xmm7,%xmm12
22886 .byte 68,15,92,229 // subps %xmm5,%xmm12
22887 .byte 68,15,89,227 // mulps %xmm3,%xmm12
22888 .byte 69,15,94,224 // divps %xmm8,%xmm12
22889 .byte 68,15,40,239 // movaps %xmm7,%xmm13
22890 .byte 69,15,93,236 // minps %xmm12,%xmm13
22891 .byte 68,15,40,231 // movaps %xmm7,%xmm12
22892 .byte 69,15,92,229 // subps %xmm13,%xmm12
22893 .byte 69,15,40,232 // movaps %xmm8,%xmm13
22894 .byte 69,15,194,234,0 // cmpeqps %xmm10,%xmm13
22895 .byte 68,15,89,227 // mulps %xmm3,%xmm12
22896 .byte 68,15,88,225 // addps %xmm1,%xmm12
22897 .byte 69,15,84,197 // andps %xmm13,%xmm8
22898 .byte 69,15,85,236 // andnps %xmm12,%xmm13
22899 .byte 69,15,86,232 // orps %xmm8,%xmm13
22900 .byte 69,15,40,193 // movaps %xmm9,%xmm8
22901 .byte 68,15,89,197 // mulps %xmm5,%xmm8
22902 .byte 15,88,205 // addps %xmm5,%xmm1
22903 .byte 69,15,88,232 // addps %xmm8,%xmm13
22904 .byte 68,15,40,197 // movaps %xmm5,%xmm8
22905 .byte 68,15,194,199,0 // cmpeqps %xmm7,%xmm8
22906 .byte 65,15,84,200 // andps %xmm8,%xmm1
22907 .byte 69,15,85,197 // andnps %xmm13,%xmm8
22908 .byte 65,15,86,200 // orps %xmm8,%xmm1
22909 .byte 68,15,40,199 // movaps %xmm7,%xmm8
22910 .byte 68,15,92,198 // subps %xmm6,%xmm8
22911 .byte 68,15,89,195 // mulps %xmm3,%xmm8
22912 .byte 68,15,94,194 // divps %xmm2,%xmm8
22913 .byte 68,15,40,231 // movaps %xmm7,%xmm12
22914 .byte 69,15,93,224 // minps %xmm8,%xmm12
22915 .byte 68,15,40,199 // movaps %xmm7,%xmm8
22916 .byte 69,15,92,196 // subps %xmm12,%xmm8
22917 .byte 68,15,89,218 // mulps %xmm2,%xmm11
22918 .byte 68,15,194,210,0 // cmpeqps %xmm2,%xmm10
22919 .byte 68,15,89,195 // mulps %xmm3,%xmm8
22920 .byte 69,15,88,195 // addps %xmm11,%xmm8
22921 .byte 65,15,84,210 // andps %xmm10,%xmm2
22922 .byte 69,15,85,208 // andnps %xmm8,%xmm10
22923 .byte 69,15,40,195 // movaps %xmm11,%xmm8
22924 .byte 68,15,88,198 // addps %xmm6,%xmm8
22925 .byte 68,15,86,210 // orps %xmm2,%xmm10
22926 .byte 65,15,40,209 // movaps %xmm9,%xmm2
22927 .byte 15,89,214 // mulps %xmm6,%xmm2
22928 .byte 68,15,88,210 // addps %xmm2,%xmm10
22929 .byte 15,40,214 // movaps %xmm6,%xmm2
22930 .byte 15,194,215,0 // cmpeqps %xmm7,%xmm2
22931 .byte 68,15,84,194 // andps %xmm2,%xmm8
22932 .byte 65,15,85,210 // andnps %xmm10,%xmm2
22933 .byte 68,15,86,194 // orps %xmm2,%xmm8
22934 .byte 68,15,89,207 // mulps %xmm7,%xmm9
22935 .byte 65,15,88,217 // addps %xmm9,%xmm3
22936 .byte 72,173 // lods %ds:(%rsi),%rax
22937 .byte 65,15,40,208 // movaps %xmm8,%xmm2
22938 .byte 255,224 // jmpq *%rax
22939
22940HIDDEN _sk_colordodge_sse2
22941.globl _sk_colordodge_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]22942FUNCTION(_sk_colordodge_sse2)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]22943_sk_colordodge_sse2:
22944 .byte 68,15,40,200 // movaps %xmm0,%xmm9
22945 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
22946 .byte 102,68,15,110,208 // movd %eax,%xmm10
22947 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
22948 .byte 69,15,40,218 // movaps %xmm10,%xmm11
22949 .byte 68,15,92,223 // subps %xmm7,%xmm11
22950 .byte 65,15,40,195 // movaps %xmm11,%xmm0
22951 .byte 65,15,89,193 // mulps %xmm9,%xmm0
22952 .byte 68,15,40,195 // movaps %xmm3,%xmm8
22953 .byte 68,15,89,196 // mulps %xmm4,%xmm8
22954 .byte 68,15,40,227 // movaps %xmm3,%xmm12
22955 .byte 69,15,92,225 // subps %xmm9,%xmm12
22956 .byte 69,15,94,196 // divps %xmm12,%xmm8
22957 .byte 68,15,40,231 // movaps %xmm7,%xmm12
22958 .byte 68,15,40,239 // movaps %xmm7,%xmm13
22959 .byte 69,15,93,232 // minps %xmm8,%xmm13
22960 .byte 69,15,40,241 // movaps %xmm9,%xmm14
22961 .byte 68,15,194,243,0 // cmpeqps %xmm3,%xmm14
22962 .byte 68,15,89,235 // mulps %xmm3,%xmm13
22963 .byte 68,15,88,232 // addps %xmm0,%xmm13
22964 .byte 69,15,84,206 // andps %xmm14,%xmm9
22965 .byte 69,15,85,245 // andnps %xmm13,%xmm14
22966 .byte 69,15,87,192 // xorps %xmm8,%xmm8
22967 .byte 68,15,92,211 // subps %xmm3,%xmm10
22968 .byte 69,15,86,241 // orps %xmm9,%xmm14
22969 .byte 69,15,40,202 // movaps %xmm10,%xmm9
22970 .byte 68,15,89,204 // mulps %xmm4,%xmm9
22971 .byte 15,88,196 // addps %xmm4,%xmm0
22972 .byte 69,15,88,241 // addps %xmm9,%xmm14
22973 .byte 68,15,40,204 // movaps %xmm4,%xmm9
22974 .byte 69,15,194,200,0 // cmpeqps %xmm8,%xmm9
22975 .byte 65,15,84,193 // andps %xmm9,%xmm0
22976 .byte 69,15,85,206 // andnps %xmm14,%xmm9
22977 .byte 65,15,86,193 // orps %xmm9,%xmm0
22978 .byte 68,15,40,235 // movaps %xmm3,%xmm13
22979 .byte 68,15,89,237 // mulps %xmm5,%xmm13
22980 .byte 68,15,40,203 // movaps %xmm3,%xmm9
22981 .byte 68,15,92,201 // subps %xmm1,%xmm9
22982 .byte 69,15,94,233 // divps %xmm9,%xmm13
22983 .byte 69,15,40,203 // movaps %xmm11,%xmm9
22984 .byte 68,15,89,201 // mulps %xmm1,%xmm9
22985 .byte 69,15,93,229 // minps %xmm13,%xmm12
22986 .byte 68,15,40,233 // movaps %xmm1,%xmm13
22987 .byte 68,15,194,235,0 // cmpeqps %xmm3,%xmm13
22988 .byte 68,15,89,227 // mulps %xmm3,%xmm12
22989 .byte 69,15,88,225 // addps %xmm9,%xmm12
22990 .byte 65,15,84,205 // andps %xmm13,%xmm1
22991 .byte 69,15,85,236 // andnps %xmm12,%xmm13
22992 .byte 68,15,86,233 // orps %xmm1,%xmm13
22993 .byte 65,15,40,202 // movaps %xmm10,%xmm1
22994 .byte 15,89,205 // mulps %xmm5,%xmm1
22995 .byte 68,15,88,205 // addps %xmm5,%xmm9
22996 .byte 68,15,88,233 // addps %xmm1,%xmm13
22997 .byte 15,40,205 // movaps %xmm5,%xmm1
22998 .byte 65,15,194,200,0 // cmpeqps %xmm8,%xmm1
22999 .byte 68,15,84,201 // andps %xmm1,%xmm9
23000 .byte 65,15,85,205 // andnps %xmm13,%xmm1
23001 .byte 68,15,86,201 // orps %xmm1,%xmm9
23002 .byte 68,15,40,227 // movaps %xmm3,%xmm12
23003 .byte 68,15,89,230 // mulps %xmm6,%xmm12
23004 .byte 15,40,203 // movaps %xmm3,%xmm1
23005 .byte 15,92,202 // subps %xmm2,%xmm1
23006 .byte 68,15,94,225 // divps %xmm1,%xmm12
23007 .byte 68,15,40,239 // movaps %xmm7,%xmm13
23008 .byte 68,15,89,218 // mulps %xmm2,%xmm11
23009 .byte 69,15,93,236 // minps %xmm12,%xmm13
23010 .byte 15,40,202 // movaps %xmm2,%xmm1
23011 .byte 15,194,203,0 // cmpeqps %xmm3,%xmm1
23012 .byte 68,15,89,235 // mulps %xmm3,%xmm13
23013 .byte 69,15,88,235 // addps %xmm11,%xmm13
23014 .byte 15,84,209 // andps %xmm1,%xmm2
23015 .byte 65,15,85,205 // andnps %xmm13,%xmm1
23016 .byte 15,86,202 // orps %xmm2,%xmm1
23017 .byte 65,15,40,210 // movaps %xmm10,%xmm2
23018 .byte 15,89,214 // mulps %xmm6,%xmm2
23019 .byte 15,88,202 // addps %xmm2,%xmm1
23020 .byte 68,15,194,198,0 // cmpeqps %xmm6,%xmm8
23021 .byte 68,15,88,222 // addps %xmm6,%xmm11
23022 .byte 69,15,84,216 // andps %xmm8,%xmm11
23023 .byte 68,15,85,193 // andnps %xmm1,%xmm8
23024 .byte 69,15,86,195 // orps %xmm11,%xmm8
23025 .byte 68,15,89,215 // mulps %xmm7,%xmm10
23026 .byte 65,15,88,218 // addps %xmm10,%xmm3
23027 .byte 72,173 // lods %ds:(%rsi),%rax
23028 .byte 65,15,40,201 // movaps %xmm9,%xmm1
23029 .byte 65,15,40,208 // movaps %xmm8,%xmm2
23030 .byte 255,224 // jmpq *%rax
23031
23032HIDDEN _sk_hardlight_sse2
23033.globl _sk_hardlight_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23034FUNCTION(_sk_hardlight_sse2)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23035_sk_hardlight_sse2:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23036 .byte 15,41,116,36,232 // movaps %xmm6,-0x18(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23037 .byte 15,40,245 // movaps %xmm5,%xmm6
23038 .byte 15,40,236 // movaps %xmm4,%xmm5
23039 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23040 .byte 102,68,15,110,216 // movd %eax,%xmm11
23041 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
23042 .byte 69,15,40,211 // movaps %xmm11,%xmm10
23043 .byte 68,15,92,215 // subps %xmm7,%xmm10
23044 .byte 69,15,40,194 // movaps %xmm10,%xmm8
23045 .byte 68,15,89,192 // mulps %xmm0,%xmm8
23046 .byte 68,15,92,219 // subps %xmm3,%xmm11
23047 .byte 69,15,40,203 // movaps %xmm11,%xmm9
23048 .byte 68,15,89,205 // mulps %xmm5,%xmm9
23049 .byte 69,15,88,200 // addps %xmm8,%xmm9
23050 .byte 68,15,40,195 // movaps %xmm3,%xmm8
23051 .byte 68,15,92,192 // subps %xmm0,%xmm8
23052 .byte 15,40,227 // movaps %xmm3,%xmm4
23053 .byte 15,89,231 // mulps %xmm7,%xmm4
23054 .byte 68,15,40,239 // movaps %xmm7,%xmm13
23055 .byte 68,15,40,247 // movaps %xmm7,%xmm14
23056 .byte 68,15,40,255 // movaps %xmm7,%xmm15
23057 .byte 68,15,92,253 // subps %xmm5,%xmm15
23058 .byte 69,15,89,248 // mulps %xmm8,%xmm15
23059 .byte 69,15,88,255 // addps %xmm15,%xmm15
23060 .byte 68,15,40,228 // movaps %xmm4,%xmm12
23061 .byte 69,15,92,231 // subps %xmm15,%xmm12
23062 .byte 68,15,40,192 // movaps %xmm0,%xmm8
23063 .byte 69,15,88,192 // addps %xmm8,%xmm8
23064 .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8
23065 .byte 15,89,197 // mulps %xmm5,%xmm0
23066 .byte 15,88,192 // addps %xmm0,%xmm0
23067 .byte 65,15,84,192 // andps %xmm8,%xmm0
23068 .byte 69,15,85,196 // andnps %xmm12,%xmm8
23069 .byte 68,15,86,192 // orps %xmm0,%xmm8
23070 .byte 69,15,40,251 // movaps %xmm11,%xmm15
23071 .byte 69,15,40,227 // movaps %xmm11,%xmm12
23072 .byte 68,15,89,223 // mulps %xmm7,%xmm11
23073 .byte 69,15,88,193 // addps %xmm9,%xmm8
23074 .byte 65,15,40,194 // movaps %xmm10,%xmm0
23075 .byte 15,89,193 // mulps %xmm1,%xmm0
23076 .byte 68,15,89,254 // mulps %xmm6,%xmm15
23077 .byte 68,15,88,248 // addps %xmm0,%xmm15
23078 .byte 15,40,195 // movaps %xmm3,%xmm0
23079 .byte 15,92,193 // subps %xmm1,%xmm0
23080 .byte 68,15,92,238 // subps %xmm6,%xmm13
23081 .byte 68,15,89,232 // mulps %xmm0,%xmm13
23082 .byte 69,15,88,237 // addps %xmm13,%xmm13
23083 .byte 15,40,196 // movaps %xmm4,%xmm0
23084 .byte 65,15,92,197 // subps %xmm13,%xmm0
23085 .byte 68,15,40,201 // movaps %xmm1,%xmm9
23086 .byte 69,15,88,201 // addps %xmm9,%xmm9
23087 .byte 68,15,194,203,2 // cmpleps %xmm3,%xmm9
23088 .byte 15,89,206 // mulps %xmm6,%xmm1
23089 .byte 15,88,201 // addps %xmm1,%xmm1
23090 .byte 65,15,84,201 // andps %xmm9,%xmm1
23091 .byte 68,15,85,200 // andnps %xmm0,%xmm9
23092 .byte 68,15,86,201 // orps %xmm1,%xmm9
23093 .byte 69,15,88,207 // addps %xmm15,%xmm9
23094 .byte 68,15,89,210 // mulps %xmm2,%xmm10
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23095 .byte 68,15,40,108,36,232 // movaps -0x18(%rsp),%xmm13
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23096 .byte 69,15,89,229 // mulps %xmm13,%xmm12
23097 .byte 69,15,88,226 // addps %xmm10,%xmm12
23098 .byte 68,15,40,210 // movaps %xmm2,%xmm10
23099 .byte 69,15,88,210 // addps %xmm10,%xmm10
23100 .byte 68,15,194,211,2 // cmpleps %xmm3,%xmm10
23101 .byte 15,40,195 // movaps %xmm3,%xmm0
23102 .byte 15,92,194 // subps %xmm2,%xmm0
23103 .byte 65,15,89,213 // mulps %xmm13,%xmm2
23104 .byte 15,88,210 // addps %xmm2,%xmm2
23105 .byte 69,15,92,245 // subps %xmm13,%xmm14
23106 .byte 68,15,89,240 // mulps %xmm0,%xmm14
23107 .byte 69,15,88,246 // addps %xmm14,%xmm14
23108 .byte 65,15,92,230 // subps %xmm14,%xmm4
23109 .byte 65,15,84,210 // andps %xmm10,%xmm2
23110 .byte 68,15,85,212 // andnps %xmm4,%xmm10
23111 .byte 68,15,86,210 // orps %xmm2,%xmm10
23112 .byte 69,15,88,212 // addps %xmm12,%xmm10
23113 .byte 65,15,88,219 // addps %xmm11,%xmm3
23114 .byte 72,173 // lods %ds:(%rsi),%rax
23115 .byte 65,15,40,192 // movaps %xmm8,%xmm0
23116 .byte 65,15,40,201 // movaps %xmm9,%xmm1
23117 .byte 65,15,40,210 // movaps %xmm10,%xmm2
23118 .byte 15,40,229 // movaps %xmm5,%xmm4
23119 .byte 15,40,238 // movaps %xmm6,%xmm5
23120 .byte 65,15,40,245 // movaps %xmm13,%xmm6
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23121 .byte 255,224 // jmpq *%rax
23122
23123HIDDEN _sk_overlay_sse2
23124.globl _sk_overlay_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23125FUNCTION(_sk_overlay_sse2)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23126_sk_overlay_sse2:
23127 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23128 .byte 102,68,15,110,192 // movd %eax,%xmm8
23129 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23130 .byte 69,15,40,200 // movaps %xmm8,%xmm9
23131 .byte 68,15,92,207 // subps %xmm7,%xmm9
23132 .byte 69,15,40,209 // movaps %xmm9,%xmm10
23133 .byte 68,15,89,208 // mulps %xmm0,%xmm10
23134 .byte 68,15,92,195 // subps %xmm3,%xmm8
23135 .byte 69,15,40,216 // movaps %xmm8,%xmm11
23136 .byte 68,15,89,220 // mulps %xmm4,%xmm11
23137 .byte 69,15,88,218 // addps %xmm10,%xmm11
23138 .byte 68,15,40,227 // movaps %xmm3,%xmm12
23139 .byte 68,15,92,224 // subps %xmm0,%xmm12
23140 .byte 15,89,196 // mulps %xmm4,%xmm0
23141 .byte 68,15,40,239 // movaps %xmm7,%xmm13
23142 .byte 68,15,92,236 // subps %xmm4,%xmm13
23143 .byte 68,15,40,244 // movaps %xmm4,%xmm14
23144 .byte 69,15,88,246 // addps %xmm14,%xmm14
23145 .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14
23146 .byte 15,88,192 // addps %xmm0,%xmm0
23147 .byte 68,15,40,211 // movaps %xmm3,%xmm10
23148 .byte 68,15,89,215 // mulps %xmm7,%xmm10
23149 .byte 69,15,89,236 // mulps %xmm12,%xmm13
23150 .byte 69,15,88,237 // addps %xmm13,%xmm13
23151 .byte 69,15,40,226 // movaps %xmm10,%xmm12
23152 .byte 69,15,92,229 // subps %xmm13,%xmm12
23153 .byte 65,15,84,198 // andps %xmm14,%xmm0
23154 .byte 69,15,85,244 // andnps %xmm12,%xmm14
23155 .byte 65,15,86,198 // orps %xmm14,%xmm0
23156 .byte 65,15,88,195 // addps %xmm11,%xmm0
23157 .byte 69,15,40,217 // movaps %xmm9,%xmm11
23158 .byte 68,15,89,217 // mulps %xmm1,%xmm11
23159 .byte 69,15,40,224 // movaps %xmm8,%xmm12
23160 .byte 68,15,89,229 // mulps %xmm5,%xmm12
23161 .byte 69,15,88,227 // addps %xmm11,%xmm12
23162 .byte 68,15,40,219 // movaps %xmm3,%xmm11
23163 .byte 68,15,92,217 // subps %xmm1,%xmm11
23164 .byte 15,89,205 // mulps %xmm5,%xmm1
23165 .byte 68,15,40,239 // movaps %xmm7,%xmm13
23166 .byte 68,15,92,237 // subps %xmm5,%xmm13
23167 .byte 68,15,40,245 // movaps %xmm5,%xmm14
23168 .byte 69,15,88,246 // addps %xmm14,%xmm14
23169 .byte 68,15,194,247,2 // cmpleps %xmm7,%xmm14
23170 .byte 15,88,201 // addps %xmm1,%xmm1
23171 .byte 69,15,89,235 // mulps %xmm11,%xmm13
23172 .byte 69,15,88,237 // addps %xmm13,%xmm13
23173 .byte 69,15,40,218 // movaps %xmm10,%xmm11
23174 .byte 69,15,92,221 // subps %xmm13,%xmm11
23175 .byte 65,15,84,206 // andps %xmm14,%xmm1
23176 .byte 69,15,85,243 // andnps %xmm11,%xmm14
23177 .byte 65,15,86,206 // orps %xmm14,%xmm1
23178 .byte 65,15,88,204 // addps %xmm12,%xmm1
23179 .byte 68,15,89,202 // mulps %xmm2,%xmm9
23180 .byte 69,15,40,216 // movaps %xmm8,%xmm11
23181 .byte 68,15,89,222 // mulps %xmm6,%xmm11
23182 .byte 69,15,88,217 // addps %xmm9,%xmm11
23183 .byte 68,15,40,203 // movaps %xmm3,%xmm9
23184 .byte 68,15,92,202 // subps %xmm2,%xmm9
23185 .byte 15,89,214 // mulps %xmm6,%xmm2
23186 .byte 68,15,40,231 // movaps %xmm7,%xmm12
23187 .byte 68,15,92,230 // subps %xmm6,%xmm12
23188 .byte 68,15,40,238 // movaps %xmm6,%xmm13
23189 .byte 69,15,88,237 // addps %xmm13,%xmm13
23190 .byte 68,15,194,239,2 // cmpleps %xmm7,%xmm13
23191 .byte 15,88,210 // addps %xmm2,%xmm2
23192 .byte 69,15,89,225 // mulps %xmm9,%xmm12
23193 .byte 69,15,88,228 // addps %xmm12,%xmm12
23194 .byte 69,15,92,212 // subps %xmm12,%xmm10
23195 .byte 65,15,84,213 // andps %xmm13,%xmm2
23196 .byte 69,15,85,234 // andnps %xmm10,%xmm13
23197 .byte 65,15,86,213 // orps %xmm13,%xmm2
23198 .byte 65,15,88,211 // addps %xmm11,%xmm2
23199 .byte 68,15,89,199 // mulps %xmm7,%xmm8
23200 .byte 65,15,88,216 // addps %xmm8,%xmm3
23201 .byte 72,173 // lods %ds:(%rsi),%rax
23202 .byte 255,224 // jmpq *%rax
23203
23204HIDDEN _sk_softlight_sse2
23205.globl _sk_softlight_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23206FUNCTION(_sk_softlight_sse2)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23207_sk_softlight_sse2:
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23208 .byte 15,41,84,36,232 // movaps %xmm2,-0x18(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23209 .byte 15,40,209 // movaps %xmm1,%xmm2
23210 .byte 68,15,40,192 // movaps %xmm0,%xmm8
23211 .byte 69,15,87,228 // xorps %xmm12,%xmm12
23212 .byte 68,15,194,231,1 // cmpltps %xmm7,%xmm12
23213 .byte 68,15,40,212 // movaps %xmm4,%xmm10
23214 .byte 68,15,94,215 // divps %xmm7,%xmm10
23215 .byte 69,15,84,212 // andps %xmm12,%xmm10
23216 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23217 .byte 102,68,15,110,200 // movd %eax,%xmm9
23218 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23219 .byte 69,15,40,241 // movaps %xmm9,%xmm14
23220 .byte 69,15,92,242 // subps %xmm10,%xmm14
23221 .byte 69,15,40,218 // movaps %xmm10,%xmm11
23222 .byte 69,15,40,234 // movaps %xmm10,%xmm13
23223 .byte 65,15,82,194 // rsqrtps %xmm10,%xmm0
23224 .byte 68,15,83,248 // rcpps %xmm0,%xmm15
23225 .byte 69,15,92,250 // subps %xmm10,%xmm15
23226 .byte 69,15,88,210 // addps %xmm10,%xmm10
23227 .byte 69,15,88,210 // addps %xmm10,%xmm10
23228 .byte 65,15,40,194 // movaps %xmm10,%xmm0
23229 .byte 15,89,192 // mulps %xmm0,%xmm0
23230 .byte 65,15,88,194 // addps %xmm10,%xmm0
23231 .byte 69,15,92,217 // subps %xmm9,%xmm11
23232 .byte 68,15,89,216 // mulps %xmm0,%xmm11
23233 .byte 184,0,0,224,64 // mov $0x40e00000,%eax
23234 .byte 102,68,15,110,208 // movd %eax,%xmm10
23235 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
23236 .byte 69,15,89,234 // mulps %xmm10,%xmm13
23237 .byte 69,15,88,235 // addps %xmm11,%xmm13
23238 .byte 68,15,40,219 // movaps %xmm3,%xmm11
23239 .byte 15,40,204 // movaps %xmm4,%xmm1
23240 .byte 68,15,89,217 // mulps %xmm1,%xmm11
23241 .byte 15,88,228 // addps %xmm4,%xmm4
23242 .byte 15,88,228 // addps %xmm4,%xmm4
23243 .byte 15,194,231,2 // cmpleps %xmm7,%xmm4
23244 .byte 68,15,84,236 // andps %xmm4,%xmm13
23245 .byte 65,15,85,231 // andnps %xmm15,%xmm4
23246 .byte 65,15,40,192 // movaps %xmm8,%xmm0
23247 .byte 15,88,192 // addps %xmm0,%xmm0
23248 .byte 65,15,86,229 // orps %xmm13,%xmm4
23249 .byte 68,15,40,232 // movaps %xmm0,%xmm13
23250 .byte 68,15,92,235 // subps %xmm3,%xmm13
23251 .byte 69,15,89,245 // mulps %xmm13,%xmm14
23252 .byte 68,15,89,239 // mulps %xmm7,%xmm13
23253 .byte 65,15,89,229 // mulps %xmm13,%xmm4
23254 .byte 65,15,88,227 // addps %xmm11,%xmm4
23255 .byte 69,15,40,217 // movaps %xmm9,%xmm11
23256 .byte 68,15,92,219 // subps %xmm3,%xmm11
23257 .byte 69,15,40,251 // movaps %xmm11,%xmm15
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23258 .byte 15,41,76,36,216 // movaps %xmm1,-0x28(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23259 .byte 68,15,89,249 // mulps %xmm1,%xmm15
23260 .byte 69,15,40,233 // movaps %xmm9,%xmm13
23261 .byte 68,15,92,239 // subps %xmm7,%xmm13
23262 .byte 69,15,89,197 // mulps %xmm13,%xmm8
23263 .byte 69,15,88,199 // addps %xmm15,%xmm8
23264 .byte 68,15,88,243 // addps %xmm3,%xmm14
23265 .byte 68,15,89,241 // mulps %xmm1,%xmm14
23266 .byte 15,194,195,2 // cmpleps %xmm3,%xmm0
23267 .byte 68,15,84,240 // andps %xmm0,%xmm14
23268 .byte 15,85,196 // andnps %xmm4,%xmm0
23269 .byte 65,15,86,198 // orps %xmm14,%xmm0
23270 .byte 65,15,88,192 // addps %xmm8,%xmm0
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23271 .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp)
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23272 .byte 68,15,40,197 // movaps %xmm5,%xmm8
23273 .byte 68,15,94,199 // divps %xmm7,%xmm8
23274 .byte 69,15,84,196 // andps %xmm12,%xmm8
23275 .byte 69,15,40,240 // movaps %xmm8,%xmm14
23276 .byte 69,15,88,246 // addps %xmm14,%xmm14
23277 .byte 69,15,88,246 // addps %xmm14,%xmm14
23278 .byte 65,15,40,230 // movaps %xmm14,%xmm4
23279 .byte 15,89,228 // mulps %xmm4,%xmm4
23280 .byte 65,15,88,230 // addps %xmm14,%xmm4
23281 .byte 69,15,40,248 // movaps %xmm8,%xmm15
23282 .byte 69,15,92,249 // subps %xmm9,%xmm15
23283 .byte 68,15,89,252 // mulps %xmm4,%xmm15
23284 .byte 69,15,40,241 // movaps %xmm9,%xmm14
23285 .byte 69,15,92,240 // subps %xmm8,%xmm14
23286 .byte 65,15,82,224 // rsqrtps %xmm8,%xmm4
23287 .byte 15,83,228 // rcpps %xmm4,%xmm4
23288 .byte 65,15,92,224 // subps %xmm8,%xmm4
23289 .byte 69,15,89,194 // mulps %xmm10,%xmm8
23290 .byte 69,15,88,199 // addps %xmm15,%xmm8
23291 .byte 68,15,40,253 // movaps %xmm5,%xmm15
23292 .byte 69,15,88,255 // addps %xmm15,%xmm15
23293 .byte 69,15,88,255 // addps %xmm15,%xmm15
23294 .byte 68,15,194,255,2 // cmpleps %xmm7,%xmm15
23295 .byte 69,15,84,199 // andps %xmm15,%xmm8
23296 .byte 68,15,85,252 // andnps %xmm4,%xmm15
23297 .byte 69,15,86,248 // orps %xmm8,%xmm15
23298 .byte 68,15,40,194 // movaps %xmm2,%xmm8
23299 .byte 69,15,88,192 // addps %xmm8,%xmm8
23300 .byte 65,15,40,224 // movaps %xmm8,%xmm4
23301 .byte 15,92,227 // subps %xmm3,%xmm4
23302 .byte 68,15,89,244 // mulps %xmm4,%xmm14
23303 .byte 15,89,231 // mulps %xmm7,%xmm4
23304 .byte 68,15,89,252 // mulps %xmm4,%xmm15
23305 .byte 15,40,227 // movaps %xmm3,%xmm4
23306 .byte 15,89,229 // mulps %xmm5,%xmm4
23307 .byte 68,15,88,252 // addps %xmm4,%xmm15
23308 .byte 65,15,40,227 // movaps %xmm11,%xmm4
23309 .byte 15,89,229 // mulps %xmm5,%xmm4
23310 .byte 65,15,89,213 // mulps %xmm13,%xmm2
23311 .byte 15,88,212 // addps %xmm4,%xmm2
23312 .byte 68,15,88,243 // addps %xmm3,%xmm14
23313 .byte 68,15,89,245 // mulps %xmm5,%xmm14
23314 .byte 68,15,194,195,2 // cmpleps %xmm3,%xmm8
23315 .byte 69,15,84,240 // andps %xmm8,%xmm14
23316 .byte 69,15,85,199 // andnps %xmm15,%xmm8
23317 .byte 69,15,86,198 // orps %xmm14,%xmm8
23318 .byte 68,15,88,194 // addps %xmm2,%xmm8
23319 .byte 68,15,40,246 // movaps %xmm6,%xmm14
23320 .byte 65,15,40,206 // movaps %xmm14,%xmm1
23321 .byte 15,94,207 // divps %xmm7,%xmm1
23322 .byte 65,15,84,204 // andps %xmm12,%xmm1
23323 .byte 15,40,225 // movaps %xmm1,%xmm4
23324 .byte 65,15,92,225 // subps %xmm9,%xmm4
23325 .byte 68,15,92,201 // subps %xmm1,%xmm9
23326 .byte 68,15,89,209 // mulps %xmm1,%xmm10
23327 .byte 15,82,241 // rsqrtps %xmm1,%xmm6
23328 .byte 15,83,246 // rcpps %xmm6,%xmm6
23329 .byte 15,92,241 // subps %xmm1,%xmm6
23330 .byte 15,88,201 // addps %xmm1,%xmm1
23331 .byte 15,88,201 // addps %xmm1,%xmm1
23332 .byte 15,40,233 // movaps %xmm1,%xmm5
23333 .byte 15,89,237 // mulps %xmm5,%xmm5
23334 .byte 15,88,233 // addps %xmm1,%xmm5
23335 .byte 15,89,236 // mulps %xmm4,%xmm5
23336 .byte 68,15,88,213 // addps %xmm5,%xmm10
23337 .byte 65,15,40,238 // movaps %xmm14,%xmm5
23338 .byte 15,40,205 // movaps %xmm5,%xmm1
23339 .byte 15,88,201 // addps %xmm1,%xmm1
23340 .byte 15,88,201 // addps %xmm1,%xmm1
23341 .byte 15,194,207,2 // cmpleps %xmm7,%xmm1
23342 .byte 68,15,84,209 // andps %xmm1,%xmm10
23343 .byte 15,85,206 // andnps %xmm6,%xmm1
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23344 .byte 15,40,84,36,232 // movaps -0x18(%rsp),%xmm2
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23345 .byte 68,15,89,234 // mulps %xmm2,%xmm13
23346 .byte 15,88,210 // addps %xmm2,%xmm2
23347 .byte 65,15,86,202 // orps %xmm10,%xmm1
23348 .byte 15,40,226 // movaps %xmm2,%xmm4
23349 .byte 15,92,227 // subps %xmm3,%xmm4
23350 .byte 68,15,89,204 // mulps %xmm4,%xmm9
23351 .byte 15,89,231 // mulps %xmm7,%xmm4
23352 .byte 15,89,204 // mulps %xmm4,%xmm1
23353 .byte 15,40,227 // movaps %xmm3,%xmm4
23354 .byte 15,89,229 // mulps %xmm5,%xmm4
23355 .byte 15,88,204 // addps %xmm4,%xmm1
23356 .byte 65,15,40,227 // movaps %xmm11,%xmm4
23357 .byte 15,89,229 // mulps %xmm5,%xmm4
23358 .byte 65,15,88,229 // addps %xmm13,%xmm4
23359 .byte 68,15,88,203 // addps %xmm3,%xmm9
23360 .byte 68,15,89,205 // mulps %xmm5,%xmm9
23361 .byte 15,40,245 // movaps %xmm5,%xmm6
23362 .byte 15,194,211,2 // cmpleps %xmm3,%xmm2
23363 .byte 68,15,84,202 // andps %xmm2,%xmm9
23364 .byte 15,85,209 // andnps %xmm1,%xmm2
23365 .byte 65,15,86,209 // orps %xmm9,%xmm2
23366 .byte 15,88,212 // addps %xmm4,%xmm2
23367 .byte 68,15,89,223 // mulps %xmm7,%xmm11
23368 .byte 65,15,88,219 // addps %xmm11,%xmm3
23369 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Klein15bf6782017-03-31 12:21:46 -0400[diff] [blame]23370 .byte 15,40,100,36,216 // movaps -0x28(%rsp),%xmm4
23371 .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23372 .byte 65,15,40,200 // movaps %xmm8,%xmm1
Mike Klein61b84162017-03-31 11:48:14 -0400[diff] [blame]23373 .byte 255,224 // jmpq *%rax
23374
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23375HIDDEN _sk_clamp_0_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23376.globl _sk_clamp_0_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23377FUNCTION(_sk_clamp_0_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23378_sk_clamp_0_sse2:
23379 .byte 69,15,87,192 // xorps %xmm8,%xmm8
23380 .byte 65,15,95,192 // maxps %xmm8,%xmm0
23381 .byte 65,15,95,200 // maxps %xmm8,%xmm1
23382 .byte 65,15,95,208 // maxps %xmm8,%xmm2
23383 .byte 65,15,95,216 // maxps %xmm8,%xmm3
23384 .byte 72,173 // lods %ds:(%rsi),%rax
23385 .byte 255,224 // jmpq *%rax
23386
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23387HIDDEN _sk_clamp_1_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23388.globl _sk_clamp_1_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23389FUNCTION(_sk_clamp_1_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23390_sk_clamp_1_sse2:
23391 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23392 .byte 102,68,15,110,192 // movd %eax,%xmm8
23393 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23394 .byte 65,15,93,192 // minps %xmm8,%xmm0
23395 .byte 65,15,93,200 // minps %xmm8,%xmm1
23396 .byte 65,15,93,208 // minps %xmm8,%xmm2
23397 .byte 65,15,93,216 // minps %xmm8,%xmm3
23398 .byte 72,173 // lods %ds:(%rsi),%rax
23399 .byte 255,224 // jmpq *%rax
23400
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23401HIDDEN _sk_clamp_a_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23402.globl _sk_clamp_a_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23403FUNCTION(_sk_clamp_a_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23404_sk_clamp_a_sse2:
23405 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23406 .byte 102,68,15,110,192 // movd %eax,%xmm8
23407 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23408 .byte 65,15,93,216 // minps %xmm8,%xmm3
23409 .byte 15,93,195 // minps %xmm3,%xmm0
23410 .byte 15,93,203 // minps %xmm3,%xmm1
23411 .byte 15,93,211 // minps %xmm3,%xmm2
23412 .byte 72,173 // lods %ds:(%rsi),%rax
23413 .byte 255,224 // jmpq *%rax
23414
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23415HIDDEN _sk_set_rgb_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23416.globl _sk_set_rgb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23417FUNCTION(_sk_set_rgb_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23418_sk_set_rgb_sse2:
23419 .byte 72,173 // lods %ds:(%rsi),%rax
23420 .byte 243,15,16,0 // movss (%rax),%xmm0
23421 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
23422 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
23423 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
23424 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
23425 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
23426 .byte 72,173 // lods %ds:(%rsi),%rax
23427 .byte 255,224 // jmpq *%rax
23428
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23429HIDDEN _sk_swap_rb_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23430.globl _sk_swap_rb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23431FUNCTION(_sk_swap_rb_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23432_sk_swap_rb_sse2:
23433 .byte 68,15,40,192 // movaps %xmm0,%xmm8
23434 .byte 72,173 // lods %ds:(%rsi),%rax
23435 .byte 15,40,194 // movaps %xmm2,%xmm0
23436 .byte 65,15,40,208 // movaps %xmm8,%xmm2
23437 .byte 255,224 // jmpq *%rax
23438
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23439HIDDEN _sk_swap_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23440.globl _sk_swap_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23441FUNCTION(_sk_swap_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23442_sk_swap_sse2:
23443 .byte 68,15,40,195 // movaps %xmm3,%xmm8
23444 .byte 68,15,40,202 // movaps %xmm2,%xmm9
23445 .byte 68,15,40,209 // movaps %xmm1,%xmm10
23446 .byte 68,15,40,216 // movaps %xmm0,%xmm11
23447 .byte 72,173 // lods %ds:(%rsi),%rax
23448 .byte 15,40,196 // movaps %xmm4,%xmm0
23449 .byte 15,40,205 // movaps %xmm5,%xmm1
23450 .byte 15,40,214 // movaps %xmm6,%xmm2
23451 .byte 15,40,223 // movaps %xmm7,%xmm3
23452 .byte 65,15,40,227 // movaps %xmm11,%xmm4
23453 .byte 65,15,40,234 // movaps %xmm10,%xmm5
23454 .byte 65,15,40,241 // movaps %xmm9,%xmm6
23455 .byte 65,15,40,248 // movaps %xmm8,%xmm7
23456 .byte 255,224 // jmpq *%rax
23457
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23458HIDDEN _sk_move_src_dst_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23459.globl _sk_move_src_dst_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23460FUNCTION(_sk_move_src_dst_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23461_sk_move_src_dst_sse2:
23462 .byte 72,173 // lods %ds:(%rsi),%rax
23463 .byte 15,40,224 // movaps %xmm0,%xmm4
23464 .byte 15,40,233 // movaps %xmm1,%xmm5
23465 .byte 15,40,242 // movaps %xmm2,%xmm6
23466 .byte 15,40,251 // movaps %xmm3,%xmm7
23467 .byte 255,224 // jmpq *%rax
23468
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23469HIDDEN _sk_move_dst_src_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23470.globl _sk_move_dst_src_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23471FUNCTION(_sk_move_dst_src_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23472_sk_move_dst_src_sse2:
23473 .byte 72,173 // lods %ds:(%rsi),%rax
23474 .byte 15,40,196 // movaps %xmm4,%xmm0
23475 .byte 15,40,205 // movaps %xmm5,%xmm1
23476 .byte 15,40,214 // movaps %xmm6,%xmm2
23477 .byte 15,40,223 // movaps %xmm7,%xmm3
23478 .byte 255,224 // jmpq *%rax
23479
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23480HIDDEN _sk_premul_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23481.globl _sk_premul_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23482FUNCTION(_sk_premul_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23483_sk_premul_sse2:
23484 .byte 15,89,195 // mulps %xmm3,%xmm0
23485 .byte 15,89,203 // mulps %xmm3,%xmm1
23486 .byte 15,89,211 // mulps %xmm3,%xmm2
23487 .byte 72,173 // lods %ds:(%rsi),%rax
23488 .byte 255,224 // jmpq *%rax
23489
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23490HIDDEN _sk_unpremul_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23491.globl _sk_unpremul_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23492FUNCTION(_sk_unpremul_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23493_sk_unpremul_sse2:
23494 .byte 69,15,87,192 // xorps %xmm8,%xmm8
23495 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23496 .byte 102,68,15,110,200 // movd %eax,%xmm9
23497 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23498 .byte 68,15,94,203 // divps %xmm3,%xmm9
23499 .byte 68,15,194,195,4 // cmpneqps %xmm3,%xmm8
23500 .byte 69,15,84,193 // andps %xmm9,%xmm8
23501 .byte 65,15,89,192 // mulps %xmm8,%xmm0
23502 .byte 65,15,89,200 // mulps %xmm8,%xmm1
23503 .byte 65,15,89,208 // mulps %xmm8,%xmm2
23504 .byte 72,173 // lods %ds:(%rsi),%rax
23505 .byte 255,224 // jmpq *%rax
23506
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23507HIDDEN _sk_from_srgb_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23508.globl _sk_from_srgb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23509FUNCTION(_sk_from_srgb_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23510_sk_from_srgb_sse2:
23511 .byte 184,145,131,158,61 // mov $0x3d9e8391,%eax
23512 .byte 102,68,15,110,192 // movd %eax,%xmm8
23513 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23514 .byte 69,15,40,232 // movaps %xmm8,%xmm13
23515 .byte 68,15,89,232 // mulps %xmm0,%xmm13
23516 .byte 68,15,40,224 // movaps %xmm0,%xmm12
23517 .byte 69,15,89,228 // mulps %xmm12,%xmm12
23518 .byte 184,154,153,153,62 // mov $0x3e99999a,%eax
23519 .byte 102,68,15,110,200 // movd %eax,%xmm9
23520 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23521 .byte 184,92,143,50,63 // mov $0x3f328f5c,%eax
23522 .byte 102,68,15,110,208 // movd %eax,%xmm10
23523 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
23524 .byte 69,15,40,241 // movaps %xmm9,%xmm14
23525 .byte 68,15,89,240 // mulps %xmm0,%xmm14
23526 .byte 69,15,88,242 // addps %xmm10,%xmm14
23527 .byte 184,10,215,35,59 // mov $0x3b23d70a,%eax
23528 .byte 102,68,15,110,216 // movd %eax,%xmm11
23529 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
23530 .byte 69,15,89,244 // mulps %xmm12,%xmm14
23531 .byte 69,15,88,243 // addps %xmm11,%xmm14
23532 .byte 184,174,71,97,61 // mov $0x3d6147ae,%eax
23533 .byte 102,68,15,110,224 // movd %eax,%xmm12
23534 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
23535 .byte 65,15,194,196,1 // cmpltps %xmm12,%xmm0
23536 .byte 68,15,84,232 // andps %xmm0,%xmm13
23537 .byte 65,15,85,198 // andnps %xmm14,%xmm0
23538 .byte 65,15,86,197 // orps %xmm13,%xmm0
23539 .byte 69,15,40,232 // movaps %xmm8,%xmm13
23540 .byte 68,15,89,233 // mulps %xmm1,%xmm13
23541 .byte 68,15,40,241 // movaps %xmm1,%xmm14
23542 .byte 69,15,89,246 // mulps %xmm14,%xmm14
23543 .byte 69,15,40,249 // movaps %xmm9,%xmm15
23544 .byte 68,15,89,249 // mulps %xmm1,%xmm15
23545 .byte 69,15,88,250 // addps %xmm10,%xmm15
23546 .byte 69,15,89,254 // mulps %xmm14,%xmm15
23547 .byte 69,15,88,251 // addps %xmm11,%xmm15
23548 .byte 65,15,194,204,1 // cmpltps %xmm12,%xmm1
23549 .byte 68,15,84,233 // andps %xmm1,%xmm13
23550 .byte 65,15,85,207 // andnps %xmm15,%xmm1
23551 .byte 65,15,86,205 // orps %xmm13,%xmm1
23552 .byte 68,15,89,194 // mulps %xmm2,%xmm8
23553 .byte 68,15,40,234 // movaps %xmm2,%xmm13
23554 .byte 69,15,89,237 // mulps %xmm13,%xmm13
23555 .byte 68,15,89,202 // mulps %xmm2,%xmm9
23556 .byte 69,15,88,202 // addps %xmm10,%xmm9
23557 .byte 69,15,89,205 // mulps %xmm13,%xmm9
23558 .byte 69,15,88,203 // addps %xmm11,%xmm9
23559 .byte 65,15,194,212,1 // cmpltps %xmm12,%xmm2
23560 .byte 68,15,84,194 // andps %xmm2,%xmm8
23561 .byte 65,15,85,209 // andnps %xmm9,%xmm2
23562 .byte 65,15,86,208 // orps %xmm8,%xmm2
23563 .byte 72,173 // lods %ds:(%rsi),%rax
23564 .byte 255,224 // jmpq *%rax
23565
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23566HIDDEN _sk_to_srgb_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23567.globl _sk_to_srgb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23568FUNCTION(_sk_to_srgb_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23569_sk_to_srgb_sse2:
23570 .byte 68,15,82,192 // rsqrtps %xmm0,%xmm8
23571 .byte 69,15,83,248 // rcpps %xmm8,%xmm15
23572 .byte 69,15,82,232 // rsqrtps %xmm8,%xmm13
23573 .byte 184,41,92,71,65 // mov $0x41475c29,%eax
23574 .byte 102,68,15,110,192 // movd %eax,%xmm8
23575 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23576 .byte 69,15,40,240 // movaps %xmm8,%xmm14
23577 .byte 68,15,89,240 // mulps %xmm0,%xmm14
23578 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23579 .byte 102,68,15,110,200 // movd %eax,%xmm9
23580 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23581 .byte 184,194,135,210,62 // mov $0x3ed287c2,%eax
23582 .byte 102,68,15,110,208 // movd %eax,%xmm10
23583 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
23584 .byte 184,206,111,48,63 // mov $0x3f306fce,%eax
23585 .byte 102,68,15,110,216 // movd %eax,%xmm11
23586 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
23587 .byte 184,168,87,202,61 // mov $0x3dca57a8,%eax
23588 .byte 53,0,0,0,128 // xor $0x80000000,%eax
23589 .byte 102,68,15,110,224 // movd %eax,%xmm12
23590 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
23591 .byte 69,15,89,251 // mulps %xmm11,%xmm15
23592 .byte 69,15,88,252 // addps %xmm12,%xmm15
23593 .byte 69,15,89,234 // mulps %xmm10,%xmm13
23594 .byte 69,15,88,239 // addps %xmm15,%xmm13
23595 .byte 69,15,40,249 // movaps %xmm9,%xmm15
23596 .byte 69,15,93,253 // minps %xmm13,%xmm15
23597 .byte 184,4,231,140,59 // mov $0x3b8ce704,%eax
23598 .byte 102,68,15,110,232 // movd %eax,%xmm13
23599 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
23600 .byte 65,15,194,197,1 // cmpltps %xmm13,%xmm0
23601 .byte 68,15,84,240 // andps %xmm0,%xmm14
23602 .byte 65,15,85,199 // andnps %xmm15,%xmm0
23603 .byte 65,15,86,198 // orps %xmm14,%xmm0
23604 .byte 68,15,82,241 // rsqrtps %xmm1,%xmm14
23605 .byte 69,15,83,254 // rcpps %xmm14,%xmm15
23606 .byte 69,15,82,246 // rsqrtps %xmm14,%xmm14
23607 .byte 69,15,89,251 // mulps %xmm11,%xmm15
23608 .byte 69,15,88,252 // addps %xmm12,%xmm15
23609 .byte 69,15,89,242 // mulps %xmm10,%xmm14
23610 .byte 69,15,88,247 // addps %xmm15,%xmm14
23611 .byte 69,15,40,249 // movaps %xmm9,%xmm15
23612 .byte 69,15,93,254 // minps %xmm14,%xmm15
23613 .byte 69,15,40,240 // movaps %xmm8,%xmm14
23614 .byte 68,15,89,241 // mulps %xmm1,%xmm14
23615 .byte 65,15,194,205,1 // cmpltps %xmm13,%xmm1
23616 .byte 68,15,84,241 // andps %xmm1,%xmm14
23617 .byte 65,15,85,207 // andnps %xmm15,%xmm1
23618 .byte 65,15,86,206 // orps %xmm14,%xmm1
23619 .byte 68,15,82,242 // rsqrtps %xmm2,%xmm14
23620 .byte 69,15,83,254 // rcpps %xmm14,%xmm15
23621 .byte 69,15,89,251 // mulps %xmm11,%xmm15
23622 .byte 69,15,88,252 // addps %xmm12,%xmm15
23623 .byte 69,15,82,222 // rsqrtps %xmm14,%xmm11
23624 .byte 69,15,89,218 // mulps %xmm10,%xmm11
23625 .byte 69,15,88,223 // addps %xmm15,%xmm11
23626 .byte 69,15,93,203 // minps %xmm11,%xmm9
23627 .byte 68,15,89,194 // mulps %xmm2,%xmm8
23628 .byte 65,15,194,213,1 // cmpltps %xmm13,%xmm2
23629 .byte 68,15,84,194 // andps %xmm2,%xmm8
23630 .byte 65,15,85,209 // andnps %xmm9,%xmm2
23631 .byte 65,15,86,208 // orps %xmm8,%xmm2
23632 .byte 72,173 // lods %ds:(%rsi),%rax
23633 .byte 255,224 // jmpq *%rax
23634
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]23635HIDDEN _sk_rgb_to_hsl_sse2
23636.globl _sk_rgb_to_hsl_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23637FUNCTION(_sk_rgb_to_hsl_sse2)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]23638_sk_rgb_to_hsl_sse2:
23639 .byte 68,15,40,201 // movaps %xmm1,%xmm9
23640 .byte 68,15,40,192 // movaps %xmm0,%xmm8
23641 .byte 69,15,40,208 // movaps %xmm8,%xmm10
23642 .byte 69,15,95,209 // maxps %xmm9,%xmm10
23643 .byte 68,15,95,210 // maxps %xmm2,%xmm10
23644 .byte 69,15,40,216 // movaps %xmm8,%xmm11
23645 .byte 69,15,93,217 // minps %xmm9,%xmm11
23646 .byte 68,15,93,218 // minps %xmm2,%xmm11
23647 .byte 65,15,40,202 // movaps %xmm10,%xmm1
23648 .byte 65,15,92,203 // subps %xmm11,%xmm1
23649 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23650 .byte 102,68,15,110,224 // movd %eax,%xmm12
23651 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
23652 .byte 68,15,94,225 // divps %xmm1,%xmm12
23653 .byte 65,184,171,170,42,62 // mov $0x3e2aaaab,%r8d
23654 .byte 65,15,40,194 // movaps %xmm10,%xmm0
23655 .byte 65,15,194,192,0 // cmpeqps %xmm8,%xmm0
23656 .byte 69,15,40,241 // movaps %xmm9,%xmm14
23657 .byte 68,15,194,242,1 // cmpltps %xmm2,%xmm14
23658 .byte 185,0,0,192,64 // mov $0x40c00000,%ecx
23659 .byte 102,68,15,110,233 // movd %ecx,%xmm13
23660 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
23661 .byte 69,15,84,238 // andps %xmm14,%xmm13
23662 .byte 69,15,40,241 // movaps %xmm9,%xmm14
23663 .byte 68,15,92,242 // subps %xmm2,%xmm14
23664 .byte 69,15,89,244 // mulps %xmm12,%xmm14
23665 .byte 69,15,88,238 // addps %xmm14,%xmm13
23666 .byte 69,15,40,242 // movaps %xmm10,%xmm14
23667 .byte 69,15,194,241,0 // cmpeqps %xmm9,%xmm14
23668 .byte 65,15,92,208 // subps %xmm8,%xmm2
23669 .byte 69,15,92,193 // subps %xmm9,%xmm8
23670 .byte 65,15,89,212 // mulps %xmm12,%xmm2
23671 .byte 185,0,0,0,64 // mov $0x40000000,%ecx
23672 .byte 69,15,89,196 // mulps %xmm12,%xmm8
23673 .byte 184,0,0,128,64 // mov $0x40800000,%eax
23674 .byte 102,68,15,110,200 // movd %eax,%xmm9
23675 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23676 .byte 69,15,88,200 // addps %xmm8,%xmm9
23677 .byte 102,68,15,110,193 // movd %ecx,%xmm8
23678 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23679 .byte 65,15,88,208 // addps %xmm8,%xmm2
23680 .byte 65,15,84,214 // andps %xmm14,%xmm2
23681 .byte 69,15,85,241 // andnps %xmm9,%xmm14
23682 .byte 68,15,86,242 // orps %xmm2,%xmm14
23683 .byte 68,15,84,232 // andps %xmm0,%xmm13
23684 .byte 65,15,85,198 // andnps %xmm14,%xmm0
23685 .byte 102,65,15,110,208 // movd %r8d,%xmm2
23686 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
23687 .byte 65,15,86,197 // orps %xmm13,%xmm0
23688 .byte 69,15,40,202 // movaps %xmm10,%xmm9
23689 .byte 69,15,194,203,4 // cmpneqps %xmm11,%xmm9
23690 .byte 65,15,84,193 // andps %xmm9,%xmm0
23691 .byte 15,89,194 // mulps %xmm2,%xmm0
23692 .byte 69,15,92,194 // subps %xmm10,%xmm8
23693 .byte 69,15,88,211 // addps %xmm11,%xmm10
23694 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
23695 .byte 102,68,15,110,224 // movd %eax,%xmm12
23696 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
23697 .byte 69,15,92,195 // subps %xmm11,%xmm8
23698 .byte 65,15,40,210 // movaps %xmm10,%xmm2
23699 .byte 65,15,89,212 // mulps %xmm12,%xmm2
23700 .byte 68,15,194,226,1 // cmpltps %xmm2,%xmm12
23701 .byte 69,15,84,196 // andps %xmm12,%xmm8
23702 .byte 69,15,85,226 // andnps %xmm10,%xmm12
23703 .byte 69,15,86,224 // orps %xmm8,%xmm12
23704 .byte 65,15,94,204 // divps %xmm12,%xmm1
23705 .byte 65,15,84,201 // andps %xmm9,%xmm1
23706 .byte 72,173 // lods %ds:(%rsi),%rax
23707 .byte 255,224 // jmpq *%rax
23708
23709HIDDEN _sk_hsl_to_rgb_sse2
23710.globl _sk_hsl_to_rgb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23711FUNCTION(_sk_hsl_to_rgb_sse2)
Mike Kleindb1cbcb2017-04-12 08:35:41 -0400[diff] [blame]23712_sk_hsl_to_rgb_sse2:
23713 .byte 15,41,124,36,232 // movaps %xmm7,-0x18(%rsp)
23714 .byte 15,41,116,36,216 // movaps %xmm6,-0x28(%rsp)
23715 .byte 15,41,108,36,200 // movaps %xmm5,-0x38(%rsp)
23716 .byte 15,41,100,36,184 // movaps %xmm4,-0x48(%rsp)
23717 .byte 15,41,92,36,168 // movaps %xmm3,-0x58(%rsp)
23718 .byte 68,15,40,210 // movaps %xmm2,%xmm10
23719 .byte 15,40,209 // movaps %xmm1,%xmm2
23720 .byte 15,40,240 // movaps %xmm0,%xmm6
23721 .byte 184,0,0,0,63 // mov $0x3f000000,%eax
23722 .byte 102,68,15,110,240 // movd %eax,%xmm14
23723 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
23724 .byte 69,15,40,202 // movaps %xmm10,%xmm9
23725 .byte 69,15,194,206,1 // cmpltps %xmm14,%xmm9
23726 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
23727 .byte 102,68,15,110,224 // movd %eax,%xmm12
23728 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
23729 .byte 65,15,40,196 // movaps %xmm12,%xmm0
23730 .byte 15,88,194 // addps %xmm2,%xmm0
23731 .byte 65,15,89,194 // mulps %xmm10,%xmm0
23732 .byte 15,40,218 // movaps %xmm2,%xmm3
23733 .byte 69,15,87,219 // xorps %xmm11,%xmm11
23734 .byte 68,15,194,218,0 // cmpeqps %xmm2,%xmm11
23735 .byte 65,15,88,210 // addps %xmm10,%xmm2
23736 .byte 65,15,89,218 // mulps %xmm10,%xmm3
23737 .byte 15,92,211 // subps %xmm3,%xmm2
23738 .byte 65,15,84,193 // andps %xmm9,%xmm0
23739 .byte 68,15,85,202 // andnps %xmm2,%xmm9
23740 .byte 68,15,86,200 // orps %xmm0,%xmm9
23741 .byte 184,0,0,0,64 // mov $0x40000000,%eax
23742 .byte 185,171,170,170,62 // mov $0x3eaaaaab,%ecx
23743 .byte 102,68,15,110,193 // movd %ecx,%xmm8
23744 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23745 .byte 68,15,41,68,36,152 // movaps %xmm8,-0x68(%rsp)
23746 .byte 68,15,88,198 // addps %xmm6,%xmm8
23747 .byte 185,0,0,0,0 // mov $0x0,%ecx
23748 .byte 102,15,110,233 // movd %ecx,%xmm5
23749 .byte 15,198,237,0 // shufps $0x0,%xmm5,%xmm5
23750 .byte 65,15,40,196 // movaps %xmm12,%xmm0
23751 .byte 65,15,194,192,1 // cmpltps %xmm8,%xmm0
23752 .byte 65,15,40,216 // movaps %xmm8,%xmm3
23753 .byte 65,15,92,220 // subps %xmm12,%xmm3
23754 .byte 15,84,216 // andps %xmm0,%xmm3
23755 .byte 65,15,85,192 // andnps %xmm8,%xmm0
23756 .byte 15,86,195 // orps %xmm3,%xmm0
23757 .byte 65,15,40,216 // movaps %xmm8,%xmm3
23758 .byte 15,194,221,1 // cmpltps %xmm5,%xmm3
23759 .byte 65,15,40,212 // movaps %xmm12,%xmm2
23760 .byte 65,15,88,208 // addps %xmm8,%xmm2
23761 .byte 15,84,211 // andps %xmm3,%xmm2
23762 .byte 15,85,216 // andnps %xmm0,%xmm3
23763 .byte 15,86,218 // orps %xmm2,%xmm3
23764 .byte 102,68,15,110,232 // movd %eax,%xmm13
23765 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
23766 .byte 69,15,89,234 // mulps %xmm10,%xmm13
23767 .byte 69,15,92,233 // subps %xmm9,%xmm13
23768 .byte 184,171,170,42,62 // mov $0x3e2aaaab,%eax
23769 .byte 65,15,40,193 // movaps %xmm9,%xmm0
23770 .byte 65,15,92,197 // subps %xmm13,%xmm0
23771 .byte 185,0,0,192,64 // mov $0x40c00000,%ecx
23772 .byte 102,68,15,110,249 // movd %ecx,%xmm15
23773 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15
23774 .byte 68,15,89,248 // mulps %xmm0,%xmm15
23775 .byte 185,171,170,42,63 // mov $0x3f2aaaab,%ecx
23776 .byte 102,15,110,225 // movd %ecx,%xmm4
23777 .byte 15,198,228,0 // shufps $0x0,%xmm4,%xmm4
23778 .byte 15,40,212 // movaps %xmm4,%xmm2
23779 .byte 15,92,211 // subps %xmm3,%xmm2
23780 .byte 15,40,203 // movaps %xmm3,%xmm1
23781 .byte 15,40,195 // movaps %xmm3,%xmm0
23782 .byte 15,194,220,1 // cmpltps %xmm4,%xmm3
23783 .byte 65,15,89,215 // mulps %xmm15,%xmm2
23784 .byte 65,15,88,213 // addps %xmm13,%xmm2
23785 .byte 15,84,211 // andps %xmm3,%xmm2
23786 .byte 65,15,85,221 // andnps %xmm13,%xmm3
23787 .byte 15,86,218 // orps %xmm2,%xmm3
23788 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
23789 .byte 65,15,40,209 // movaps %xmm9,%xmm2
23790 .byte 15,84,208 // andps %xmm0,%xmm2
23791 .byte 15,85,195 // andnps %xmm3,%xmm0
23792 .byte 15,86,194 // orps %xmm2,%xmm0
23793 .byte 102,15,110,248 // movd %eax,%xmm7
23794 .byte 15,198,255,0 // shufps $0x0,%xmm7,%xmm7
23795 .byte 15,194,207,1 // cmpltps %xmm7,%xmm1
23796 .byte 69,15,89,199 // mulps %xmm15,%xmm8
23797 .byte 69,15,88,197 // addps %xmm13,%xmm8
23798 .byte 68,15,84,193 // andps %xmm1,%xmm8
23799 .byte 15,85,200 // andnps %xmm0,%xmm1
23800 .byte 65,15,86,200 // orps %xmm8,%xmm1
23801 .byte 69,15,40,195 // movaps %xmm11,%xmm8
23802 .byte 68,15,85,193 // andnps %xmm1,%xmm8
23803 .byte 65,15,40,196 // movaps %xmm12,%xmm0
23804 .byte 15,194,198,1 // cmpltps %xmm6,%xmm0
23805 .byte 15,40,206 // movaps %xmm6,%xmm1
23806 .byte 65,15,92,204 // subps %xmm12,%xmm1
23807 .byte 15,84,200 // andps %xmm0,%xmm1
23808 .byte 15,85,198 // andnps %xmm6,%xmm0
23809 .byte 15,86,193 // orps %xmm1,%xmm0
23810 .byte 15,40,206 // movaps %xmm6,%xmm1
23811 .byte 15,194,205,1 // cmpltps %xmm5,%xmm1
23812 .byte 65,15,40,212 // movaps %xmm12,%xmm2
23813 .byte 15,88,214 // addps %xmm6,%xmm2
23814 .byte 15,84,209 // andps %xmm1,%xmm2
23815 .byte 15,85,200 // andnps %xmm0,%xmm1
23816 .byte 15,86,202 // orps %xmm2,%xmm1
23817 .byte 15,40,196 // movaps %xmm4,%xmm0
23818 .byte 15,92,193 // subps %xmm1,%xmm0
23819 .byte 15,40,217 // movaps %xmm1,%xmm3
23820 .byte 15,40,209 // movaps %xmm1,%xmm2
23821 .byte 15,194,204,1 // cmpltps %xmm4,%xmm1
23822 .byte 65,15,89,199 // mulps %xmm15,%xmm0
23823 .byte 65,15,88,197 // addps %xmm13,%xmm0
23824 .byte 15,84,193 // andps %xmm1,%xmm0
23825 .byte 65,15,85,205 // andnps %xmm13,%xmm1
23826 .byte 15,86,200 // orps %xmm0,%xmm1
23827 .byte 65,15,194,214,1 // cmpltps %xmm14,%xmm2
23828 .byte 65,15,40,193 // movaps %xmm9,%xmm0
23829 .byte 15,84,194 // andps %xmm2,%xmm0
23830 .byte 15,85,209 // andnps %xmm1,%xmm2
23831 .byte 15,86,208 // orps %xmm0,%xmm2
23832 .byte 15,194,223,1 // cmpltps %xmm7,%xmm3
23833 .byte 65,15,40,199 // movaps %xmm15,%xmm0
23834 .byte 15,89,198 // mulps %xmm6,%xmm0
23835 .byte 65,15,88,197 // addps %xmm13,%xmm0
23836 .byte 15,84,195 // andps %xmm3,%xmm0
23837 .byte 15,85,218 // andnps %xmm2,%xmm3
23838 .byte 15,86,216 // orps %xmm0,%xmm3
23839 .byte 65,15,40,203 // movaps %xmm11,%xmm1
23840 .byte 15,85,203 // andnps %xmm3,%xmm1
23841 .byte 15,92,116,36,152 // subps -0x68(%rsp),%xmm6
23842 .byte 15,40,198 // movaps %xmm6,%xmm0
23843 .byte 15,194,197,1 // cmpltps %xmm5,%xmm0
23844 .byte 15,40,214 // movaps %xmm6,%xmm2
23845 .byte 65,15,92,212 // subps %xmm12,%xmm2
23846 .byte 65,15,40,220 // movaps %xmm12,%xmm3
23847 .byte 68,15,194,230,1 // cmpltps %xmm6,%xmm12
23848 .byte 65,15,84,212 // andps %xmm12,%xmm2
23849 .byte 68,15,85,230 // andnps %xmm6,%xmm12
23850 .byte 68,15,86,226 // orps %xmm2,%xmm12
23851 .byte 15,88,222 // addps %xmm6,%xmm3
23852 .byte 15,84,216 // andps %xmm0,%xmm3
23853 .byte 65,15,85,196 // andnps %xmm12,%xmm0
23854 .byte 15,86,195 // orps %xmm3,%xmm0
23855 .byte 15,40,232 // movaps %xmm0,%xmm5
23856 .byte 15,194,239,1 // cmpltps %xmm7,%xmm5
23857 .byte 15,40,208 // movaps %xmm0,%xmm2
23858 .byte 15,194,212,1 // cmpltps %xmm4,%xmm2
23859 .byte 15,92,224 // subps %xmm0,%xmm4
23860 .byte 65,15,194,198,1 // cmpltps %xmm14,%xmm0
23861 .byte 65,15,89,247 // mulps %xmm15,%xmm6
23862 .byte 65,15,89,231 // mulps %xmm15,%xmm4
23863 .byte 65,15,88,245 // addps %xmm13,%xmm6
23864 .byte 65,15,88,229 // addps %xmm13,%xmm4
23865 .byte 15,84,226 // andps %xmm2,%xmm4
23866 .byte 65,15,85,213 // andnps %xmm13,%xmm2
23867 .byte 15,86,212 // orps %xmm4,%xmm2
23868 .byte 68,15,84,200 // andps %xmm0,%xmm9
23869 .byte 15,85,194 // andnps %xmm2,%xmm0
23870 .byte 65,15,86,193 // orps %xmm9,%xmm0
23871 .byte 15,84,245 // andps %xmm5,%xmm6
23872 .byte 15,85,232 // andnps %xmm0,%xmm5
23873 .byte 15,86,238 // orps %xmm6,%xmm5
23874 .byte 69,15,84,211 // andps %xmm11,%xmm10
23875 .byte 68,15,85,221 // andnps %xmm5,%xmm11
23876 .byte 69,15,86,194 // orps %xmm10,%xmm8
23877 .byte 65,15,86,202 // orps %xmm10,%xmm1
23878 .byte 69,15,86,211 // orps %xmm11,%xmm10
23879 .byte 72,173 // lods %ds:(%rsi),%rax
23880 .byte 65,15,40,192 // movaps %xmm8,%xmm0
23881 .byte 65,15,40,210 // movaps %xmm10,%xmm2
23882 .byte 15,40,92,36,168 // movaps -0x58(%rsp),%xmm3
23883 .byte 15,40,100,36,184 // movaps -0x48(%rsp),%xmm4
23884 .byte 15,40,108,36,200 // movaps -0x38(%rsp),%xmm5
23885 .byte 15,40,116,36,216 // movaps -0x28(%rsp),%xmm6
23886 .byte 15,40,124,36,232 // movaps -0x18(%rsp),%xmm7
23887 .byte 255,224 // jmpq *%rax
23888
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23889HIDDEN _sk_scale_1_float_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23890.globl _sk_scale_1_float_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23891FUNCTION(_sk_scale_1_float_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23892_sk_scale_1_float_sse2:
23893 .byte 72,173 // lods %ds:(%rsi),%rax
23894 .byte 243,68,15,16,0 // movss (%rax),%xmm8
23895 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23896 .byte 65,15,89,192 // mulps %xmm8,%xmm0
23897 .byte 65,15,89,200 // mulps %xmm8,%xmm1
23898 .byte 65,15,89,208 // mulps %xmm8,%xmm2
23899 .byte 65,15,89,216 // mulps %xmm8,%xmm3
23900 .byte 72,173 // lods %ds:(%rsi),%rax
23901 .byte 255,224 // jmpq *%rax
23902
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23903HIDDEN _sk_scale_u8_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23904.globl _sk_scale_u8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23905FUNCTION(_sk_scale_u8_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23906_sk_scale_u8_sse2:
23907 .byte 72,173 // lods %ds:(%rsi),%rax
23908 .byte 72,139,0 // mov (%rax),%rax
23909 .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8
23910 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
23911 .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8
23912 .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8
23913 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
23914 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
23915 .byte 102,68,15,110,200 // movd %eax,%xmm9
23916 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23917 .byte 69,15,89,200 // mulps %xmm8,%xmm9
23918 .byte 65,15,89,193 // mulps %xmm9,%xmm0
23919 .byte 65,15,89,201 // mulps %xmm9,%xmm1
23920 .byte 65,15,89,209 // mulps %xmm9,%xmm2
23921 .byte 65,15,89,217 // mulps %xmm9,%xmm3
23922 .byte 72,173 // lods %ds:(%rsi),%rax
23923 .byte 255,224 // jmpq *%rax
23924
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23925HIDDEN _sk_lerp_1_float_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23926.globl _sk_lerp_1_float_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23927FUNCTION(_sk_lerp_1_float_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23928_sk_lerp_1_float_sse2:
23929 .byte 72,173 // lods %ds:(%rsi),%rax
23930 .byte 243,68,15,16,0 // movss (%rax),%xmm8
23931 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
23932 .byte 15,92,196 // subps %xmm4,%xmm0
23933 .byte 65,15,89,192 // mulps %xmm8,%xmm0
23934 .byte 15,88,196 // addps %xmm4,%xmm0
23935 .byte 15,92,205 // subps %xmm5,%xmm1
23936 .byte 65,15,89,200 // mulps %xmm8,%xmm1
23937 .byte 15,88,205 // addps %xmm5,%xmm1
23938 .byte 15,92,214 // subps %xmm6,%xmm2
23939 .byte 65,15,89,208 // mulps %xmm8,%xmm2
23940 .byte 15,88,214 // addps %xmm6,%xmm2
23941 .byte 15,92,223 // subps %xmm7,%xmm3
23942 .byte 65,15,89,216 // mulps %xmm8,%xmm3
23943 .byte 15,88,223 // addps %xmm7,%xmm3
23944 .byte 72,173 // lods %ds:(%rsi),%rax
23945 .byte 255,224 // jmpq *%rax
23946
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23947HIDDEN _sk_lerp_u8_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23948.globl _sk_lerp_u8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23949FUNCTION(_sk_lerp_u8_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23950_sk_lerp_u8_sse2:
23951 .byte 72,173 // lods %ds:(%rsi),%rax
23952 .byte 72,139,0 // mov (%rax),%rax
23953 .byte 102,68,15,110,4,56 // movd (%rax,%rdi,1),%xmm8
23954 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
23955 .byte 102,69,15,96,193 // punpcklbw %xmm9,%xmm8
23956 .byte 102,69,15,97,193 // punpcklwd %xmm9,%xmm8
23957 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
23958 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
23959 .byte 102,68,15,110,200 // movd %eax,%xmm9
23960 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
23961 .byte 69,15,89,200 // mulps %xmm8,%xmm9
23962 .byte 15,92,196 // subps %xmm4,%xmm0
23963 .byte 65,15,89,193 // mulps %xmm9,%xmm0
23964 .byte 15,88,196 // addps %xmm4,%xmm0
23965 .byte 15,92,205 // subps %xmm5,%xmm1
23966 .byte 65,15,89,201 // mulps %xmm9,%xmm1
23967 .byte 15,88,205 // addps %xmm5,%xmm1
23968 .byte 15,92,214 // subps %xmm6,%xmm2
23969 .byte 65,15,89,209 // mulps %xmm9,%xmm2
23970 .byte 15,88,214 // addps %xmm6,%xmm2
23971 .byte 15,92,223 // subps %xmm7,%xmm3
23972 .byte 65,15,89,217 // mulps %xmm9,%xmm3
23973 .byte 15,88,223 // addps %xmm7,%xmm3
23974 .byte 72,173 // lods %ds:(%rsi),%rax
23975 .byte 255,224 // jmpq *%rax
23976
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]23977HIDDEN _sk_lerp_565_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23978.globl _sk_lerp_565_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]23979FUNCTION(_sk_lerp_565_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]23980_sk_lerp_565_sse2:
23981 .byte 72,173 // lods %ds:(%rsi),%rax
23982 .byte 72,139,0 // mov (%rax),%rax
23983 .byte 243,68,15,126,4,120 // movq (%rax,%rdi,2),%xmm8
23984 .byte 102,15,239,219 // pxor %xmm3,%xmm3
23985 .byte 102,68,15,97,195 // punpcklwd %xmm3,%xmm8
23986 .byte 184,0,248,0,0 // mov $0xf800,%eax
23987 .byte 102,15,110,216 // movd %eax,%xmm3
23988 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
23989 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
23990 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
23991 .byte 184,8,33,132,55 // mov $0x37842108,%eax
23992 .byte 102,68,15,110,208 // movd %eax,%xmm10
23993 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
23994 .byte 69,15,89,209 // mulps %xmm9,%xmm10
23995 .byte 184,224,7,0,0 // mov $0x7e0,%eax
23996 .byte 102,15,110,216 // movd %eax,%xmm3
23997 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
23998 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
23999 .byte 68,15,91,203 // cvtdq2ps %xmm3,%xmm9
24000 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
24001 .byte 102,68,15,110,216 // movd %eax,%xmm11
24002 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24003 .byte 69,15,89,217 // mulps %xmm9,%xmm11
24004 .byte 184,31,0,0,0 // mov $0x1f,%eax
24005 .byte 102,15,110,216 // movd %eax,%xmm3
24006 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
24007 .byte 102,65,15,219,216 // pand %xmm8,%xmm3
24008 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
24009 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
24010 .byte 102,15,110,216 // movd %eax,%xmm3
24011 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
24012 .byte 65,15,89,216 // mulps %xmm8,%xmm3
24013 .byte 15,92,196 // subps %xmm4,%xmm0
24014 .byte 65,15,89,194 // mulps %xmm10,%xmm0
24015 .byte 15,88,196 // addps %xmm4,%xmm0
24016 .byte 15,92,205 // subps %xmm5,%xmm1
24017 .byte 65,15,89,203 // mulps %xmm11,%xmm1
24018 .byte 15,88,205 // addps %xmm5,%xmm1
24019 .byte 15,92,214 // subps %xmm6,%xmm2
24020 .byte 15,89,211 // mulps %xmm3,%xmm2
24021 .byte 15,88,214 // addps %xmm6,%xmm2
24022 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24023 .byte 102,15,110,216 // movd %eax,%xmm3
24024 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
24025 .byte 72,173 // lods %ds:(%rsi),%rax
24026 .byte 255,224 // jmpq *%rax
24027
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]24028HIDDEN _sk_load_tables_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]24029.globl _sk_load_tables_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]24030FUNCTION(_sk_load_tables_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]24031_sk_load_tables_sse2:
24032 .byte 72,173 // lods %ds:(%rsi),%rax
24033 .byte 72,139,8 // mov (%rax),%rcx
24034 .byte 76,139,64,8 // mov 0x8(%rax),%r8
24035 .byte 243,68,15,111,4,185 // movdqu (%rcx,%rdi,4),%xmm8
24036 .byte 185,255,0,0,0 // mov $0xff,%ecx
24037 .byte 102,15,110,193 // movd %ecx,%xmm0
24038 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
24039 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
24040 .byte 102,65,15,114,209,8 // psrld $0x8,%xmm9
24041 .byte 102,68,15,219,200 // pand %xmm0,%xmm9
24042 .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
24043 .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10
24044 .byte 102,68,15,219,208 // pand %xmm0,%xmm10
24045 .byte 102,65,15,219,192 // pand %xmm8,%xmm0
24046 .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3
24047 .byte 102,72,15,126,217 // movq %xmm3,%rcx
24048 .byte 65,137,201 // mov %ecx,%r9d
24049 .byte 72,193,233,32 // shr $0x20,%rcx
24050 .byte 102,73,15,126,194 // movq %xmm0,%r10
24051 .byte 69,137,211 // mov %r10d,%r11d
24052 .byte 73,193,234,32 // shr $0x20,%r10
24053 .byte 243,67,15,16,28,144 // movss (%r8,%r10,4),%xmm3
24054 .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0
24055 .byte 15,20,216 // unpcklps %xmm0,%xmm3
24056 .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0
24057 .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1
24058 .byte 15,20,193 // unpcklps %xmm1,%xmm0
24059 .byte 15,20,195 // unpcklps %xmm3,%xmm0
24060 .byte 76,139,64,16 // mov 0x10(%rax),%r8
24061 .byte 102,65,15,112,201,78 // pshufd $0x4e,%xmm9,%xmm1
24062 .byte 102,73,15,126,202 // movq %xmm1,%r10
24063 .byte 77,137,209 // mov %r10,%r9
24064 .byte 73,193,233,32 // shr $0x20,%r9
24065 .byte 102,76,15,126,201 // movq %xmm9,%rcx
24066 .byte 65,137,203 // mov %ecx,%r11d
24067 .byte 65,129,227,255,255,255,0 // and $0xffffff,%r11d
24068 .byte 72,193,233,30 // shr $0x1e,%rcx
24069 .byte 65,129,226,255,255,255,0 // and $0xffffff,%r10d
24070 .byte 243,65,15,16,28,8 // movss (%r8,%rcx,1),%xmm3
24071 .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1
24072 .byte 15,20,217 // unpcklps %xmm1,%xmm3
24073 .byte 243,67,15,16,12,152 // movss (%r8,%r11,4),%xmm1
24074 .byte 243,67,15,16,20,144 // movss (%r8,%r10,4),%xmm2
24075 .byte 15,20,202 // unpcklps %xmm2,%xmm1
24076 .byte 15,20,203 // unpcklps %xmm3,%xmm1
24077 .byte 76,139,72,24 // mov 0x18(%rax),%r9
24078 .byte 102,65,15,112,210,78 // pshufd $0x4e,%xmm10,%xmm2
24079 .byte 102,72,15,126,209 // movq %xmm2,%rcx
24080 .byte 68,15,183,193 // movzwl %cx,%r8d
24081 .byte 72,193,233,32 // shr $0x20,%rcx
24082 .byte 102,76,15,126,208 // movq %xmm10,%rax
24083 .byte 68,15,183,208 // movzwl %ax,%r10d
24084 .byte 72,193,232,30 // shr $0x1e,%rax
24085 .byte 243,69,15,16,12,1 // movss (%r9,%rax,1),%xmm9
24086 .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2
24087 .byte 68,15,20,202 // unpcklps %xmm2,%xmm9
24088 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
24089 .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3
24090 .byte 15,20,211 // unpcklps %xmm3,%xmm2
24091 .byte 65,15,20,209 // unpcklps %xmm9,%xmm2
24092 .byte 102,65,15,114,208,24 // psrld $0x18,%xmm8
24093 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24094 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
24095 .byte 102,15,110,216 // movd %eax,%xmm3
24096 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
24097 .byte 65,15,89,216 // mulps %xmm8,%xmm3
24098 .byte 72,173 // lods %ds:(%rsi),%rax
24099 .byte 255,224 // jmpq *%rax
24100
Mike Kleina3735cd2017-04-17 13:19:05 -0400[diff] [blame]24101HIDDEN _sk_load_tables_u16_be_sse2
24102.globl _sk_load_tables_u16_be_sse2
24103FUNCTION(_sk_load_tables_u16_be_sse2)
24104_sk_load_tables_u16_be_sse2:
24105 .byte 72,173 // lods %ds:(%rsi),%rax
24106 .byte 72,139,8 // mov (%rax),%rcx
24107 .byte 76,139,64,8 // mov 0x8(%rax),%r8
24108 .byte 243,15,111,4,249 // movdqu (%rcx,%rdi,8),%xmm0
24109 .byte 243,15,111,76,249,16 // movdqu 0x10(%rcx,%rdi,8),%xmm1
24110 .byte 102,68,15,111,208 // movdqa %xmm0,%xmm10
24111 .byte 102,68,15,97,209 // punpcklwd %xmm1,%xmm10
24112 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
24113 .byte 102,65,15,111,202 // movdqa %xmm10,%xmm1
24114 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
24115 .byte 102,68,15,105,208 // punpckhwd %xmm0,%xmm10
24116 .byte 102,69,15,239,192 // pxor %xmm8,%xmm8
24117 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
24118 .byte 102,65,15,97,192 // punpcklwd %xmm8,%xmm0
24119 .byte 185,255,0,0,0 // mov $0xff,%ecx
24120 .byte 102,15,110,217 // movd %ecx,%xmm3
24121 .byte 102,68,15,112,203,0 // pshufd $0x0,%xmm3,%xmm9
24122 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
24123 .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3
24124 .byte 102,72,15,126,217 // movq %xmm3,%rcx
24125 .byte 68,15,183,201 // movzwl %cx,%r9d
24126 .byte 72,193,233,32 // shr $0x20,%rcx
24127 .byte 102,73,15,126,194 // movq %xmm0,%r10
24128 .byte 69,15,183,218 // movzwl %r10w,%r11d
24129 .byte 73,193,234,30 // shr $0x1e,%r10
24130 .byte 243,67,15,16,28,16 // movss (%r8,%r10,1),%xmm3
24131 .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0
24132 .byte 15,20,216 // unpcklps %xmm0,%xmm3
24133 .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0
24134 .byte 243,67,15,16,20,136 // movss (%r8,%r9,4),%xmm2
24135 .byte 15,20,194 // unpcklps %xmm2,%xmm0
24136 .byte 15,20,195 // unpcklps %xmm3,%xmm0
24137 .byte 76,139,80,16 // mov 0x10(%rax),%r10
24138 .byte 102,15,115,217,8 // psrldq $0x8,%xmm1
24139 .byte 102,65,15,97,200 // punpcklwd %xmm8,%xmm1
24140 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
24141 .byte 102,15,112,209,78 // pshufd $0x4e,%xmm1,%xmm2
24142 .byte 102,73,15,126,208 // movq %xmm2,%r8
24143 .byte 69,15,183,200 // movzwl %r8w,%r9d
24144 .byte 73,193,232,32 // shr $0x20,%r8
24145 .byte 102,72,15,126,201 // movq %xmm1,%rcx
24146 .byte 68,15,183,217 // movzwl %cx,%r11d
24147 .byte 72,193,233,30 // shr $0x1e,%rcx
24148 .byte 243,65,15,16,20,10 // movss (%r10,%rcx,1),%xmm2
24149 .byte 243,67,15,16,12,130 // movss (%r10,%r8,4),%xmm1
24150 .byte 15,20,209 // unpcklps %xmm1,%xmm2
24151 .byte 243,67,15,16,12,154 // movss (%r10,%r11,4),%xmm1
24152 .byte 243,67,15,16,28,138 // movss (%r10,%r9,4),%xmm3
24153 .byte 15,20,203 // unpcklps %xmm3,%xmm1
24154 .byte 15,20,202 // unpcklps %xmm2,%xmm1
24155 .byte 76,139,72,24 // mov 0x18(%rax),%r9
24156 .byte 102,69,15,112,218,78 // pshufd $0x4e,%xmm10,%xmm11
24157 .byte 102,69,15,97,208 // punpcklwd %xmm8,%xmm10
24158 .byte 102,69,15,219,209 // pand %xmm9,%xmm10
24159 .byte 102,65,15,112,210,78 // pshufd $0x4e,%xmm10,%xmm2
24160 .byte 102,72,15,126,209 // movq %xmm2,%rcx
24161 .byte 68,15,183,193 // movzwl %cx,%r8d
24162 .byte 72,193,233,32 // shr $0x20,%rcx
24163 .byte 102,76,15,126,208 // movq %xmm10,%rax
24164 .byte 68,15,183,208 // movzwl %ax,%r10d
24165 .byte 72,193,232,30 // shr $0x1e,%rax
24166 .byte 243,69,15,16,12,1 // movss (%r9,%rax,1),%xmm9
24167 .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2
24168 .byte 68,15,20,202 // unpcklps %xmm2,%xmm9
24169 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
24170 .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3
24171 .byte 15,20,211 // unpcklps %xmm3,%xmm2
24172 .byte 65,15,20,209 // unpcklps %xmm9,%xmm2
24173 .byte 184,128,0,128,55 // mov $0x37800080,%eax
24174 .byte 102,68,15,110,200 // movd %eax,%xmm9
24175 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24176 .byte 102,65,15,111,219 // movdqa %xmm11,%xmm3
24177 .byte 102,15,113,243,8 // psllw $0x8,%xmm3
24178 .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11
24179 .byte 102,68,15,235,219 // por %xmm3,%xmm11
24180 .byte 102,69,15,97,216 // punpcklwd %xmm8,%xmm11
24181 .byte 65,15,91,219 // cvtdq2ps %xmm11,%xmm3
24182 .byte 65,15,89,217 // mulps %xmm9,%xmm3
24183 .byte 72,173 // lods %ds:(%rsi),%rax
24184 .byte 255,224 // jmpq *%rax
24185
24186HIDDEN _sk_load_tables_rgb_u16_be_sse2
24187.globl _sk_load_tables_rgb_u16_be_sse2
24188FUNCTION(_sk_load_tables_rgb_u16_be_sse2)
24189_sk_load_tables_rgb_u16_be_sse2:
24190 .byte 72,173 // lods %ds:(%rsi),%rax
24191 .byte 76,141,12,127 // lea (%rdi,%rdi,2),%r9
24192 .byte 72,139,8 // mov (%rax),%rcx
24193 .byte 76,139,64,8 // mov 0x8(%rax),%r8
24194 .byte 243,70,15,111,28,73 // movdqu (%rcx,%r9,2),%xmm11
24195 .byte 243,66,15,111,68,73,8 // movdqu 0x8(%rcx,%r9,2),%xmm0
24196 .byte 102,15,115,216,4 // psrldq $0x4,%xmm0
24197 .byte 102,65,15,111,219 // movdqa %xmm11,%xmm3
24198 .byte 102,15,115,219,6 // psrldq $0x6,%xmm3
24199 .byte 102,68,15,97,216 // punpcklwd %xmm0,%xmm11
24200 .byte 102,15,115,216,6 // psrldq $0x6,%xmm0
24201 .byte 102,15,97,216 // punpcklwd %xmm0,%xmm3
24202 .byte 102,65,15,111,195 // movdqa %xmm11,%xmm0
24203 .byte 102,15,97,195 // punpcklwd %xmm3,%xmm0
24204 .byte 102,68,15,112,208,78 // pshufd $0x4e,%xmm0,%xmm10
24205 .byte 102,68,15,105,219 // punpckhwd %xmm3,%xmm11
24206 .byte 102,69,15,239,192 // pxor %xmm8,%xmm8
24207 .byte 102,65,15,97,192 // punpcklwd %xmm8,%xmm0
24208 .byte 185,255,0,0,0 // mov $0xff,%ecx
24209 .byte 102,15,110,217 // movd %ecx,%xmm3
24210 .byte 102,68,15,112,203,0 // pshufd $0x0,%xmm3,%xmm9
24211 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
24212 .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3
24213 .byte 102,72,15,126,217 // movq %xmm3,%rcx
24214 .byte 68,15,183,201 // movzwl %cx,%r9d
24215 .byte 72,193,233,32 // shr $0x20,%rcx
24216 .byte 102,73,15,126,194 // movq %xmm0,%r10
24217 .byte 69,15,183,218 // movzwl %r10w,%r11d
24218 .byte 73,193,234,30 // shr $0x1e,%r10
24219 .byte 243,67,15,16,28,16 // movss (%r8,%r10,1),%xmm3
24220 .byte 243,65,15,16,4,136 // movss (%r8,%rcx,4),%xmm0
24221 .byte 15,20,216 // unpcklps %xmm0,%xmm3
24222 .byte 243,67,15,16,4,152 // movss (%r8,%r11,4),%xmm0
24223 .byte 243,67,15,16,12,136 // movss (%r8,%r9,4),%xmm1
24224 .byte 15,20,193 // unpcklps %xmm1,%xmm0
24225 .byte 15,20,195 // unpcklps %xmm3,%xmm0
24226 .byte 76,139,80,16 // mov 0x10(%rax),%r10
24227 .byte 102,69,15,97,208 // punpcklwd %xmm8,%xmm10
24228 .byte 102,69,15,219,209 // pand %xmm9,%xmm10
24229 .byte 102,65,15,112,202,78 // pshufd $0x4e,%xmm10,%xmm1
24230 .byte 102,73,15,126,200 // movq %xmm1,%r8
24231 .byte 69,15,183,200 // movzwl %r8w,%r9d
24232 .byte 73,193,232,32 // shr $0x20,%r8
24233 .byte 102,76,15,126,209 // movq %xmm10,%rcx
24234 .byte 68,15,183,217 // movzwl %cx,%r11d
24235 .byte 72,193,233,30 // shr $0x1e,%rcx
24236 .byte 243,65,15,16,28,10 // movss (%r10,%rcx,1),%xmm3
24237 .byte 243,67,15,16,12,130 // movss (%r10,%r8,4),%xmm1
24238 .byte 15,20,217 // unpcklps %xmm1,%xmm3
24239 .byte 243,67,15,16,12,154 // movss (%r10,%r11,4),%xmm1
24240 .byte 243,67,15,16,20,138 // movss (%r10,%r9,4),%xmm2
24241 .byte 15,20,202 // unpcklps %xmm2,%xmm1
24242 .byte 15,20,203 // unpcklps %xmm3,%xmm1
24243 .byte 76,139,72,24 // mov 0x18(%rax),%r9
24244 .byte 102,69,15,97,216 // punpcklwd %xmm8,%xmm11
24245 .byte 102,69,15,219,217 // pand %xmm9,%xmm11
24246 .byte 102,65,15,112,211,78 // pshufd $0x4e,%xmm11,%xmm2
24247 .byte 102,72,15,126,209 // movq %xmm2,%rcx
24248 .byte 68,15,183,193 // movzwl %cx,%r8d
24249 .byte 72,193,233,32 // shr $0x20,%rcx
24250 .byte 102,76,15,126,216 // movq %xmm11,%rax
24251 .byte 68,15,183,208 // movzwl %ax,%r10d
24252 .byte 72,193,232,30 // shr $0x1e,%rax
24253 .byte 243,69,15,16,4,1 // movss (%r9,%rax,1),%xmm8
24254 .byte 243,65,15,16,20,137 // movss (%r9,%rcx,4),%xmm2
24255 .byte 68,15,20,194 // unpcklps %xmm2,%xmm8
24256 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
24257 .byte 243,67,15,16,28,129 // movss (%r9,%r8,4),%xmm3
24258 .byte 15,20,211 // unpcklps %xmm3,%xmm2
24259 .byte 65,15,20,208 // unpcklps %xmm8,%xmm2
24260 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24261 .byte 102,15,110,216 // movd %eax,%xmm3
24262 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
24263 .byte 72,173 // lods %ds:(%rsi),%rax
24264 .byte 255,224 // jmpq *%rax
24265
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]24266HIDDEN _sk_byte_tables_sse2
24267.globl _sk_byte_tables_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]24268FUNCTION(_sk_byte_tables_sse2)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]24269_sk_byte_tables_sse2:
24270 .byte 65,86 // push %r14
24271 .byte 83 // push %rbx
24272 .byte 72,173 // lods %ds:(%rsi),%rax
24273 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
24274 .byte 102,68,15,110,193 // movd %ecx,%xmm8
24275 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24276 .byte 65,15,89,192 // mulps %xmm8,%xmm0
24277 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0
24278 .byte 102,72,15,126,193 // movq %xmm0,%rcx
24279 .byte 65,137,200 // mov %ecx,%r8d
24280 .byte 72,193,233,32 // shr $0x20,%rcx
24281 .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0
24282 .byte 102,73,15,126,193 // movq %xmm0,%r9
24283 .byte 69,137,202 // mov %r9d,%r10d
24284 .byte 77,137,203 // mov %r9,%r11
24285 .byte 73,193,235,32 // shr $0x20,%r11
24286 .byte 76,139,48 // mov (%rax),%r14
24287 .byte 76,139,72,8 // mov 0x8(%rax),%r9
24288 .byte 71,15,182,20,22 // movzbl (%r14,%r10,1),%r10d
24289 .byte 67,15,182,28,30 // movzbl (%r14,%r11,1),%ebx
24290 .byte 193,227,8 // shl $0x8,%ebx
24291 .byte 68,9,211 // or %r10d,%ebx
24292 .byte 71,15,182,4,6 // movzbl (%r14,%r8,1),%r8d
24293 .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx
24294 .byte 193,225,8 // shl $0x8,%ecx
24295 .byte 68,9,193 // or %r8d,%ecx
24296 .byte 102,15,196,193,0 // pinsrw $0x0,%ecx,%xmm0
24297 .byte 102,15,196,195,1 // pinsrw $0x1,%ebx,%xmm0
24298 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
24299 .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0
24300 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0
24301 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
24302 .byte 185,129,128,128,59 // mov $0x3b808081,%ecx
24303 .byte 102,68,15,110,209 // movd %ecx,%xmm10
24304 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24305 .byte 65,15,89,194 // mulps %xmm10,%xmm0
24306 .byte 65,15,89,200 // mulps %xmm8,%xmm1
24307 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
24308 .byte 102,72,15,126,201 // movq %xmm1,%rcx
24309 .byte 65,137,200 // mov %ecx,%r8d
24310 .byte 72,193,233,32 // shr $0x20,%rcx
24311 .byte 102,15,112,201,78 // pshufd $0x4e,%xmm1,%xmm1
24312 .byte 102,72,15,126,203 // movq %xmm1,%rbx
24313 .byte 65,137,218 // mov %ebx,%r10d
24314 .byte 72,193,235,32 // shr $0x20,%rbx
24315 .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d
24316 .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx
24317 .byte 193,227,8 // shl $0x8,%ebx
24318 .byte 68,9,211 // or %r10d,%ebx
24319 .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d
24320 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
24321 .byte 193,225,8 // shl $0x8,%ecx
24322 .byte 68,9,193 // or %r8d,%ecx
24323 .byte 102,15,196,201,0 // pinsrw $0x0,%ecx,%xmm1
24324 .byte 102,15,196,203,1 // pinsrw $0x1,%ebx,%xmm1
24325 .byte 102,65,15,96,201 // punpcklbw %xmm9,%xmm1
24326 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1
24327 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
24328 .byte 65,15,89,202 // mulps %xmm10,%xmm1
24329 .byte 76,139,72,16 // mov 0x10(%rax),%r9
24330 .byte 65,15,89,208 // mulps %xmm8,%xmm2
24331 .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2
24332 .byte 102,72,15,126,211 // movq %xmm2,%rbx
24333 .byte 65,137,216 // mov %ebx,%r8d
24334 .byte 72,193,235,32 // shr $0x20,%rbx
24335 .byte 102,15,112,210,78 // pshufd $0x4e,%xmm2,%xmm2
24336 .byte 102,72,15,126,209 // movq %xmm2,%rcx
24337 .byte 65,137,202 // mov %ecx,%r10d
24338 .byte 72,193,233,32 // shr $0x20,%rcx
24339 .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d
24340 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
24341 .byte 193,225,8 // shl $0x8,%ecx
24342 .byte 68,9,209 // or %r10d,%ecx
24343 .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d
24344 .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx
24345 .byte 193,227,8 // shl $0x8,%ebx
24346 .byte 68,9,195 // or %r8d,%ebx
24347 .byte 102,15,196,211,0 // pinsrw $0x0,%ebx,%xmm2
24348 .byte 102,15,196,209,1 // pinsrw $0x1,%ecx,%xmm2
24349 .byte 102,65,15,96,209 // punpcklbw %xmm9,%xmm2
24350 .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2
24351 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
24352 .byte 65,15,89,210 // mulps %xmm10,%xmm2
24353 .byte 72,139,64,24 // mov 0x18(%rax),%rax
24354 .byte 68,15,89,195 // mulps %xmm3,%xmm8
24355 .byte 102,65,15,91,216 // cvtps2dq %xmm8,%xmm3
24356 .byte 102,72,15,126,217 // movq %xmm3,%rcx
24357 .byte 65,137,200 // mov %ecx,%r8d
24358 .byte 72,193,233,32 // shr $0x20,%rcx
24359 .byte 102,15,112,219,78 // pshufd $0x4e,%xmm3,%xmm3
24360 .byte 102,72,15,126,219 // movq %xmm3,%rbx
24361 .byte 65,137,217 // mov %ebx,%r9d
24362 .byte 72,193,235,32 // shr $0x20,%rbx
24363 .byte 70,15,182,12,8 // movzbl (%rax,%r9,1),%r9d
24364 .byte 15,182,28,24 // movzbl (%rax,%rbx,1),%ebx
24365 .byte 193,227,8 // shl $0x8,%ebx
24366 .byte 68,9,203 // or %r9d,%ebx
24367 .byte 70,15,182,4,0 // movzbl (%rax,%r8,1),%r8d
24368 .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax
24369 .byte 193,224,8 // shl $0x8,%eax
24370 .byte 68,9,192 // or %r8d,%eax
24371 .byte 102,15,196,216,0 // pinsrw $0x0,%eax,%xmm3
24372 .byte 102,15,196,219,1 // pinsrw $0x1,%ebx,%xmm3
24373 .byte 102,65,15,96,217 // punpcklbw %xmm9,%xmm3
24374 .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3
24375 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
24376 .byte 65,15,89,218 // mulps %xmm10,%xmm3
24377 .byte 72,173 // lods %ds:(%rsi),%rax
24378 .byte 91 // pop %rbx
24379 .byte 65,94 // pop %r14
24380 .byte 255,224 // jmpq *%rax
24381
24382HIDDEN _sk_byte_tables_rgb_sse2
24383.globl _sk_byte_tables_rgb_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]24384FUNCTION(_sk_byte_tables_rgb_sse2)
Mike Klein40de6da2017-04-07 13:09:29 -0400[diff] [blame]24385_sk_byte_tables_rgb_sse2:
24386 .byte 65,86 // push %r14
24387 .byte 83 // push %rbx
24388 .byte 72,173 // lods %ds:(%rsi),%rax
24389 .byte 139,72,24 // mov 0x18(%rax),%ecx
24390 .byte 255,201 // dec %ecx
24391 .byte 102,68,15,110,193 // movd %ecx,%xmm8
24392 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
24393 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24394 .byte 65,15,89,192 // mulps %xmm8,%xmm0
24395 .byte 102,15,91,192 // cvtps2dq %xmm0,%xmm0
24396 .byte 102,72,15,126,193 // movq %xmm0,%rcx
24397 .byte 65,137,200 // mov %ecx,%r8d
24398 .byte 72,193,233,32 // shr $0x20,%rcx
24399 .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0
24400 .byte 102,73,15,126,193 // movq %xmm0,%r9
24401 .byte 69,137,202 // mov %r9d,%r10d
24402 .byte 77,137,203 // mov %r9,%r11
24403 .byte 73,193,235,32 // shr $0x20,%r11
24404 .byte 76,139,48 // mov (%rax),%r14
24405 .byte 76,139,72,8 // mov 0x8(%rax),%r9
24406 .byte 71,15,182,20,22 // movzbl (%r14,%r10,1),%r10d
24407 .byte 67,15,182,28,30 // movzbl (%r14,%r11,1),%ebx
24408 .byte 193,227,8 // shl $0x8,%ebx
24409 .byte 68,9,211 // or %r10d,%ebx
24410 .byte 71,15,182,4,6 // movzbl (%r14,%r8,1),%r8d
24411 .byte 65,15,182,12,14 // movzbl (%r14,%rcx,1),%ecx
24412 .byte 193,225,8 // shl $0x8,%ecx
24413 .byte 68,9,193 // or %r8d,%ecx
24414 .byte 102,15,196,193,0 // pinsrw $0x0,%ecx,%xmm0
24415 .byte 102,15,196,195,1 // pinsrw $0x1,%ebx,%xmm0
24416 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
24417 .byte 102,65,15,96,193 // punpcklbw %xmm9,%xmm0
24418 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0
24419 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
24420 .byte 185,129,128,128,59 // mov $0x3b808081,%ecx
24421 .byte 102,68,15,110,209 // movd %ecx,%xmm10
24422 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24423 .byte 65,15,89,194 // mulps %xmm10,%xmm0
24424 .byte 65,15,89,200 // mulps %xmm8,%xmm1
24425 .byte 102,15,91,201 // cvtps2dq %xmm1,%xmm1
24426 .byte 102,72,15,126,201 // movq %xmm1,%rcx
24427 .byte 65,137,200 // mov %ecx,%r8d
24428 .byte 72,193,233,32 // shr $0x20,%rcx
24429 .byte 102,15,112,201,78 // pshufd $0x4e,%xmm1,%xmm1
24430 .byte 102,72,15,126,203 // movq %xmm1,%rbx
24431 .byte 65,137,218 // mov %ebx,%r10d
24432 .byte 72,193,235,32 // shr $0x20,%rbx
24433 .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d
24434 .byte 65,15,182,28,25 // movzbl (%r9,%rbx,1),%ebx
24435 .byte 193,227,8 // shl $0x8,%ebx
24436 .byte 68,9,211 // or %r10d,%ebx
24437 .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d
24438 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
24439 .byte 193,225,8 // shl $0x8,%ecx
24440 .byte 68,9,193 // or %r8d,%ecx
24441 .byte 102,15,196,201,0 // pinsrw $0x0,%ecx,%xmm1
24442 .byte 102,15,196,203,1 // pinsrw $0x1,%ebx,%xmm1
24443 .byte 102,65,15,96,201 // punpcklbw %xmm9,%xmm1
24444 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1
24445 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
24446 .byte 65,15,89,202 // mulps %xmm10,%xmm1
24447 .byte 72,139,64,16 // mov 0x10(%rax),%rax
24448 .byte 65,15,89,208 // mulps %xmm8,%xmm2
24449 .byte 102,15,91,210 // cvtps2dq %xmm2,%xmm2
24450 .byte 102,72,15,126,209 // movq %xmm2,%rcx
24451 .byte 65,137,200 // mov %ecx,%r8d
24452 .byte 72,193,233,32 // shr $0x20,%rcx
24453 .byte 102,15,112,210,78 // pshufd $0x4e,%xmm2,%xmm2
24454 .byte 102,72,15,126,211 // movq %xmm2,%rbx
24455 .byte 65,137,217 // mov %ebx,%r9d
24456 .byte 72,193,235,32 // shr $0x20,%rbx
24457 .byte 70,15,182,12,8 // movzbl (%rax,%r9,1),%r9d
24458 .byte 15,182,28,24 // movzbl (%rax,%rbx,1),%ebx
24459 .byte 193,227,8 // shl $0x8,%ebx
24460 .byte 68,9,203 // or %r9d,%ebx
24461 .byte 70,15,182,4,0 // movzbl (%rax,%r8,1),%r8d
24462 .byte 15,182,4,8 // movzbl (%rax,%rcx,1),%eax
24463 .byte 193,224,8 // shl $0x8,%eax
24464 .byte 68,9,192 // or %r8d,%eax
24465 .byte 102,15,196,208,0 // pinsrw $0x0,%eax,%xmm2
24466 .byte 102,15,196,211,1 // pinsrw $0x1,%ebx,%xmm2
24467 .byte 102,65,15,96,209 // punpcklbw %xmm9,%xmm2
24468 .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2
24469 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
24470 .byte 65,15,89,210 // mulps %xmm10,%xmm2
24471 .byte 72,173 // lods %ds:(%rsi),%rax
24472 .byte 91 // pop %rbx
24473 .byte 65,94 // pop %r14
24474 .byte 255,224 // jmpq *%rax
24475
Mike Kleinc7d9c0b2017-04-17 14:43:59 -0400[diff] [blame]24476HIDDEN _sk_table_r_sse2
24477.globl _sk_table_r_sse2
24478FUNCTION(_sk_table_r_sse2)
24479_sk_table_r_sse2:
24480 .byte 72,173 // lods %ds:(%rsi),%rax
24481 .byte 76,139,8 // mov (%rax),%r9
24482 .byte 139,64,8 // mov 0x8(%rax),%eax
24483 .byte 255,200 // dec %eax
24484 .byte 102,68,15,110,192 // movd %eax,%xmm8
24485 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
24486 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24487 .byte 68,15,89,192 // mulps %xmm0,%xmm8
24488 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
24489 .byte 102,65,15,112,192,78 // pshufd $0x4e,%xmm8,%xmm0
24490 .byte 102,72,15,126,192 // movq %xmm0,%rax
24491 .byte 65,137,192 // mov %eax,%r8d
24492 .byte 72,193,232,32 // shr $0x20,%rax
24493 .byte 102,76,15,126,193 // movq %xmm8,%rcx
24494 .byte 65,137,202 // mov %ecx,%r10d
24495 .byte 72,193,233,32 // shr $0x20,%rcx
24496 .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8
24497 .byte 243,65,15,16,4,129 // movss (%r9,%rax,4),%xmm0
24498 .byte 68,15,20,192 // unpcklps %xmm0,%xmm8
24499 .byte 243,67,15,16,4,145 // movss (%r9,%r10,4),%xmm0
24500 .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9
24501 .byte 65,15,20,193 // unpcklps %xmm9,%xmm0
24502 .byte 65,15,20,192 // unpcklps %xmm8,%xmm0
24503 .byte 72,173 // lods %ds:(%rsi),%rax
24504 .byte 255,224 // jmpq *%rax
24505
24506HIDDEN _sk_table_g_sse2
24507.globl _sk_table_g_sse2
24508FUNCTION(_sk_table_g_sse2)
24509_sk_table_g_sse2:
24510 .byte 72,173 // lods %ds:(%rsi),%rax
24511 .byte 76,139,8 // mov (%rax),%r9
24512 .byte 139,64,8 // mov 0x8(%rax),%eax
24513 .byte 255,200 // dec %eax
24514 .byte 102,68,15,110,192 // movd %eax,%xmm8
24515 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
24516 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24517 .byte 68,15,89,193 // mulps %xmm1,%xmm8
24518 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
24519 .byte 102,65,15,112,200,78 // pshufd $0x4e,%xmm8,%xmm1
24520 .byte 102,72,15,126,200 // movq %xmm1,%rax
24521 .byte 65,137,192 // mov %eax,%r8d
24522 .byte 72,193,232,32 // shr $0x20,%rax
24523 .byte 102,76,15,126,193 // movq %xmm8,%rcx
24524 .byte 65,137,202 // mov %ecx,%r10d
24525 .byte 72,193,233,32 // shr $0x20,%rcx
24526 .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8
24527 .byte 243,65,15,16,12,129 // movss (%r9,%rax,4),%xmm1
24528 .byte 68,15,20,193 // unpcklps %xmm1,%xmm8
24529 .byte 243,67,15,16,12,145 // movss (%r9,%r10,4),%xmm1
24530 .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9
24531 .byte 65,15,20,201 // unpcklps %xmm9,%xmm1
24532 .byte 65,15,20,200 // unpcklps %xmm8,%xmm1
24533 .byte 72,173 // lods %ds:(%rsi),%rax
24534 .byte 255,224 // jmpq *%rax
24535
24536HIDDEN _sk_table_b_sse2
24537.globl _sk_table_b_sse2
24538FUNCTION(_sk_table_b_sse2)
24539_sk_table_b_sse2:
24540 .byte 72,173 // lods %ds:(%rsi),%rax
24541 .byte 76,139,8 // mov (%rax),%r9
24542 .byte 139,64,8 // mov 0x8(%rax),%eax
24543 .byte 255,200 // dec %eax
24544 .byte 102,68,15,110,192 // movd %eax,%xmm8
24545 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
24546 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24547 .byte 68,15,89,194 // mulps %xmm2,%xmm8
24548 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
24549 .byte 102,65,15,112,208,78 // pshufd $0x4e,%xmm8,%xmm2
24550 .byte 102,72,15,126,208 // movq %xmm2,%rax
24551 .byte 65,137,192 // mov %eax,%r8d
24552 .byte 72,193,232,32 // shr $0x20,%rax
24553 .byte 102,76,15,126,193 // movq %xmm8,%rcx
24554 .byte 65,137,202 // mov %ecx,%r10d
24555 .byte 72,193,233,32 // shr $0x20,%rcx
24556 .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8
24557 .byte 243,65,15,16,20,129 // movss (%r9,%rax,4),%xmm2
24558 .byte 68,15,20,194 // unpcklps %xmm2,%xmm8
24559 .byte 243,67,15,16,20,145 // movss (%r9,%r10,4),%xmm2
24560 .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9
24561 .byte 65,15,20,209 // unpcklps %xmm9,%xmm2
24562 .byte 65,15,20,208 // unpcklps %xmm8,%xmm2
24563 .byte 72,173 // lods %ds:(%rsi),%rax
24564 .byte 255,224 // jmpq *%rax
24565
24566HIDDEN _sk_table_a_sse2
24567.globl _sk_table_a_sse2
24568FUNCTION(_sk_table_a_sse2)
24569_sk_table_a_sse2:
24570 .byte 72,173 // lods %ds:(%rsi),%rax
24571 .byte 76,139,8 // mov (%rax),%r9
24572 .byte 139,64,8 // mov 0x8(%rax),%eax
24573 .byte 255,200 // dec %eax
24574 .byte 102,68,15,110,192 // movd %eax,%xmm8
24575 .byte 102,69,15,112,192,0 // pshufd $0x0,%xmm8,%xmm8
24576 .byte 69,15,91,192 // cvtdq2ps %xmm8,%xmm8
24577 .byte 68,15,89,195 // mulps %xmm3,%xmm8
24578 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
24579 .byte 102,65,15,112,216,78 // pshufd $0x4e,%xmm8,%xmm3
24580 .byte 102,72,15,126,216 // movq %xmm3,%rax
24581 .byte 65,137,192 // mov %eax,%r8d
24582 .byte 72,193,232,32 // shr $0x20,%rax
24583 .byte 102,76,15,126,193 // movq %xmm8,%rcx
24584 .byte 65,137,202 // mov %ecx,%r10d
24585 .byte 72,193,233,32 // shr $0x20,%rcx
24586 .byte 243,69,15,16,4,137 // movss (%r9,%rcx,4),%xmm8
24587 .byte 243,65,15,16,28,129 // movss (%r9,%rax,4),%xmm3
24588 .byte 68,15,20,195 // unpcklps %xmm3,%xmm8
24589 .byte 243,67,15,16,28,145 // movss (%r9,%r10,4),%xmm3
24590 .byte 243,71,15,16,12,129 // movss (%r9,%r8,4),%xmm9
24591 .byte 65,15,20,217 // unpcklps %xmm9,%xmm3
24592 .byte 65,15,20,216 // unpcklps %xmm8,%xmm3
24593 .byte 72,173 // lods %ds:(%rsi),%rax
24594 .byte 255,224 // jmpq *%rax
24595
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24596HIDDEN _sk_parametric_r_sse2
24597.globl _sk_parametric_r_sse2
24598FUNCTION(_sk_parametric_r_sse2)
24599_sk_parametric_r_sse2:
24600 .byte 72,173 // lods %ds:(%rsi),%rax
24601 .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9
24602 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24603 .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8
24604 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24605 .byte 68,15,89,192 // mulps %xmm0,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24606 .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10
24607 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24608 .byte 68,15,89,208 // mulps %xmm0,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24609 .byte 65,15,194,193,2 // cmpleps %xmm9,%xmm0
24610 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9
24611 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24612 .byte 69,15,88,193 // addps %xmm9,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24613 .byte 243,68,15,16,24 // movss (%rax),%xmm11
24614 .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24615 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24616 .byte 69,15,88,209 // addps %xmm9,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24617 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24618 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9
24619 .byte 68,15,89,13,131,36,0,0 // mulps 0x2483(%rip),%xmm9 # 4370 <_sk_callback_sse2+0x100>
24620 .byte 68,15,84,21,139,36,0,0 // andps 0x248b(%rip),%xmm10 # 4380 <_sk_callback_sse2+0x110>
24621 .byte 68,15,86,21,147,36,0,0 // orps 0x2493(%rip),%xmm10 # 4390 <_sk_callback_sse2+0x120>
24622 .byte 68,15,88,13,155,36,0,0 // addps 0x249b(%rip),%xmm9 # 43a0 <_sk_callback_sse2+0x130>
24623 .byte 68,15,40,37,163,36,0,0 // movaps 0x24a3(%rip),%xmm12 # 43b0 <_sk_callback_sse2+0x140>
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24624 .byte 69,15,89,226 // mulps %xmm10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24625 .byte 69,15,92,204 // subps %xmm12,%xmm9
24626 .byte 68,15,88,21,163,36,0,0 // addps 0x24a3(%rip),%xmm10 # 43c0 <_sk_callback_sse2+0x150>
24627 .byte 68,15,40,37,171,36,0,0 // movaps 0x24ab(%rip),%xmm12 # 43d0 <_sk_callback_sse2+0x160>
24628 .byte 69,15,94,226 // divps %xmm10,%xmm12
24629 .byte 69,15,92,204 // subps %xmm12,%xmm9
24630 .byte 69,15,89,203 // mulps %xmm11,%xmm9
24631 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
24632 .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11
24633 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24634 .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12
24635 .byte 68,15,84,37,149,36,0,0 // andps 0x2495(%rip),%xmm12 # 43e0 <_sk_callback_sse2+0x170>
24636 .byte 69,15,87,210 // xorps %xmm10,%xmm10
24637 .byte 69,15,92,220 // subps %xmm12,%xmm11
24638 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24639 .byte 69,15,92,227 // subps %xmm11,%xmm12
24640 .byte 68,15,88,13,141,36,0,0 // addps 0x248d(%rip),%xmm9 # 43f0 <_sk_callback_sse2+0x180>
24641 .byte 68,15,40,29,149,36,0,0 // movaps 0x2495(%rip),%xmm11 # 4400 <_sk_callback_sse2+0x190>
24642 .byte 69,15,89,220 // mulps %xmm12,%xmm11
24643 .byte 69,15,92,203 // subps %xmm11,%xmm9
24644 .byte 68,15,40,29,149,36,0,0 // movaps 0x2495(%rip),%xmm11 # 4410 <_sk_callback_sse2+0x1a0>
24645 .byte 69,15,92,220 // subps %xmm12,%xmm11
24646 .byte 68,15,40,37,153,36,0,0 // movaps 0x2499(%rip),%xmm12 # 4420 <_sk_callback_sse2+0x1b0>
24647 .byte 69,15,94,227 // divps %xmm11,%xmm12
24648 .byte 69,15,88,225 // addps %xmm9,%xmm12
24649 .byte 68,15,89,37,153,36,0,0 // mulps 0x2499(%rip),%xmm12 # 4430 <_sk_callback_sse2+0x1c0>
24650 .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9
24651 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
24652 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24653 .byte 69,15,88,217 // addps %xmm9,%xmm11
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24654 .byte 68,15,84,192 // andps %xmm0,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24655 .byte 65,15,85,195 // andnps %xmm11,%xmm0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24656 .byte 65,15,86,192 // orps %xmm8,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24657 .byte 65,15,95,194 // maxps %xmm10,%xmm0
24658 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24659 .byte 102,68,15,110,192 // movd %eax,%xmm8
24660 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24661 .byte 65,15,93,192 // minps %xmm8,%xmm0
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24662 .byte 72,173 // lods %ds:(%rsi),%rax
24663 .byte 255,224 // jmpq *%rax
24664
24665HIDDEN _sk_parametric_g_sse2
24666.globl _sk_parametric_g_sse2
24667FUNCTION(_sk_parametric_g_sse2)
24668_sk_parametric_g_sse2:
24669 .byte 72,173 // lods %ds:(%rsi),%rax
24670 .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9
24671 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24672 .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8
24673 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24674 .byte 68,15,89,193 // mulps %xmm1,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24675 .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10
24676 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24677 .byte 68,15,89,209 // mulps %xmm1,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24678 .byte 65,15,194,201,2 // cmpleps %xmm9,%xmm1
24679 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9
24680 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24681 .byte 69,15,88,193 // addps %xmm9,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24682 .byte 243,68,15,16,24 // movss (%rax),%xmm11
24683 .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24684 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24685 .byte 69,15,88,209 // addps %xmm9,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24686 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24687 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9
24688 .byte 68,15,89,13,10,36,0,0 // mulps 0x240a(%rip),%xmm9 # 4440 <_sk_callback_sse2+0x1d0>
24689 .byte 68,15,84,21,18,36,0,0 // andps 0x2412(%rip),%xmm10 # 4450 <_sk_callback_sse2+0x1e0>
24690 .byte 68,15,86,21,26,36,0,0 // orps 0x241a(%rip),%xmm10 # 4460 <_sk_callback_sse2+0x1f0>
24691 .byte 68,15,88,13,34,36,0,0 // addps 0x2422(%rip),%xmm9 # 4470 <_sk_callback_sse2+0x200>
24692 .byte 68,15,40,37,42,36,0,0 // movaps 0x242a(%rip),%xmm12 # 4480 <_sk_callback_sse2+0x210>
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24693 .byte 69,15,89,226 // mulps %xmm10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24694 .byte 69,15,92,204 // subps %xmm12,%xmm9
24695 .byte 68,15,88,21,42,36,0,0 // addps 0x242a(%rip),%xmm10 # 4490 <_sk_callback_sse2+0x220>
24696 .byte 68,15,40,37,50,36,0,0 // movaps 0x2432(%rip),%xmm12 # 44a0 <_sk_callback_sse2+0x230>
24697 .byte 69,15,94,226 // divps %xmm10,%xmm12
24698 .byte 69,15,92,204 // subps %xmm12,%xmm9
24699 .byte 69,15,89,203 // mulps %xmm11,%xmm9
24700 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
24701 .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11
24702 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24703 .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12
24704 .byte 68,15,84,37,28,36,0,0 // andps 0x241c(%rip),%xmm12 # 44b0 <_sk_callback_sse2+0x240>
24705 .byte 69,15,87,210 // xorps %xmm10,%xmm10
24706 .byte 69,15,92,220 // subps %xmm12,%xmm11
24707 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24708 .byte 69,15,92,227 // subps %xmm11,%xmm12
24709 .byte 68,15,88,13,20,36,0,0 // addps 0x2414(%rip),%xmm9 # 44c0 <_sk_callback_sse2+0x250>
24710 .byte 68,15,40,29,28,36,0,0 // movaps 0x241c(%rip),%xmm11 # 44d0 <_sk_callback_sse2+0x260>
24711 .byte 69,15,89,220 // mulps %xmm12,%xmm11
24712 .byte 69,15,92,203 // subps %xmm11,%xmm9
24713 .byte 68,15,40,29,28,36,0,0 // movaps 0x241c(%rip),%xmm11 # 44e0 <_sk_callback_sse2+0x270>
24714 .byte 69,15,92,220 // subps %xmm12,%xmm11
24715 .byte 68,15,40,37,32,36,0,0 // movaps 0x2420(%rip),%xmm12 # 44f0 <_sk_callback_sse2+0x280>
24716 .byte 69,15,94,227 // divps %xmm11,%xmm12
24717 .byte 69,15,88,225 // addps %xmm9,%xmm12
24718 .byte 68,15,89,37,32,36,0,0 // mulps 0x2420(%rip),%xmm12 # 4500 <_sk_callback_sse2+0x290>
24719 .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9
24720 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
24721 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24722 .byte 69,15,88,217 // addps %xmm9,%xmm11
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24723 .byte 68,15,84,193 // andps %xmm1,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24724 .byte 65,15,85,203 // andnps %xmm11,%xmm1
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24725 .byte 65,15,86,200 // orps %xmm8,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24726 .byte 65,15,95,202 // maxps %xmm10,%xmm1
24727 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24728 .byte 102,68,15,110,192 // movd %eax,%xmm8
24729 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24730 .byte 65,15,93,200 // minps %xmm8,%xmm1
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24731 .byte 72,173 // lods %ds:(%rsi),%rax
24732 .byte 255,224 // jmpq *%rax
24733
24734HIDDEN _sk_parametric_b_sse2
24735.globl _sk_parametric_b_sse2
24736FUNCTION(_sk_parametric_b_sse2)
24737_sk_parametric_b_sse2:
24738 .byte 72,173 // lods %ds:(%rsi),%rax
24739 .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9
24740 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24741 .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8
24742 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24743 .byte 68,15,89,194 // mulps %xmm2,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24744 .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10
24745 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24746 .byte 68,15,89,210 // mulps %xmm2,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24747 .byte 65,15,194,209,2 // cmpleps %xmm9,%xmm2
24748 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9
24749 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24750 .byte 69,15,88,193 // addps %xmm9,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24751 .byte 243,68,15,16,24 // movss (%rax),%xmm11
24752 .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24753 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24754 .byte 69,15,88,209 // addps %xmm9,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24755 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24756 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9
24757 .byte 68,15,89,13,145,35,0,0 // mulps 0x2391(%rip),%xmm9 # 4510 <_sk_callback_sse2+0x2a0>
24758 .byte 68,15,84,21,153,35,0,0 // andps 0x2399(%rip),%xmm10 # 4520 <_sk_callback_sse2+0x2b0>
24759 .byte 68,15,86,21,161,35,0,0 // orps 0x23a1(%rip),%xmm10 # 4530 <_sk_callback_sse2+0x2c0>
24760 .byte 68,15,88,13,169,35,0,0 // addps 0x23a9(%rip),%xmm9 # 4540 <_sk_callback_sse2+0x2d0>
24761 .byte 68,15,40,37,177,35,0,0 // movaps 0x23b1(%rip),%xmm12 # 4550 <_sk_callback_sse2+0x2e0>
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24762 .byte 69,15,89,226 // mulps %xmm10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24763 .byte 69,15,92,204 // subps %xmm12,%xmm9
24764 .byte 68,15,88,21,177,35,0,0 // addps 0x23b1(%rip),%xmm10 # 4560 <_sk_callback_sse2+0x2f0>
24765 .byte 68,15,40,37,185,35,0,0 // movaps 0x23b9(%rip),%xmm12 # 4570 <_sk_callback_sse2+0x300>
24766 .byte 69,15,94,226 // divps %xmm10,%xmm12
24767 .byte 69,15,92,204 // subps %xmm12,%xmm9
24768 .byte 69,15,89,203 // mulps %xmm11,%xmm9
24769 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
24770 .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11
24771 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24772 .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12
24773 .byte 68,15,84,37,163,35,0,0 // andps 0x23a3(%rip),%xmm12 # 4580 <_sk_callback_sse2+0x310>
24774 .byte 69,15,87,210 // xorps %xmm10,%xmm10
24775 .byte 69,15,92,220 // subps %xmm12,%xmm11
24776 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24777 .byte 69,15,92,227 // subps %xmm11,%xmm12
24778 .byte 68,15,88,13,155,35,0,0 // addps 0x239b(%rip),%xmm9 # 4590 <_sk_callback_sse2+0x320>
24779 .byte 68,15,40,29,163,35,0,0 // movaps 0x23a3(%rip),%xmm11 # 45a0 <_sk_callback_sse2+0x330>
24780 .byte 69,15,89,220 // mulps %xmm12,%xmm11
24781 .byte 69,15,92,203 // subps %xmm11,%xmm9
24782 .byte 68,15,40,29,163,35,0,0 // movaps 0x23a3(%rip),%xmm11 # 45b0 <_sk_callback_sse2+0x340>
24783 .byte 69,15,92,220 // subps %xmm12,%xmm11
24784 .byte 68,15,40,37,167,35,0,0 // movaps 0x23a7(%rip),%xmm12 # 45c0 <_sk_callback_sse2+0x350>
24785 .byte 69,15,94,227 // divps %xmm11,%xmm12
24786 .byte 69,15,88,225 // addps %xmm9,%xmm12
24787 .byte 68,15,89,37,167,35,0,0 // mulps 0x23a7(%rip),%xmm12 # 45d0 <_sk_callback_sse2+0x360>
24788 .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9
24789 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
24790 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24791 .byte 69,15,88,217 // addps %xmm9,%xmm11
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24792 .byte 68,15,84,194 // andps %xmm2,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24793 .byte 65,15,85,211 // andnps %xmm11,%xmm2
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24794 .byte 65,15,86,208 // orps %xmm8,%xmm2
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24795 .byte 65,15,95,210 // maxps %xmm10,%xmm2
24796 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24797 .byte 102,68,15,110,192 // movd %eax,%xmm8
24798 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24799 .byte 65,15,93,208 // minps %xmm8,%xmm2
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24800 .byte 72,173 // lods %ds:(%rsi),%rax
24801 .byte 255,224 // jmpq *%rax
24802
24803HIDDEN _sk_parametric_a_sse2
24804.globl _sk_parametric_a_sse2
24805FUNCTION(_sk_parametric_a_sse2)
24806_sk_parametric_a_sse2:
24807 .byte 72,173 // lods %ds:(%rsi),%rax
24808 .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9
24809 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24810 .byte 243,68,15,16,64,12 // movss 0xc(%rax),%xmm8
24811 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24812 .byte 68,15,89,195 // mulps %xmm3,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24813 .byte 243,68,15,16,80,4 // movss 0x4(%rax),%xmm10
24814 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
24815 .byte 68,15,89,211 // mulps %xmm3,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24816 .byte 65,15,194,217,2 // cmpleps %xmm9,%xmm3
24817 .byte 243,68,15,16,72,24 // movss 0x18(%rax),%xmm9
24818 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24819 .byte 69,15,88,193 // addps %xmm9,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24820 .byte 243,68,15,16,24 // movss (%rax),%xmm11
24821 .byte 243,68,15,16,72,8 // movss 0x8(%rax),%xmm9
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24822 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24823 .byte 69,15,88,209 // addps %xmm9,%xmm10
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24824 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24825 .byte 69,15,91,202 // cvtdq2ps %xmm10,%xmm9
24826 .byte 68,15,89,13,24,35,0,0 // mulps 0x2318(%rip),%xmm9 # 45e0 <_sk_callback_sse2+0x370>
24827 .byte 68,15,84,21,32,35,0,0 // andps 0x2320(%rip),%xmm10 # 45f0 <_sk_callback_sse2+0x380>
24828 .byte 68,15,86,21,40,35,0,0 // orps 0x2328(%rip),%xmm10 # 4600 <_sk_callback_sse2+0x390>
24829 .byte 68,15,88,13,48,35,0,0 // addps 0x2330(%rip),%xmm9 # 4610 <_sk_callback_sse2+0x3a0>
24830 .byte 68,15,40,37,56,35,0,0 // movaps 0x2338(%rip),%xmm12 # 4620 <_sk_callback_sse2+0x3b0>
Mike Kleinda164342017-04-19 14:33:58 -0400[diff] [blame]24831 .byte 69,15,89,226 // mulps %xmm10,%xmm12
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24832 .byte 69,15,92,204 // subps %xmm12,%xmm9
24833 .byte 68,15,88,21,56,35,0,0 // addps 0x2338(%rip),%xmm10 # 4630 <_sk_callback_sse2+0x3c0>
24834 .byte 68,15,40,37,64,35,0,0 // movaps 0x2340(%rip),%xmm12 # 4640 <_sk_callback_sse2+0x3d0>
24835 .byte 69,15,94,226 // divps %xmm10,%xmm12
24836 .byte 69,15,92,204 // subps %xmm12,%xmm9
24837 .byte 69,15,89,203 // mulps %xmm11,%xmm9
24838 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
24839 .byte 69,15,91,218 // cvtdq2ps %xmm10,%xmm11
24840 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24841 .byte 69,15,194,227,1 // cmpltps %xmm11,%xmm12
24842 .byte 68,15,84,37,42,35,0,0 // andps 0x232a(%rip),%xmm12 # 4650 <_sk_callback_sse2+0x3e0>
24843 .byte 69,15,87,210 // xorps %xmm10,%xmm10
24844 .byte 69,15,92,220 // subps %xmm12,%xmm11
24845 .byte 69,15,40,225 // movaps %xmm9,%xmm12
24846 .byte 69,15,92,227 // subps %xmm11,%xmm12
24847 .byte 68,15,88,13,34,35,0,0 // addps 0x2322(%rip),%xmm9 # 4660 <_sk_callback_sse2+0x3f0>
24848 .byte 68,15,40,29,42,35,0,0 // movaps 0x232a(%rip),%xmm11 # 4670 <_sk_callback_sse2+0x400>
24849 .byte 69,15,89,220 // mulps %xmm12,%xmm11
24850 .byte 69,15,92,203 // subps %xmm11,%xmm9
24851 .byte 68,15,40,29,42,35,0,0 // movaps 0x232a(%rip),%xmm11 # 4680 <_sk_callback_sse2+0x410>
24852 .byte 69,15,92,220 // subps %xmm12,%xmm11
24853 .byte 68,15,40,37,46,35,0,0 // movaps 0x232e(%rip),%xmm12 # 4690 <_sk_callback_sse2+0x420>
24854 .byte 69,15,94,227 // divps %xmm11,%xmm12
24855 .byte 69,15,88,225 // addps %xmm9,%xmm12
24856 .byte 68,15,89,37,46,35,0,0 // mulps 0x232e(%rip),%xmm12 # 46a0 <_sk_callback_sse2+0x430>
24857 .byte 102,69,15,91,204 // cvtps2dq %xmm12,%xmm9
24858 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
24859 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24860 .byte 69,15,88,217 // addps %xmm9,%xmm11
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24861 .byte 68,15,84,195 // andps %xmm3,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24862 .byte 65,15,85,219 // andnps %xmm11,%xmm3
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24863 .byte 65,15,86,216 // orps %xmm8,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]24864 .byte 65,15,95,218 // maxps %xmm10,%xmm3
24865 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
24866 .byte 102,68,15,110,192 // movd %eax,%xmm8
24867 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24868 .byte 65,15,93,216 // minps %xmm8,%xmm3
Mike Klein44375172017-04-17 19:32:05 -0400[diff] [blame]24869 .byte 72,173 // lods %ds:(%rsi),%rax
24870 .byte 255,224 // jmpq *%rax
24871
Mike Klein4e3e9f82017-04-20 11:04:29 -0400[diff] [blame]24872HIDDEN _sk_lab_to_xyz_sse2
24873.globl _sk_lab_to_xyz_sse2
24874FUNCTION(_sk_lab_to_xyz_sse2)
24875_sk_lab_to_xyz_sse2:
24876 .byte 184,0,0,200,66 // mov $0x42c80000,%eax
24877 .byte 102,68,15,110,200 // movd %eax,%xmm9
24878 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
24879 .byte 68,15,89,200 // mulps %xmm0,%xmm9
24880 .byte 184,0,0,127,67 // mov $0x437f0000,%eax
24881 .byte 102,68,15,110,192 // movd %eax,%xmm8
24882 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24883 .byte 65,15,89,200 // mulps %xmm8,%xmm1
24884 .byte 184,0,0,0,67 // mov $0x43000000,%eax
24885 .byte 102,15,110,192 // movd %eax,%xmm0
24886 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
24887 .byte 15,92,200 // subps %xmm0,%xmm1
24888 .byte 68,15,89,194 // mulps %xmm2,%xmm8
24889 .byte 68,15,92,192 // subps %xmm0,%xmm8
24890 .byte 184,0,0,128,65 // mov $0x41800000,%eax
24891 .byte 102,15,110,208 // movd %eax,%xmm2
24892 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
24893 .byte 65,15,88,209 // addps %xmm9,%xmm2
24894 .byte 184,203,61,13,60 // mov $0x3c0d3dcb,%eax
24895 .byte 102,15,110,192 // movd %eax,%xmm0
24896 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
24897 .byte 15,89,194 // mulps %xmm2,%xmm0
24898 .byte 184,111,18,3,59 // mov $0x3b03126f,%eax
24899 .byte 102,15,110,208 // movd %eax,%xmm2
24900 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
24901 .byte 15,89,209 // mulps %xmm1,%xmm2
24902 .byte 15,88,208 // addps %xmm0,%xmm2
24903 .byte 184,10,215,163,59 // mov $0x3ba3d70a,%eax
24904 .byte 102,15,110,200 // movd %eax,%xmm1
24905 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
24906 .byte 65,15,89,200 // mulps %xmm8,%xmm1
24907 .byte 68,15,40,208 // movaps %xmm0,%xmm10
24908 .byte 68,15,92,209 // subps %xmm1,%xmm10
24909 .byte 15,40,202 // movaps %xmm2,%xmm1
24910 .byte 15,89,201 // mulps %xmm1,%xmm1
24911 .byte 15,89,202 // mulps %xmm2,%xmm1
24912 .byte 184,194,24,17,60 // mov $0x3c1118c2,%eax
24913 .byte 102,68,15,110,192 // movd %eax,%xmm8
24914 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
24915 .byte 69,15,40,200 // movaps %xmm8,%xmm9
24916 .byte 68,15,194,201,1 // cmpltps %xmm1,%xmm9
24917 .byte 184,203,61,13,62 // mov $0x3e0d3dcb,%eax
24918 .byte 102,68,15,110,216 // movd %eax,%xmm11
24919 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
24920 .byte 65,15,92,211 // subps %xmm11,%xmm2
24921 .byte 184,80,128,3,62 // mov $0x3e038050,%eax
24922 .byte 102,68,15,110,224 // movd %eax,%xmm12
24923 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
24924 .byte 65,15,89,212 // mulps %xmm12,%xmm2
24925 .byte 65,15,84,201 // andps %xmm9,%xmm1
24926 .byte 68,15,85,202 // andnps %xmm2,%xmm9
24927 .byte 68,15,86,201 // orps %xmm1,%xmm9
24928 .byte 15,40,208 // movaps %xmm0,%xmm2
24929 .byte 15,89,210 // mulps %xmm2,%xmm2
24930 .byte 15,89,208 // mulps %xmm0,%xmm2
24931 .byte 65,15,40,200 // movaps %xmm8,%xmm1
24932 .byte 15,194,202,1 // cmpltps %xmm2,%xmm1
24933 .byte 65,15,92,195 // subps %xmm11,%xmm0
24934 .byte 65,15,89,196 // mulps %xmm12,%xmm0
24935 .byte 15,84,209 // andps %xmm1,%xmm2
24936 .byte 15,85,200 // andnps %xmm0,%xmm1
24937 .byte 15,86,202 // orps %xmm2,%xmm1
24938 .byte 65,15,40,194 // movaps %xmm10,%xmm0
24939 .byte 15,89,192 // mulps %xmm0,%xmm0
24940 .byte 65,15,89,194 // mulps %xmm10,%xmm0
24941 .byte 68,15,194,192,1 // cmpltps %xmm0,%xmm8
24942 .byte 69,15,92,211 // subps %xmm11,%xmm10
24943 .byte 69,15,89,212 // mulps %xmm12,%xmm10
24944 .byte 65,15,84,192 // andps %xmm8,%xmm0
24945 .byte 69,15,85,194 // andnps %xmm10,%xmm8
24946 .byte 68,15,86,192 // orps %xmm0,%xmm8
24947 .byte 184,31,215,118,63 // mov $0x3f76d71f,%eax
24948 .byte 102,15,110,192 // movd %eax,%xmm0
24949 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
24950 .byte 65,15,89,193 // mulps %xmm9,%xmm0
24951 .byte 184,246,64,83,63 // mov $0x3f5340f6,%eax
24952 .byte 102,15,110,208 // movd %eax,%xmm2
24953 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
24954 .byte 65,15,89,208 // mulps %xmm8,%xmm2
24955 .byte 72,173 // lods %ds:(%rsi),%rax
24956 .byte 255,224 // jmpq *%rax
24957
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]24958HIDDEN _sk_load_a8_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]24959.globl _sk_load_a8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]24960FUNCTION(_sk_load_a8_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]24961_sk_load_a8_sse2:
24962 .byte 72,173 // lods %ds:(%rsi),%rax
24963 .byte 72,139,0 // mov (%rax),%rax
24964 .byte 102,15,110,4,56 // movd (%rax,%rdi,1),%xmm0
24965 .byte 102,15,239,201 // pxor %xmm1,%xmm1
24966 .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
24967 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
24968 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
24969 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
24970 .byte 102,15,110,216 // movd %eax,%xmm3
24971 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
24972 .byte 15,89,216 // mulps %xmm0,%xmm3
24973 .byte 72,173 // lods %ds:(%rsi),%rax
24974 .byte 15,87,192 // xorps %xmm0,%xmm0
24975 .byte 102,15,239,201 // pxor %xmm1,%xmm1
24976 .byte 15,87,210 // xorps %xmm2,%xmm2
24977 .byte 255,224 // jmpq *%rax
24978
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]24979HIDDEN _sk_gather_a8_sse2
24980.globl _sk_gather_a8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]24981FUNCTION(_sk_gather_a8_sse2)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]24982_sk_gather_a8_sse2:
24983 .byte 72,173 // lods %ds:(%rsi),%rax
24984 .byte 76,139,8 // mov (%rax),%r9
24985 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
24986 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
24987 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
24988 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
24989 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
24990 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
24991 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
24992 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
24993 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
24994 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
24995 .byte 102,15,254,193 // paddd %xmm1,%xmm0
24996 .byte 102,72,15,126,192 // movq %xmm0,%rax
24997 .byte 65,137,192 // mov %eax,%r8d
24998 .byte 72,193,232,32 // shr $0x20,%rax
24999 .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0
25000 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25001 .byte 65,137,202 // mov %ecx,%r10d
25002 .byte 72,193,233,32 // shr $0x20,%rcx
25003 .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d
25004 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
25005 .byte 193,225,8 // shl $0x8,%ecx
25006 .byte 68,9,209 // or %r10d,%ecx
25007 .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d
25008 .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax
25009 .byte 193,224,8 // shl $0x8,%eax
25010 .byte 68,9,192 // or %r8d,%eax
25011 .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0
25012 .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%xmm0
25013 .byte 102,15,239,201 // pxor %xmm1,%xmm1
25014 .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
25015 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
25016 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
25017 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25018 .byte 102,15,110,216 // movd %eax,%xmm3
25019 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25020 .byte 15,89,216 // mulps %xmm0,%xmm3
25021 .byte 72,173 // lods %ds:(%rsi),%rax
25022 .byte 15,87,192 // xorps %xmm0,%xmm0
25023 .byte 102,15,239,201 // pxor %xmm1,%xmm1
25024 .byte 102,15,239,210 // pxor %xmm2,%xmm2
25025 .byte 255,224 // jmpq *%rax
25026
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25027HIDDEN _sk_store_a8_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25028.globl _sk_store_a8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25029FUNCTION(_sk_store_a8_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25030_sk_store_a8_sse2:
25031 .byte 72,173 // lods %ds:(%rsi),%rax
25032 .byte 72,139,0 // mov (%rax),%rax
25033 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
25034 .byte 102,68,15,110,193 // movd %ecx,%xmm8
25035 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25036 .byte 68,15,89,195 // mulps %xmm3,%xmm8
25037 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
25038 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
25039 .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
25040 .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
25041 .byte 102,69,15,103,192 // packuswb %xmm8,%xmm8
25042 .byte 102,68,15,126,4,56 // movd %xmm8,(%rax,%rdi,1)
25043 .byte 72,173 // lods %ds:(%rsi),%rax
25044 .byte 255,224 // jmpq *%rax
25045
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25046HIDDEN _sk_load_g8_sse2
25047.globl _sk_load_g8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25048FUNCTION(_sk_load_g8_sse2)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25049_sk_load_g8_sse2:
25050 .byte 72,173 // lods %ds:(%rsi),%rax
25051 .byte 72,139,0 // mov (%rax),%rax
25052 .byte 102,15,110,4,56 // movd (%rax,%rdi,1),%xmm0
25053 .byte 102,15,239,201 // pxor %xmm1,%xmm1
25054 .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
25055 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
25056 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25057 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25058 .byte 102,15,110,192 // movd %eax,%xmm0
25059 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25060 .byte 15,89,193 // mulps %xmm1,%xmm0
25061 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
25062 .byte 102,15,110,216 // movd %eax,%xmm3
25063 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25064 .byte 72,173 // lods %ds:(%rsi),%rax
25065 .byte 15,40,200 // movaps %xmm0,%xmm1
25066 .byte 15,40,208 // movaps %xmm0,%xmm2
25067 .byte 255,224 // jmpq *%rax
25068
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25069HIDDEN _sk_gather_g8_sse2
25070.globl _sk_gather_g8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25071FUNCTION(_sk_gather_g8_sse2)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25072_sk_gather_g8_sse2:
25073 .byte 72,173 // lods %ds:(%rsi),%rax
25074 .byte 76,139,8 // mov (%rax),%r9
25075 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25076 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25077 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25078 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25079 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25080 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25081 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25082 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25083 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25084 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25085 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25086 .byte 102,72,15,126,192 // movq %xmm0,%rax
25087 .byte 65,137,192 // mov %eax,%r8d
25088 .byte 72,193,232,32 // shr $0x20,%rax
25089 .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0
25090 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25091 .byte 65,137,202 // mov %ecx,%r10d
25092 .byte 72,193,233,32 // shr $0x20,%rcx
25093 .byte 71,15,182,20,17 // movzbl (%r9,%r10,1),%r10d
25094 .byte 65,15,182,12,9 // movzbl (%r9,%rcx,1),%ecx
25095 .byte 193,225,8 // shl $0x8,%ecx
25096 .byte 68,9,209 // or %r10d,%ecx
25097 .byte 71,15,182,4,1 // movzbl (%r9,%r8,1),%r8d
25098 .byte 65,15,182,4,1 // movzbl (%r9,%rax,1),%eax
25099 .byte 193,224,8 // shl $0x8,%eax
25100 .byte 68,9,192 // or %r8d,%eax
25101 .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0
25102 .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%xmm0
25103 .byte 102,15,239,201 // pxor %xmm1,%xmm1
25104 .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
25105 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
25106 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25107 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25108 .byte 102,15,110,192 // movd %eax,%xmm0
25109 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25110 .byte 15,89,193 // mulps %xmm1,%xmm0
25111 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
25112 .byte 102,15,110,216 // movd %eax,%xmm3
25113 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25114 .byte 72,173 // lods %ds:(%rsi),%rax
25115 .byte 15,40,200 // movaps %xmm0,%xmm1
25116 .byte 15,40,208 // movaps %xmm0,%xmm2
25117 .byte 255,224 // jmpq *%rax
25118
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]25119HIDDEN _sk_gather_i8_sse2
25120.globl _sk_gather_i8_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25121FUNCTION(_sk_gather_i8_sse2)
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]25122_sk_gather_i8_sse2:
25123 .byte 72,173 // lods %ds:(%rsi),%rax
25124 .byte 73,137,192 // mov %rax,%r8
25125 .byte 77,133,192 // test %r8,%r8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25126 .byte 116,5 // je 2718 <_sk_gather_i8_sse2+0xf>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]25127 .byte 76,137,192 // mov %r8,%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25128 .byte 235,2 // jmp 271a <_sk_gather_i8_sse2+0x11>
Mike Klein7d3d8722017-04-06 17:53:18 -0400[diff] [blame]25129 .byte 72,173 // lods %ds:(%rsi),%rax
25130 .byte 76,139,16 // mov (%rax),%r10
25131 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25132 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25133 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25134 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25135 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25136 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25137 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25138 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25139 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25140 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25141 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25142 .byte 102,72,15,126,192 // movq %xmm0,%rax
25143 .byte 65,137,193 // mov %eax,%r9d
25144 .byte 72,193,232,32 // shr $0x20,%rax
25145 .byte 102,15,112,192,78 // pshufd $0x4e,%xmm0,%xmm0
25146 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25147 .byte 65,137,203 // mov %ecx,%r11d
25148 .byte 72,193,233,32 // shr $0x20,%rcx
25149 .byte 71,15,182,28,26 // movzbl (%r10,%r11,1),%r11d
25150 .byte 65,15,182,12,10 // movzbl (%r10,%rcx,1),%ecx
25151 .byte 193,225,8 // shl $0x8,%ecx
25152 .byte 68,9,217 // or %r11d,%ecx
25153 .byte 71,15,182,12,10 // movzbl (%r10,%r9,1),%r9d
25154 .byte 65,15,182,4,2 // movzbl (%r10,%rax,1),%eax
25155 .byte 193,224,8 // shl $0x8,%eax
25156 .byte 68,9,200 // or %r9d,%eax
25157 .byte 102,15,196,192,0 // pinsrw $0x0,%eax,%xmm0
25158 .byte 102,15,196,193,1 // pinsrw $0x1,%ecx,%xmm0
25159 .byte 102,15,239,201 // pxor %xmm1,%xmm1
25160 .byte 102,15,96,193 // punpcklbw %xmm1,%xmm0
25161 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
25162 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1
25163 .byte 102,72,15,126,200 // movq %xmm1,%rax
25164 .byte 68,15,182,200 // movzbl %al,%r9d
25165 .byte 72,193,232,32 // shr $0x20,%rax
25166 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25167 .byte 77,139,64,8 // mov 0x8(%r8),%r8
25168 .byte 68,15,182,209 // movzbl %cl,%r10d
25169 .byte 72,193,233,30 // shr $0x1e,%rcx
25170 .byte 102,65,15,110,4,8 // movd (%r8,%rcx,1),%xmm0
25171 .byte 102,65,15,110,12,128 // movd (%r8,%rax,4),%xmm1
25172 .byte 102,15,98,193 // punpckldq %xmm1,%xmm0
25173 .byte 102,67,15,110,28,144 // movd (%r8,%r10,4),%xmm3
25174 .byte 102,67,15,110,12,136 // movd (%r8,%r9,4),%xmm1
25175 .byte 102,15,98,217 // punpckldq %xmm1,%xmm3
25176 .byte 102,15,98,216 // punpckldq %xmm0,%xmm3
25177 .byte 184,255,0,0,0 // mov $0xff,%eax
25178 .byte 102,15,110,192 // movd %eax,%xmm0
25179 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25180 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25181 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
25182 .byte 102,15,219,200 // pand %xmm0,%xmm1
25183 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
25184 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
25185 .byte 102,15,219,208 // pand %xmm0,%xmm2
25186 .byte 102,15,219,195 // pand %xmm3,%xmm0
25187 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
25188 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25189 .byte 102,68,15,110,192 // movd %eax,%xmm8
25190 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25191 .byte 65,15,89,192 // mulps %xmm8,%xmm0
25192 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
25193 .byte 65,15,89,200 // mulps %xmm8,%xmm1
25194 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
25195 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25196 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
25197 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25198 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25199 .byte 72,173 // lods %ds:(%rsi),%rax
25200 .byte 255,224 // jmpq *%rax
25201
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25202HIDDEN _sk_load_565_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25203.globl _sk_load_565_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25204FUNCTION(_sk_load_565_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25205_sk_load_565_sse2:
25206 .byte 72,173 // lods %ds:(%rsi),%rax
25207 .byte 72,139,0 // mov (%rax),%rax
25208 .byte 243,15,126,20,120 // movq (%rax,%rdi,2),%xmm2
25209 .byte 102,15,239,192 // pxor %xmm0,%xmm0
25210 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2
25211 .byte 184,0,248,0,0 // mov $0xf800,%eax
25212 .byte 102,15,110,192 // movd %eax,%xmm0
25213 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25214 .byte 102,15,219,194 // pand %xmm2,%xmm0
25215 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25216 .byte 184,8,33,132,55 // mov $0x37842108,%eax
25217 .byte 102,15,110,192 // movd %eax,%xmm0
25218 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25219 .byte 15,89,193 // mulps %xmm1,%xmm0
25220 .byte 184,224,7,0,0 // mov $0x7e0,%eax
25221 .byte 102,15,110,200 // movd %eax,%xmm1
25222 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
25223 .byte 102,15,219,202 // pand %xmm2,%xmm1
25224 .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3
25225 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
25226 .byte 102,15,110,200 // movd %eax,%xmm1
25227 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
25228 .byte 15,89,203 // mulps %xmm3,%xmm1
25229 .byte 184,31,0,0,0 // mov $0x1f,%eax
25230 .byte 102,15,110,216 // movd %eax,%xmm3
25231 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
25232 .byte 102,15,219,218 // pand %xmm2,%xmm3
25233 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25234 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
25235 .byte 102,15,110,208 // movd %eax,%xmm2
25236 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
25237 .byte 15,89,211 // mulps %xmm3,%xmm2
25238 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
25239 .byte 102,15,110,216 // movd %eax,%xmm3
25240 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25241 .byte 72,173 // lods %ds:(%rsi),%rax
25242 .byte 255,224 // jmpq *%rax
25243
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25244HIDDEN _sk_gather_565_sse2
25245.globl _sk_gather_565_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25246FUNCTION(_sk_gather_565_sse2)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25247_sk_gather_565_sse2:
25248 .byte 72,173 // lods %ds:(%rsi),%rax
25249 .byte 76,139,8 // mov (%rax),%r9
25250 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25251 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25252 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25253 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25254 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25255 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25256 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25257 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25258 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25259 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25260 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25261 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1
25262 .byte 102,72,15,126,200 // movq %xmm1,%rax
25263 .byte 65,137,192 // mov %eax,%r8d
25264 .byte 72,193,232,32 // shr $0x20,%rax
25265 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25266 .byte 65,137,202 // mov %ecx,%r10d
25267 .byte 72,193,233,32 // shr $0x20,%rcx
25268 .byte 102,67,15,196,20,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm2
25269 .byte 102,65,15,196,20,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm2
25270 .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx
25271 .byte 102,15,196,209,2 // pinsrw $0x2,%ecx,%xmm2
25272 .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax
25273 .byte 102,15,196,208,3 // pinsrw $0x3,%eax,%xmm2
25274 .byte 102,15,239,192 // pxor %xmm0,%xmm0
25275 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2
25276 .byte 184,0,248,0,0 // mov $0xf800,%eax
25277 .byte 102,15,110,192 // movd %eax,%xmm0
25278 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25279 .byte 102,15,219,194 // pand %xmm2,%xmm0
25280 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25281 .byte 184,8,33,132,55 // mov $0x37842108,%eax
25282 .byte 102,15,110,192 // movd %eax,%xmm0
25283 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25284 .byte 15,89,193 // mulps %xmm1,%xmm0
25285 .byte 184,224,7,0,0 // mov $0x7e0,%eax
25286 .byte 102,15,110,200 // movd %eax,%xmm1
25287 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
25288 .byte 102,15,219,202 // pand %xmm2,%xmm1
25289 .byte 15,91,217 // cvtdq2ps %xmm1,%xmm3
25290 .byte 184,33,8,2,58 // mov $0x3a020821,%eax
25291 .byte 102,15,110,200 // movd %eax,%xmm1
25292 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
25293 .byte 15,89,203 // mulps %xmm3,%xmm1
25294 .byte 184,31,0,0,0 // mov $0x1f,%eax
25295 .byte 102,15,110,216 // movd %eax,%xmm3
25296 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
25297 .byte 102,15,219,218 // pand %xmm2,%xmm3
25298 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25299 .byte 184,8,33,4,61 // mov $0x3d042108,%eax
25300 .byte 102,15,110,208 // movd %eax,%xmm2
25301 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
25302 .byte 15,89,211 // mulps %xmm3,%xmm2
25303 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
25304 .byte 102,15,110,216 // movd %eax,%xmm3
25305 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25306 .byte 72,173 // lods %ds:(%rsi),%rax
25307 .byte 255,224 // jmpq *%rax
25308
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25309HIDDEN _sk_store_565_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25310.globl _sk_store_565_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25311FUNCTION(_sk_store_565_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25312_sk_store_565_sse2:
25313 .byte 72,173 // lods %ds:(%rsi),%rax
25314 .byte 72,139,0 // mov (%rax),%rax
25315 .byte 185,0,0,248,65 // mov $0x41f80000,%ecx
25316 .byte 102,68,15,110,193 // movd %ecx,%xmm8
25317 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25318 .byte 69,15,40,200 // movaps %xmm8,%xmm9
25319 .byte 68,15,89,200 // mulps %xmm0,%xmm9
25320 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25321 .byte 102,65,15,114,241,11 // pslld $0xb,%xmm9
25322 .byte 185,0,0,124,66 // mov $0x427c0000,%ecx
25323 .byte 102,68,15,110,209 // movd %ecx,%xmm10
25324 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
25325 .byte 68,15,89,209 // mulps %xmm1,%xmm10
25326 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
25327 .byte 102,65,15,114,242,5 // pslld $0x5,%xmm10
25328 .byte 102,69,15,235,209 // por %xmm9,%xmm10
25329 .byte 68,15,89,194 // mulps %xmm2,%xmm8
25330 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
25331 .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
25332 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
25333 .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
25334 .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
25335 .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
25336 .byte 72,173 // lods %ds:(%rsi),%rax
25337 .byte 255,224 // jmpq *%rax
25338
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25339HIDDEN _sk_load_4444_sse2
25340.globl _sk_load_4444_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25341FUNCTION(_sk_load_4444_sse2)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25342_sk_load_4444_sse2:
25343 .byte 72,173 // lods %ds:(%rsi),%rax
25344 .byte 72,139,0 // mov (%rax),%rax
25345 .byte 243,68,15,126,12,120 // movq (%rax,%rdi,2),%xmm9
25346 .byte 102,15,239,192 // pxor %xmm0,%xmm0
25347 .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9
25348 .byte 184,0,240,0,0 // mov $0xf000,%eax
25349 .byte 102,15,110,192 // movd %eax,%xmm0
25350 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25351 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
25352 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25353 .byte 184,137,136,136,55 // mov $0x37888889,%eax
25354 .byte 102,15,110,192 // movd %eax,%xmm0
25355 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25356 .byte 15,89,193 // mulps %xmm1,%xmm0
25357 .byte 184,0,15,0,0 // mov $0xf00,%eax
25358 .byte 102,15,110,200 // movd %eax,%xmm1
25359 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
25360 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
25361 .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
25362 .byte 184,137,136,136,57 // mov $0x39888889,%eax
25363 .byte 102,15,110,200 // movd %eax,%xmm1
25364 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
25365 .byte 15,89,202 // mulps %xmm2,%xmm1
25366 .byte 184,240,0,0,0 // mov $0xf0,%eax
25367 .byte 102,15,110,208 // movd %eax,%xmm2
25368 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25369 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
25370 .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
25371 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
25372 .byte 102,15,110,208 // movd %eax,%xmm2
25373 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
25374 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25375 .byte 184,15,0,0,0 // mov $0xf,%eax
25376 .byte 102,15,110,216 // movd %eax,%xmm3
25377 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
25378 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
25379 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
25380 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
25381 .byte 102,15,110,216 // movd %eax,%xmm3
25382 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25383 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25384 .byte 72,173 // lods %ds:(%rsi),%rax
25385 .byte 255,224 // jmpq *%rax
25386
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25387HIDDEN _sk_gather_4444_sse2
25388.globl _sk_gather_4444_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25389FUNCTION(_sk_gather_4444_sse2)
Mike Klein21bd3e42017-04-06 16:32:29 -0400[diff] [blame]25390_sk_gather_4444_sse2:
25391 .byte 72,173 // lods %ds:(%rsi),%rax
25392 .byte 76,139,8 // mov (%rax),%r9
25393 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25394 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25395 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25396 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25397 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25398 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25399 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25400 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25401 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25402 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25403 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25404 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1
25405 .byte 102,72,15,126,200 // movq %xmm1,%rax
25406 .byte 65,137,192 // mov %eax,%r8d
25407 .byte 72,193,232,32 // shr $0x20,%rax
25408 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25409 .byte 65,137,202 // mov %ecx,%r10d
25410 .byte 72,193,233,32 // shr $0x20,%rcx
25411 .byte 102,71,15,196,12,81,0 // pinsrw $0x0,(%r9,%r10,2),%xmm9
25412 .byte 102,69,15,196,12,73,1 // pinsrw $0x1,(%r9,%rcx,2),%xmm9
25413 .byte 67,15,183,12,65 // movzwl (%r9,%r8,2),%ecx
25414 .byte 102,68,15,196,201,2 // pinsrw $0x2,%ecx,%xmm9
25415 .byte 65,15,183,4,65 // movzwl (%r9,%rax,2),%eax
25416 .byte 102,68,15,196,200,3 // pinsrw $0x3,%eax,%xmm9
25417 .byte 102,15,239,192 // pxor %xmm0,%xmm0
25418 .byte 102,68,15,97,200 // punpcklwd %xmm0,%xmm9
25419 .byte 184,0,240,0,0 // mov $0xf000,%eax
25420 .byte 102,15,110,192 // movd %eax,%xmm0
25421 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25422 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
25423 .byte 15,91,200 // cvtdq2ps %xmm0,%xmm1
25424 .byte 184,137,136,136,55 // mov $0x37888889,%eax
25425 .byte 102,15,110,192 // movd %eax,%xmm0
25426 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
25427 .byte 15,89,193 // mulps %xmm1,%xmm0
25428 .byte 184,0,15,0,0 // mov $0xf00,%eax
25429 .byte 102,15,110,200 // movd %eax,%xmm1
25430 .byte 102,15,112,201,0 // pshufd $0x0,%xmm1,%xmm1
25431 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
25432 .byte 15,91,209 // cvtdq2ps %xmm1,%xmm2
25433 .byte 184,137,136,136,57 // mov $0x39888889,%eax
25434 .byte 102,15,110,200 // movd %eax,%xmm1
25435 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
25436 .byte 15,89,202 // mulps %xmm2,%xmm1
25437 .byte 184,240,0,0,0 // mov $0xf0,%eax
25438 .byte 102,15,110,208 // movd %eax,%xmm2
25439 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25440 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
25441 .byte 68,15,91,194 // cvtdq2ps %xmm2,%xmm8
25442 .byte 184,137,136,136,59 // mov $0x3b888889,%eax
25443 .byte 102,15,110,208 // movd %eax,%xmm2
25444 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
25445 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25446 .byte 184,15,0,0,0 // mov $0xf,%eax
25447 .byte 102,15,110,216 // movd %eax,%xmm3
25448 .byte 102,15,112,219,0 // pshufd $0x0,%xmm3,%xmm3
25449 .byte 102,65,15,219,217 // pand %xmm9,%xmm3
25450 .byte 68,15,91,195 // cvtdq2ps %xmm3,%xmm8
25451 .byte 184,137,136,136,61 // mov $0x3d888889,%eax
25452 .byte 102,15,110,216 // movd %eax,%xmm3
25453 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25454 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25455 .byte 72,173 // lods %ds:(%rsi),%rax
25456 .byte 255,224 // jmpq *%rax
25457
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25458HIDDEN _sk_store_4444_sse2
25459.globl _sk_store_4444_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25460FUNCTION(_sk_store_4444_sse2)
Mike Kleinf809fef2017-03-31 13:52:45 -0400[diff] [blame]25461_sk_store_4444_sse2:
25462 .byte 72,173 // lods %ds:(%rsi),%rax
25463 .byte 72,139,0 // mov (%rax),%rax
25464 .byte 185,0,0,112,65 // mov $0x41700000,%ecx
25465 .byte 102,68,15,110,193 // movd %ecx,%xmm8
25466 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25467 .byte 69,15,40,200 // movaps %xmm8,%xmm9
25468 .byte 68,15,89,200 // mulps %xmm0,%xmm9
25469 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25470 .byte 102,65,15,114,241,12 // pslld $0xc,%xmm9
25471 .byte 69,15,40,208 // movaps %xmm8,%xmm10
25472 .byte 68,15,89,209 // mulps %xmm1,%xmm10
25473 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
25474 .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
25475 .byte 102,69,15,235,209 // por %xmm9,%xmm10
25476 .byte 69,15,40,200 // movaps %xmm8,%xmm9
25477 .byte 68,15,89,202 // mulps %xmm2,%xmm9
25478 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25479 .byte 102,65,15,114,241,4 // pslld $0x4,%xmm9
25480 .byte 68,15,89,195 // mulps %xmm3,%xmm8
25481 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
25482 .byte 102,69,15,86,193 // orpd %xmm9,%xmm8
25483 .byte 102,69,15,86,194 // orpd %xmm10,%xmm8
25484 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
25485 .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
25486 .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
25487 .byte 102,68,15,214,4,120 // movq %xmm8,(%rax,%rdi,2)
25488 .byte 72,173 // lods %ds:(%rsi),%rax
25489 .byte 255,224 // jmpq *%rax
25490
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25491HIDDEN _sk_load_8888_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25492.globl _sk_load_8888_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25493FUNCTION(_sk_load_8888_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25494_sk_load_8888_sse2:
25495 .byte 72,173 // lods %ds:(%rsi),%rax
25496 .byte 72,139,0 // mov (%rax),%rax
25497 .byte 243,15,111,28,184 // movdqu (%rax,%rdi,4),%xmm3
25498 .byte 184,255,0,0,0 // mov $0xff,%eax
25499 .byte 102,15,110,192 // movd %eax,%xmm0
25500 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25501 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25502 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
25503 .byte 102,15,219,200 // pand %xmm0,%xmm1
25504 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
25505 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
25506 .byte 102,15,219,208 // pand %xmm0,%xmm2
25507 .byte 102,15,219,195 // pand %xmm3,%xmm0
25508 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
25509 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25510 .byte 102,68,15,110,192 // movd %eax,%xmm8
25511 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25512 .byte 65,15,89,192 // mulps %xmm8,%xmm0
25513 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
25514 .byte 65,15,89,200 // mulps %xmm8,%xmm1
25515 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
25516 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25517 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
25518 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25519 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25520 .byte 72,173 // lods %ds:(%rsi),%rax
25521 .byte 255,224 // jmpq *%rax
25522
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]25523HIDDEN _sk_gather_8888_sse2
25524.globl _sk_gather_8888_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25525FUNCTION(_sk_gather_8888_sse2)
Mike Kleindec4ea82017-04-06 15:04:05 -0400[diff] [blame]25526_sk_gather_8888_sse2:
25527 .byte 72,173 // lods %ds:(%rsi),%rax
25528 .byte 76,139,8 // mov (%rax),%r9
25529 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25530 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25531 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25532 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25533 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25534 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25535 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25536 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25537 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25538 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25539 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25540 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1
25541 .byte 102,72,15,126,200 // movq %xmm1,%rax
25542 .byte 65,137,192 // mov %eax,%r8d
25543 .byte 72,193,232,32 // shr $0x20,%rax
25544 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25545 .byte 65,137,202 // mov %ecx,%r10d
25546 .byte 72,193,233,32 // shr $0x20,%rcx
25547 .byte 102,65,15,110,4,137 // movd (%r9,%rcx,4),%xmm0
25548 .byte 102,65,15,110,12,129 // movd (%r9,%rax,4),%xmm1
25549 .byte 102,15,98,193 // punpckldq %xmm1,%xmm0
25550 .byte 102,67,15,110,28,145 // movd (%r9,%r10,4),%xmm3
25551 .byte 102,67,15,110,12,129 // movd (%r9,%r8,4),%xmm1
25552 .byte 102,15,98,217 // punpckldq %xmm1,%xmm3
25553 .byte 102,15,98,216 // punpckldq %xmm0,%xmm3
25554 .byte 184,255,0,0,0 // mov $0xff,%eax
25555 .byte 102,15,110,192 // movd %eax,%xmm0
25556 .byte 102,15,112,192,0 // pshufd $0x0,%xmm0,%xmm0
25557 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25558 .byte 102,15,114,209,8 // psrld $0x8,%xmm1
25559 .byte 102,15,219,200 // pand %xmm0,%xmm1
25560 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
25561 .byte 102,15,114,210,16 // psrld $0x10,%xmm2
25562 .byte 102,15,219,208 // pand %xmm0,%xmm2
25563 .byte 102,15,219,195 // pand %xmm3,%xmm0
25564 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
25565 .byte 184,129,128,128,59 // mov $0x3b808081,%eax
25566 .byte 102,68,15,110,192 // movd %eax,%xmm8
25567 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25568 .byte 65,15,89,192 // mulps %xmm8,%xmm0
25569 .byte 15,91,201 // cvtdq2ps %xmm1,%xmm1
25570 .byte 65,15,89,200 // mulps %xmm8,%xmm1
25571 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
25572 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25573 .byte 102,15,114,211,24 // psrld $0x18,%xmm3
25574 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25575 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25576 .byte 72,173 // lods %ds:(%rsi),%rax
25577 .byte 255,224 // jmpq *%rax
25578
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25579HIDDEN _sk_store_8888_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25580.globl _sk_store_8888_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25581FUNCTION(_sk_store_8888_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25582_sk_store_8888_sse2:
25583 .byte 72,173 // lods %ds:(%rsi),%rax
25584 .byte 72,139,0 // mov (%rax),%rax
25585 .byte 185,0,0,127,67 // mov $0x437f0000,%ecx
25586 .byte 102,68,15,110,193 // movd %ecx,%xmm8
25587 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25588 .byte 69,15,40,200 // movaps %xmm8,%xmm9
25589 .byte 68,15,89,200 // mulps %xmm0,%xmm9
25590 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25591 .byte 69,15,40,208 // movaps %xmm8,%xmm10
25592 .byte 68,15,89,209 // mulps %xmm1,%xmm10
25593 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
25594 .byte 102,65,15,114,242,8 // pslld $0x8,%xmm10
25595 .byte 102,69,15,235,209 // por %xmm9,%xmm10
25596 .byte 69,15,40,200 // movaps %xmm8,%xmm9
25597 .byte 68,15,89,202 // mulps %xmm2,%xmm9
25598 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25599 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
25600 .byte 68,15,89,195 // mulps %xmm3,%xmm8
25601 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
25602 .byte 102,65,15,114,240,24 // pslld $0x18,%xmm8
25603 .byte 102,69,15,235,193 // por %xmm9,%xmm8
25604 .byte 102,69,15,235,194 // por %xmm10,%xmm8
25605 .byte 243,68,15,127,4,184 // movdqu %xmm8,(%rax,%rdi,4)
25606 .byte 72,173 // lods %ds:(%rsi),%rax
25607 .byte 255,224 // jmpq *%rax
25608
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25609HIDDEN _sk_load_f16_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25610.globl _sk_load_f16_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25611FUNCTION(_sk_load_f16_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25612_sk_load_f16_sse2:
25613 .byte 72,173 // lods %ds:(%rsi),%rax
25614 .byte 72,139,0 // mov (%rax),%rax
25615 .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0
25616 .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1
Mike Klein114e6b32017-04-03 22:21:15 -0400[diff] [blame]25617 .byte 102,68,15,111,192 // movdqa %xmm0,%xmm8
25618 .byte 102,68,15,97,193 // punpcklwd %xmm1,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25619 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25620 .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14
25621 .byte 102,68,15,97,240 // punpcklwd %xmm0,%xmm14
Mike Klein114e6b32017-04-03 22:21:15 -0400[diff] [blame]25622 .byte 102,68,15,105,192 // punpckhwd %xmm0,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25623 .byte 102,69,15,239,210 // pxor %xmm10,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25624 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1
25625 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1
25626 .byte 102,68,15,111,13,29,24,0,0 // movdqa 0x181d(%rip),%xmm9 # 46b0 <_sk_callback_sse2+0x440>
25627 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
25628 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
25629 .byte 102,15,239,200 // pxor %xmm0,%xmm1
25630 .byte 102,15,114,240,16 // pslld $0x10,%xmm0
25631 .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25632 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25633 .byte 102,68,15,235,232 // por %xmm0,%xmm13
25634 .byte 102,68,15,111,29,2,24,0,0 // movdqa 0x1802(%rip),%xmm11 # 46c0 <_sk_callback_sse2+0x450>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25635 .byte 102,69,15,254,235 // paddd %xmm11,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25636 .byte 102,68,15,111,37,4,24,0,0 // movdqa 0x1804(%rip),%xmm12 # 46d0 <_sk_callback_sse2+0x460>
25637 .byte 102,65,15,239,204 // pxor %xmm12,%xmm1
25638 .byte 102,15,111,29,7,24,0,0 // movdqa 0x1807(%rip),%xmm3 # 46e0 <_sk_callback_sse2+0x470>
25639 .byte 102,15,111,195 // movdqa %xmm3,%xmm0
25640 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0
25641 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0
25642 .byte 102,65,15,115,222,8 // psrldq $0x8,%xmm14
25643 .byte 102,69,15,97,242 // punpcklwd %xmm10,%xmm14
25644 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1
25645 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
25646 .byte 102,68,15,239,241 // pxor %xmm1,%xmm14
25647 .byte 102,15,114,241,16 // pslld $0x10,%xmm1
25648 .byte 102,65,15,111,214 // movdqa %xmm14,%xmm2
25649 .byte 102,15,114,242,13 // pslld $0xd,%xmm2
25650 .byte 102,15,235,209 // por %xmm1,%xmm2
25651 .byte 102,65,15,254,211 // paddd %xmm11,%xmm2
25652 .byte 102,69,15,239,244 // pxor %xmm12,%xmm14
25653 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25654 .byte 102,65,15,102,206 // pcmpgtd %xmm14,%xmm1
25655 .byte 102,15,223,202 // pandn %xmm2,%xmm1
25656 .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13
25657 .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13
25658 .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2
25659 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
25660 .byte 102,68,15,239,234 // pxor %xmm2,%xmm13
25661 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
25662 .byte 102,69,15,111,245 // movdqa %xmm13,%xmm14
25663 .byte 102,65,15,114,246,13 // pslld $0xd,%xmm14
25664 .byte 102,68,15,235,242 // por %xmm2,%xmm14
25665 .byte 102,69,15,254,243 // paddd %xmm11,%xmm14
25666 .byte 102,69,15,239,236 // pxor %xmm12,%xmm13
25667 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
25668 .byte 102,65,15,102,213 // pcmpgtd %xmm13,%xmm2
25669 .byte 102,65,15,223,214 // pandn %xmm14,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25670 .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25671 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
25672 .byte 102,69,15,219,200 // pand %xmm8,%xmm9
25673 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8
25674 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25675 .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
25676 .byte 102,65,15,114,242,13 // pslld $0xd,%xmm10
25677 .byte 102,69,15,235,209 // por %xmm9,%xmm10
25678 .byte 102,69,15,254,211 // paddd %xmm11,%xmm10
25679 .byte 102,69,15,239,196 // pxor %xmm12,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25680 .byte 102,65,15,102,216 // pcmpgtd %xmm8,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25681 .byte 102,65,15,223,218 // pandn %xmm10,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25682 .byte 72,173 // lods %ds:(%rsi),%rax
25683 .byte 255,224 // jmpq *%rax
25684
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]25685HIDDEN _sk_gather_f16_sse2
25686.globl _sk_gather_f16_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25687FUNCTION(_sk_gather_f16_sse2)
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]25688_sk_gather_f16_sse2:
25689 .byte 72,173 // lods %ds:(%rsi),%rax
25690 .byte 76,139,8 // mov (%rax),%r9
25691 .byte 243,15,91,201 // cvttps2dq %xmm1,%xmm1
25692 .byte 102,15,110,80,16 // movd 0x10(%rax),%xmm2
25693 .byte 102,15,112,210,0 // pshufd $0x0,%xmm2,%xmm2
25694 .byte 102,15,112,217,245 // pshufd $0xf5,%xmm1,%xmm3
25695 .byte 102,15,244,218 // pmuludq %xmm2,%xmm3
25696 .byte 102,15,112,219,232 // pshufd $0xe8,%xmm3,%xmm3
25697 .byte 102,15,244,209 // pmuludq %xmm1,%xmm2
25698 .byte 102,15,112,202,232 // pshufd $0xe8,%xmm2,%xmm1
25699 .byte 102,15,98,203 // punpckldq %xmm3,%xmm1
25700 .byte 243,15,91,192 // cvttps2dq %xmm0,%xmm0
25701 .byte 102,15,254,193 // paddd %xmm1,%xmm0
25702 .byte 102,15,112,200,78 // pshufd $0x4e,%xmm0,%xmm1
25703 .byte 102,72,15,126,200 // movq %xmm1,%rax
25704 .byte 65,137,192 // mov %eax,%r8d
25705 .byte 72,193,232,32 // shr $0x20,%rax
25706 .byte 102,72,15,126,193 // movq %xmm0,%rcx
25707 .byte 65,137,202 // mov %ecx,%r10d
25708 .byte 72,193,233,32 // shr $0x20,%rcx
25709 .byte 243,65,15,126,4,201 // movq (%r9,%rcx,8),%xmm0
25710 .byte 243,67,15,126,12,209 // movq (%r9,%r10,8),%xmm1
25711 .byte 102,15,108,200 // punpcklqdq %xmm0,%xmm1
25712 .byte 243,65,15,126,4,193 // movq (%r9,%rax,8),%xmm0
25713 .byte 243,67,15,126,20,193 // movq (%r9,%r8,8),%xmm2
25714 .byte 102,15,108,208 // punpcklqdq %xmm0,%xmm2
25715 .byte 102,68,15,111,193 // movdqa %xmm1,%xmm8
25716 .byte 102,68,15,97,194 // punpcklwd %xmm2,%xmm8
25717 .byte 102,15,105,202 // punpckhwd %xmm2,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25718 .byte 102,69,15,111,240 // movdqa %xmm8,%xmm14
25719 .byte 102,68,15,97,241 // punpcklwd %xmm1,%xmm14
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]25720 .byte 102,68,15,105,193 // punpckhwd %xmm1,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25721 .byte 102,69,15,239,210 // pxor %xmm10,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25722 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1
25723 .byte 102,65,15,97,202 // punpcklwd %xmm10,%xmm1
25724 .byte 102,68,15,111,13,149,22,0,0 // movdqa 0x1695(%rip),%xmm9 # 46f0 <_sk_callback_sse2+0x480>
25725 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
25726 .byte 102,65,15,219,193 // pand %xmm9,%xmm0
25727 .byte 102,15,239,200 // pxor %xmm0,%xmm1
25728 .byte 102,15,114,240,16 // pslld $0x10,%xmm0
25729 .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25730 .byte 102,65,15,114,245,13 // pslld $0xd,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25731 .byte 102,68,15,235,232 // por %xmm0,%xmm13
25732 .byte 102,68,15,111,29,122,22,0,0 // movdqa 0x167a(%rip),%xmm11 # 4700 <_sk_callback_sse2+0x490>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25733 .byte 102,69,15,254,235 // paddd %xmm11,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25734 .byte 102,68,15,111,37,124,22,0,0 // movdqa 0x167c(%rip),%xmm12 # 4710 <_sk_callback_sse2+0x4a0>
25735 .byte 102,65,15,239,204 // pxor %xmm12,%xmm1
25736 .byte 102,15,111,29,127,22,0,0 // movdqa 0x167f(%rip),%xmm3 # 4720 <_sk_callback_sse2+0x4b0>
25737 .byte 102,15,111,195 // movdqa %xmm3,%xmm0
25738 .byte 102,15,102,193 // pcmpgtd %xmm1,%xmm0
25739 .byte 102,65,15,223,197 // pandn %xmm13,%xmm0
25740 .byte 102,65,15,115,222,8 // psrldq $0x8,%xmm14
25741 .byte 102,69,15,97,242 // punpcklwd %xmm10,%xmm14
25742 .byte 102,65,15,111,206 // movdqa %xmm14,%xmm1
25743 .byte 102,65,15,219,201 // pand %xmm9,%xmm1
25744 .byte 102,68,15,239,241 // pxor %xmm1,%xmm14
25745 .byte 102,15,114,241,16 // pslld $0x10,%xmm1
25746 .byte 102,65,15,111,214 // movdqa %xmm14,%xmm2
25747 .byte 102,15,114,242,13 // pslld $0xd,%xmm2
25748 .byte 102,15,235,209 // por %xmm1,%xmm2
25749 .byte 102,65,15,254,211 // paddd %xmm11,%xmm2
25750 .byte 102,69,15,239,244 // pxor %xmm12,%xmm14
25751 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25752 .byte 102,65,15,102,206 // pcmpgtd %xmm14,%xmm1
25753 .byte 102,15,223,202 // pandn %xmm2,%xmm1
25754 .byte 102,69,15,111,232 // movdqa %xmm8,%xmm13
25755 .byte 102,69,15,97,234 // punpcklwd %xmm10,%xmm13
25756 .byte 102,65,15,111,213 // movdqa %xmm13,%xmm2
25757 .byte 102,65,15,219,209 // pand %xmm9,%xmm2
25758 .byte 102,68,15,239,234 // pxor %xmm2,%xmm13
25759 .byte 102,15,114,242,16 // pslld $0x10,%xmm2
25760 .byte 102,69,15,111,245 // movdqa %xmm13,%xmm14
25761 .byte 102,65,15,114,246,13 // pslld $0xd,%xmm14
25762 .byte 102,68,15,235,242 // por %xmm2,%xmm14
25763 .byte 102,69,15,254,243 // paddd %xmm11,%xmm14
25764 .byte 102,69,15,239,236 // pxor %xmm12,%xmm13
25765 .byte 102,15,111,211 // movdqa %xmm3,%xmm2
25766 .byte 102,65,15,102,213 // pcmpgtd %xmm13,%xmm2
25767 .byte 102,65,15,223,214 // pandn %xmm14,%xmm2
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25768 .byte 102,65,15,115,216,8 // psrldq $0x8,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25769 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
25770 .byte 102,69,15,219,200 // pand %xmm8,%xmm9
25771 .byte 102,69,15,239,193 // pxor %xmm9,%xmm8
25772 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25773 .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
25774 .byte 102,65,15,114,242,13 // pslld $0xd,%xmm10
25775 .byte 102,69,15,235,209 // por %xmm9,%xmm10
25776 .byte 102,69,15,254,211 // paddd %xmm11,%xmm10
25777 .byte 102,69,15,239,196 // pxor %xmm12,%xmm8
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25778 .byte 102,65,15,102,216 // pcmpgtd %xmm8,%xmm3
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25779 .byte 102,65,15,223,218 // pandn %xmm10,%xmm3
Mike Klein5f055f02017-04-06 20:02:11 -0400[diff] [blame]25780 .byte 72,173 // lods %ds:(%rsi),%rax
25781 .byte 255,224 // jmpq *%rax
25782
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]25783HIDDEN _sk_store_f16_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25784.globl _sk_store_f16_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25785FUNCTION(_sk_store_f16_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25786_sk_store_f16_sse2:
25787 .byte 72,173 // lods %ds:(%rsi),%rax
25788 .byte 72,139,0 // mov (%rax),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25789 .byte 102,68,15,111,21,167,21,0,0 // movdqa 0x15a7(%rip),%xmm10 # 4730 <_sk_callback_sse2+0x4c0>
25790 .byte 102,68,15,111,224 // movdqa %xmm0,%xmm12
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25791 .byte 102,68,15,111,232 // movdqa %xmm0,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25792 .byte 102,69,15,219,234 // pand %xmm10,%xmm13
25793 .byte 102,69,15,239,229 // pxor %xmm13,%xmm12
25794 .byte 102,68,15,111,13,154,21,0,0 // movdqa 0x159a(%rip),%xmm9 # 4740 <_sk_callback_sse2+0x4d0>
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25795 .byte 102,65,15,114,213,16 // psrld $0x10,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25796 .byte 102,69,15,111,193 // movdqa %xmm9,%xmm8
25797 .byte 102,69,15,102,196 // pcmpgtd %xmm12,%xmm8
25798 .byte 102,65,15,114,212,13 // psrld $0xd,%xmm12
25799 .byte 102,68,15,111,29,139,21,0,0 // movdqa 0x158b(%rip),%xmm11 # 4750 <_sk_callback_sse2+0x4e0>
25800 .byte 102,69,15,235,235 // por %xmm11,%xmm13
25801 .byte 102,69,15,254,236 // paddd %xmm12,%xmm13
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25802 .byte 102,65,15,114,245,16 // pslld $0x10,%xmm13
25803 .byte 102,65,15,114,229,16 // psrad $0x10,%xmm13
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25804 .byte 102,69,15,223,197 // pandn %xmm13,%xmm8
25805 .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
25806 .byte 102,68,15,111,233 // movdqa %xmm1,%xmm13
25807 .byte 102,68,15,111,241 // movdqa %xmm1,%xmm14
25808 .byte 102,69,15,219,242 // pand %xmm10,%xmm14
25809 .byte 102,69,15,239,238 // pxor %xmm14,%xmm13
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25810 .byte 102,65,15,114,214,16 // psrld $0x10,%xmm14
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25811 .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12
25812 .byte 102,69,15,102,229 // pcmpgtd %xmm13,%xmm12
25813 .byte 102,65,15,114,213,13 // psrld $0xd,%xmm13
25814 .byte 102,69,15,235,243 // por %xmm11,%xmm14
25815 .byte 102,69,15,254,245 // paddd %xmm13,%xmm14
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25816 .byte 102,65,15,114,246,16 // pslld $0x10,%xmm14
25817 .byte 102,65,15,114,230,16 // psrad $0x10,%xmm14
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25818 .byte 102,69,15,223,230 // pandn %xmm14,%xmm12
25819 .byte 102,69,15,107,228 // packssdw %xmm12,%xmm12
25820 .byte 102,68,15,111,242 // movdqa %xmm2,%xmm14
25821 .byte 102,68,15,111,250 // movdqa %xmm2,%xmm15
25822 .byte 102,69,15,219,250 // pand %xmm10,%xmm15
25823 .byte 102,69,15,239,247 // pxor %xmm15,%xmm14
25824 .byte 102,65,15,114,215,16 // psrld $0x10,%xmm15
25825 .byte 102,69,15,111,233 // movdqa %xmm9,%xmm13
25826 .byte 102,69,15,102,238 // pcmpgtd %xmm14,%xmm13
Mike Klein097d0932017-04-20 09:11:53 -0400[diff] [blame]25827 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25828 .byte 102,69,15,235,251 // por %xmm11,%xmm15
25829 .byte 102,69,15,254,254 // paddd %xmm14,%xmm15
25830 .byte 102,65,15,114,247,16 // pslld $0x10,%xmm15
25831 .byte 102,65,15,114,231,16 // psrad $0x10,%xmm15
25832 .byte 102,69,15,223,239 // pandn %xmm15,%xmm13
25833 .byte 102,69,15,107,237 // packssdw %xmm13,%xmm13
25834 .byte 102,68,15,219,211 // pand %xmm3,%xmm10
25835 .byte 102,68,15,111,243 // movdqa %xmm3,%xmm14
25836 .byte 102,69,15,239,242 // pxor %xmm10,%xmm14
25837 .byte 102,65,15,114,210,16 // psrld $0x10,%xmm10
25838 .byte 102,69,15,102,206 // pcmpgtd %xmm14,%xmm9
25839 .byte 102,65,15,114,214,13 // psrld $0xd,%xmm14
25840 .byte 102,69,15,235,211 // por %xmm11,%xmm10
25841 .byte 102,69,15,254,214 // paddd %xmm14,%xmm10
25842 .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10
25843 .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10
25844 .byte 102,69,15,223,202 // pandn %xmm10,%xmm9
25845 .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25846 .byte 102,69,15,97,196 // punpcklwd %xmm12,%xmm8
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]25847 .byte 102,69,15,97,233 // punpcklwd %xmm9,%xmm13
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]25848 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25849 .byte 102,69,15,98,205 // punpckldq %xmm13,%xmm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25850 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
Mike Kleind0ce1482017-04-19 17:19:30 -0400[diff] [blame]25851 .byte 102,69,15,106,197 // punpckhdq %xmm13,%xmm8
Mike Klein95f53be2017-04-04 10:24:56 -0400[diff] [blame]25852 .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]25853 .byte 72,173 // lods %ds:(%rsi),%rax
25854 .byte 255,224 // jmpq *%rax
25855
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]25856HIDDEN _sk_load_u16_be_sse2
25857.globl _sk_load_u16_be_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25858FUNCTION(_sk_load_u16_be_sse2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]25859_sk_load_u16_be_sse2:
25860 .byte 72,173 // lods %ds:(%rsi),%rax
25861 .byte 72,139,0 // mov (%rax),%rax
25862 .byte 243,15,111,4,248 // movdqu (%rax,%rdi,8),%xmm0
25863 .byte 243,15,111,76,248,16 // movdqu 0x10(%rax,%rdi,8),%xmm1
25864 .byte 102,15,111,208 // movdqa %xmm0,%xmm2
25865 .byte 102,15,97,209 // punpcklwd %xmm1,%xmm2
25866 .byte 102,15,105,193 // punpckhwd %xmm1,%xmm0
25867 .byte 102,15,111,202 // movdqa %xmm2,%xmm1
25868 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
25869 .byte 102,15,105,208 // punpckhwd %xmm0,%xmm2
25870 .byte 184,128,0,128,55 // mov $0x37800080,%eax
25871 .byte 102,68,15,110,192 // movd %eax,%xmm8
25872 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25873 .byte 102,15,111,193 // movdqa %xmm1,%xmm0
25874 .byte 102,15,113,240,8 // psllw $0x8,%xmm0
25875 .byte 102,15,112,217,78 // pshufd $0x4e,%xmm1,%xmm3
25876 .byte 102,15,113,209,8 // psrlw $0x8,%xmm1
25877 .byte 102,15,235,200 // por %xmm0,%xmm1
25878 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
25879 .byte 102,65,15,97,201 // punpcklwd %xmm9,%xmm1
25880 .byte 15,91,193 // cvtdq2ps %xmm1,%xmm0
25881 .byte 65,15,89,192 // mulps %xmm8,%xmm0
25882 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25883 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
25884 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
25885 .byte 102,15,235,217 // por %xmm1,%xmm3
25886 .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3
25887 .byte 15,91,203 // cvtdq2ps %xmm3,%xmm1
25888 .byte 65,15,89,200 // mulps %xmm8,%xmm1
25889 .byte 102,68,15,111,210 // movdqa %xmm2,%xmm10
25890 .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10
25891 .byte 102,15,112,218,78 // pshufd $0x4e,%xmm2,%xmm3
25892 .byte 102,15,113,210,8 // psrlw $0x8,%xmm2
25893 .byte 102,65,15,235,210 // por %xmm10,%xmm2
25894 .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2
25895 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
25896 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25897 .byte 102,68,15,111,211 // movdqa %xmm3,%xmm10
25898 .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10
25899 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
25900 .byte 102,65,15,235,218 // por %xmm10,%xmm3
25901 .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3
25902 .byte 15,91,219 // cvtdq2ps %xmm3,%xmm3
25903 .byte 65,15,89,216 // mulps %xmm8,%xmm3
25904 .byte 72,173 // lods %ds:(%rsi),%rax
25905 .byte 255,224 // jmpq *%rax
25906
Mike Kleinb3821732017-04-17 10:58:05 -0400[diff] [blame]25907HIDDEN _sk_load_rgb_u16_be_sse2
25908.globl _sk_load_rgb_u16_be_sse2
25909FUNCTION(_sk_load_rgb_u16_be_sse2)
25910_sk_load_rgb_u16_be_sse2:
25911 .byte 72,173 // lods %ds:(%rsi),%rax
25912 .byte 72,139,0 // mov (%rax),%rax
25913 .byte 72,141,12,127 // lea (%rdi,%rdi,2),%rcx
25914 .byte 243,15,111,20,72 // movdqu (%rax,%rcx,2),%xmm2
25915 .byte 243,15,111,68,72,8 // movdqu 0x8(%rax,%rcx,2),%xmm0
25916 .byte 102,15,115,216,4 // psrldq $0x4,%xmm0
25917 .byte 102,15,111,202 // movdqa %xmm2,%xmm1
25918 .byte 102,15,115,217,6 // psrldq $0x6,%xmm1
25919 .byte 102,15,97,208 // punpcklwd %xmm0,%xmm2
25920 .byte 102,15,115,216,6 // psrldq $0x6,%xmm0
25921 .byte 102,15,97,200 // punpcklwd %xmm0,%xmm1
25922 .byte 102,15,111,194 // movdqa %xmm2,%xmm0
25923 .byte 102,15,97,193 // punpcklwd %xmm1,%xmm0
25924 .byte 102,15,112,216,78 // pshufd $0x4e,%xmm0,%xmm3
25925 .byte 102,15,105,209 // punpckhwd %xmm1,%xmm2
25926 .byte 184,128,0,128,55 // mov $0x37800080,%eax
25927 .byte 102,68,15,110,192 // movd %eax,%xmm8
25928 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
25929 .byte 102,15,111,200 // movdqa %xmm0,%xmm1
25930 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
25931 .byte 102,15,113,208,8 // psrlw $0x8,%xmm0
25932 .byte 102,15,235,193 // por %xmm1,%xmm0
25933 .byte 102,69,15,239,201 // pxor %xmm9,%xmm9
25934 .byte 102,65,15,97,193 // punpcklwd %xmm9,%xmm0
25935 .byte 15,91,192 // cvtdq2ps %xmm0,%xmm0
25936 .byte 65,15,89,192 // mulps %xmm8,%xmm0
25937 .byte 102,15,111,203 // movdqa %xmm3,%xmm1
25938 .byte 102,15,113,241,8 // psllw $0x8,%xmm1
25939 .byte 102,15,113,211,8 // psrlw $0x8,%xmm3
25940 .byte 102,15,235,217 // por %xmm1,%xmm3
25941 .byte 102,65,15,97,217 // punpcklwd %xmm9,%xmm3
25942 .byte 15,91,203 // cvtdq2ps %xmm3,%xmm1
25943 .byte 65,15,89,200 // mulps %xmm8,%xmm1
25944 .byte 102,15,111,218 // movdqa %xmm2,%xmm3
25945 .byte 102,15,113,243,8 // psllw $0x8,%xmm3
25946 .byte 102,15,113,210,8 // psrlw $0x8,%xmm2
25947 .byte 102,15,235,211 // por %xmm3,%xmm2
25948 .byte 102,65,15,97,209 // punpcklwd %xmm9,%xmm2
25949 .byte 15,91,210 // cvtdq2ps %xmm2,%xmm2
25950 .byte 65,15,89,208 // mulps %xmm8,%xmm2
25951 .byte 184,0,0,128,63 // mov $0x3f800000,%eax
25952 .byte 102,15,110,216 // movd %eax,%xmm3
25953 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
25954 .byte 72,173 // lods %ds:(%rsi),%rax
25955 .byte 255,224 // jmpq *%rax
25956
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]25957HIDDEN _sk_store_u16_be_sse2
25958.globl _sk_store_u16_be_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]25959FUNCTION(_sk_store_u16_be_sse2)
Mike Klein3146bb92017-04-05 14:45:02 -0400[diff] [blame]25960_sk_store_u16_be_sse2:
25961 .byte 72,173 // lods %ds:(%rsi),%rax
25962 .byte 72,139,0 // mov (%rax),%rax
25963 .byte 185,0,255,127,71 // mov $0x477fff00,%ecx
25964 .byte 102,68,15,110,201 // movd %ecx,%xmm9
25965 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
25966 .byte 69,15,40,193 // movaps %xmm9,%xmm8
25967 .byte 68,15,89,192 // mulps %xmm0,%xmm8
25968 .byte 102,69,15,91,192 // cvtps2dq %xmm8,%xmm8
25969 .byte 102,65,15,114,240,16 // pslld $0x10,%xmm8
25970 .byte 102,65,15,114,224,16 // psrad $0x10,%xmm8
25971 .byte 102,69,15,107,192 // packssdw %xmm8,%xmm8
25972 .byte 102,69,15,111,208 // movdqa %xmm8,%xmm10
25973 .byte 102,65,15,113,242,8 // psllw $0x8,%xmm10
25974 .byte 102,65,15,113,208,8 // psrlw $0x8,%xmm8
25975 .byte 102,69,15,235,194 // por %xmm10,%xmm8
25976 .byte 69,15,40,209 // movaps %xmm9,%xmm10
25977 .byte 68,15,89,209 // mulps %xmm1,%xmm10
25978 .byte 102,69,15,91,210 // cvtps2dq %xmm10,%xmm10
25979 .byte 102,65,15,114,242,16 // pslld $0x10,%xmm10
25980 .byte 102,65,15,114,226,16 // psrad $0x10,%xmm10
25981 .byte 102,69,15,107,210 // packssdw %xmm10,%xmm10
25982 .byte 102,69,15,111,218 // movdqa %xmm10,%xmm11
25983 .byte 102,65,15,113,243,8 // psllw $0x8,%xmm11
25984 .byte 102,65,15,113,210,8 // psrlw $0x8,%xmm10
25985 .byte 102,69,15,235,211 // por %xmm11,%xmm10
25986 .byte 69,15,40,217 // movaps %xmm9,%xmm11
25987 .byte 68,15,89,218 // mulps %xmm2,%xmm11
25988 .byte 102,69,15,91,219 // cvtps2dq %xmm11,%xmm11
25989 .byte 102,65,15,114,243,16 // pslld $0x10,%xmm11
25990 .byte 102,65,15,114,227,16 // psrad $0x10,%xmm11
25991 .byte 102,69,15,107,219 // packssdw %xmm11,%xmm11
25992 .byte 102,69,15,111,227 // movdqa %xmm11,%xmm12
25993 .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12
25994 .byte 102,65,15,113,211,8 // psrlw $0x8,%xmm11
25995 .byte 102,69,15,235,220 // por %xmm12,%xmm11
25996 .byte 68,15,89,203 // mulps %xmm3,%xmm9
25997 .byte 102,69,15,91,201 // cvtps2dq %xmm9,%xmm9
25998 .byte 102,65,15,114,241,16 // pslld $0x10,%xmm9
25999 .byte 102,65,15,114,225,16 // psrad $0x10,%xmm9
26000 .byte 102,69,15,107,201 // packssdw %xmm9,%xmm9
26001 .byte 102,69,15,111,225 // movdqa %xmm9,%xmm12
26002 .byte 102,65,15,113,244,8 // psllw $0x8,%xmm12
26003 .byte 102,65,15,113,209,8 // psrlw $0x8,%xmm9
26004 .byte 102,69,15,235,204 // por %xmm12,%xmm9
26005 .byte 102,69,15,97,194 // punpcklwd %xmm10,%xmm8
26006 .byte 102,69,15,97,217 // punpcklwd %xmm9,%xmm11
26007 .byte 102,69,15,111,200 // movdqa %xmm8,%xmm9
26008 .byte 102,69,15,98,203 // punpckldq %xmm11,%xmm9
26009 .byte 243,68,15,127,12,248 // movdqu %xmm9,(%rax,%rdi,8)
26010 .byte 102,69,15,106,195 // punpckhdq %xmm11,%xmm8
26011 .byte 243,68,15,127,68,248,16 // movdqu %xmm8,0x10(%rax,%rdi,8)
26012 .byte 72,173 // lods %ds:(%rsi),%rax
26013 .byte 255,224 // jmpq *%rax
26014
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]26015HIDDEN _sk_load_f32_sse2
26016.globl _sk_load_f32_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26017FUNCTION(_sk_load_f32_sse2)
Mike Klein14987eb2017-04-06 10:22:26 -0400[diff] [blame]26018_sk_load_f32_sse2:
26019 .byte 72,173 // lods %ds:(%rsi),%rax
26020 .byte 72,139,0 // mov (%rax),%rax
26021 .byte 72,137,249 // mov %rdi,%rcx
26022 .byte 72,193,225,4 // shl $0x4,%rcx
26023 .byte 68,15,16,4,8 // movups (%rax,%rcx,1),%xmm8
26024 .byte 15,16,68,8,16 // movups 0x10(%rax,%rcx,1),%xmm0
26025 .byte 15,16,92,8,32 // movups 0x20(%rax,%rcx,1),%xmm3
26026 .byte 68,15,16,76,8,48 // movups 0x30(%rax,%rcx,1),%xmm9
26027 .byte 65,15,40,208 // movaps %xmm8,%xmm2
26028 .byte 15,20,208 // unpcklps %xmm0,%xmm2
26029 .byte 15,40,203 // movaps %xmm3,%xmm1
26030 .byte 65,15,20,201 // unpcklps %xmm9,%xmm1
26031 .byte 68,15,21,192 // unpckhps %xmm0,%xmm8
26032 .byte 65,15,21,217 // unpckhps %xmm9,%xmm3
26033 .byte 15,40,194 // movaps %xmm2,%xmm0
26034 .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0
26035 .byte 15,18,202 // movhlps %xmm2,%xmm1
26036 .byte 65,15,40,208 // movaps %xmm8,%xmm2
26037 .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2
26038 .byte 65,15,18,216 // movhlps %xmm8,%xmm3
26039 .byte 72,173 // lods %ds:(%rsi),%rax
26040 .byte 255,224 // jmpq *%rax
26041
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26042HIDDEN _sk_store_f32_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26043.globl _sk_store_f32_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26044FUNCTION(_sk_store_f32_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26045_sk_store_f32_sse2:
26046 .byte 72,173 // lods %ds:(%rsi),%rax
26047 .byte 72,139,0 // mov (%rax),%rax
26048 .byte 72,137,249 // mov %rdi,%rcx
26049 .byte 72,193,225,4 // shl $0x4,%rcx
26050 .byte 68,15,40,192 // movaps %xmm0,%xmm8
26051 .byte 68,15,40,200 // movaps %xmm0,%xmm9
26052 .byte 68,15,20,201 // unpcklps %xmm1,%xmm9
26053 .byte 68,15,40,210 // movaps %xmm2,%xmm10
26054 .byte 68,15,40,218 // movaps %xmm2,%xmm11
26055 .byte 68,15,20,219 // unpcklps %xmm3,%xmm11
26056 .byte 68,15,21,193 // unpckhps %xmm1,%xmm8
26057 .byte 68,15,21,211 // unpckhps %xmm3,%xmm10
26058 .byte 69,15,40,225 // movaps %xmm9,%xmm12
26059 .byte 102,69,15,20,227 // unpcklpd %xmm11,%xmm12
26060 .byte 69,15,18,217 // movhlps %xmm9,%xmm11
26061 .byte 69,15,40,200 // movaps %xmm8,%xmm9
26062 .byte 102,69,15,20,202 // unpcklpd %xmm10,%xmm9
26063 .byte 69,15,18,208 // movhlps %xmm8,%xmm10
26064 .byte 102,68,15,17,36,8 // movupd %xmm12,(%rax,%rcx,1)
26065 .byte 68,15,17,92,8,16 // movups %xmm11,0x10(%rax,%rcx,1)
26066 .byte 102,68,15,17,76,8,32 // movupd %xmm9,0x20(%rax,%rcx,1)
26067 .byte 68,15,17,84,8,48 // movups %xmm10,0x30(%rax,%rcx,1)
26068 .byte 72,173 // lods %ds:(%rsi),%rax
26069 .byte 255,224 // jmpq *%rax
26070
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26071HIDDEN _sk_clamp_x_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26072.globl _sk_clamp_x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26073FUNCTION(_sk_clamp_x_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26074_sk_clamp_x_sse2:
26075 .byte 72,173 // lods %ds:(%rsi),%rax
26076 .byte 69,15,87,192 // xorps %xmm8,%xmm8
26077 .byte 68,15,95,192 // maxps %xmm0,%xmm8
26078 .byte 243,68,15,16,8 // movss (%rax),%xmm9
26079 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26080 .byte 102,15,118,192 // pcmpeqd %xmm0,%xmm0
26081 .byte 102,65,15,254,193 // paddd %xmm9,%xmm0
26082 .byte 68,15,93,192 // minps %xmm0,%xmm8
26083 .byte 72,173 // lods %ds:(%rsi),%rax
26084 .byte 65,15,40,192 // movaps %xmm8,%xmm0
26085 .byte 255,224 // jmpq *%rax
26086
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26087HIDDEN _sk_clamp_y_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26088.globl _sk_clamp_y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26089FUNCTION(_sk_clamp_y_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26090_sk_clamp_y_sse2:
26091 .byte 72,173 // lods %ds:(%rsi),%rax
26092 .byte 69,15,87,192 // xorps %xmm8,%xmm8
26093 .byte 68,15,95,193 // maxps %xmm1,%xmm8
26094 .byte 243,68,15,16,8 // movss (%rax),%xmm9
26095 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26096 .byte 102,15,118,201 // pcmpeqd %xmm1,%xmm1
26097 .byte 102,65,15,254,201 // paddd %xmm9,%xmm1
26098 .byte 68,15,93,193 // minps %xmm1,%xmm8
26099 .byte 72,173 // lods %ds:(%rsi),%rax
26100 .byte 65,15,40,200 // movaps %xmm8,%xmm1
26101 .byte 255,224 // jmpq *%rax
26102
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26103HIDDEN _sk_repeat_x_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26104.globl _sk_repeat_x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26105FUNCTION(_sk_repeat_x_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26106_sk_repeat_x_sse2:
26107 .byte 72,173 // lods %ds:(%rsi),%rax
26108 .byte 243,68,15,16,0 // movss (%rax),%xmm8
26109 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26110 .byte 68,15,40,200 // movaps %xmm0,%xmm9
26111 .byte 69,15,94,200 // divps %xmm8,%xmm9
26112 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
26113 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10
26114 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26115 .byte 68,15,84,13,188,16,0,0 // andps 0x10bc(%rip),%xmm9 # 4760 <_sk_callback_sse2+0x4f0>
26116 .byte 69,15,92,209 // subps %xmm9,%xmm10
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26117 .byte 69,15,89,208 // mulps %xmm8,%xmm10
26118 .byte 65,15,92,194 // subps %xmm10,%xmm0
26119 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9
26120 .byte 102,69,15,254,200 // paddd %xmm8,%xmm9
26121 .byte 65,15,93,193 // minps %xmm9,%xmm0
26122 .byte 72,173 // lods %ds:(%rsi),%rax
26123 .byte 255,224 // jmpq *%rax
26124
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26125HIDDEN _sk_repeat_y_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26126.globl _sk_repeat_y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26127FUNCTION(_sk_repeat_y_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26128_sk_repeat_y_sse2:
26129 .byte 72,173 // lods %ds:(%rsi),%rax
26130 .byte 243,68,15,16,0 // movss (%rax),%xmm8
26131 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26132 .byte 68,15,40,201 // movaps %xmm1,%xmm9
26133 .byte 69,15,94,200 // divps %xmm8,%xmm9
26134 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
26135 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10
26136 .byte 69,15,194,202,1 // cmpltps %xmm10,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26137 .byte 68,15,84,13,132,16,0,0 // andps 0x1084(%rip),%xmm9 # 4770 <_sk_callback_sse2+0x500>
26138 .byte 69,15,92,209 // subps %xmm9,%xmm10
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26139 .byte 69,15,89,208 // mulps %xmm8,%xmm10
26140 .byte 65,15,92,202 // subps %xmm10,%xmm1
26141 .byte 102,69,15,118,201 // pcmpeqd %xmm9,%xmm9
26142 .byte 102,69,15,254,200 // paddd %xmm8,%xmm9
26143 .byte 65,15,93,201 // minps %xmm9,%xmm1
26144 .byte 72,173 // lods %ds:(%rsi),%rax
26145 .byte 255,224 // jmpq *%rax
26146
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26147HIDDEN _sk_mirror_x_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26148.globl _sk_mirror_x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26149FUNCTION(_sk_mirror_x_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26150_sk_mirror_x_sse2:
26151 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26152 .byte 243,68,15,16,0 // movss (%rax),%xmm8
26153 .byte 69,15,40,200 // movaps %xmm8,%xmm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26154 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26155 .byte 65,15,92,193 // subps %xmm9,%xmm0
26156 .byte 243,69,15,88,192 // addss %xmm8,%xmm8
26157 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26158 .byte 68,15,40,208 // movaps %xmm0,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26159 .byte 69,15,94,208 // divps %xmm8,%xmm10
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26160 .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11
26161 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11
26162 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26163 .byte 68,15,84,21,58,16,0,0 // andps 0x103a(%rip),%xmm10 # 4780 <_sk_callback_sse2+0x510>
26164 .byte 69,15,87,228 // xorps %xmm12,%xmm12
26165 .byte 69,15,92,218 // subps %xmm10,%xmm11
26166 .byte 69,15,89,216 // mulps %xmm8,%xmm11
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26167 .byte 65,15,92,195 // subps %xmm11,%xmm0
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26168 .byte 65,15,92,193 // subps %xmm9,%xmm0
26169 .byte 68,15,92,224 // subps %xmm0,%xmm12
26170 .byte 65,15,84,196 // andps %xmm12,%xmm0
26171 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8
26172 .byte 102,69,15,254,193 // paddd %xmm9,%xmm8
26173 .byte 65,15,93,192 // minps %xmm8,%xmm0
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26174 .byte 72,173 // lods %ds:(%rsi),%rax
26175 .byte 255,224 // jmpq *%rax
26176
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26177HIDDEN _sk_mirror_y_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26178.globl _sk_mirror_y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26179FUNCTION(_sk_mirror_y_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26180_sk_mirror_y_sse2:
26181 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26182 .byte 243,68,15,16,0 // movss (%rax),%xmm8
26183 .byte 69,15,40,200 // movaps %xmm8,%xmm9
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26184 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26185 .byte 65,15,92,201 // subps %xmm9,%xmm1
26186 .byte 243,69,15,88,192 // addss %xmm8,%xmm8
26187 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26188 .byte 68,15,40,209 // movaps %xmm1,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26189 .byte 69,15,94,208 // divps %xmm8,%xmm10
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26190 .byte 243,69,15,91,218 // cvttps2dq %xmm10,%xmm11
26191 .byte 69,15,91,219 // cvtdq2ps %xmm11,%xmm11
26192 .byte 69,15,194,211,1 // cmpltps %xmm11,%xmm10
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26193 .byte 68,15,84,21,224,15,0,0 // andps 0xfe0(%rip),%xmm10 # 4790 <_sk_callback_sse2+0x520>
26194 .byte 69,15,87,228 // xorps %xmm12,%xmm12
26195 .byte 69,15,92,218 // subps %xmm10,%xmm11
26196 .byte 69,15,89,216 // mulps %xmm8,%xmm11
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26197 .byte 65,15,92,203 // subps %xmm11,%xmm1
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26198 .byte 65,15,92,201 // subps %xmm9,%xmm1
26199 .byte 68,15,92,225 // subps %xmm1,%xmm12
26200 .byte 65,15,84,204 // andps %xmm12,%xmm1
26201 .byte 102,69,15,118,192 // pcmpeqd %xmm8,%xmm8
26202 .byte 102,69,15,254,193 // paddd %xmm9,%xmm8
26203 .byte 65,15,93,200 // minps %xmm8,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26204 .byte 72,173 // lods %ds:(%rsi),%rax
26205 .byte 255,224 // jmpq *%rax
26206
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26207HIDDEN _sk_luminance_to_alpha_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26208.globl _sk_luminance_to_alpha_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26209FUNCTION(_sk_luminance_to_alpha_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26210_sk_luminance_to_alpha_sse2:
26211 .byte 184,208,179,89,62 // mov $0x3e59b3d0,%eax
26212 .byte 102,15,110,216 // movd %eax,%xmm3
26213 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
26214 .byte 15,89,216 // mulps %xmm0,%xmm3
26215 .byte 184,89,23,55,63 // mov $0x3f371759,%eax
26216 .byte 102,15,110,192 // movd %eax,%xmm0
26217 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
26218 .byte 15,89,193 // mulps %xmm1,%xmm0
26219 .byte 15,88,195 // addps %xmm3,%xmm0
26220 .byte 184,152,221,147,61 // mov $0x3d93dd98,%eax
26221 .byte 102,15,110,216 // movd %eax,%xmm3
26222 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
26223 .byte 15,89,218 // mulps %xmm2,%xmm3
26224 .byte 15,88,216 // addps %xmm0,%xmm3
26225 .byte 72,173 // lods %ds:(%rsi),%rax
26226 .byte 15,87,192 // xorps %xmm0,%xmm0
26227 .byte 15,87,201 // xorps %xmm1,%xmm1
26228 .byte 15,87,210 // xorps %xmm2,%xmm2
26229 .byte 255,224 // jmpq *%rax
26230
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26231HIDDEN _sk_matrix_2x3_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26232.globl _sk_matrix_2x3_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26233FUNCTION(_sk_matrix_2x3_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26234_sk_matrix_2x3_sse2:
26235 .byte 68,15,40,201 // movaps %xmm1,%xmm9
26236 .byte 68,15,40,192 // movaps %xmm0,%xmm8
26237 .byte 72,173 // lods %ds:(%rsi),%rax
26238 .byte 243,15,16,0 // movss (%rax),%xmm0
26239 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
26240 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
26241 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
26242 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26243 .byte 243,68,15,16,88,16 // movss 0x10(%rax),%xmm11
26244 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26245 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26246 .byte 69,15,88,211 // addps %xmm11,%xmm10
26247 .byte 65,15,89,192 // mulps %xmm8,%xmm0
26248 .byte 65,15,88,194 // addps %xmm10,%xmm0
26249 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
26250 .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10
26251 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26252 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
26253 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26254 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26255 .byte 69,15,88,211 // addps %xmm11,%xmm10
26256 .byte 65,15,89,200 // mulps %xmm8,%xmm1
26257 .byte 65,15,88,202 // addps %xmm10,%xmm1
26258 .byte 72,173 // lods %ds:(%rsi),%rax
26259 .byte 255,224 // jmpq *%rax
26260
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26261HIDDEN _sk_matrix_3x4_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26262.globl _sk_matrix_3x4_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26263FUNCTION(_sk_matrix_3x4_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26264_sk_matrix_3x4_sse2:
26265 .byte 68,15,40,201 // movaps %xmm1,%xmm9
26266 .byte 68,15,40,192 // movaps %xmm0,%xmm8
26267 .byte 72,173 // lods %ds:(%rsi),%rax
26268 .byte 243,15,16,0 // movss (%rax),%xmm0
26269 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
26270 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
26271 .byte 243,68,15,16,80,12 // movss 0xc(%rax),%xmm10
26272 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26273 .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11
26274 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26275 .byte 243,68,15,16,96,36 // movss 0x24(%rax),%xmm12
26276 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26277 .byte 68,15,89,218 // mulps %xmm2,%xmm11
26278 .byte 69,15,88,220 // addps %xmm12,%xmm11
26279 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26280 .byte 69,15,88,211 // addps %xmm11,%xmm10
26281 .byte 65,15,89,192 // mulps %xmm8,%xmm0
26282 .byte 65,15,88,194 // addps %xmm10,%xmm0
26283 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
26284 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
26285 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26286 .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11
26287 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26288 .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12
26289 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26290 .byte 68,15,89,218 // mulps %xmm2,%xmm11
26291 .byte 69,15,88,220 // addps %xmm12,%xmm11
26292 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26293 .byte 69,15,88,211 // addps %xmm11,%xmm10
26294 .byte 65,15,89,200 // mulps %xmm8,%xmm1
26295 .byte 65,15,88,202 // addps %xmm10,%xmm1
26296 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
26297 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26298 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
26299 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26300 .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12
26301 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26302 .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13
26303 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26304 .byte 68,15,89,226 // mulps %xmm2,%xmm12
26305 .byte 69,15,88,229 // addps %xmm13,%xmm12
26306 .byte 69,15,89,217 // mulps %xmm9,%xmm11
26307 .byte 69,15,88,220 // addps %xmm12,%xmm11
26308 .byte 69,15,89,208 // mulps %xmm8,%xmm10
26309 .byte 69,15,88,211 // addps %xmm11,%xmm10
26310 .byte 72,173 // lods %ds:(%rsi),%rax
26311 .byte 65,15,40,210 // movaps %xmm10,%xmm2
26312 .byte 255,224 // jmpq *%rax
26313
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26314HIDDEN _sk_matrix_4x5_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26315.globl _sk_matrix_4x5_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26316FUNCTION(_sk_matrix_4x5_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26317_sk_matrix_4x5_sse2:
26318 .byte 68,15,40,201 // movaps %xmm1,%xmm9
26319 .byte 68,15,40,192 // movaps %xmm0,%xmm8
26320 .byte 72,173 // lods %ds:(%rsi),%rax
26321 .byte 243,15,16,0 // movss (%rax),%xmm0
26322 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
26323 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
26324 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
26325 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26326 .byte 243,68,15,16,88,32 // movss 0x20(%rax),%xmm11
26327 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26328 .byte 243,68,15,16,96,48 // movss 0x30(%rax),%xmm12
26329 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26330 .byte 243,68,15,16,104,64 // movss 0x40(%rax),%xmm13
26331 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26332 .byte 68,15,89,227 // mulps %xmm3,%xmm12
26333 .byte 69,15,88,229 // addps %xmm13,%xmm12
26334 .byte 68,15,89,218 // mulps %xmm2,%xmm11
26335 .byte 69,15,88,220 // addps %xmm12,%xmm11
26336 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26337 .byte 69,15,88,211 // addps %xmm11,%xmm10
26338 .byte 65,15,89,192 // mulps %xmm8,%xmm0
26339 .byte 65,15,88,194 // addps %xmm10,%xmm0
26340 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
26341 .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10
26342 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26343 .byte 243,68,15,16,88,36 // movss 0x24(%rax),%xmm11
26344 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26345 .byte 243,68,15,16,96,52 // movss 0x34(%rax),%xmm12
26346 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26347 .byte 243,68,15,16,104,68 // movss 0x44(%rax),%xmm13
26348 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26349 .byte 68,15,89,227 // mulps %xmm3,%xmm12
26350 .byte 69,15,88,229 // addps %xmm13,%xmm12
26351 .byte 68,15,89,218 // mulps %xmm2,%xmm11
26352 .byte 69,15,88,220 // addps %xmm12,%xmm11
26353 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26354 .byte 69,15,88,211 // addps %xmm11,%xmm10
26355 .byte 65,15,89,200 // mulps %xmm8,%xmm1
26356 .byte 65,15,88,202 // addps %xmm10,%xmm1
26357 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
26358 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26359 .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11
26360 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26361 .byte 243,68,15,16,96,40 // movss 0x28(%rax),%xmm12
26362 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26363 .byte 243,68,15,16,104,56 // movss 0x38(%rax),%xmm13
26364 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26365 .byte 243,68,15,16,112,72 // movss 0x48(%rax),%xmm14
26366 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
26367 .byte 68,15,89,235 // mulps %xmm3,%xmm13
26368 .byte 69,15,88,238 // addps %xmm14,%xmm13
26369 .byte 68,15,89,226 // mulps %xmm2,%xmm12
26370 .byte 69,15,88,229 // addps %xmm13,%xmm12
26371 .byte 69,15,89,217 // mulps %xmm9,%xmm11
26372 .byte 69,15,88,220 // addps %xmm12,%xmm11
26373 .byte 69,15,89,208 // mulps %xmm8,%xmm10
26374 .byte 69,15,88,211 // addps %xmm11,%xmm10
26375 .byte 243,68,15,16,88,12 // movss 0xc(%rax),%xmm11
26376 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26377 .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12
26378 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26379 .byte 243,68,15,16,104,44 // movss 0x2c(%rax),%xmm13
26380 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26381 .byte 243,68,15,16,112,60 // movss 0x3c(%rax),%xmm14
26382 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
26383 .byte 243,68,15,16,120,76 // movss 0x4c(%rax),%xmm15
26384 .byte 69,15,198,255,0 // shufps $0x0,%xmm15,%xmm15
26385 .byte 68,15,89,243 // mulps %xmm3,%xmm14
26386 .byte 69,15,88,247 // addps %xmm15,%xmm14
26387 .byte 68,15,89,234 // mulps %xmm2,%xmm13
26388 .byte 69,15,88,238 // addps %xmm14,%xmm13
26389 .byte 69,15,89,225 // mulps %xmm9,%xmm12
26390 .byte 69,15,88,229 // addps %xmm13,%xmm12
26391 .byte 69,15,89,216 // mulps %xmm8,%xmm11
26392 .byte 69,15,88,220 // addps %xmm12,%xmm11
26393 .byte 72,173 // lods %ds:(%rsi),%rax
26394 .byte 65,15,40,210 // movaps %xmm10,%xmm2
26395 .byte 65,15,40,219 // movaps %xmm11,%xmm3
26396 .byte 255,224 // jmpq *%rax
26397
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26398HIDDEN _sk_matrix_perspective_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26399.globl _sk_matrix_perspective_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26400FUNCTION(_sk_matrix_perspective_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26401_sk_matrix_perspective_sse2:
26402 .byte 68,15,40,192 // movaps %xmm0,%xmm8
26403 .byte 72,173 // lods %ds:(%rsi),%rax
26404 .byte 243,15,16,0 // movss (%rax),%xmm0
26405 .byte 243,68,15,16,72,4 // movss 0x4(%rax),%xmm9
26406 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
26407 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26408 .byte 243,68,15,16,80,8 // movss 0x8(%rax),%xmm10
26409 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26410 .byte 68,15,89,201 // mulps %xmm1,%xmm9
26411 .byte 69,15,88,202 // addps %xmm10,%xmm9
26412 .byte 65,15,89,192 // mulps %xmm8,%xmm0
26413 .byte 65,15,88,193 // addps %xmm9,%xmm0
26414 .byte 243,68,15,16,72,12 // movss 0xc(%rax),%xmm9
26415 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26416 .byte 243,68,15,16,80,16 // movss 0x10(%rax),%xmm10
26417 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26418 .byte 243,68,15,16,88,20 // movss 0x14(%rax),%xmm11
26419 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26420 .byte 68,15,89,209 // mulps %xmm1,%xmm10
26421 .byte 69,15,88,211 // addps %xmm11,%xmm10
26422 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26423 .byte 69,15,88,202 // addps %xmm10,%xmm9
26424 .byte 243,68,15,16,80,24 // movss 0x18(%rax),%xmm10
26425 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26426 .byte 243,68,15,16,88,28 // movss 0x1c(%rax),%xmm11
26427 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26428 .byte 243,68,15,16,96,32 // movss 0x20(%rax),%xmm12
26429 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26430 .byte 68,15,89,217 // mulps %xmm1,%xmm11
26431 .byte 69,15,88,220 // addps %xmm12,%xmm11
26432 .byte 69,15,89,208 // mulps %xmm8,%xmm10
26433 .byte 69,15,88,211 // addps %xmm11,%xmm10
26434 .byte 65,15,83,202 // rcpps %xmm10,%xmm1
26435 .byte 15,89,193 // mulps %xmm1,%xmm0
26436 .byte 68,15,89,201 // mulps %xmm1,%xmm9
26437 .byte 72,173 // lods %ds:(%rsi),%rax
26438 .byte 65,15,40,201 // movaps %xmm9,%xmm1
26439 .byte 255,224 // jmpq *%rax
26440
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26441HIDDEN _sk_linear_gradient_sse2
26442.globl _sk_linear_gradient_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26443FUNCTION(_sk_linear_gradient_sse2)
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26444_sk_linear_gradient_sse2:
26445 .byte 72,173 // lods %ds:(%rsi),%rax
26446 .byte 243,68,15,16,72,16 // movss 0x10(%rax),%xmm9
26447 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26448 .byte 243,68,15,16,80,20 // movss 0x14(%rax),%xmm10
26449 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26450 .byte 243,68,15,16,88,24 // movss 0x18(%rax),%xmm11
26451 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26452 .byte 243,68,15,16,96,28 // movss 0x1c(%rax),%xmm12
26453 .byte 69,15,198,228,0 // shufps $0x0,%xmm12,%xmm12
26454 .byte 72,139,8 // mov (%rax),%rcx
26455 .byte 72,133,201 // test %rcx,%rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26456 .byte 15,132,15,1,0,0 // je 3cc4 <_sk_linear_gradient_sse2+0x149>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26457 .byte 72,139,64,8 // mov 0x8(%rax),%rax
26458 .byte 72,131,192,32 // add $0x20,%rax
26459 .byte 69,15,87,192 // xorps %xmm8,%xmm8
26460 .byte 15,87,219 // xorps %xmm3,%xmm3
26461 .byte 15,87,210 // xorps %xmm2,%xmm2
26462 .byte 15,87,201 // xorps %xmm1,%xmm1
26463 .byte 243,68,15,16,112,224 // movss -0x20(%rax),%xmm14
26464 .byte 243,68,15,16,104,228 // movss -0x1c(%rax),%xmm13
26465 .byte 69,15,198,246,0 // shufps $0x0,%xmm14,%xmm14
26466 .byte 69,15,40,252 // movaps %xmm12,%xmm15
26467 .byte 68,15,40,224 // movaps %xmm0,%xmm12
26468 .byte 69,15,194,230,1 // cmpltps %xmm14,%xmm12
26469 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26470 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26471 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26472 .byte 69,15,84,196 // andps %xmm12,%xmm8
26473 .byte 69,15,86,198 // orps %xmm14,%xmm8
26474 .byte 243,68,15,16,104,232 // movss -0x18(%rax),%xmm13
26475 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26476 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26477 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26478 .byte 65,15,84,204 // andps %xmm12,%xmm1
26479 .byte 65,15,86,206 // orps %xmm14,%xmm1
26480 .byte 243,68,15,16,104,236 // movss -0x14(%rax),%xmm13
26481 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26482 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26483 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26484 .byte 65,15,84,212 // andps %xmm12,%xmm2
26485 .byte 65,15,86,214 // orps %xmm14,%xmm2
26486 .byte 243,68,15,16,104,240 // movss -0x10(%rax),%xmm13
26487 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26488 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26489 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26490 .byte 65,15,84,220 // andps %xmm12,%xmm3
26491 .byte 65,15,86,222 // orps %xmm14,%xmm3
26492 .byte 243,68,15,16,104,244 // movss -0xc(%rax),%xmm13
26493 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26494 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26495 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26496 .byte 69,15,84,204 // andps %xmm12,%xmm9
26497 .byte 69,15,86,206 // orps %xmm14,%xmm9
26498 .byte 243,68,15,16,104,248 // movss -0x8(%rax),%xmm13
26499 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26500 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26501 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26502 .byte 69,15,84,212 // andps %xmm12,%xmm10
26503 .byte 69,15,86,214 // orps %xmm14,%xmm10
26504 .byte 243,68,15,16,104,252 // movss -0x4(%rax),%xmm13
26505 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26506 .byte 69,15,40,244 // movaps %xmm12,%xmm14
26507 .byte 69,15,85,245 // andnps %xmm13,%xmm14
26508 .byte 69,15,84,220 // andps %xmm12,%xmm11
26509 .byte 69,15,86,222 // orps %xmm14,%xmm11
26510 .byte 243,68,15,16,40 // movss (%rax),%xmm13
26511 .byte 69,15,198,237,0 // shufps $0x0,%xmm13,%xmm13
26512 .byte 69,15,84,252 // andps %xmm12,%xmm15
26513 .byte 69,15,85,229 // andnps %xmm13,%xmm12
26514 .byte 69,15,86,231 // orps %xmm15,%xmm12
26515 .byte 72,131,192,36 // add $0x24,%rax
26516 .byte 72,255,201 // dec %rcx
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26517 .byte 15,133,8,255,255,255 // jne 3bca <_sk_linear_gradient_sse2+0x4f>
26518 .byte 235,13 // jmp 3cd1 <_sk_linear_gradient_sse2+0x156>
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26519 .byte 15,87,201 // xorps %xmm1,%xmm1
26520 .byte 15,87,210 // xorps %xmm2,%xmm2
26521 .byte 15,87,219 // xorps %xmm3,%xmm3
26522 .byte 69,15,87,192 // xorps %xmm8,%xmm8
26523 .byte 68,15,89,192 // mulps %xmm0,%xmm8
26524 .byte 69,15,88,193 // addps %xmm9,%xmm8
26525 .byte 15,89,200 // mulps %xmm0,%xmm1
26526 .byte 65,15,88,202 // addps %xmm10,%xmm1
26527 .byte 15,89,208 // mulps %xmm0,%xmm2
26528 .byte 65,15,88,211 // addps %xmm11,%xmm2
26529 .byte 15,89,216 // mulps %xmm0,%xmm3
26530 .byte 65,15,88,220 // addps %xmm12,%xmm3
26531 .byte 72,173 // lods %ds:(%rsi),%rax
26532 .byte 65,15,40,192 // movaps %xmm8,%xmm0
26533 .byte 255,224 // jmpq *%rax
26534
Mike Klein7c4af542017-03-29 18:21:14 -0400[diff] [blame]26535HIDDEN _sk_linear_gradient_2stops_sse2
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26536.globl _sk_linear_gradient_2stops_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26537FUNCTION(_sk_linear_gradient_2stops_sse2)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26538_sk_linear_gradient_2stops_sse2:
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26539 .byte 68,15,40,192 // movaps %xmm0,%xmm8
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26540 .byte 72,173 // lods %ds:(%rsi),%rax
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26541 .byte 243,15,16,0 // movss (%rax),%xmm0
26542 .byte 243,15,16,72,4 // movss 0x4(%rax),%xmm1
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26543 .byte 15,198,192,0 // shufps $0x0,%xmm0,%xmm0
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26544 .byte 243,15,16,80,16 // movss 0x10(%rax),%xmm2
26545 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26546 .byte 65,15,89,192 // mulps %xmm8,%xmm0
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26547 .byte 15,88,194 // addps %xmm2,%xmm0
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26548 .byte 15,198,201,0 // shufps $0x0,%xmm1,%xmm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26549 .byte 243,15,16,80,20 // movss 0x14(%rax),%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26550 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
26551 .byte 65,15,89,200 // mulps %xmm8,%xmm1
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26552 .byte 15,88,202 // addps %xmm2,%xmm1
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26553 .byte 243,15,16,80,8 // movss 0x8(%rax),%xmm2
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26554 .byte 15,198,210,0 // shufps $0x0,%xmm2,%xmm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26555 .byte 243,15,16,88,24 // movss 0x18(%rax),%xmm3
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26556 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
26557 .byte 65,15,89,208 // mulps %xmm8,%xmm2
26558 .byte 15,88,211 // addps %xmm3,%xmm2
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26559 .byte 243,15,16,88,12 // movss 0xc(%rax),%xmm3
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26560 .byte 15,198,219,0 // shufps $0x0,%xmm3,%xmm3
Herb Derby7b4202d2017-04-10 10:52:34 -0400[diff] [blame]26561 .byte 243,68,15,16,72,28 // movss 0x1c(%rax),%xmm9
Mike Klein8a823fa2017-04-05 17:29:26 -0400[diff] [blame]26562 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26563 .byte 65,15,89,216 // mulps %xmm8,%xmm3
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26564 .byte 65,15,88,217 // addps %xmm9,%xmm3
26565 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]26566 .byte 255,224 // jmpq *%rax
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26567
26568HIDDEN _sk_save_xy_sse2
26569.globl _sk_save_xy_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26570FUNCTION(_sk_save_xy_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26571_sk_save_xy_sse2:
26572 .byte 72,173 // lods %ds:(%rsi),%rax
26573 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26574 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26575 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26576 .byte 69,15,40,200 // movaps %xmm8,%xmm9
26577 .byte 68,15,88,200 // addps %xmm0,%xmm9
26578 .byte 243,69,15,91,209 // cvttps2dq %xmm9,%xmm10
26579 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10
26580 .byte 69,15,40,217 // movaps %xmm9,%xmm11
26581 .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]26582 .byte 68,15,40,37,7,10,0,0 // movaps 0xa07(%rip),%xmm12 # 47a0 <_sk_callback_sse2+0x530>
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26583 .byte 69,15,84,220 // andps %xmm12,%xmm11
26584 .byte 69,15,92,211 // subps %xmm11,%xmm10
26585 .byte 69,15,92,202 // subps %xmm10,%xmm9
26586 .byte 68,15,88,193 // addps %xmm1,%xmm8
26587 .byte 243,69,15,91,208 // cvttps2dq %xmm8,%xmm10
26588 .byte 69,15,91,210 // cvtdq2ps %xmm10,%xmm10
26589 .byte 69,15,40,216 // movaps %xmm8,%xmm11
26590 .byte 69,15,194,218,1 // cmpltps %xmm10,%xmm11
26591 .byte 69,15,84,220 // andps %xmm12,%xmm11
26592 .byte 69,15,92,211 // subps %xmm11,%xmm10
26593 .byte 69,15,92,194 // subps %xmm10,%xmm8
26594 .byte 15,17,0 // movups %xmm0,(%rax)
26595 .byte 15,17,72,32 // movups %xmm1,0x20(%rax)
26596 .byte 68,15,17,72,64 // movups %xmm9,0x40(%rax)
26597 .byte 68,15,17,64,96 // movups %xmm8,0x60(%rax)
26598 .byte 72,173 // lods %ds:(%rsi),%rax
26599 .byte 255,224 // jmpq *%rax
26600
26601HIDDEN _sk_accumulate_sse2
26602.globl _sk_accumulate_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26603FUNCTION(_sk_accumulate_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26604_sk_accumulate_sse2:
26605 .byte 72,173 // lods %ds:(%rsi),%rax
26606 .byte 68,15,16,128,128,0,0,0 // movups 0x80(%rax),%xmm8
26607 .byte 68,15,16,136,160,0,0,0 // movups 0xa0(%rax),%xmm9
26608 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26609 .byte 69,15,40,193 // movaps %xmm9,%xmm8
26610 .byte 68,15,89,192 // mulps %xmm0,%xmm8
26611 .byte 65,15,88,224 // addps %xmm8,%xmm4
26612 .byte 69,15,40,193 // movaps %xmm9,%xmm8
26613 .byte 68,15,89,193 // mulps %xmm1,%xmm8
26614 .byte 65,15,88,232 // addps %xmm8,%xmm5
26615 .byte 69,15,40,193 // movaps %xmm9,%xmm8
26616 .byte 68,15,89,194 // mulps %xmm2,%xmm8
26617 .byte 65,15,88,240 // addps %xmm8,%xmm6
26618 .byte 68,15,89,203 // mulps %xmm3,%xmm9
26619 .byte 65,15,88,249 // addps %xmm9,%xmm7
26620 .byte 72,173 // lods %ds:(%rsi),%rax
26621 .byte 255,224 // jmpq *%rax
26622
26623HIDDEN _sk_bilinear_nx_sse2
26624.globl _sk_bilinear_nx_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26625FUNCTION(_sk_bilinear_nx_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26626_sk_bilinear_nx_sse2:
26627 .byte 72,173 // lods %ds:(%rsi),%rax
26628 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
26629 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26630 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26631 .byte 15,16,0 // movups (%rax),%xmm0
26632 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26633 .byte 65,15,88,192 // addps %xmm8,%xmm0
26634 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26635 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26636 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26637 .byte 69,15,92,193 // subps %xmm9,%xmm8
26638 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax)
26639 .byte 72,173 // lods %ds:(%rsi),%rax
26640 .byte 255,224 // jmpq *%rax
26641
26642HIDDEN _sk_bilinear_px_sse2
26643.globl _sk_bilinear_px_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26644FUNCTION(_sk_bilinear_px_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26645_sk_bilinear_px_sse2:
26646 .byte 72,173 // lods %ds:(%rsi),%rax
26647 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26648 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26649 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26650 .byte 15,16,0 // movups (%rax),%xmm0
26651 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26652 .byte 65,15,88,192 // addps %xmm8,%xmm0
26653 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
26654 .byte 72,173 // lods %ds:(%rsi),%rax
26655 .byte 255,224 // jmpq *%rax
26656
26657HIDDEN _sk_bilinear_ny_sse2
26658.globl _sk_bilinear_ny_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26659FUNCTION(_sk_bilinear_ny_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26660_sk_bilinear_ny_sse2:
26661 .byte 72,173 // lods %ds:(%rsi),%rax
26662 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
26663 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26664 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26665 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26666 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26667 .byte 65,15,88,200 // addps %xmm8,%xmm1
26668 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26669 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26670 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26671 .byte 69,15,92,193 // subps %xmm9,%xmm8
26672 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax)
26673 .byte 72,173 // lods %ds:(%rsi),%rax
26674 .byte 255,224 // jmpq *%rax
26675
26676HIDDEN _sk_bilinear_py_sse2
26677.globl _sk_bilinear_py_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26678FUNCTION(_sk_bilinear_py_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26679_sk_bilinear_py_sse2:
26680 .byte 72,173 // lods %ds:(%rsi),%rax
26681 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26682 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26683 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26684 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26685 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26686 .byte 65,15,88,200 // addps %xmm8,%xmm1
26687 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
26688 .byte 72,173 // lods %ds:(%rsi),%rax
26689 .byte 255,224 // jmpq *%rax
26690
26691HIDDEN _sk_bicubic_n3x_sse2
26692.globl _sk_bicubic_n3x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26693FUNCTION(_sk_bicubic_n3x_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26694_sk_bicubic_n3x_sse2:
26695 .byte 72,173 // lods %ds:(%rsi),%rax
26696 .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx
26697 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26698 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26699 .byte 15,16,0 // movups (%rax),%xmm0
26700 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26701 .byte 65,15,88,192 // addps %xmm8,%xmm0
26702 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26703 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26704 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26705 .byte 69,15,92,193 // subps %xmm9,%xmm8
26706 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
26707 .byte 102,68,15,110,201 // movd %ecx,%xmm9
26708 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26709 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
26710 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26711 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26712 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26713 .byte 69,15,89,192 // mulps %xmm8,%xmm8
26714 .byte 69,15,88,202 // addps %xmm10,%xmm9
26715 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26716 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
26717 .byte 72,173 // lods %ds:(%rsi),%rax
26718 .byte 255,224 // jmpq *%rax
26719
26720HIDDEN _sk_bicubic_n1x_sse2
26721.globl _sk_bicubic_n1x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26722FUNCTION(_sk_bicubic_n1x_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26723_sk_bicubic_n1x_sse2:
26724 .byte 72,173 // lods %ds:(%rsi),%rax
26725 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
26726 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26727 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26728 .byte 15,16,0 // movups (%rax),%xmm0
26729 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26730 .byte 65,15,88,192 // addps %xmm8,%xmm0
26731 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26732 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26733 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26734 .byte 69,15,92,193 // subps %xmm9,%xmm8
26735 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
26736 .byte 102,68,15,110,201 // movd %ecx,%xmm9
26737 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26738 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26739 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26740 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26741 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26742 .byte 69,15,88,202 // addps %xmm10,%xmm9
26743 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26744 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26745 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26746 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26747 .byte 69,15,88,202 // addps %xmm10,%xmm9
26748 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
26749 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26750 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26751 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26752 .byte 69,15,88,202 // addps %xmm10,%xmm9
26753 .byte 68,15,17,136,128,0,0,0 // movups %xmm9,0x80(%rax)
26754 .byte 72,173 // lods %ds:(%rsi),%rax
26755 .byte 255,224 // jmpq *%rax
26756
26757HIDDEN _sk_bicubic_p1x_sse2
26758.globl _sk_bicubic_p1x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26759FUNCTION(_sk_bicubic_p1x_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26760_sk_bicubic_p1x_sse2:
26761 .byte 72,173 // lods %ds:(%rsi),%rax
26762 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26763 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26764 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26765 .byte 15,16,0 // movups (%rax),%xmm0
26766 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26767 .byte 65,15,88,192 // addps %xmm8,%xmm0
26768 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
26769 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26770 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26771 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26772 .byte 102,68,15,110,217 // movd %ecx,%xmm11
26773 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26774 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26775 .byte 69,15,88,211 // addps %xmm11,%xmm10
26776 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26777 .byte 69,15,88,208 // addps %xmm8,%xmm10
26778 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
26779 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26780 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26781 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26782 .byte 69,15,88,208 // addps %xmm8,%xmm10
26783 .byte 68,15,17,144,128,0,0,0 // movups %xmm10,0x80(%rax)
26784 .byte 72,173 // lods %ds:(%rsi),%rax
26785 .byte 255,224 // jmpq *%rax
26786
26787HIDDEN _sk_bicubic_p3x_sse2
26788.globl _sk_bicubic_p3x_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26789FUNCTION(_sk_bicubic_p3x_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26790_sk_bicubic_p3x_sse2:
26791 .byte 72,173 // lods %ds:(%rsi),%rax
26792 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26793 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26794 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26795 .byte 15,16,0 // movups (%rax),%xmm0
26796 .byte 68,15,16,72,64 // movups 0x40(%rax),%xmm9
26797 .byte 65,15,88,192 // addps %xmm8,%xmm0
26798 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
26799 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26800 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26801 .byte 69,15,89,193 // mulps %xmm9,%xmm8
26802 .byte 69,15,89,201 // mulps %xmm9,%xmm9
26803 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
26804 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26805 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26806 .byte 69,15,88,194 // addps %xmm10,%xmm8
26807 .byte 69,15,89,193 // mulps %xmm9,%xmm8
26808 .byte 68,15,17,128,128,0,0,0 // movups %xmm8,0x80(%rax)
26809 .byte 72,173 // lods %ds:(%rsi),%rax
26810 .byte 255,224 // jmpq *%rax
26811
26812HIDDEN _sk_bicubic_n3y_sse2
26813.globl _sk_bicubic_n3y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26814FUNCTION(_sk_bicubic_n3y_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26815_sk_bicubic_n3y_sse2:
26816 .byte 72,173 // lods %ds:(%rsi),%rax
26817 .byte 185,0,0,192,191 // mov $0xbfc00000,%ecx
26818 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26819 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26820 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26821 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26822 .byte 65,15,88,200 // addps %xmm8,%xmm1
26823 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26824 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26825 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26826 .byte 69,15,92,193 // subps %xmm9,%xmm8
26827 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
26828 .byte 102,68,15,110,201 // movd %ecx,%xmm9
26829 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26830 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
26831 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26832 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26833 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26834 .byte 69,15,89,192 // mulps %xmm8,%xmm8
26835 .byte 69,15,88,202 // addps %xmm10,%xmm9
26836 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26837 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
26838 .byte 72,173 // lods %ds:(%rsi),%rax
26839 .byte 255,224 // jmpq *%rax
26840
26841HIDDEN _sk_bicubic_n1y_sse2
26842.globl _sk_bicubic_n1y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26843FUNCTION(_sk_bicubic_n1y_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26844_sk_bicubic_n1y_sse2:
26845 .byte 72,173 // lods %ds:(%rsi),%rax
26846 .byte 185,0,0,0,191 // mov $0xbf000000,%ecx
26847 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26848 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26849 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26850 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26851 .byte 65,15,88,200 // addps %xmm8,%xmm1
26852 .byte 185,0,0,128,63 // mov $0x3f800000,%ecx
26853 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26854 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26855 .byte 69,15,92,193 // subps %xmm9,%xmm8
26856 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
26857 .byte 102,68,15,110,201 // movd %ecx,%xmm9
26858 .byte 69,15,198,201,0 // shufps $0x0,%xmm9,%xmm9
26859 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26860 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26861 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26862 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26863 .byte 69,15,88,202 // addps %xmm10,%xmm9
26864 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26865 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26866 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26867 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26868 .byte 69,15,88,202 // addps %xmm10,%xmm9
26869 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
26870 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26871 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26872 .byte 69,15,89,200 // mulps %xmm8,%xmm9
26873 .byte 69,15,88,202 // addps %xmm10,%xmm9
26874 .byte 68,15,17,136,160,0,0,0 // movups %xmm9,0xa0(%rax)
26875 .byte 72,173 // lods %ds:(%rsi),%rax
26876 .byte 255,224 // jmpq *%rax
26877
26878HIDDEN _sk_bicubic_p1y_sse2
26879.globl _sk_bicubic_p1y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26880FUNCTION(_sk_bicubic_p1y_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26881_sk_bicubic_p1y_sse2:
26882 .byte 72,173 // lods %ds:(%rsi),%rax
26883 .byte 185,0,0,0,63 // mov $0x3f000000,%ecx
26884 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26885 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26886 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26887 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26888 .byte 65,15,88,200 // addps %xmm8,%xmm1
26889 .byte 185,85,85,149,191 // mov $0xbf955555,%ecx
26890 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26891 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26892 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26893 .byte 102,68,15,110,217 // movd %ecx,%xmm11
26894 .byte 69,15,198,219,0 // shufps $0x0,%xmm11,%xmm11
26895 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26896 .byte 69,15,88,211 // addps %xmm11,%xmm10
26897 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26898 .byte 69,15,88,208 // addps %xmm8,%xmm10
26899 .byte 185,57,142,99,61 // mov $0x3d638e39,%ecx
26900 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26901 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26902 .byte 69,15,89,209 // mulps %xmm9,%xmm10
26903 .byte 69,15,88,208 // addps %xmm8,%xmm10
26904 .byte 68,15,17,144,160,0,0,0 // movups %xmm10,0xa0(%rax)
26905 .byte 72,173 // lods %ds:(%rsi),%rax
26906 .byte 255,224 // jmpq *%rax
26907
26908HIDDEN _sk_bicubic_p3y_sse2
26909.globl _sk_bicubic_p3y_sse2
Mike Klein86714282017-04-13 17:37:38 -0400[diff] [blame]26910FUNCTION(_sk_bicubic_p3y_sse2)
Mike Klein0a904492017-04-12 12:52:48 -0400[diff] [blame]26911_sk_bicubic_p3y_sse2:
26912 .byte 72,173 // lods %ds:(%rsi),%rax
26913 .byte 185,0,0,192,63 // mov $0x3fc00000,%ecx
26914 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26915 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26916 .byte 15,16,72,32 // movups 0x20(%rax),%xmm1
26917 .byte 68,15,16,72,96 // movups 0x60(%rax),%xmm9
26918 .byte 65,15,88,200 // addps %xmm8,%xmm1
26919 .byte 185,114,28,199,62 // mov $0x3ec71c72,%ecx
26920 .byte 102,68,15,110,193 // movd %ecx,%xmm8
26921 .byte 69,15,198,192,0 // shufps $0x0,%xmm8,%xmm8
26922 .byte 69,15,89,193 // mulps %xmm9,%xmm8
26923 .byte 69,15,89,201 // mulps %xmm9,%xmm9
26924 .byte 185,171,170,170,190 // mov $0xbeaaaaab,%ecx
26925 .byte 102,68,15,110,209 // movd %ecx,%xmm10
26926 .byte 69,15,198,210,0 // shufps $0x0,%xmm10,%xmm10
26927 .byte 69,15,88,194 // addps %xmm10,%xmm8
26928 .byte 69,15,89,193 // mulps %xmm9,%xmm8
26929 .byte 68,15,17,128,160,0,0,0 // movups %xmm8,0xa0(%rax)
26930 .byte 72,173 // lods %ds:(%rsi),%rax
26931 .byte 255,224 // jmpq *%rax
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26932
26933HIDDEN _sk_callback_sse2
26934.globl _sk_callback_sse2
26935FUNCTION(_sk_callback_sse2)
26936_sk_callback_sse2:
26937 .byte 65,87 // push %r15
26938 .byte 65,86 // push %r14
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26939 .byte 65,84 // push %r12
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26940 .byte 83 // push %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26941 .byte 72,131,236,72 // sub $0x48,%rsp
26942 .byte 15,41,124,36,48 // movaps %xmm7,0x30(%rsp)
26943 .byte 15,41,116,36,32 // movaps %xmm6,0x20(%rsp)
26944 .byte 15,41,108,36,16 // movaps %xmm5,0x10(%rsp)
26945 .byte 15,41,36,36 // movaps %xmm4,(%rsp)
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26946 .byte 73,137,214 // mov %rdx,%r14
26947 .byte 73,137,255 // mov %rdi,%r15
26948 .byte 72,173 // lods %ds:(%rsi),%rax
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26949 .byte 72,137,195 // mov %rax,%rbx
26950 .byte 73,137,244 // mov %rsi,%r12
26951 .byte 15,40,224 // movaps %xmm0,%xmm4
26952 .byte 15,20,225 // unpcklps %xmm1,%xmm4
26953 .byte 15,40,234 // movaps %xmm2,%xmm5
26954 .byte 15,20,235 // unpcklps %xmm3,%xmm5
26955 .byte 15,21,193 // unpckhps %xmm1,%xmm0
26956 .byte 15,21,211 // unpckhps %xmm3,%xmm2
26957 .byte 15,40,204 // movaps %xmm4,%xmm1
26958 .byte 102,15,20,205 // unpcklpd %xmm5,%xmm1
26959 .byte 15,18,236 // movhlps %xmm4,%xmm5
26960 .byte 15,40,216 // movaps %xmm0,%xmm3
26961 .byte 102,15,20,218 // unpcklpd %xmm2,%xmm3
26962 .byte 15,18,208 // movhlps %xmm0,%xmm2
26963 .byte 102,15,17,75,8 // movupd %xmm1,0x8(%rbx)
26964 .byte 15,17,107,24 // movups %xmm5,0x18(%rbx)
26965 .byte 102,15,17,91,40 // movupd %xmm3,0x28(%rbx)
26966 .byte 15,17,83,56 // movups %xmm2,0x38(%rbx)
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26967 .byte 190,4,0,0,0 // mov $0x4,%esi
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26968 .byte 72,137,223 // mov %rbx,%rdi
26969 .byte 255,19 // callq *(%rbx)
26970 .byte 72,139,131,136,0,0,0 // mov 0x88(%rbx),%rax
26971 .byte 15,16,32 // movups (%rax),%xmm4
26972 .byte 15,16,64,16 // movups 0x10(%rax),%xmm0
26973 .byte 15,16,88,32 // movups 0x20(%rax),%xmm3
26974 .byte 15,16,80,48 // movups 0x30(%rax),%xmm2
26975 .byte 15,40,236 // movaps %xmm4,%xmm5
26976 .byte 15,20,232 // unpcklps %xmm0,%xmm5
26977 .byte 15,40,203 // movaps %xmm3,%xmm1
26978 .byte 15,20,202 // unpcklps %xmm2,%xmm1
26979 .byte 15,21,224 // unpckhps %xmm0,%xmm4
26980 .byte 15,21,218 // unpckhps %xmm2,%xmm3
26981 .byte 15,40,197 // movaps %xmm5,%xmm0
26982 .byte 102,15,20,193 // unpcklpd %xmm1,%xmm0
26983 .byte 15,18,205 // movhlps %xmm5,%xmm1
26984 .byte 15,40,212 // movaps %xmm4,%xmm2
26985 .byte 102,15,20,211 // unpcklpd %xmm3,%xmm2
26986 .byte 15,18,220 // movhlps %xmm4,%xmm3
26987 .byte 76,137,230 // mov %r12,%rsi
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26988 .byte 72,173 // lods %ds:(%rsi),%rax
26989 .byte 76,137,255 // mov %r15,%rdi
26990 .byte 76,137,242 // mov %r14,%rdx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26991 .byte 15,40,36,36 // movaps (%rsp),%xmm4
26992 .byte 15,40,108,36,16 // movaps 0x10(%rsp),%xmm5
26993 .byte 15,40,116,36,32 // movaps 0x20(%rsp),%xmm6
26994 .byte 15,40,124,36,48 // movaps 0x30(%rsp),%xmm7
26995 .byte 72,131,196,72 // add $0x48,%rsp
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26996 .byte 91 // pop %rbx
Mike Kleinc17dc242017-04-20 16:21:57 -0400[diff] [blame]26997 .byte 65,92 // pop %r12
Mike Klein7fee90c2017-04-07 16:55:09 -0400[diff] [blame]26998 .byte 65,94 // pop %r14
26999 .byte 65,95 // pop %r15
27000 .byte 255,224 // jmpq *%rax
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]27001
27002BALIGN16
27003 .byte 0,0 // add %al,(%rax)
27004 .byte 0,63 // add %bh,(%rdi)
27005 .byte 0,0 // add %al,(%rax)
27006 .byte 0,63 // add %bh,(%rdi)
27007 .byte 0,0 // add %al,(%rax)
27008 .byte 0,63 // add %bh,(%rdi)
27009 .byte 0,0 // add %al,(%rax)
27010 .byte 0,63 // add %bh,(%rdi)
27011 .byte 0,0 // add %al,(%rax)
27012 .byte 128,63,0 // cmpb $0x0,(%rdi)
27013 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27014 .byte 63 // (bad)
27015 .byte 0,0 // add %al,(%rax)
Mike Kleinb4bbc642017-04-27 08:59:55 -0400[diff] [blame^]27016 .byte 128,63,0 // cmpb $0x0,(%rdi)
27017 .byte 0,0 // add %al,(%rax)
27018 .byte 52,0 // xor $0x0,%al
27019 .byte 0,0 // add %al,(%rax)
27020 .byte 52,0 // xor $0x0,%al
27021 .byte 0,0 // add %al,(%rax)
27022 .byte 52,0 // xor $0x0,%al
27023 .byte 0,0 // add %al,(%rax)
27024 .byte 52,255 // xor $0xff,%al
27025 .byte 255 // (bad)
27026 .byte 127,0 // jg 4384 <.literal16+0x34>
27027 .byte 255 // (bad)
27028 .byte 255 // (bad)
27029 .byte 127,0 // jg 4388 <.literal16+0x38>
27030 .byte 255 // (bad)
27031 .byte 255 // (bad)
27032 .byte 127,0 // jg 438c <.literal16+0x3c>
27033 .byte 255 // (bad)
27034 .byte 255 // (bad)
27035 .byte 127,0 // jg 4390 <.literal16+0x40>
27036 .byte 0,0 // add %al,(%rax)
27037 .byte 0,63 // add %bh,(%rdi)
27038 .byte 0,0 // add %al,(%rax)
27039 .byte 0,63 // add %bh,(%rdi)
27040 .byte 0,0 // add %al,(%rax)
27041 .byte 0,63 // add %bh,(%rdi)
27042 .byte 0,0 // add %al,(%rax)
27043 .byte 0,63 // add %bh,(%rdi)
27044 .byte 119,115 // ja 4415 <.literal16+0xc5>
27045 .byte 248 // clc
27046 .byte 194,119,115 // retq $0x7377
27047 .byte 248 // clc
27048 .byte 194,119,115 // retq $0x7377
27049 .byte 248 // clc
27050 .byte 194,119,115 // retq $0x7377
27051 .byte 248 // clc
27052 .byte 194,117,191 // retq $0xbf75
27053 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27054 .byte 63 // (bad)
27055 .byte 117,191 // jne 4379 <.literal16+0x29>
27056 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27057 .byte 63 // (bad)
27058 .byte 249 // stc
27059 .byte 68,180,62 // rex.R mov $0x3e,%spl
27060 .byte 249 // stc
27061 .byte 68,180,62 // rex.R mov $0x3e,%spl
27062 .byte 249 // stc
27063 .byte 68,180,62 // rex.R mov $0x3e,%spl
27064 .byte 249 // stc
27065 .byte 68,180,62 // rex.R mov $0x3e,%spl
27066 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
27067 .byte 233,220,63,163,233 // jmpq ffffffffe9a383ba <_sk_callback_sse2+0xffffffffe9a3414a>
27068 .byte 220,63 // fdivrl (%rdi)
27069 .byte 0,0 // add %al,(%rax)
27070 .byte 128,63,0 // cmpb $0x0,(%rdi)
27071 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27072 .byte 63 // (bad)
27073 .byte 0,0 // add %al,(%rax)
27074 .byte 128,63,81 // cmpb $0x51,(%rdi)
27075 .byte 140,242 // mov %?,%edx
27076 .byte 66,81 // rex.X push %rcx
27077 .byte 140,242 // mov %?,%edx
27078 .byte 66,81 // rex.X push %rcx
27079 .byte 140,242 // mov %?,%edx
27080 .byte 66,81 // rex.X push %rcx
27081 .byte 140,242 // mov %?,%edx
27082 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
27083 .byte 63 // (bad)
27084 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
27085 .byte 63 // (bad)
27086 .byte 248 // clc
27087 .byte 245 // cmc
27088 .byte 154 // (bad)
27089 .byte 64,248 // rex clc
27090 .byte 245 // cmc
27091 .byte 154 // (bad)
27092 .byte 64,248 // rex clc
27093 .byte 245 // cmc
27094 .byte 154 // (bad)
27095 .byte 64,248 // rex clc
27096 .byte 245 // cmc
27097 .byte 154 // (bad)
27098 .byte 64,254 // rex (bad)
27099 .byte 210,221 // rcr %cl,%ch
27100 .byte 65,254 // rex.B (bad)
27101 .byte 210,221 // rcr %cl,%ch
27102 .byte 65,254 // rex.B (bad)
27103 .byte 210,221 // rcr %cl,%ch
27104 .byte 65,254 // rex.B (bad)
27105 .byte 210,221 // rcr %cl,%ch
27106 .byte 65,0,0 // add %al,(%r8)
27107 .byte 0,75,0 // add %cl,0x0(%rbx)
27108 .byte 0,0 // add %al,(%rax)
27109 .byte 75,0,0 // rex.WXB add %al,(%r8)
27110 .byte 0,75,0 // add %cl,0x0(%rbx)
27111 .byte 0,0 // add %al,(%rax)
27112 .byte 75,0,0 // rex.WXB add %al,(%r8)
27113 .byte 0,52,0 // add %dh,(%rax,%rax,1)
27114 .byte 0,0 // add %al,(%rax)
27115 .byte 52,0 // xor $0x0,%al
27116 .byte 0,0 // add %al,(%rax)
27117 .byte 52,0 // xor $0x0,%al
27118 .byte 0,0 // add %al,(%rax)
27119 .byte 52,255 // xor $0xff,%al
27120 .byte 255 // (bad)
27121 .byte 127,0 // jg 4454 <.literal16+0x104>
27122 .byte 255 // (bad)
27123 .byte 255 // (bad)
27124 .byte 127,0 // jg 4458 <.literal16+0x108>
27125 .byte 255 // (bad)
27126 .byte 255 // (bad)
27127 .byte 127,0 // jg 445c <.literal16+0x10c>
27128 .byte 255 // (bad)
27129 .byte 255 // (bad)
27130 .byte 127,0 // jg 4460 <.literal16+0x110>
27131 .byte 0,0 // add %al,(%rax)
27132 .byte 0,63 // add %bh,(%rdi)
27133 .byte 0,0 // add %al,(%rax)
27134 .byte 0,63 // add %bh,(%rdi)
27135 .byte 0,0 // add %al,(%rax)
27136 .byte 0,63 // add %bh,(%rdi)
27137 .byte 0,0 // add %al,(%rax)
27138 .byte 0,63 // add %bh,(%rdi)
27139 .byte 119,115 // ja 44e5 <.literal16+0x195>
27140 .byte 248 // clc
27141 .byte 194,119,115 // retq $0x7377
27142 .byte 248 // clc
27143 .byte 194,119,115 // retq $0x7377
27144 .byte 248 // clc
27145 .byte 194,119,115 // retq $0x7377
27146 .byte 248 // clc
27147 .byte 194,117,191 // retq $0xbf75
27148 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27149 .byte 63 // (bad)
27150 .byte 117,191 // jne 4449 <.literal16+0xf9>
27151 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27152 .byte 63 // (bad)
27153 .byte 249 // stc
27154 .byte 68,180,62 // rex.R mov $0x3e,%spl
27155 .byte 249 // stc
27156 .byte 68,180,62 // rex.R mov $0x3e,%spl
27157 .byte 249 // stc
27158 .byte 68,180,62 // rex.R mov $0x3e,%spl
27159 .byte 249 // stc
27160 .byte 68,180,62 // rex.R mov $0x3e,%spl
27161 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
27162 .byte 233,220,63,163,233 // jmpq ffffffffe9a3848a <_sk_callback_sse2+0xffffffffe9a3421a>
27163 .byte 220,63 // fdivrl (%rdi)
27164 .byte 0,0 // add %al,(%rax)
27165 .byte 128,63,0 // cmpb $0x0,(%rdi)
27166 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27167 .byte 63 // (bad)
27168 .byte 0,0 // add %al,(%rax)
27169 .byte 128,63,81 // cmpb $0x51,(%rdi)
27170 .byte 140,242 // mov %?,%edx
27171 .byte 66,81 // rex.X push %rcx
27172 .byte 140,242 // mov %?,%edx
27173 .byte 66,81 // rex.X push %rcx
27174 .byte 140,242 // mov %?,%edx
27175 .byte 66,81 // rex.X push %rcx
27176 .byte 140,242 // mov %?,%edx
27177 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
27178 .byte 63 // (bad)
27179 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
27180 .byte 63 // (bad)
27181 .byte 248 // clc
27182 .byte 245 // cmc
27183 .byte 154 // (bad)
27184 .byte 64,248 // rex clc
27185 .byte 245 // cmc
27186 .byte 154 // (bad)
27187 .byte 64,248 // rex clc
27188 .byte 245 // cmc
27189 .byte 154 // (bad)
27190 .byte 64,248 // rex clc
27191 .byte 245 // cmc
27192 .byte 154 // (bad)
27193 .byte 64,254 // rex (bad)
27194 .byte 210,221 // rcr %cl,%ch
27195 .byte 65,254 // rex.B (bad)
27196 .byte 210,221 // rcr %cl,%ch
27197 .byte 65,254 // rex.B (bad)
27198 .byte 210,221 // rcr %cl,%ch
27199 .byte 65,254 // rex.B (bad)
27200 .byte 210,221 // rcr %cl,%ch
27201 .byte 65,0,0 // add %al,(%r8)
27202 .byte 0,75,0 // add %cl,0x0(%rbx)
27203 .byte 0,0 // add %al,(%rax)
27204 .byte 75,0,0 // rex.WXB add %al,(%r8)
27205 .byte 0,75,0 // add %cl,0x0(%rbx)
27206 .byte 0,0 // add %al,(%rax)
27207 .byte 75,0,0 // rex.WXB add %al,(%r8)
27208 .byte 0,52,0 // add %dh,(%rax,%rax,1)
27209 .byte 0,0 // add %al,(%rax)
27210 .byte 52,0 // xor $0x0,%al
27211 .byte 0,0 // add %al,(%rax)
27212 .byte 52,0 // xor $0x0,%al
27213 .byte 0,0 // add %al,(%rax)
27214 .byte 52,255 // xor $0xff,%al
27215 .byte 255 // (bad)
27216 .byte 127,0 // jg 4524 <.literal16+0x1d4>
27217 .byte 255 // (bad)
27218 .byte 255 // (bad)
27219 .byte 127,0 // jg 4528 <.literal16+0x1d8>
27220 .byte 255 // (bad)
27221 .byte 255 // (bad)
27222 .byte 127,0 // jg 452c <.literal16+0x1dc>
27223 .byte 255 // (bad)
27224 .byte 255 // (bad)
27225 .byte 127,0 // jg 4530 <.literal16+0x1e0>
27226 .byte 0,0 // add %al,(%rax)
27227 .byte 0,63 // add %bh,(%rdi)
27228 .byte 0,0 // add %al,(%rax)
27229 .byte 0,63 // add %bh,(%rdi)
27230 .byte 0,0 // add %al,(%rax)
27231 .byte 0,63 // add %bh,(%rdi)
27232 .byte 0,0 // add %al,(%rax)
27233 .byte 0,63 // add %bh,(%rdi)
27234 .byte 119,115 // ja 45b5 <.literal16+0x265>
27235 .byte 248 // clc
27236 .byte 194,119,115 // retq $0x7377
27237 .byte 248 // clc
27238 .byte 194,119,115 // retq $0x7377
27239 .byte 248 // clc
27240 .byte 194,119,115 // retq $0x7377
27241 .byte 248 // clc
27242 .byte 194,117,191 // retq $0xbf75
27243 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27244 .byte 63 // (bad)
27245 .byte 117,191 // jne 4519 <.literal16+0x1c9>
27246 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27247 .byte 63 // (bad)
27248 .byte 249 // stc
27249 .byte 68,180,62 // rex.R mov $0x3e,%spl
27250 .byte 249 // stc
27251 .byte 68,180,62 // rex.R mov $0x3e,%spl
27252 .byte 249 // stc
27253 .byte 68,180,62 // rex.R mov $0x3e,%spl
27254 .byte 249 // stc
27255 .byte 68,180,62 // rex.R mov $0x3e,%spl
27256 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
27257 .byte 233,220,63,163,233 // jmpq ffffffffe9a3855a <_sk_callback_sse2+0xffffffffe9a342ea>
27258 .byte 220,63 // fdivrl (%rdi)
27259 .byte 0,0 // add %al,(%rax)
27260 .byte 128,63,0 // cmpb $0x0,(%rdi)
27261 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27262 .byte 63 // (bad)
27263 .byte 0,0 // add %al,(%rax)
27264 .byte 128,63,81 // cmpb $0x51,(%rdi)
27265 .byte 140,242 // mov %?,%edx
27266 .byte 66,81 // rex.X push %rcx
27267 .byte 140,242 // mov %?,%edx
27268 .byte 66,81 // rex.X push %rcx
27269 .byte 140,242 // mov %?,%edx
27270 .byte 66,81 // rex.X push %rcx
27271 .byte 140,242 // mov %?,%edx
27272 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
27273 .byte 63 // (bad)
27274 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
27275 .byte 63 // (bad)
27276 .byte 248 // clc
27277 .byte 245 // cmc
27278 .byte 154 // (bad)
27279 .byte 64,248 // rex clc
27280 .byte 245 // cmc
27281 .byte 154 // (bad)
27282 .byte 64,248 // rex clc
27283 .byte 245 // cmc
27284 .byte 154 // (bad)
27285 .byte 64,248 // rex clc
27286 .byte 245 // cmc
27287 .byte 154 // (bad)
27288 .byte 64,254 // rex (bad)
27289 .byte 210,221 // rcr %cl,%ch
27290 .byte 65,254 // rex.B (bad)
27291 .byte 210,221 // rcr %cl,%ch
27292 .byte 65,254 // rex.B (bad)
27293 .byte 210,221 // rcr %cl,%ch
27294 .byte 65,254 // rex.B (bad)
27295 .byte 210,221 // rcr %cl,%ch
27296 .byte 65,0,0 // add %al,(%r8)
27297 .byte 0,75,0 // add %cl,0x0(%rbx)
27298 .byte 0,0 // add %al,(%rax)
27299 .byte 75,0,0 // rex.WXB add %al,(%r8)
27300 .byte 0,75,0 // add %cl,0x0(%rbx)
27301 .byte 0,0 // add %al,(%rax)
27302 .byte 75,0,0 // rex.WXB add %al,(%r8)
27303 .byte 0,52,0 // add %dh,(%rax,%rax,1)
27304 .byte 0,0 // add %al,(%rax)
27305 .byte 52,0 // xor $0x0,%al
27306 .byte 0,0 // add %al,(%rax)
27307 .byte 52,0 // xor $0x0,%al
27308 .byte 0,0 // add %al,(%rax)
27309 .byte 52,255 // xor $0xff,%al
27310 .byte 255 // (bad)
27311 .byte 127,0 // jg 45f4 <.literal16+0x2a4>
27312 .byte 255 // (bad)
27313 .byte 255 // (bad)
27314 .byte 127,0 // jg 45f8 <.literal16+0x2a8>
27315 .byte 255 // (bad)
27316 .byte 255 // (bad)
27317 .byte 127,0 // jg 45fc <.literal16+0x2ac>
27318 .byte 255 // (bad)
27319 .byte 255 // (bad)
27320 .byte 127,0 // jg 4600 <.literal16+0x2b0>
27321 .byte 0,0 // add %al,(%rax)
27322 .byte 0,63 // add %bh,(%rdi)
27323 .byte 0,0 // add %al,(%rax)
27324 .byte 0,63 // add %bh,(%rdi)
27325 .byte 0,0 // add %al,(%rax)
27326 .byte 0,63 // add %bh,(%rdi)
27327 .byte 0,0 // add %al,(%rax)
27328 .byte 0,63 // add %bh,(%rdi)
27329 .byte 119,115 // ja 4685 <.literal16+0x335>
27330 .byte 248 // clc
27331 .byte 194,119,115 // retq $0x7377
27332 .byte 248 // clc
27333 .byte 194,119,115 // retq $0x7377
27334 .byte 248 // clc
27335 .byte 194,119,115 // retq $0x7377
27336 .byte 248 // clc
27337 .byte 194,117,191 // retq $0xbf75
27338 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27339 .byte 63 // (bad)
27340 .byte 117,191 // jne 45e9 <.literal16+0x299>
27341 .byte 191,63,117,191,191 // mov $0xbfbf753f,%edi
27342 .byte 63 // (bad)
27343 .byte 249 // stc
27344 .byte 68,180,62 // rex.R mov $0x3e,%spl
27345 .byte 249 // stc
27346 .byte 68,180,62 // rex.R mov $0x3e,%spl
27347 .byte 249 // stc
27348 .byte 68,180,62 // rex.R mov $0x3e,%spl
27349 .byte 249 // stc
27350 .byte 68,180,62 // rex.R mov $0x3e,%spl
27351 .byte 163,233,220,63,163,233,220,63,163 // movabs %eax,0xa33fdce9a33fdce9
27352 .byte 233,220,63,163,233 // jmpq ffffffffe9a3862a <_sk_callback_sse2+0xffffffffe9a343ba>
27353 .byte 220,63 // fdivrl (%rdi)
27354 .byte 0,0 // add %al,(%rax)
27355 .byte 128,63,0 // cmpb $0x0,(%rdi)
27356 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27357 .byte 63 // (bad)
27358 .byte 0,0 // add %al,(%rax)
27359 .byte 128,63,81 // cmpb $0x51,(%rdi)
27360 .byte 140,242 // mov %?,%edx
27361 .byte 66,81 // rex.X push %rcx
27362 .byte 140,242 // mov %?,%edx
27363 .byte 66,81 // rex.X push %rcx
27364 .byte 140,242 // mov %?,%edx
27365 .byte 66,81 // rex.X push %rcx
27366 .byte 140,242 // mov %?,%edx
27367 .byte 66,141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%r15,4),%edi
27368 .byte 63 // (bad)
27369 .byte 141,188,190,63,141,188,190 // lea -0x414372c1(%rsi,%rdi,4),%edi
27370 .byte 63 // (bad)
27371 .byte 248 // clc
27372 .byte 245 // cmc
27373 .byte 154 // (bad)
27374 .byte 64,248 // rex clc
27375 .byte 245 // cmc
27376 .byte 154 // (bad)
27377 .byte 64,248 // rex clc
27378 .byte 245 // cmc
27379 .byte 154 // (bad)
27380 .byte 64,248 // rex clc
27381 .byte 245 // cmc
27382 .byte 154 // (bad)
27383 .byte 64,254 // rex (bad)
27384 .byte 210,221 // rcr %cl,%ch
27385 .byte 65,254 // rex.B (bad)
27386 .byte 210,221 // rcr %cl,%ch
27387 .byte 65,254 // rex.B (bad)
27388 .byte 210,221 // rcr %cl,%ch
27389 .byte 65,254 // rex.B (bad)
27390 .byte 210,221 // rcr %cl,%ch
27391 .byte 65,0,0 // add %al,(%r8)
27392 .byte 0,75,0 // add %cl,0x0(%rbx)
27393 .byte 0,0 // add %al,(%rax)
27394 .byte 75,0,0 // rex.WXB add %al,(%r8)
27395 .byte 0,75,0 // add %cl,0x0(%rbx)
27396 .byte 0,0 // add %al,(%rax)
27397 .byte 75,0,128,0,0,0,128 // rex.WXB add %al,-0x80000000(%r8)
27398 .byte 0,0 // add %al,(%rax)
27399 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27400 .byte 0,0 // add %al,(%rax)
27401 .byte 0,0 // add %al,(%rax)
27402 .byte 0,56 // add %bh,(%rax)
27403 .byte 0,0 // add %al,(%rax)
27404 .byte 0,56 // add %bh,(%rax)
27405 .byte 0,0 // add %al,(%rax)
27406 .byte 0,56 // add %bh,(%rax)
27407 .byte 0,0 // add %al,(%rax)
27408 .byte 0,56 // add %bh,(%rax)
27409 .byte 0,0 // add %al,(%rax)
27410 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27411 .byte 0,0 // add %al,(%rax)
27412 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27413 .byte 0,4,0 // add %al,(%rax,%rax,1)
27414 .byte 128,0,4 // addb $0x4,(%rax)
27415 .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax)
27416 .byte 0,4,0 // add %al,(%rax,%rax,1)
27417 .byte 128,0,128 // addb $0x80,(%rax)
27418 .byte 0,0 // add %al,(%rax)
27419 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27420 .byte 0,0 // add %al,(%rax)
27421 .byte 0,128,0,0,0,0 // add %al,0x0(%rax)
27422 .byte 0,56 // add %bh,(%rax)
27423 .byte 0,0 // add %al,(%rax)
27424 .byte 0,56 // add %bh,(%rax)
27425 .byte 0,0 // add %al,(%rax)
27426 .byte 0,56 // add %bh,(%rax)
27427 .byte 0,0 // add %al,(%rax)
27428 .byte 0,56 // add %bh,(%rax)
27429 .byte 0,0 // add %al,(%rax)
27430 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27431 .byte 0,0 // add %al,(%rax)
27432 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27433 .byte 0,4,0 // add %al,(%rax,%rax,1)
27434 .byte 128,0,4 // addb $0x4,(%rax)
27435 .byte 0,128,0,4,0,128 // add %al,-0x7ffffc00(%rax)
27436 .byte 0,4,0 // add %al,(%rax,%rax,1)
27437 .byte 128,0,0 // addb $0x0,(%rax)
27438 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27439 .byte 0,0 // add %al,(%rax)
27440 .byte 0,128,0,0,0,128 // add %al,-0x80000000(%rax)
27441 .byte 0,0 // add %al,(%rax)
27442 .byte 128,56,0 // cmpb $0x0,(%rax)
27443 .byte 0,128,56,0,0,128 // add %al,-0x7fffffc8(%rax)
27444 .byte 56,0 // cmp %al,(%rax)
27445 .byte 0,128,56,0,64,254 // add %al,-0x1bfffc8(%rax)
27446 .byte 255,0 // incl (%rax)
27447 .byte 64,254 // rex (bad)
27448 .byte 255,0 // incl (%rax)
27449 .byte 64,254 // rex (bad)
27450 .byte 255,0 // incl (%rax)
27451 .byte 64,254 // rex (bad)
27452 .byte 255,0 // incl (%rax)
27453 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27454 .byte 63 // (bad)
27455 .byte 0,0 // add %al,(%rax)
27456 .byte 128,63,0 // cmpb $0x0,(%rdi)
27457 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27458 .byte 63 // (bad)
27459 .byte 0,0 // add %al,(%rax)
27460 .byte 128,63,0 // cmpb $0x0,(%rdi)
27461 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27462 .byte 63 // (bad)
27463 .byte 0,0 // add %al,(%rax)
27464 .byte 128,63,0 // cmpb $0x0,(%rdi)
27465 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27466 .byte 63 // (bad)
27467 .byte 0,0 // add %al,(%rax)
27468 .byte 128,63,0 // cmpb $0x0,(%rdi)
27469 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27470 .byte 63 // (bad)
27471 .byte 0,0 // add %al,(%rax)
27472 .byte 128,63,0 // cmpb $0x0,(%rdi)
27473 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
27474 .byte 63 // (bad)
27475 .byte 0,0 // add %al,(%rax)
27476 .byte 128,63,0 // cmpb $0x0,(%rdi)
27477 .byte 0,128,63,0,0,128 // add %al,-0x7fffffc1(%rax)
Mike Klein2229b572017-04-21 10:30:29 -0400[diff] [blame]27478 .byte 63 // (bad)
Mike Kleind7e06ae2017-03-29 16:33:06 -0400[diff] [blame]27479#endif